Full text of Review (Federal Reserve Bank of St. Louis) : January/February 2010, Vol. 92, No. 1

View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

The Effects of Recessions
Across Demographic Groups
Kristie M. Engemann and Howard J. Wall
The burdens of a recession are not spread evenly across demographic groups. As the public and
media noticed, from the start of the current recession in December 2007 through June 2009 men
accounted for more than three-quarters of net job losses. Other differences have garnered less attention but are just as interesting. During the same period, the employment of single people fell at more
than twice the rate that it did for married people and the decline for black workers was one and a
half times that for white workers. To provide a more complete understanding of the effect of recessions, this paper examines the different effects of this and previous recessions across a range of
demographic categories: sex, marital status, race, age, and education level. (JEL E32, J20, R12)
Federal Reserve Bank of St. Louis Review, January/February 2010, 92(1), pp. 1-26.

ince the U.S. economy entered its current recession in December 2007, most
demographic groups and industries
have seen steep job losses. By standard
measures of overall labor-market performance,
the news has been dire: Between 2007:Q4 and
2009:Q3, U.S. nonfarm employment fell by about
6.8 million jobs while the national unemployment rate rose from 4.8 percent to 9.6 percent.
Although the picture has been bleak overall,
the recession’s ill effects have not been distributed
evenly across demographic groups. The difference in job losses between men and women has
garnered the most attention; by 2009:Q2, men
accounted for 76 percent of net job losses despite
having only a slim majority (51 percent) of nonfarm employment at the start of the recession. In
light of the disproportionate effects on men, some
commentators in the press and elsewhere have
labeled the current recession a “man-cession” or
even the “Great Man-Cession.”
This paper takes the different effects on men
and women as a starting point and examines the

employment experiences across a range of other
demographic categories—marital status, race, age,
and education. The purpose is to understand more
about what recessions mean for people. Such
information will, hopefully, give us an idea of
what needs to be done to help policymakers
address the effects of the current recession and
better prepare for future ones.

WHY LOOK AT DEMOGRAPHIC
DIFFERENCES?
The dominant explanation for the current
man-cession is that it follows from differences
in the severity of the recession across industries.
According to Hoff Sommers (2009), men “are
bearing the brunt of the current economic crisis
because they predominate in manufacturing and
construction, the hardest-hit sectors” and that
women “are a majority in recession-resistant
fields, such as education and health care.” Greg
Mankiw (2009) echoes this in his blog: “[A] large
part of the explanation is the sectoral mix of this

Kristie M. Engemann is a senior research associate and Howard J. Wall is a vice president and economist at the Federal Reserve Bank of St. Louis.

views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced,
published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts,
synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

particular downturn in economic activity, including a significant slump in residential construction.”
Job losses have indeed been steepest in the
goods-producing industries—natural resources
and mining, construction, and manufacturing—
which accounted for about half of total losses from
2007:Q4 to 2009:Q3. Job losses have not been the
rule across all industries, however, as the education and health services sector actually saw an
increase of 768,000 and the government sector
added 115,000 jobs.
Despite the current interest in the phenomenon, the large difference in the relative effects of
the recession on the employment of men and
women is not unusual. Men always bear the brunt
of job losses during recessions; and, compared
with previous recessions, men have actually been
bearing a smaller proportion during this one. During the five recessions between 1969 and 1991,
male employment fell by an average of 3.1 percent
per recession, whereas female employment actually tended to rise by 0.3 percent per recession.1
Women have a much larger presence in the workforce now than between 1969 and 1991, however,
so a more relevant comparison is to the 2001
recession. In that recession, employment peaked
in 2001:Q1 and bottomed out in 2003:Q3, with
a total loss of a little over 2.6 million jobs. Men
accounted for 78 percent of those job losses, similar to the 76 percent in the current recession.2
So, in terms of job losses, the current recession
has hit men in roughly the same proportion as
did the previous recession, but by a much smaller
proportion than in earlier recessions.
The difference in employment between the
sexes is only one of the interesting and significant
differences across demographic groups. Nonfarm
(or payroll) employment data are not broken down
by demographic categories other than sex, however. Fortunately, the Bureau of Labor Statistics
(BLS) also conducts a separate monthly “household” survey that includes several demographic
categories. Employment measures from the payroll and household surveys differ in that they
1

See Goodman, Antczak, and Freeman (1993).

Note that data splitting nonfarm employment by sex were available
only up through 2009:Q2 at the time this paper was prepared.

J A N UA RY / F E B R UA RY

2010

cover different types of employment. For example, payroll employment does not include farm
employment or self-employment. Nevertheless,
the two employment measures capture the same
broad patterns in male-female employment. In
fact, by fortunate coincidence, the household
survey indicates the same 76/24 split in the male/
female employment losses between 2007:Q4 and
2009:Q2 that appeared in the nonfarm employment
data discussed above. From this point forward,
the data we refer to come from this household
survey.
The differences in household employment
by sex, marital status, and race from 2007:Q4 to
2009:Q3 are illustrated by Figure 1. Whereas total
employment losses amounted to 4.7 percent,
male employment fell by 6.4 percent and female
employment fell by 2.9 percent. Similarly, large
differences in employment losses have occurred
according to marital status and race: Employment
of single adults fell at nearly twice the rate as it
did for married adults, and white employment
fell by only about two-thirds as much as black
employment.
Figure 2 shows employment changes by age
groups, indicating much larger-than-average
employment losses for those 16 to 19, 20 to 24,
and 35 to 44 years of age. In contrast, employment
among those 55 years and older actually rose by
4 percent. Unsurprisingly, changes in employment
across education levels also have been significant.
For example, Figure 3 shows that employment of
those without a high school diploma fell by 7.5
percent while employment for those with at least
a bachelor’s degree actually rose by 0.4 percent.
What accounts for the variation in the employment changes across these demographic groups?
The oft-cited reason for the difference between the
sexes is that it is a reflection of what has happened
to industries. As discussed below, this is not a
terribly satisfying explanation, but it does make
some sense. Analogous explanations are not likely
to fit the other demographic categories, however.
For example, perhaps single people are more
heavily concentrated in industries hit hardest by
the recession, but it is difficult to imagine why this
would be so. It is much easier to imagine instead
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 1
Employment Changes by Selected Demographic Categories (2007:Q4–2009:Q3)
Percent Change in Employment
0.0
–1.0
–2.0

Female
–3.0

–2.9

Married
–3.5

–4.0
–5.0

White
–4.4

–4.7

–6.0
–7.0

Male

Single

–6.4

–6.3

Black
–7.0

–8.0
Total

Sex

Marital Status

Race

Figure 2
Employment Changes by Age Groups (2007:Q4–2009:Q3)
Percent Change in Employment
10.0
4.0

5.0
0.0
–5.0

–3.5

–4.7

–5.5

–10.0

–8.1

–8.4

–15.0
–20.0
–20.2
–25.0
Total

Ages
16-19

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Ages
20-24

Ages
25-34

Ages
35-44

Ages
45-54

Ages
55+

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Figure 3
Employment Changes by Education Levels (2007:Q4–2009:Q3)
Percent Change in Employment
1.0

0.4

0.0
–1.0
–2.0
–3.0
–4.0
–5.0

–4.3
–4.7

–6.0
–7.0

–6.8

–8.0
Total

–7.5
No High School
Diploma

High School
Diploma Only

that single people might have lost proportionally
more jobs because the average single person is
younger and, therefore, less experienced and
less educated than the average married person.
Because of these differences, we would expect
that, within a given industry, single people would
bear disproportionate job losses.
The industry story might not even be a good
causal explanation for the differences between
sexes. As discussed by Wall (2009), because men
have tended to be affected disproportionately
across all industries, the man-cession cannot be
explained by industry mix alone but must have
some relation to demographic differences. For
example, men are less likely than women to have
finished high school, a fact that is consistent with
their relative job losses.
More generally, it is not a simple matter to
separate the role of industry from the role of demographics. For example, is the decrease in employment larger for manufacturing than for other
industries because it experienced a larger external
shock? Or was the shock the same across sectors,
but job losses in manufacturing were greater
4

J A N UA RY / F E B R UA RY

2010

Some College Bachelor’s Degree
(Including
or Higher
Associate’s Degree)

because its workers, on average, have lower education levels than do workers in other sectors?
Put another way, would there have been fewer job
losses in manufacturing if workers in the sector
had higher education levels? There are no simple
answers to these questions because there is no
proximate cause for what happened in manufacturing that was different from, say, professional
and business services. The recession experience
may have differed between the two sectors because
they experienced different external shocks; or
perhaps they experienced the same external shock,
but the demographic differences of their workforce
led to different outcomes. Most likely, some combination of the two explanations accounts for the
different employment outcomes.
The questions can be turned around to refer
to demographic groups: Are the different impacts
of the recession across demographic groups attributable to the industries in which the groups are
employed or to the differences in the groups’
characteristics? Again, the most likely explanation is some combination of the two.
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

WHAT ARE THE EMPLOYMENT
EFFECTS OF A RECESSION?
When the word “recession” is used to describe
specific periods of economic weakness, it refers
most often to the official recession dates determined by the business cycle–dating committee of
the National Bureau of Economic Research (NBER).
When weighing their decisions whether to label
a period a recession, NBER committee members
take into account a wide variety of economic indicators. As a result, NBER dates for recessions tend
to coincide most closely with periods in which
the broadest measure of economic activity, real
gross domestic product (GDP), is contracting. It
used to be that NBER recession periods coincided
with periods of falling employment. Beginning
with the 1990-91 recession, however, this link
was broken, and the economy experienced a prolonged period of job losses well after the end of
the official recession. Such a so-called jobless
recovery also occurred in the wake of the 2001
recession.
This disconnect between official recessions
and falling employment means that it is not possible to use NBER recession dates to compare the
effects of recent recessions with earlier ones. For
pre-1990 recessions, one could measure the
change in employment from the beginning to the
end of an official recession and obtain a reasonably complete picture of the recession’s employment effects. For post-1990 recessions, however,
the full effects of a recession on employment were
not realized until after the recession ended, and
at times even began before the onset of the official
recession.
Therefore, an alternative metric is needed to
determine the period during which recessions
affected employment. Keep in mind that using
this different metric means that estimates of the
effects of the current recession on the various
demographic groups will differ somewhat from
those in Figures 1 through 3. Nonetheless, the
scale of the effects and the comparisons across
categories within demographic groups are the
same with either set of numbers.
Fortunately, there is a fairly straightforward
statistical method for determining the timing of
recessions: a Markov-switching model. Briefly,
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

the model takes any data series, which in our case
is household employment, and estimates growth
rates that are typical for expansionary and recessionary phases. At the same time, the model
decides for each data point the phase that best
describes that period, taking into account the
periods immediately prior. For example, positive
employment growth that has persisted for many
periods will be called an expansionary period,
while negative growth that has persisted for many
periods will be called a recessionary period. The
tougher job is deciding on the more ambiguous
periods—such as when growth is positive for one
period following several periods of negative
growth or when a period has middling growth—
so that it is not obvious if the period should be
labeled part of a recession. We will leave it to
the model to decide these tough questions so that
there will be a consistent application across
recessions.3
Appendix A provides the estimated periods
for which household employment was in recession surrounding each of the six official recessions
since 1974. Figure A1 compares the growth rate
of household employment with the official NBER
recession dates, showing that employment growth
first dipped below zero in early 2007, months
before the start of the official recession, and
remained weak thereafter. As a consequence, the
last three quarters of 2007 are classified as recessionary, meaning that household employment
was in recession three quarters earlier than the
start of the official recession.4

TOTAL EFFECTS OF RECESSIONS
ON TOTAL EMPLOYMENT
Once the timing of official recessions is disentangled from the periods during which they are
3

See Owyang, Piger, and Wall (2008) for a technical description of
the statistical methodology and for results using aggregate payroll
employment. A quarter is designated as recessionary if the probability of recession exceeds 50 percent.

Note that the disjoint between official recessions and household
employment recessions is not as severe as might have been expected.
This is because household employment tends to recover earlier
than payroll employment, which is the measure most often used
in discussions of jobless recoveries.

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Table 1
Total Percent Effects of Recessions on Employment
Recession

Employment Change

Forgone Employment

Total Effect

1974-75

–2.0

–1.9

–3.9

1980

–1.0

–1.3

–2.2

1981-82

–1.7

–4.4

–6.0

1990-91

–1.2

–2.3

–3.5

2001

–1.2

–1.3

–2.5

2007-09

–4.6

–3.3

–7.9

Average

–1.9

–2.4

–4.3

NOTE: The recession dates and the employment data are for the household employment series produced by the BLS.

affecting employment, the total employment
changes related to the current recession can be
calculated and compared with those of earlier
recessions. The percent changes in total employment during each of the estimated recession
periods are provided in the second column of
Table 1. The most notable result in the column is
that the 4.6 percent employment loss from the
current recession dwarfs those of the other five,
which were in the 1 to 2 percent range.
Typically, the effects of a recession on employment are seen as simply the difference between
the levels of employment at the start and end of
a recessionary period, as in the second column
of Table 1. This assumes, though, that there would
have been zero employment growth even if there
had been no recession. However, a recession not
only causes a drop in employment from the prerecession level, it also prevents employment
growth that would have occurred. This “forgone”
employment is also an effect of the recession and
needs to be accounted for in an analysis of the
recession’s total effects on employment. Figure 4
provides a diagrammatic explanation of the total
costs of the recession on employment.
In the figure, the solid line is the actual path
that employment followed over time, including
a recession with falling employment. The dotted
line is the path that employment would have followed if the recession had not occurred. This is
an extremely stylized diagram that assumes that
employment growth is constant and positive dur6

J A N UA RY / F E B R UA RY

2010

ing expansionary periods and constant and negative during recessions. The direct change in the
level of employment is C-B, the difference between
the levels of employment at the end and at the
beginning of the recession. If the recession had
not occurred, the level of employment would
have continued to rise and would have reached
level A when the recession ended. Thus, the total
effect of the recession on employment is C-A,
which includes forgone employment (B-A) and
the change in employment (C-B).
The most straightforward way to account for
forgone employment is to assume that employment would have grown at some typical rate if the
recession had not occurred. We also must account
for differences in growth rates before and after the
mid-1980s, when the so-called Great Moderation
meant significantly less variability in the growth
of a wide range of economic variables. Specifically,
we assume as relevant that, during a recession,
employment would have grown at its median
growth rate for the periods 1972-84 and 1985-2009.
The third column of Table 1 shows estimates
of the employment growth that was forgone during
each of the six recessions, which for the current
recession is not particularly onerous. Although
forgone employment has been above average, it is
much smaller than it was for the 1981-82 recession, primarily because median employment
growth before 1985 was higher than after 1985.
Nonetheless, by combining forgone employment
with the employment decline, the total effect of
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 4
The Two Effects of Recession on Employment
Employment
Employment Path
Without Recession

Actual Path
of Employment

B
C

Recession
Time

the current recession is the highest among the six
recessions, with only the 1981-82 recession coming close.
In the subsequent section, a similar exercise
is performed for a variety of demographic categories. Specifically, the exercise shows the effects
of the current recession by sex and compares them
with previous recessions. It then does this, in turn,
for marital status, race, and age, with extra attention paid to the differences between men and
women for each category.

RECESSIONS ACROSS
DEMOGRAPHIC CATEGORIES
When calculating forgone employment, one
must also consider the sometimes large differences
in typical growth across demographic categories.
For reference, the different employment trends
are summarized in Appendix B, which provides
employment-to-population ratios for 1972-2009
for the demographic categories examined below.
As with total employment in the previous section,
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

we assume, as relevant, that, during a recession
employment for each demographic category would
have grown at its median growth rate for the
periods 1972-84 and 1985-2009 (see Appendix C).5

Sex
As already mentioned, men always bear the
brunt of employment losses during recessions,
and the current recession has been no different.
This is true whether one looks at payroll employment, as earlier studies have, or at household
employment, as this study does. As reported in
Table 2, male household employment has fallen
2.46 times the rate that female employment has
(–6.4 percent vs. –2.6 percent) during the current
recession. Looking at earlier recessions, it is clear
that the current one is actually in the lower half
in terms of the relative effects on men. During
the two recessions in the 1980s, male and female
5

This breakpoint will also take account of the significant decrease
in female employment growth that occurred after 1990 as the rapid
increases in female labor-force participation wound down.

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Table 2
Percent Effects of Recessions on Employment
Recession

Men

Women

Men/Women

1974-75

–2.8

–0.9

3.10

1980

–1.7

0.0

–58.12

1981-82

–3.3

0.5

–6.76

1990-91

–2.0

–0.3

6.90

2001

–1.2

–1.1

1.13

2007-09

–6.4

–2.6

2.46

Average

–2.9

–0.7

3.99

NOTE: The recession dates and the employment data are for the household employment series produced by the BLS.

employment moved in opposite directions, while
during the 1990-91 recession, male employment
fell nearly seven times the rate that female employment did. The 1974-75 recession was somewhat
comparable to the current recession in the relative
employment loss for men, but the 2001 recession
saw male employment fall only slightly more
than female employment.
The story of the current recession changes a
great deal when forgone employment is considered. As reported in Figure 5, male forgone employment has been only 62 percent that of women.
This is because employment growth for women
has tended to be higher than that for men during
the entire sample period—meaning that, for every
quarter of recession, more female than male
employment is forgone. Adding the two effects
together reveals that men as a whole have still
borne a much larger effect of the recession, but
it is 1.33 times the effect for women rather than
2.46 times, as suggested by the employment
changes alone.
Now that we know the total employment
effects of the current recession, how does it
compare with earlier ones? Has it been the Great
Man-Cession? Figure 6 shows the total effects of
the six recessions since 1974 on male and female
employment, along with the relative effect for men
and women. For both sexes, this has been the most
costly recession in terms of employment. Male
employment is 8.9 percent lower than it would
have been without a recession, which is rivaled
8

J A N UA RY / F E B R UA RY

2010

only by the total effect of the 1981-82 recession.
For women, the current recession is somewhat
similar to the 1981-82 recession, when female
employment actually rose (recall Table 2). However, because female employment growth was
much higher before than after 1985, they experienced a comparatively higher percentage of forgone employment during the 1981-82 recession.6
The male-to-female ratio for the current recession, 1.33, is surpassed only by the 1980 recession
and is much higher than for all other recessions. So,
even though it’s not quite the Great Man-Cession,
it’s still been relatively more severe for men than
is usual. Interestingly, the estimates also indicate
that the total effects of the 2001 recession were
actually higher for women than for men. Recall
from Table 2 that employment losses for men and
women did not differ by much, so the higher forgone employment for women means a higher
total effect.

Marital Status
Over the course of the current recession, the
employment of married people has fallen at 76
percent of the rate that employment of single
people has fallen (Figure 7). Married employment
fell by 4 percent while single employment fell
by 5.3 percent. Because single employment has
6

A recent paper by DiCecio et al. (2008) reviews changes in laborforce participation, separating trends from the changes due to
economic conditions.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 5
Total Effects of 2007-09 Recession: Men versus Women
Percent Change in Employment
4.0
2.46
2.0

1.33

0.62

0.0
–2.0
–2.5

–2.6
–4.0

–4.1
–6.0
–6.4

–6.7

–8.0
–10.0

–8.9

Employment Change

Forgone Employment
Men/Women

Women

Men

Total Effect

Figure 6
Total Effects of Recessions: Men versus Women
Percent Change in Employment
4.0
2.0

1.61

1.23

1.18

1.33

0.82

0.0
–2.0

–1.6

–2.3

–2.7
–4.0

–3.4

–3.8

–4.2

–3.2

–2.8

–5.4

–6.0
–6.6

–6.7

–8.0
–8.9

–10.0

1974-75

1980

1981-82
Men

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Women

1990-91

2001

2007-09

Men/Women

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Figure 7
Total Effects of 2007-09 Recession: Married versus Single
Percent Change in Employment
2.0
0.76

0.67

0.57

0.0
–2.0
–2.4
–4.0

–4.0

–4.3
–5.3

–6.0

–6.5
–8.0
–10.0

–9.6

–12.0

Employment Change

Forgone Employment

Married/Single

Single

Married

Total Effect

Figure 8
Total Effects of Recessions: Married versus Single
Percent Change in Employment
4.0
2.0

1.46

0.78

0.73

0.67

0.51

0.0
–2.0
–2.4

–1.7

–1.7
–2.9

–4.0 –3.4
–4.4
–6.0

–3.4
–4.3

–5.1
–6.5

–7.0

–8.0
–10.0

–9.6

–12.0

1974-75

1980
Married

J A N UA RY / F E B R UA RY

2010

1981-82
Single

1990-91

2001

2007-09

Married/Single

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 9
Total Effects of 2007-09 Recession: Marital Status and Men versus Women
Percent Change in Employment
3.0

1.82

1.61

0.0
–3.0
–4.4

–6.0

–6.5

–7.1
–9.0
–12.0

–11.8

–15.0

Married
Men

Single
Women

tended to grow much faster than married employment since 1985, the forgone employment for
singles during the recession has been much larger.
Adding up the two effects, the total effect of the
recession for married people has been 67 percent
as large as the total effect for single people.
The relative effects of this recession on married
and single workers are typical of those for the previous five recessions (Figure 8). Single people have
almost always borne a greater total effect, although,
because the median employment growth for singles is lower than it was before 1985, the forgone
employment for singles was relatively less important for the past two recessions. With the exception of the 1980 recession, married people bore
between 50 and 80 percent of the total effect that
single people did. For the 1980 recession, employment for singles, particularly single women, was
higher at the end of the recession than at the beginning. As we have seen, that recession was really
one that hit men the hardest relative to other
categories.
An interesting difference is found in the comparison of men and women in the married and
single categories (Figure 9). In the current recesF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Men/Women

sion, married men and women saw smaller job
losses than did their single counterparts, meaning
that married women saw the smallest reduction
in employment of the four groups. In part, this
can be explained by what has been called the
“added-worker effect.”7 According to this effect,
some married women enter the labor force during
recessions following their husbands’ job losses.
The added-worker effect can account for the fact
that the number of women in the labor force, either
employed or looking for employment, has actually
risen during the current recession, whereas the
male labor force has fallen.
Another explanation for the difference
between married and single people is that married
people are more likely to have children to support
and are, therefore, more likely to take a new job
at lower pay after they lose a job. Also, many of
the differences for marital status are reflections
of other demographic differences that make them
more likely to be affected by a recession: Compared with married people, single people tend to
7

See, for example, Stephens (2002). DeRiviere (2008) has estimated
the size of a related effect called the “pin-money” hypothesis.

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

be younger (i.e., have less work experience) and
have lower education levels.

White and Black
As with all demographic groups, the differences across racial categories are intertwined
with differences across other categories as well.
For example, black men, for whom average education is lower than for black women or whites,
saw the largest decrease in employment. Black
women, on the other hand, have seen the most
forgone employment of any of these sex-race categories. Underlying these differences is the longterm trend of women, especially black women,
becoming more likely to be employed (see
Appendices B and C).
The white-black employment effects of the
current recession are illustrated by Figure 10,
which indicates that white employment has fallen
at 58 percent of the rate that black employment
has (–4.4 percent vs. –7.5 percent). Because black
employment has tended to grow faster than white
employment, white forgone employment has been
only 79 percent of that for blacks.
Figure 11 shows the relative total effects of
the past six recessions on black and white employment. Recent recessions have actually tended to
affect black employment relatively more than they
used to, even as blacks have become more successful in the labor market. For the past three recessions, the ratio of white-to-black total effects has
been between 0.65 and 0.74, after it had been
above 1 for the two recessions of the 1980s, indicating that white employment had been more
adversely affected. In part, this change over time
is because the gap between white and black
employment growth has reversed.
It is worth breaking out the two employment
effects (employment change and forgone employment) for all six recessions to see how the whiteto-black ratios have changed over time (Figure 12).
Before 1985, white employment grew at a median
rate of 2.5 percent per year, whereas the analogous
number for black employment was 2.1 percent.
Thus, there was more forgone white employment
for each quarter of recession. Since 1985, however,
median white employment growth has fallen by
12

J A N UA RY / F E B R UA RY

2010

half, whereas median black employment growth
has fallen by only one quarter. The ratio of direct
employment changes has also fallen over time,
meaning that the direct employment change
used to be relatively smaller for blacks than it
has become. (The 1974-75 recession, however,
hit black employment much harder than white
employment.) As a consequence, blacks tend to
bear a relatively larger burden during recessions
now than they used to.
In a sense, the recent success in labor markets
has made the total effects of recessions on blacks
greater than in the past. As already discussed,
black employment has been growing faster than
white employment, so each quarter of recession
means a greater loss of employment for blacks.
Also, because blacks now have higher participation in the labor market than in the past and their
education and work experience still lags those of
whites, more blacks are vulnerable to the effects
of recession than had been the case earlier.
We have alluded to white/black differences
in the relative effects of recessions on men and
women. This is illustrated for the current recession by Figure 13, which shows that the total effect
on white men and women is smaller than that on
black men and women, respectively. Also, the total
effect on white men is 59 percent greater than that
on white women, while the total effect on black
men is 18 percent higher than that on black women.
This difference is because black women have
seen a much larger decline in employment than
have white women (–5.3 percent vs. –2.3 percent)
while also seeing more forgone employment
because black women’s median growth rate is
nearly twice that of white women. Just as described
for black employment overall, this story is really
a side effect of the labor market success of black
women, who have seen rapid employment growth
relative to black men and white women.

White and Other
The race category “Other” captures all who
are neither white nor black and has become an
increasingly important category in the labor market: In 1972, the Other category accounted for
1.2 percent of total employment, but by 2009:Q3
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

TOTAL EFFECTS OF 2007-09 RECESSION: HISPANICS
Over the course of the current recession, Hispanic employment has not fallen by as much as
overall employment (–3.4 percent vs. –4.6 percent), and, for both men and women, Hispanic
employment has fallen by at least 1 percentage point less than has overall employment. On the
other hand, because Hispanic employment has tended to grow at almost twice the rate of overall
employment, these simple employment changes do not capture the whole story. Specifically,
whereas overall forgone employment has been 3.3 percent, Hispanic forgone employment has
been 6.1 percent, with similar numbers for men and women. In total, the recession has hit Hispanic
employment relatively hard, resulting in employment that is 9.5 percent lower than it would have
been if the recession had not occurred. As with overall employment, the effects of the recession
have been more severe for Hispanic men, who have borne about a 40 percent larger total effect
than have Hispanic women. For Hispanics, however, the difference between men and women
comes from the employment change rather than forgone employment.1

Figure
Total Effects of 2007-09 Recession: Hispanics
Percent Change in Employment
6.0
3.83

4.0
2.0

1.40

0.90

0.0
–2.0
–4.0
–6.0

–1.3
–3.4
–4.9
–6.1

–8.0

–5.7

–6.3
–7.6

–10.0

–9.5
–10.6

–12.0

Employment Change
Total

Forgone Employment
Men

Women

Total Effect

Men/Women

Note that it is not possible to do simple comparisons of the Hispanic experience across recessions because the data have been subject
to extremely large spikes following new estimates of the Hispanic population.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Figure 10
Total Effects of 2007-09 Recession: White versus Black
Percent Change in Employment
2.0
0.58

0.79

0.65

0.0
–2.0
–3.0

–4.0

–3.8

–4.4
–6.0
–8.0

–7.4

–7.5

–10.0
–11.3

–12.0

Employment Change

Forgone Employment
White/Black

Black

White

Total Effect

Figure 11
Total Effects of Recessions: White versus Black
Percent Change in Employment
2.0

1.29

0.57

1.09

0.74

0.65

0.0
–2.0
–2.3

–1.8

–2.4

–4.0 –3.7

–3.2

–3.5
–4.7

–6.0
–6.3

–6.5

–5.8
–7.4

–8.0
–10.0

–11.3

–12.0

1974-75

1980

1981-82
White

J A N UA RY / F E B R UA RY

2010

Black

1990-91

2001

2007-09

White/Black

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 12
The Two Employment Effects of Recessions: White versus Black
Relative Percent Change in Employment
1.60
1.39

1.40
1.21

1.20

1.21

1.00

0.89
0.79

0.80

0.79
0.58

0.60
0.40

0.79

0.69

0.68

0.37

0.20
0.00

1974-75

1980

1981-82

White/Black Employment Changes

1990-91

2001

2007-09

White/Black Forgone Employment

Figure 13
Total Effects of 2007-09 Recession: White and Black and Men versus Women
Percent Change in Employment
3.0

1.59

1.18

0.0
–3.0
–6.0
–9.0

–5.4
–8.5
–10.8

–12.0
–12.7
–15.0

White

Black
Men

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Women

Men/Women

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Figure 14
Total Effects of 2007-09 Recession: White versus Other
Percent Change in Employment
4.0
1.69

2.0

0.68

0.37

0.0
–2.0
–2.6

–4.0

–3.0

–4.4
–6.0
–8.0

–7.4
–8.3

–10.0
–10.9

–12.0

Employment Change

Forgone Employment
White/Other

Other

White

Total Effect

Figure 15
Total Effects of Recessions: White versus Other
Percent Change in Employment
6.0
3.60

4.0
2.11

2.0

2.01

1.33

0.69

0.68

0.0
–0.7

–2.0

–1.8

–1.7

–2.3

–4.0 –3.7

–2.4
–3.5

–3.5
–4.7

–6.0
–6.3
–8.0

–7.4

–10.0
–10.9

–12.0

1974-75

1980

1981-82
White

J A N UA RY / F E B R UA RY

2010

Other

1990-91

2001

2007-09

White/Other

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 16
Total Effects of 2007-09 Recession: White and Other and Men versus Women
Percent Change in Employment
3.0

1.59

1.09

0.0
–3.0
–6.0
–9.0

–5.4
–8.5

–12.0

–11.6
–12.6

–15.0

White

Other
Men

Women

it had risen to 7 percent.8 Over that period, the
composition of the category changed a great deal,
reflecting large influxes of immigrants from China,
India, and other Asian countries. In 2007, the
average education level of the group was much
higher than for the population as a whole, which is
reflected in the group’s employment performance
during the recession.
As depicted by Figure 14, the Other group
has seen a drop in employment about half that
of whites. On the other hand, because median
employment growth for the group is nearly three
times that of whites, the group’s forgone employment during the current recession has been almost
triple that of whites. In total then, employment
for the group is estimated to be 10.9 percent lower
than if the recession had not occurred. This effect
is of roughly the same magnitude as for blacks,
but for very different reasons. The bulk of the
effect for blacks was from a drop in employment,
whereas for people in the Other category the bulk
8

At the start of 1972, the white and black shares of employment
were 89.4 percent and 9.4 percent, respectively. By 2009:Q3 the
shares were 82.3 percent and 10.7 percent, respectively.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Men/Women

of the effect was from forgone employment.
It has only been in the past two recessions
that the Other category experienced a larger total
effect than did whites (Figure 15). During the four
earlier recessions, employment of this group rose
by between 3.3 percent and 9.3 percent, whereas
negative employment changes are the current
norm. So, despite large forgone employment during recessions, the total effects of recessions on
the group used to be relatively small.
Unlike the other two race categories, men and
women in the Other category have seen similar
total effects from the current recession (Figure 16).
Just as with the total effects over time, this equality of the sexes is a recent phenomenon. For example, for the earliest three recessions in our sample
period, men saw much larger negative total effects
during recessions.

Age Groups
The different effects of the current recession
are stark when they are broken down by age
groups. Teen employment has fallen by 23.8 percent during the recession, whereas employment
of those 55 years and older has risen by 7.4 percent
J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Figure 17
Total Effects of 2007-09 Recession: Age Groups
Percent Change in Employment
10.0

7.4

5.0
0.1

0.0

1.4

–2.8

–5.0

1.4

0.4
–2.9

–5.0

–10.0

–10.6

–7.6

–9.5

–4.6

–6.1
–9.2

–12.4

–15.0

–10.4

–20.0
–25.0

–23.7

–23.8

–30.0

Employment Change
16-19

Forgone Employment
20-24

25-34

(Figure 17). The 20- to 24-year and 35- to 44-year
age groups also have experienced significant
employment declines, while the employment
drop for the 45- to 54-year age group has been
relatively minor.
One reason the 55-plus age group has seen
increased employment during the current recession is the effect of the recession on the decision
to retire. A dominant feature of the recession has
been a significant collapse of stock prices and the
resulting devaluation of many people’s retirement
savings. So, instead of retiring, large numbers of
this age group have elected to remain employed,
thereby suppressing the normal effect that the
recession would have had. In fact, employment
of this age group was higher than it would have
been without a recession: It grew by 7.4 percent
during the recession, but, without a recession, it
would have grown by 6.1 percent. This leaves a
total effect of an increase in employment of 1.4
percent. From these numbers, it is not possible to
determine the number of people who were pushed
into employment because of the collapse of retirement savings. The push effect is something greater
than 1.4 percent because that number is the push
18

J A N UA RY / F E B R UA RY

2010

35-44

Total Effect
45-54

55+

effect minus the decrease in the demand for these
workers that resulted from the recession.
At the other end of the age spectrum, the total
effect on employment for the 16- to 19-year group
was the same as its employment change. Forgone
employment was almost zero because effectively
there has been no trend employment growth for
this age group. The share of the population of this
group that is employed has been falling steadily
over time, even when the economy was not in
recession (see Appendix B).
As the group with the lowest average education and the least experience, it is not surprising
that teenagers have borne a much bigger-thanaverage burden of the recession. We need to be
careful, however, before attributing the entire
change in employment to the recession. The federal minimum wage rose in the middle of the
recession in 2008 and would have had its largest
negative employment impacts on the two youngest
age groups. A majority of those working at or below
the minimum wage in 2008 were younger than
25 years, and almost half of those were teenagers.
The age breakdown also provides interesting
insights into the nature of the relatively large effect
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 18
Total Effects of 2007-09 Recession: Age Groups and Men versus Women
Percent Change in Employment
5.0
1.78
1.41

2.46

1.16

1.07

1.3 1.7 0.80

0.0
–2.5

–5.0
–6.6

–6.2

–10.0
–11.7

–11.9

–15.0

–10.2 –9.6

–13.9
–18.5

–20.0
–25.0
–26.0
–30.0

16-19

20-24

25-34
Men

Women

that the recession has had on men. The three oldest groups saw relatively similar effects on men
and women (Figure 18). For the 25- to 34-year age
group, on the other hand, the total effect on men
has been 2.46 times the total effect on women.
Therefore, any explanation of the man-cession
must include a discussion of the role of age.

THE ROLE OF EDUCATIONAL
ATTAINMENT
The final demographic category is educational
attainment, which, because of its importance as
a causal factor in the results across all other categories, warrants its own section. Figure 19 breaks
down the effect of the current recession according to educational attainment. Keep in mind that
the employment data by educational attainment
include only those aged 25 and older. This gives
a better idea of the employment effects once people achieve their highest education level.
The employment change during the recession
has been greatest for those without a high school
diploma, followed by those who have completed
high school but have not attended any college.
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

35-44

45-54

55+

Men/Women

Employment for those with some college fell
slightly, while employment for those with a bachelor’s degree actually rose during the recession.
Because trends across these groups differ a great
deal, so do the estimates of their forgone employment. Specifically, employment for those without
a high school diploma has been trending down
for many years, so part of the decrease in employment during the recession would have occurred
anyway. Correcting for this, the total effect of the
recession on the employment of those without a
high school diploma has been a drop of 13.2 percent. Above-average effects have also been experienced by those with a high school diploma but
no college. The total effect on those with at least
a bachelor’s degree has also been higher than average because forgone employment for this group
was the highest among the four categories.
Using Figure 20, it is possible to map the
results for educational attainment onto the results
across other demographic groups. Specifically,
recall that the effect of the current recession on
men has been 1.33 times its effect on women.
Figure 20 shows that men are less likely to have
completed high school, whereas women are much
more likely to have some college (particularly an
J A N UA RY / F E B R UA RY

2010

Engemann and Wall

Figure 19
Total Effects of 2007-09 Recession: Education Level
Percent Change in Employment
4.0
2.0
2.0
0.8
0.0

No High School Diploma
High School Diploma

0.0

Some College (including
Associate’s Degree)

–0.9

–2.0

Bachelor’s Degree or Higher

–4.0

–4.9 –5.1

–6.0
–8.0

–7.2

–7.4

–10.0
–12.0
–14.0

–13.2

–14.0

–16.0
Employment Change

Forgone Employment

associate’s degree in the nearly recession-proof
nursing profession). Recall also that the effect of
the recession on single people has been much
greater than it has been on married people. From
Figure 20, we can see that single people 25 years
and older are much more likely to not have a high
school diploma or to have only a high school
diploma. They are also much less likely to have
a bachelor’s degree.
Educational attainment across racial categories
maps just as easily onto the employment effects
described in previous sections. Relative to white
employment, the effect of the current recession
on black employment is larger primarily because
of larger direct decreases in employment. In contrast, the effect on the employment of those in
the Other category is also larger than for whites,
but this is primarily because of higher forgone

J A N UA RY / F E B R UA RY

2010

Total Effect

employment. Relative to whites, blacks are much
less likely to have a bachelor’s degree and more
likely to have a high school diploma or less
(Figure 21). For those in the Other category,
those 25 years and older are much more likely to
have a bachelor’s degree or higher.
It is not possible to conclude from the analysis
here that educational attainment is the primary
determinant of the extent to which a recession
affects employment across demographic groups.
Other factors—such as the industries in which
people tend to be employed, job experience, cultural differences, etc.—clearly matter, also.
Nevertheless, any discussion of the effects of a
recession across demographic groups should have
educational attainment as one of the first, if not
the first, factor that is considered.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

Figure 20
Educational Attainment, 25 Years and Older: 2007 (Difference from the Total: Sex and
Marital Status)
Percentage Point Difference
3

No High School Diploma

High School Diploma

Some College (including
Associate’s Degree)

Bachelor’s Degree or Higher

–1
–2
–3
–4
–5
–6

Male

Female

Married

Single

Figure 21
Educational Attainment, 25 Years and Older: 2007 (Difference from the Total: Race)
Percentage Point Difference
15

No High School Diploma
High School Diploma

Some College (including
Associate’s Degree)

Bachelor’s Degree or Higher

–5

–10

–15

White

Black

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Other

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

REFERENCES
DeRiviere, Linda. “Have We Come a Long Way? Using the Survey of Labour and Income Dynamics to Revisit
the ‘Pin Money’ Theory.” Journal of Socio-Economics, December 2008, 37(6), pp. 2340-67.
DiCecio, Riccardo; Engemann, Kristie M.; Owyang, Michael T. and Wheeler, Christopher H. “Changing Trends
in the Labor Force: A Survey.” Federal Reserve Bank of St. Louis Review, January/February 2008, 90(1),
pp. 47-62; http://research.stlouisfed.org/publications/review/08/01/DiCecio.pdf.
Goodman, William; Antczak, Stephen and Freeman, Laura. “Women and Jobs in Recessions: 1969-92.”
Monthly Labor Review, July 1993, 116(7), pp. 26-35.
Hoff Sommers, Christina. “No Country for Burly Men.” The Weekly Standard, June 29, 2009, 14(39);
www.weeklystandard.com/Content/Public/Articles/000/000/016/659dkrod.asp.
Mankiw, Greg. “This Recession’s Gender Gap.” June 7, 2009; http://gregmankiw.blogspot.com.
Owyang, Michael T.; Piger, Jeremy and Wall, Howard J. “A State-Level Analysis of the Great Moderation.”
Regional Science and Urban Economics, November 2008, 38(6), pp. 578-89.
Stephens, Melvin Jr. “Worker Displacement and the Added Worker Effect.” Journal of Labor Economics, July
2002, 20(3), pp. 504-37.
Wall, Howard J. “The ‘Man-Cession’ of 2008-09: It’s Big, but It’s Not Great.” Federal Reserve Bank of St. Louis
The Regional Economist, October 2009, pp. 4-9;
http://research.stlouisfed.org/publications/regional/09/10/mancession.pdf.

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Engemann and Wall

APPENDIX A
Table A1
The Timing of Official Recessions and Employment Recessions
Recession

NBER Dates

Household Employment Dates

1974-75

1974:Q1–1975:Q1

1974:Q4–1975:Q2

1980

1980:Q2–1980:Q3

1981-82

1981:Q4–1982:Q4

1981:Q3–1983:Q1

1990-91

1990:Q4–1991:Q1

1990:Q2–1991:Q4

2001

2001:Q2–2001:Q4

2001:Q2–2002:Q1

2008:Q1–?

2007:Q2–?

2007-09

NOTE: The official recession dates are determined by the NBER. The dates for household employment recessions are estimated with
a Markov-switching model.

Figure A1
Household Employment Growth Rate (1972-2009) Quarterly, Seasonally Adjusted
7.5

7.5

5.0

2.5

0.0

–2.5

–5.0

–7.5
1972

–7.5
1976

1980

1984

1988

1992

1996

2000

2004

2008

NOTE: Gray bars indicate official NBER recessions.
SOURCE: BLS/Haver Analytics.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

100
90
80
70
60
50
40
30
20
10
0

Married

White

J A N UA RY / F E B R UA RY

2010

19
7
19 2
74
19
7
19 6
78
19
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08

19
7
19 2
74
19
7
19 6
7
19 8
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08
19
7
19 2
74
19
7
19 6
78
19
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08

100
90
80
70
60
50
40
30
20
10
0

19
7
19 2
74
19
7
19 6
78
19
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08

19
7
19 2
74
19
7
19 6
7
19 8
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08

Engemann and Wall

APPENDIX B

Figure B1

Employment-to-Population Ratio for the Sexes (1972-2009) Overall, by Marital Status, Race,
and Age
Total Across Groups

Men
Women
Total

100
90
80
70
60
50
40
30
20
10
0

Single

Black

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

100
90
80
70
60
50
40
30
20
10
0

Ages 25-34

Ages 45-54

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

19
7
19 2
74
19
7
19 6
78
19
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08

Ages 16-19

19
7
19 2
74
19
7
19 6
78
19
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08

100
90
80
70
60
50
40
30
20
10
0

19
7
19 2
74
19
7
19 6
78
19
8
19 0
8
19 2
8
19 4
86
19
8
19 8
9
19 0
9
19 2
94
19
9
19 6
9
20 8
0
20 0
02
20
0
20 4
06
20
08

Engemann and Wall

APPENDIX B
Figure B1, cont’d

Employment-to-Population Ratio for the Sexes (1972-2009) Overall, by Marital Status, Race,
and Age
100
90
80
70
60
50
40
30
20
10
0

100
90
80
70
60
50
40
30
20
10
0

Ages 20-24

Ages 35-44

Ages 55+

J A N UA RY / F E B R UA RY

2010

Engemann and Wall

APPENDIX C
Table C1
Median Employment Growth Rates (Annual)
A. Aggregate and by marital status
Aggregate
1972-84

Married

1985-2009

1972-84

Single

1985-2009

1972-84

1985-2009

Total

2.5

1.3

1.6

1.0

3.7

1.7

Men

1.9

1.0

0.7

4.7

1.6

Women

3.4

1.7

3.4

0.9

3.7

1.4

B. By race
White

Total

Black

Other

1972-84

1985-2009

1972-84

1985-2009

1972-84

1985-2009

2.5

1.2

2.1

1.5

8.0

3.3

Men

1.7

1.0

2.0

1.0

9.0

3.3

Women

3.1

1.2

3.0

2.2

4.6

4.4

C. By age group
16-19 Years

20-24 Years

25-34 Years

35-44 Years

45-54 Years

55+ Years

1972-84 1985-2009 1972-84 1985-2009 1972-84 1985-2009 1972-84 1985-2009 1972-84 1985-2009 1972-84 1985-2009

Total

–0.4

0.0

2.4

–0.6

5.3

–0.2

3.6

1.2

0.1

3.1

–0.1

2.4

Men

–0.4

–0.5

2.2

–0.8

4.1

–0.4

2.5

1.1

–0.4

2.5

–0.6

2.0

Women

–0.2

–0.8

2.6

–0.3

5.9

0.1

5.0

1.7

0.7

3.2

0.4

2.8

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Community Colleges and Economic Mobility
Natalia A. Kolesnikova
This paper examines the role of community colleges in the U.S. higher education system and their
advantages and shortcomings. In particular, it discusses the population of community college
students and economic returns to community college education for various demographic groups.
It offers new evidence on the returns to an associate’s degree. Furthermore, the paper uses data
from the National Survey of College Graduates to compare educational objectives, progress, and
labor market outcomes of individuals who start their postsecondary education at community
colleges with those who start at four-year colleges. Particular attention is paid to the Federal
Reserve’s Eighth District, the geographic area served by the Federal Reserve Bank of St. Louis.
(JEL I20, I21, J30)
Federal Reserve Bank of St. Louis Review, January/February 2010, 92(1), 27-53.

oliet Junior College (Joliet, Illinois), the
oldest community college in the nation,
was founded in 1901. Since then, community colleges have become increasingly
important for the U.S. education and training system. Today, 11.5 million students (6.5 million of whom are studying for college credits) are
enrolled in almost 1,200 community colleges,
according to the American Association of
Community Colleges. Community college students constitute a remarkable 46 percent of all
U.S. undergraduates.
The term “junior college” originally referred
to any two-year, postsecondary school. Over the
last few decades, the term “community college”
became more popular to describe public two-year
institutions as it better conveys the mission of
these colleges to serve their local communities.
This distinction was not prevalent before the
1980s and the two terms are still often used interchangeably. However, in 1992 the American

Association of Junior Colleges did change its name
to the American Association of Community
Colleges.
The original goal of two-year colleges was to
prepare students, through an associate’s degree
(AD) program, to transfer to a four-year college.
Over time, the purpose evolved to include
workforce training programs, schooling toward
certification in areas such as nursing and other
professions, and adult continuing education
classes. A more recent development is that some
community colleges now offer bachelor’s degrees
in a number of fields.
However, there are big differences across
states in how the community college system is
used. Rouse (1998) found evidence suggesting
that states tend to focus their resources on either
a community college or a four-year college system.
California has the largest network of the former;
66 percent of the state’s current undergraduates
attend community colleges. In contrast, only 16

Natalia A. Kolesnikova is an economist at the Federal Reserve Bank in St. Louis. Luke Shimek and Yang Liu provided research assistance.
Portions of this paper previously appeared in The Regional Economist (Kolesnikova and Shimek, 2008, and Kolesnikova, 2009a) and as a
Federal Reserve Bank of St. Louis Community Development Research Report (Kolesnikova, 2009b).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 1
College Enrollment Statistics for the Federal Reserve Eighth District
Region/State

Enrollment in community college
(Fall 2005)

Percent of all undergraduates
(Fall 2005)

United States

6,184,000

47,771

Eighth District states
Arkansas
Illinois

352,824

Indiana

59,969

Kentucky

84,669

Mississippi

66,298

Missouri

86,742

Tennessee

74,829

SOURCE: U.S. Department of Education, National Center for Education Statistics.

percent of undergraduates in Nevada and
Vermont are enrolled in community colleges.1
Among the states within the Federal Reserve
System’s Eighth District (which consists of all
of Arkansas and parts of Missouri, Mississippi,
Illinois, Indiana, Tennessee, and Kentucky)
Illinois and Mississippi have the largest proportion
of undergraduates—about half—in community
colleges. Indiana has the lowest percentage—19
percent. Table 1 summarizes enrollment statistics
for the Eighth District states.
For many individuals, community colleges
represent a unique opportunity to receive a postsecondary education and improve their economic
status. Community colleges thus serve as a path
to upward economic mobility for a large part of
the population. Given the significant role community colleges in U.S. higher education, it is important to have as much information as possible about
community college students, their goals, educational choices, and outcomes. This paper concentrates on several of these topics and attempts to
present a comprehensive picture of community
college education. In particular, it addresses the
following questions:
1

These are the 2005 state-level statistics from the National Center
for Education (U.S. Department of Education), the latest information available when this paper was written.

J A N UA RY / F E B R UA RY

2010

• What are the advantages of community
colleges compared with traditional fouryear colleges?
• Do students attending community colleges
differ from students at traditional four-year
colleges?
• What are the economic returns of attending
a community college?
• What are the intentions of community
college students with regard to their educational objectives?
• Does starting postsecondary education at a
community college affect a person’s chances
of obtaining a bachelor’s degree and postgraduate education?
• Do students who attended a community
college and received an AD before obtaining
a bachelor’s degree have different educational and labor market outcomes than those
who did not have an AD before obtaining
a bachelor’s degree?
The paper reviews the existing literature on
community college education. In addition, it
offers new evidence on the returns to attaining
an AD and uses the National Survey of College
Graduates (NSCG) to carefully analyze the differences in a variety of educational and economic
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Table 2
Comparisons of College Tuition and Fees for the Federal Reserve Eighth District
Average tuition and required fees (2006-07)
Region/State

Four-year
public college (in-state) ($)

Four-year
private college ($)

Two-year
community college ($)

United States

5,685

20,492

2,017

Arkansas

4,937

13,396

1,890

Illinois

8,038

20,181

2,252

Indiana

6,284

22,060

2,713

Kentucky

5,821

14,739

2,633

Mississippi

4,457

12,300

1,709

Eighth District states

Missouri

6,320

16,539

2,284

Tennessee

5,009

17,576

2,474

SOURCE: U.S. Department of Education, National Center for Education Statistics.

outcomes between individuals who started their
postsecondary education at community colleges
and those who started at four-year institutions.2

ADVANTAGES OF COMMUNITY
COLLEGES
Compared with a traditional four-year college,
a community college has several important advantages for students. To begin, the open admission
policy makes it easier for students to enroll regardless of their prior academic record.
Attending community colleges costs less
because of lower tuition and other fees than those
at four-year colleges. Community college students
on average paid $2,017 in tuition and fees for the
2006-07 academic year, which is less than half
the amount for students in public four-year universities ($5,685) and only about one-tenth of
the tuition and fees for students in private fouryear universities ($20,492), according to the U.S.
Department of Education.
Table 2 presents a comparison of tuition costs
and other fees for the Federal Reserve’s Eighth
2

The latest available data are used throughout the paper, which
means that time periods may vary between different sections of
the paper.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

District. Mississippi has the lowest tuition among
the states of the Eighth District. Attending a fouryear private college in Mississippi costs $12,300
per year on average. Attending a four-year public
college costs significantly less: $4,457 per year.
Community college tuition in Mississippi is
$1,709 per year. Even the state with the highest
community college tuition in the District, Indiana,
charges only $2,713 per year. In comparison,
tuition at a private four-year college in Indiana
costs on average $22,060 per year. Illinois has the
highest tuition for four-year public universities
in the District ($8,038).
In addition, most community college students
live at home, thus saving the added room and
board expenses incurred by students at other
institutions. Finally, community colleges offer a
more flexible curriculum, and their schedules
include evening and weekend classes, which
gives students an opportunity to attend college
while working full-time.

Community College Students
The population of community college students
is diverse and differs from the typical population
at four-year colleges. Community college populations have 60 percent white, 15 percent black,
J A N UA RY / F E B R UA RY

2010

Kolesnikova

and 14 percent Hispanic students.3 Forty-one
percent of community college students are males.
In comparison, students attending four-year colleges are more likely to be white (70 percent) and
male (45 percent).
Because of the flexibility they offer and the
relatively low monetary and time costs of attending, community colleges have more so-called
nontraditional students than four-year colleges.
Community college students are more likely to
be older: 35 percent are 30 years old or older compared with 16 percent in four-year colleges. The
average community college student is 28 years
old, with a median age of 24. The corresponding
ages for students in four-year colleges are 24 and
21 years.
Only 31 percent of community college students are enrolled full-time, in part because students attending community colleges are more
likely to also be working. In contrast, 63 percent
of students at four-year colleges are enrolled fulltime. Only 21.4 percent of all community college
students do not work, compared with 30.5 percent
at four-year colleges. Furthermore, 40.8 percent
of community college students work full-time,
compared with 22.8 percent of their four-year
college counterparts.
More students in community colleges are firstgeneration college students than are students
attending four-year colleges. More than 40 percent
of the former have parents with only a high school
education or less. In contrast, only 27 percent of
four-year college students have parents with a
high school education or less.
Not surprisingly, most community college
students attend an institution close to their home.
They live on average 40 miles away from the college they attend. In comparison, students at fouryear institutions attend colleges on average 230
miles away from their home. More than 95 percent
of community college students attend colleges in
their home states compared with 83 percent of
students at four-year colleges.
3

Unless noted otherwise, the data in this section are from the Center
for Education Statistics, U.S. Department of Education, 2003-04
as presented in Horn and Nevill (2006).

J A N UA RY / F E B R UA RY

2010

LABOR MARKET RETURNS
What is the economic payoff to attending
community college? The answer to this question
is rather complicated, partly because of the lack
of available data. Until 1990, the U.S. Census
Bureau recorded only the number of years of education, making it impossible to identify individuals attending community college specifically.
In the 1990 and 2000 U.S. censuses, the highest
educational attainment was recorded instead of
years of education. This makes it possible to focus
on individuals with a completed AD. Still, this
information does not make it possible to identify
an institution students attended if they did not
complete a degree.
Several available studies use different longitudinal survey data instead. Most of the surveys
record data on various characteristics of respondents, starting with their teenage years and following them through the years.4 One limitation
of these studies is that, given the timeline of surveys, they include only students who enrolled in
community college soon after graduating from
high school.
Most studies found that students who
attended community college, but did not complete
a degree, earn 9 to 13 percent more than those
with only a high school diploma. The estimation
technique usually attempts to control for differences in academic preparation between the two
groups as measured by test scores and class rank.
Furthermore, researchers found an increase in
annual earnings of 5 percent to 8 percent associated with each year of education at a community
college. This finding is particularly interesting
because it is similar to the return to a year of
schooling in a four-year college.
Jacobson, LaLonde, and Sullivan (2005)
looked at a very different group—older, hightenure, displaced workers. Most retraining efforts
for this group take place at community colleges.
These researchers found that one year of community college schooling increases the long-term
earnings of displaced workers by about 9 percent
for men and about 13 percent for women com4

See Kane and Rouse (1999) for a survey of these studies and a
more detailed description of data.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

pared with earnings for similar workers who did
not attend community college. Another important
fact reported by these authors is that technically
oriented and math and science courses lead to a
higher earnings payoff—about 14 percent for men
and 29 percent for women—but less technically
oriented courses yield very low and possibly zero
returns.

RETURNS TO AN ASSOCIATE’S
DEGREE
Another way to assess the value of a community college education is to determine how much
more a person with an AD earns compared with
a similar person with only a high school diploma.
Separate studies by Kane and Rouse (1995) and
Leigh and Gill (1997) estimated the labor market
return to an AD is about 16 to 27 percent.
The much-larger dataset from the U.S. 2000
Census5 affords answers to more detailed questions. For instance, are there differences in labor
market returns to an AD between different demographic groups? Are the returns the same across
different cities? Data also allow looking at the
differentials in hourly wages rather than annual
earnings.
The sample consists of men and women 25
to 55 years of age with an AD or a high school
diploma who live either in the 20 largest metropolitan areas of the United States (including
St. Louis) or in large metropolitan areas of the
Eighth District (Memphis, Little Rock, and
Louisville).
A simple matching estimator was used to
calculate, for each metropolitan area j, the rate of
return to an AD. Intuitively, people who have an
AD were matched with those who do not but who
have otherwise similar demographic characteristics. We can ask, then, how their wages differ. It
is assumed that productivity, which translates into
wages, is a function of education and age, since
older workers tend to have more work experience.
5

Data are from 2000 Public Use Micro Sample of the U.S. Census
(see Ruggles et al., 2004).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

More precisely, for an individual with age
x = X in metropolitan area j, we would like to
estimate the causal effect of an AD (AD = 1),

(

Δ ( X , j ) = E y 1 x = X , AD = 1, j

(

)

− E y 0 x = X , AD = 1, j ,
the difference between the wage of an individual
with an AD and his or her potential wage if formal
education stopped at the high school level. Here,
y1 is the logarithm of the worker’s wage if the
individual has an AD, and y0 is the logarithm of
the worker’s wage if the individual stops his or
her education at high school. Of course, we cannot
directly observe the second term in the above
equation; we never observe what a person with
an AD would have earned with only a high school
education.
If, however, we are willing to eliminate selection problems by assumption (including the issue
of ability bias that has received close attention in
the literature), we have

(

)

(

)

E y 0 x = X , AD = 1, j = E y 0 x = X , AD = 0, j ,
This equation simply means that the wages of a
person with an AD, if he or she did not receive
it, would have been the same as the wages of a
similar person with a high school diploma. Thus,
the mean return to an AD in a particular metropolitan area j, denoted as Δ共j 兲, is
Δ ( j ) = ∫ Δ ( x j ) dF ( x j ),
where dF 共x | j 兲 is the distribution of x in the
metropolitan area.
In principle, Δ共j 兲 might vary across cities
simply because of differences in the age distributions in these cities. Such differences would be
of little interest, so to “standardize” the estimates,
I use the national cumulative distribution function of x and calculate
Δ n ( j ) = ∫ Δ ( x j )dFn ( x ),
where dFn共x 兲 is derived from the national data.6
6

For more on this approach to a nonparametric estimation of
returns to schooling, see Black, Kolesnikova, and Taylor (2009).

J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 3
Labor Market Returns* to Associate’s Degree (Relative to High School) for Women by Race
Region/City

White

Black

Hispanic

0.29

0.30

0.29

Atlanta

0.27

0.29

0.53

Baltimore

0.28

0.20

Boston

0.29

0.33

0.31

Chicago

0.25

0.23

0.21

Dallas

0.30

0.27

0.24

Detroit

0.32

0.19

0.25

Houston

0.24

0.45

0.20

Los Angeles

0.20

0.26

0.30

Miami

0.25

0.30

0.33

United States
20 largest metropolitan areas

Minneapolis

0.23

0.28

0.24

New York

0.26

0.35

0.28

Philadelphia

0.28

0.24

0.38

Phoenix

0.24

0.33

0.18

Pittsburgh

0.29

0.19

—

Riverside-San Bernardino

0.31

0.40

0.36

San Diego

0.23

0.21

0.28

San Francisco

0.26

0.21

0.30

Seattle

0.25

0.29

0.39

St. Louis

0.24

0.43

—

Washington

0.23

0.26

0.37

Memphis

0.23

0.31

—

Little Rock

0.37

—

Louisville

0.32

—

Eighth District large metropolitan areas

NOTE: *The numbers can be interpreted as percentage increases in wages. (See footnote 7 for more information.)
SOURCE: Author’s calculations. Data are from 2000 Public Use Micro Sample (PUMS) of the U.S. Census. Results are missing if data
were insufficient because of small sample size.

This estimation is performed separately for men
and women and for different racial groups.
One immediate feature of the results is that,
though the estimated average returns to an AD
are consistent with other researchers’ findings,
there are significant differences among demographic groups (Tables 3 and 4). Women of all
races have higher returns to an AD than men do,
which might be due to the fact that women are
more likely to major in nursing and related
32

J A N UA RY / F E B R UA RY

2010

health fields. The return to an AD also varies
among racial groups. Hourly wages of white men
with an AD are 18 percent higher than wages of
white men who stopped their formal education
at high school.7 The same returns are much
7

Tables 3 and 4 report differences in mean log wages between
holders of ADs and high school graduates. Differences in mean
log wages, called log points differences, approximate percentage
differences.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Table 4
Labor Market Returns* to Associate’s Degree (Relative to High School) for Men by Race
Region/City

White

Black

Hispanic

0.18

0.25

0.27

Atlanta

0.21

0.26

0.39

Baltimore

0.15

0.26

0.19

Boston

0.17

0.06

0.25

Chicago

0.10

0.21

0.19

Dallas

0.24

0.28

0.29

Detroit

0.21

0.22

0.34

Houston

0.19

0.21

0.27

Los Angeles

0.16

0.35

0.30

Miami

0.30

0.25

0.30

United States
20 largest metropolitan areas

Minneapolis

0.17

0.27

0.32

New York

0.11

0.24

0.21

Philadelphia

0.15

0.17

0.32

Phoenix

0.18

0.42

0.24

Pittsburgh

0.16

0.17

—

Riverside-San Bernardino

0.20

0.15

0.24

San Diego

0.15

0.36

0.24

San Francisco

0.12

0.48

0.23

Seattle

0.04

0.22

0.17

St. Louis

0.11

0.13

—

Washington

0.18

0.22

0.16

Memphis

0.16

0.22

—

Little Rock

0.22

—

Louisville

0.18

0.17

—

Eighth District large metropolitan areas

higher for black and Hispanic men—25 and 27
percent higher, respectively.
Furthermore, the return to an AD is not the
same across different cities in the United States.
For example, white men with ADs are paid only
4 percent more than white high school graduates in
Seattle but as much as 30 percent more in Miami.
For Hispanic men, the return to an AD is 16 percent in Washington, D.C., but it is more than twice
as much—39 percent—in Atlanta. Cross-city difF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

ferentials for white women are not as large, but
they are significant for minority women.
Tables 3 and 4 also present estimated returns
to an AD in four large metropolitan areas of the
Eighth District. White men with an AD earn on
average 11 percent more in St. Louis, 16 percent
more in Memphis, 22 percent more in Little Rock,
and 18 percent more in Louisville than similar
men with only a high school diploma. For black
men, returns to an AD are 13 percent in St. Louis,
J A N UA RY / F E B R UA RY

2010

Kolesnikova

22 percent in Memphis, and 17 percent in
Louisville. Consistent with the rest of the country,
women’s returns are higher than men’s. For example, black women in St. Louis with an AD earn
43 percent more than black women with only a
high school education.
Why is there such a large variation in returns to
an AD across cities? Although no formal research
has been done on this topic, possible explanations
might be locational differences in labor market
conditions and industrial composition.

DIFFERENT EDUCATIONAL PATHS
Community college students have various
educational goals and intentions when they enter
college. Although many plan to obtain an AD,
some students enroll to take just a few classes to
improve their skills or to become certified in a
certain field. Some intend to transfer to a fouryear institution without any formal community
college credentials.
This ability of community colleges to offer
students many options provides a unique opportunity to obtain postsecondary education for many
students who would not have it otherwise. On
the other hand, because the educational objectives
of students—and, thus, their paths—are so different, it is difficult to track their progress through
college and to assess the effect of community
college education on their educational attainment
and labor market outcomes. The fact that most
students attend community colleges part-time
and take longer to complete their program makes
the task even more complicated.
Critics of the community college system often
point out that a significant proportion of community college students earn relatively few college
credits. Kane and Rouse (1999) calculated that
the majority of community college students complete one year or less and 35 percent complete
only one semester of study or less. The study
also showed that fewer than half of community
college students complete any degrees. In particular, about 15 percent receive a certificate, 16
percent complete an AD, and another 16 percent
eventually receive a bachelor’s degree or higher.
Kane and Rouse (1999) point out that, unlike their
34

J A N UA RY / F E B R UA RY

2010

community college counterparts, almost 60 percent of four-year college entrants receive at least
a bachelor’s degree.
Does this mean that enrolling in a two-year
college somehow reduces an individual’s educational attainment? One view is that easy access
to community college sidetracks students from a
four-year college, where they are more likely to
obtain a bachelor’s degree. On the other hand,
many nontraditional students who attend community college would not attend four-year colleges. For them, community colleges provide a
chance for a postsecondary education they would
not have had otherwise. Therefore, researchers
argue, even if attending a community college
instead of a four-year college might lower educational attainment for some students, more students
have access to higher education, which raises
overall educational attainment in society.
To better answer questions about the effect of
community colleges on educational attainment,
it is necessary to consider students’ intentions
toward their educational objectives together with
their outcomes. The problem is a lack of reliable
data that measure students’ goals and preparation.
The U.S. Department of Education has
attempted to study educational outcomes of
community college students. Its report used data
from several sources, including those tracking
students over time (Hoachlande et al., 2003). The
study found that about 90 percent of students
entering community college intended to obtain a
formal credential or to transfer to a four-year college. One could argue that it is more reasonable
to consider completion rates only for those who
intended to obtain a degree in the first place. The
report estimated that, depending on the data used,
between 51 percent and 63 percent of these students had fulfilled their expectations within six
to eight years after initial enrollment. In particular,
about 11 percent had earned a certificate, 17 percent to 18 percent had earned an AD, 11 percent
to 28 percent (depending on the data used) had
attained a bachelor’s degree or higher, and 12 percent to 13 percent had transferred to a four-year
college without attaining a formal degree.
Keeping in mind that a primary goal of twoyear colleges is to prepare students to continue
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

their studies at four-year institutions, it is particularly important to evaluate their transfer rates.
The U.S. Department of Education (Hoachlande
et al., 2003) report indicated that, overall, about
29 percent of community college students had
transferred to four-year colleges. Interestingly, 51
percent of those who intended to complete a
bachelor’s degree when they first started had
transferred. At the time data were collected, 80
percent of those who transferred either obtained
a bachelor’s degree or were still working toward it.
What about the students who left community
college without any formal credential? This
amounts to more than half of those who started
classes. According to the report, about one-third
of this group said that postsecondary education
improved their salary. For 47 percent, attending
community college led to increased job opportunities. About 43 percent reported improvement
in job performance, and 47 percent said they had
more job responsibilities.
Students who did receive a certificate or a
degree were more likely to be satisfied with their
outcomes. About 80 percent of them said their
salaries had increased. Almost 85 percent reported
having a better job or more responsibilities.

FROM A COMMUNITY COLLEGE
TO A BACHELOR’S DEGREE
As discussed previously, even though community colleges initially were introduced to help
prepare students for four-year colleges, fewer than
a third of community college students transfer to
four-year colleges. Still, it is important to compare the outcomes of students who transfer to a
four-year institution with the outcomes of their
counterparts who began at a four-year institution.
A recent study by Long and Kurlaender (2008)
evaluates whether there is what the authors term
a “community college penalty.” The study uses a
unique longitudinal dataset that includes everyone who entered Ohio public institutions of higher
education in the fall of 1998 with follow-up over
the next nine years. It provides information on
students’ high school preparation, entrance examinations, degree intentions, family background,
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

college performance and, finally, degree completion. As long as students transfer between Ohio’s
public colleges and universities, they remain in
the dataset. This makes it possible to track most
students’ progress from starting postsecondary
education at a community college to receiving a
bachelor’s degree from a four-year college.
The study finds that there is indeed a “penalty”
resulting from beginning postsecondary education
at a community college. The rates of dropping
out or “stopping out” without a degree are much
higher for those who start at community colleges
than for those who start at four-year institutions.
For example, community college students were
36 percent less likely to obtain a bachelor’s degree
than similar students who started at four-year
colleges.
One possible explanation for this result is that
four-year college students start with an intention
to graduate while community college students
have different educational objectives. The study
finds, however, that even community college
students who expressed an intention to obtain a
four-year bachelor’s degree are significantly less
likely to do so within nine years of starting their
postsecondary studies. Only 26 percent of this
group have a bachelor’s degree nine years after
starting their postsecondary education. To put it
in perspective, 50 percent and 73 percent of those
who start at nonselective and selective four-year
institutions, respectively, obtain a bachelor’s
degree. In addition, students who start at community colleges have fewer total earned credits
than students who start at four-year colleges.
The observed differences in educational outcomes may occur because of the differences
between the students at two-year and four-year
institutions. Demographic, family, and other
characteristics of students who begin at community colleges differ from those of students who
begin at four-year institutions. Such differences
might lead to a selection bias of the estimates.
However, the negative effect of starting postsecondary education at a community college remains
even after adjusting for selection bias by controlling for students’ race, gender, age, ability (measured by ACT scores), and family income. The
authors find “a persistent community college
J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 5
Proportion of Bachelor’s Degree Holders with Associate’s Degrees by Region of Residence
Proportion of bachelor’s degree holders
Region

With AD (%)

Who attended community college (%)

New England

Middle Atlantic

East North Central

West North Central

South Atlantic

East South Central

West South Central

Mountain

Pacific

SOURCE: Author’s calculations. Data are from NSCG (2003).

penalty,” but they suggest that “it is worth comparing the size of the penalty to the difference in
costs at two-year versus four-year institutions.”

LONG-TERM EDUCATIONAL
OUTCOMES
Few community college students go on to
receive a bachelor’s degree. Still, some successfully transfer to four-year colleges and obtain a
bachelor’s degree or higher. This section compares
these individuals with those who start postsecondary education at traditional four-year colleges
and analyzes their long-term educational outcomes.
The NSCG is a joint project of the U.S Census
Bureau and the National Science Foundation.
The 2003 survey included a sample of respondents
to the 2000 Decennial Census long form who indicated they have a bachelor’s degree or higher in
any field of study. The survey collected detailed
information about their education, current and
past employment, current salary, and demographic
characteristics. In particular, the dataset reports
educational background characteristics, such as
type of college attended, major field of study, number of degrees, and the highest degree received.
Most importantly, for my purposes, it identifies
respondents who have an AD or attended a com36

J A N UA RY / F E B R UA RY

2010

munity college. One shortcoming of these data,
however, is a lack of family background information and ability measures.
Among people who have at least a bachelor’s
degree, 17 percent report having received an AD.
(I assume here that they started their postsecondary education at a community college and,
after receiving an AD, continued their education
at a four-year college.) The rest of this section
compares this group with the rest of the respondents with at least a bachelor’s degree.8 I start
the comparison of the two groups by presenting
some descriptive statistics.
Table 5 reports the proportion of respondents
with a bachelor’s degree who either attended a
community college or have an AD; this group is
classified according to region of residence.9
Between 14 and 20 percent of four-year college
graduates have an AD, depending on the region.
Bachelor’s degree holders in the Pacific and South
Atlantic regions are most likely to have an AD
8

The dataset also identifies individuals who attended a community
college but does not identify what they were studying. It is impossible to know whether a person took classes for credit in preparation
for college or not. Because of this limitation, I ignore these individuals’ community college experience.

The definition of the region in this context is provided in
Appendix A. “Region” is the smallest geographic unit of analysis
available in the NSCG dataset.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Table 6
Proportion of Bachelor’s Degree Holders with Associate’s Degree by Region of Birth
Proportion of bachelor’s degree holders
Region

With AD (%)

Who attended community college (%)

New England

Middle Atlantic

East North Central

West North Central

South Atlantic

East South Central

West South Central

Mountain

Pacific

SOURCE: Author’s calculations. Data are from NSCG (2003).

(20 percent and 18 percent, respectively), while
New England residents with a bachelor’s degree
are least likely to have an AD (14 percent). As
many as 58 percent of bachelor’s degree holders
attended a community college at some point in
the Pacific region, but only 31 percent did in
New England.
Table 6 reports similar statistics by region of
birth. People with a bachelor’s degree who were
born in the Pacific region are significantly more
likely to attend community college (60 percent)
or have an AD (24 percent) than people who were
born in other regions. This is not surprising given
that the Pacific region includes California, the state
with the highest community college enrollment.
Figure 1 presents a distribution of parental
education. Consistent with other studies, I find
that AD holders are much more likely to be firstgeneration college students than those who do
not have an AD. They are also more likely to have
parents with a level of education less than a college degree.
Next, I examine whether there are differences
in educational choices between those who obtained
an AD before enrolling in a four-year college and
those who did not. Table 7 summarizes the types
of four-year institutions that respondents attended.
The Carnegie Foundation Classification of InstiF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

tutions of Higher Education is used to categorize
universities as Research Universities (I and II),
Doctorate Granting (I and II), Master’s Granting (I
and II), Liberal Arts (I and II), and Associates of
Art Colleges that include community colleges.10
While direct comparison of the quality of
education provided by different universities is
difficult, traditionally research and doctorategranting universities are perceived as being more
selective and having better resources than master’sgranting and liberal arts colleges.
Table 7 shows that people with a prior AD
were significantly less likely to attend Research I
universities (18 percent vs. 26 percent) and slightly
less likely to attend Doctorate-Granting universities (6 percent vs. 7 percent). On the other hand,
a much higher proportion attended Master’sGranting universities (36 percent vs. 27 percent).
It also seems that people with a prior AD were
much less likely to attend more selective Liberal
Arts I colleges than their counterparts (1 percent
vs. 6 percent). To sum up, it appears that AD
recipients attended less-selective (and perhaps
less-expensive) institutions for their bachelor’s
studies. Figure 2 shows that students with an
AD are also more likely to be enrolled in public
10

See Appendix B for definitions of the Carnegie Foundation
Classification of Institutions of Higher Education categories.

J A N UA RY / F E B R UA RY

2010

Kolesnikova

Figure 1
Parents’ Education
Father’s Education

Percent
35

No Associate’s Degree

With Associate’s Degree
25
20
15
10
5

at
e
or
D
oc
t

re
e

re
e
’s
D
eg

Pr
of
es
sio

as
te
r

’s
D
ch

ig
h

eg
re
e

l
oo
Sc
h
ig
h
Le
ss
th
a

lle
ge

Mother’s Education

Percent
45
40
35
30
25
20
15
10
5

J A N UA RY / F E B R UA RY

te
or
a

eg
D
na
l

D
oc
t

re
e

re
e
D
eg
Pr
of
es
sio

as
te
M

or
ch
el

m
So

r ’s

’s
D

Co
e

h
ig
H

Ba
2010

eg
re
e

ol
Sc
ho

ol
Sc
ho
ig
h
H
n
ha
Le
ss
t
38

lle
ge

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Figure 2
Public or Private Institution of Bachelor’s Degree
Percent
70
No Associate’s Degree

With Associate’s Degree

50
40
30
20
10
0
Public

Private

Unknown

Table 7
Institution Awarding First Bachelor’s Degree
All bachelor’s
degree holders (%)

No AD (%)

With AD (%)

Research University I

24.57

25.88

18.29

Research University II

7.39

7.52

6.74

Doctorate-Granting I

6.68

6.89

5.72

Carnegie Classification of Institution

Doctorate-Granting II
Master’s I

5.78
28.5

5.57

6.79

26.95

36.02

Master’s II

2.24

2.03

3.23

Baccalaureate (Liberal Arts) I

5.08

5.86

1.32

Baccalaureate (Liberal Arts) II

7.62

7.51

8.18

Associate of Art Colleges

0.27

0.15

0.84

Other

2.62

2.49

3.35

Missing information

9.23

9.17

9.54

NOTE: See Appendix B for descriptions of the classifications.
SOURCE: Author’s calculations. Data are from NSCG (2003).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 8
Distribution of Major Fields of Study of First Bachelor’s Degree
Major field of study

All bachelor’s
degree holders (%)

No AD (%)

3.86

3.87

3.82

1.99

1.86

2.66

Computer and math sciences
Computer and information sciences
Mathematics and statistics
Biological, agricultural, environmental

With AD (%)

1.87

2.01

1.16

6.2

6.48

4.87

Agricultural and food sciences

0.8

0.83

0.69

Biological sciences

4.97

5.27

3.54

0.43

0.38

0.64

2.9

3.14

1.71

Environmental life sciences
Physical and related sciences
Chemistry, except biochemistry

1.51

1.67

0.73

Earth, atmospheric, and ocean sciences

0.57

0.59

0.46

Physics and astronomy

0.62

0.68

0.32

Other physical sciences

0.2

Social and related sciences

0.2

14.27

12.69

Economics

2.16

2.4

1.04

Political and related sciences

3.11

3.34

2.01

Psychology

4.61

4.55

4.9

Sociology and anthropology

2.76

2.64

3.32
1.42

Other social sciences

1.36

1.34

7.7

7.99

6.35

Aerospace, aeronautical, and astronautical

0.29

0.31

0.22

Chemical engineering

0.6

0.68

0.22

Civil and architectural engineering

1.16

1.19

1.02

Electrical and computer engineering

2.39

2.43

2.2

Industrial engineering

0.48

0.5

0.41

Mechanical engineering

1.76

1.81

1.48

Other engineering

1.02

1.07

0.8

Engineering

Health, science education, technology

9.57

9.46

10.21

Health

6.6

6.51

7.08

Science and mathematics teacher education

1.15

1.19

0.97

Technology and technical fields

0.88

1.62

Other science- and education-related fields

0.82

0.88

0.54

Business, management, art

55.75

54.79

60.32

Management and administration fields

17.61

16.6

22.45

Education, except science and math teacher education

13.51

13.39

14.08

2.28

2.19

2.71

Social service and related fields
Sales and marketing fields

2.61

2.56

2.86

Art and humanities fields

12.69

13.22

10.15

7.05

6.83

8.07

Other non-science and education fields
SOURCE: Author’s calculations. Data are from NSCG (2003).

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Table 9
Proportion of Males by Major Field of Study
Major field of study

All bachelor’s
degree holders (%)

Computer and math sciences
Computer and information sciences
Mathematics and statistics
Biological, agricultural, environmental

No AD (%)

With AD (%)

Agricultural and food sciences

Biological sciences

Environmental life sciences

Physical and related sciences

Chemistry, except biochemistry

Earth, atmospheric, and ocean sciences

Physics and astronomy

Other physical sciences

Social and related sciences

Economics

Political and related sciences

Psychology

Sociology and anthropology

Other social sciences

Aerospace, aeronautical, and astronautical

Chemical engineering

Civil and architectural engineering

Electrical and computer engineering

Industrial engineering

Mechanical engineering

Other engineering

Engineering

Health, science education, technology

Health

Science and mathematics teacher education

Technology and technical fields

Other science- and education-related fields

Business, management, art

Management and administration fields

Education, except science and math teacher education

Social service and related fields

Sales and marketing fields

Art and humanities fields

Other non-science and education fields

SOURCE: Author’s calculations. Data are from NSCG (2003).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 10
Age at Attaining Degrees
All

No AD

With AD

26.8

26.2

29.5

Age at first bachelor’s degree (years)
Mean
Standard deviation (SD)

5.3

Minimum

4.6
15

7.3
16

Maximum

10 percent

25 percent

Median

75 percent

90 percent

29.7

29.3

31.7

Age at highest degree (years)
Mean
SD

7.2

Minimum

6.8
15

8.3
16

Maximum

10 percent

25 percent

Median

75 percent

90 percent

SOURCE: Author’s calculations. Data are from NSCG (2003).

colleges than students who do not have an AD
and less likely to attend private colleges.
Are there differences in major fields of study
between the two groups? One of the main objectives of community colleges is to prepare students
for four-year college studies. Do students who
have taken classes at a community college choose
different fields of study than students who did
not go to community college before attending a
four-year institution?
Fortunately, NSCG data provide detailed
information on respondents’ degree majors. As
shown in Table 8, fewer people with an AD major
in sciences and engineering than people with no
AD. Instead, people with an AD are more likely
to major in health, technology, and management
than their counterparts. Preference for the health
42

J A N UA RY / F E B R UA RY

2010

and technology fields is expected, given that community colleges often focus more on these disciplines. It is somewhat surprising that so many more
AD holders choose to major in management than
people without an AD (22 percent vs. 17 percent).
Interestingly, there is little difference in gender
distribution across major fields of study between
the two groups (Table 9). There are some exceptions, however. More women with ADs choose
to major in computer and information sciences,
economics, aerospace engineering, chemical
engineering, and marketing than women without
an AD; and more men with ADs choose to major
in mathematics and statistics, agriculture, environmental and earth sciences, sociology, and
industrial engineering than men without an AD.
It is hard to know whether this is a result of stuF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Figure 3
Number of Degrees (Bachelor’s and Higher)
Percent
80
No Associate’s Degree

With Associate’s Degree

60
50
40
30
20
10
0
1

4 or More

Figure 4
Highest Degree Attained
Percent
80
No Associate’s Degree

With Associate’s Degree

60
50
40
30
20
10
0
Bachelor’s

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Master’s

Doctorate

Professional

J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 11
Years from Bachelor’s Degree to Advanced Degree
All

No AD

With AD

Mean

7.8

7.9

7.4

Standard deviation (SD)

6.8

6.9

6.6

11.1

10.9

12.8

7.0

6.9

8.5

Mean

5.4

5.3

4.4

4.8

Years from bachelor’s degree to master’s degree

Years from bachelor’s degree to doctorate
Mean
SD
Years from bachelor’s degree to professional degree

SOURCE: Author’s calculations. Data are from NSCG (2003).

Table 12
Proportion of Associate’s Degree Holders
by Highest Degree
Highest degree

With AD (%)

Bachelor’s degree

20.7

Master’s degree

14.3

Doctorate

5.8

Professional

9.5

SOURCE: Author’s calculations. Data are from NSCG (2003).

dents’ exposure to some subjects before entering
a four-year institution or other effects on some
students’ choice of a major field of study.
Not surprisingly, AD holders are older on
average when they obtain a bachelor’s degree.
Their mean age is 29.5 years, compared with the
mean age of 26.2 years of those who obtain a
bachelor’s degree without an AD (Table 10).
Almost 70 percent of bachelor’s degree holders
with an AD do not continue their education
beyond their first bachelor’s degree. This contrasts
with the fewer than 60 percent of their counterparts without an AD (Figure 3). A higher proportion of people with no AD go on to receive a
master’s degree, a doctorate, or a professional
degree (e.g., J.D. or M.D.) (Figure 4). Table 11
44

J A N UA RY / F E B R UA RY

2010

shows that, for those who continued beyond a
bachelor’s degree, it took slightly less time on
average to obtain a master’s or a professional
degree if a person had an AD but longer to finish
a Ph.D. program.
Table 12 presents another way to compare the
highest education levels of people with and without an AD. Among people with only a bachelor’s
degree, about 21 percent have a prior AD. Among
those who received a master’s degree, only 14.3
percent have an AD. The proportion of people
with an AD is even smaller among those with a
doctorate or a professional degree (5.8 and 9.5
percent, respectively).

LONG-TERM LABOR MARKET
OUTCOMES
This section compares labor market outcomes
of people with an AD who proceeded to receive
a bachelor’s degree or higher and the labor market outcomes of their counterparts with no AD.
In particular, it concentrates on one measure of
labor market outcome—annual salary.
This analysis considers only individuals of
prime age (23 to 55 years old) who are employed.
Since salaries are top-coded in the NSCG dataset,
those above the 95th percentile of salary distribution are omitted from the sample, as are those
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Table 13
Salaries (in US$) by Education Level
All

No AD

With AD

Mean

57,686

58,559

53,696

Standard deviation (SD)

29,660

30,218

26,597

A. All

B. Bachelor’s degree
Mean

54,126

54,667

52,022

28,319

28,855

26,029

Mean

60,676

61,323

56,997

28,663

29,030

26,185

C. Master’s degree

D. Doctorate
Mean

70,711

71,246

62,906

29,837

29,832

28,851

Mean

78,705

79,491

70,349

36,711

36,793

34,799

E. Professional degree

SOURCE: Author’s calculations. Data are from NSCG (2003).

below the 5th percentile, to maintain distribution
symmetry. Thus, individuals who earn less than
$10,000 or more than $150,000 per year are not
included.
Table 13 shows the average annual salary by
education level for the full sample and then separately for individuals with and without an AD.
As expected and well documented in many other
studies, people with a higher level of education
have, on average, higher earnings. Bachelor’s
degree holders earn $54,126 per year; people with
master’s degrees earn $60,676 per year; people
with a doctorate earn $70,711 per year, and people
with professional degrees earn $78,705 per year,
on average. Remarkably, annual salaries for individuals with an AD differ from those without an
AD for all education levels. Regardless of the highest degree, people who started their postsecondary
education with an AD earn less on average than
those who started at a four-year college. The difference is particularly large for those with a doctorate or a professional degree.
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

A regression analysis can be used to better
understand this phenomenon. In particular, I
estimate the following equation:

S = β0 + β1 ∗ X + β1 ∗ I AD + ε ,
where S is an individual’s annual salary in dollars,
X is a vector of various characteristics that will
be defined shortly, and IAD is an indicator of
whether a person has an AD, in which case it is
equal to 1; otherwise, it is 0. The goal is to compare individuals with the same characteristics X
but different values of an indicator IAD , 0 or 1.
The question is how an AD affects one’s salary.
Relevant characteristics include age, gender, race,
major field of study, and highest degree attained.
The estimation results of the above equation
are reported in Table 14, panel A. The dependent
variable is salary S. The results indicate that an
annual salary increases by about $542 per year
as people age and accumulate more work experience. Women, on average, earn $12,137 per year
less than men with similar characteristics. Minority groups earn less compared with whites. The
J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 14
Regression Analysis: Effects of Various Factors on Salary
Coefficient

Standard error

t-Statistic

541.9

13.7

39.6

A. All
Age
Woman

–12136.5

233.5

–52.0

Black

–4943.0

410.5

–12.0

Hispanic

–5768.8

460.1

–12.5

Asian

–2558.6

416.9

–6.1

Associate’s degree

–3854.1

283.1

–13.6

Controls
Major field of study

Yes

Highest degree

Yes

Number of observations

59,346

Adjusted R 2

0.22

B. Bachelor’s degree
Age
Woman

487.6

17.6

27.73

–12724.9

300.1

–42.4

Black

–6017.6

522.7

–11.5

Hispanic

–6807.9

577.8

–11.8

Asian

–3267.3

565.7

–5.78

Associate’s degree

–3620.8

346.0

–10.46

574.1

24.3

23.6

Controls: Major field of study
Number of observations

Yes
34,067

Adjusted R 2

0.19

C. Master’s degree
Age
Woman

–11460.2

421.3

–27.2

Black

–2198.1

716.5

–3.1

Hispanic

–3549.5

865.6

–4.1

–980.5

707.4

–1.4

–3379.1

536.7

–6.3

Asian
Associate’s degree
Controls: Major field of study
Number of observations

17,803

Adjusted R 2

J A N UA RY / F E B R UA RY

Yes
0.23

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Table 14, cont’d
Regression Analysis: Effects of Various Factors on Salary
Coefficient

Standard error

t-Statistic

1078.1

58.4

18.5

D. Doctorate
Age
Woman

–8176.1

884.1

–9.3

Black

–7725.4

1839.1

–4.2

Hispanic

–3055.8

1937.8

–1.6

Asian

–3544.6

1116.6

–3.17

Associate’s degree

–9565.3

1679.5

–5.7

984.4

81.7

12.1

Woman

–7949.2

1349.9

–5.9

Black

–2325.0

2921.2

–0.8

Hispanic

–3006.2

2775.5

–1.1

Asian

–2473.6

2393.2

–1.0

Associate’s degree

–9423.2

2416.5

–3.9

Controls: Major field of study

Yes

Number of observations

4,521

Adjusted R 2

0.21

E. Professional degree
Age

Controls: Major field of study

Yes

Number of observations

2,955

Adjusted R 2

0.08

annual salary of blacks is $4,943 lower on average
than that of comparable whites. The corresponding difference for Hispanics is $5,769, and it is
$2,559 for Asians. These facts are well documented
in the economics literature. The most striking
finding, however, is that even when other factors
are controlled, people with an AD earn $3,854
less per year than their counterparts with no AD.
All coefficients are statistically significant at a 5
percent level or better.
The same equation is also estimated separately
for each education-level group: bachelor’s degree,
master’s degree, doctoral degree, and professional
degree. Panels B through E of Table 14 show the
results of the estimations. The same pattern is
observed for each education-level group: Older
workers earn more; women and minorities earn
less. More importantly, those who earn an AD and
then a more-advanced degree have lower earnings
than those who earn a bachelor’s degree or higher
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

but no AD. For example, bachelor’s degree holders
earn $3,621 less per year when they have a prior
AD. Strikingly, earning gaps are observed even
for those community college students who receive
a doctorate or a professional degree. Their salaries
are $9,565 and $9,423 lower, respectively, than
salaries of their counterparts who started at a traditional four-year college.
One possible explanation for the salary disparity is that the quality of education differs for
the two groups. For example, labor markets might
assign an additional premium for a degree from
an elite college. Controls were included for the
type of institution awarding a bachelor’s degree
to test this possibility. Results remain virtually
unchanged, which allows rejection of this
explanation.11
11

These results are not reported here but are available from the
author upon request.

J A N UA RY / F E B R UA RY

2010

Kolesnikova

Table 15
Regression Analysis: Effects of Work Experience
Coefficient

Standard error

t-Statistic

605.2

13.4

45.2

A. All
Experience
Woman

–12100.4

232.5

–52.0

Black

–4342.6

409.1

–10.6

Hispanic

–5412.6

458.6

–11.8

Asian

–3018.7

414.9

–7.3

Associate’s degree

–2426.1

281.2

–8.6

574.8

17.2

33.4

Woman

12681.1

298.6

–42.5

Black

–5583.1

520.2

–10.7

Hispanic

–6345.7

575.4

–11.0

Asian

–3627.2

562.9

–6.4

Associate’s degree

–2268.7

342.8

–6.6

532.9

23.7

22.5

Controls
Major field of study

Yes

Highest degree

Yes

Number of observations

59,346

Adjusted R 2

0.23

B. Bachelor’s degree
Experience

Controls: Major field of study
Number of observations

Yes
34,067

Adjusted R 2

0.20

C. Master’s degree
Experience
Woman

–11671.3

421.2

–27.7

Black

–1349.4

718.7

–1.9

Hispanic

–3534.4

866.9

–4.1

Asian

–1836.0

705.9

–2.6

Associate’s degree

–2117.2

537.3

–3.9

Controls: Major field of study
Number of observations

17,803

Adjusted R 2

J A N UA RY / F E B R UA RY

Yes
0.23

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

Table 15, cont’d
Regression Analysis: Effects of Work Experience
Coefficient

Standard error

t-Statistic

1374.1

55.5

24.8

D. Doctorate
Experience
Woman

–7583.2

860.7

–8.8

Black

–6014.6

1791.7

–3.4

Hispanic

–2556.2

1885.4

–1.4

Asian

–3012.9

1086.6

–2.8

Associate’s degree

–6883.8

1625.8

–4.2

1185.5

81.5

14.6

Woman

–7061.9

1340.0

–5.3

Black

–2025.1

2890.2

–0.7

Hispanic

–2899.1

2745.9

–1.1

Asian

–2455.6

2362.8

–1.0

Associate’s degree

–7767.6

2392.3

–3.3

Controls: Major field of study

Yes

Number of observations

4,521

Adjusted R 2

0.25

E. Professional degree
Experience

Controls: Major field of study

Yes

Number of observations

2,955

Adjusted R 2

0.10

One might also be concerned that when we
compare people of the same age with and without
an AD, we in fact compare people with different
levels of experience. People who start at a community college take longer, on average, to graduate
with a bachelor’s degree, so they have less work
experience after receiving a bachelor’s degree. It
could be argued, however, that these people are
accumulating work experience while in school if
they study part-time and continue to work. Still,
to check the robustness of the results, I replaced
the age variable in the analysis with the experience
variable. “Experience” is defined as the number of
years from the time a person received the highest
degree until the time of the survey. It is assumed
that a person has been working continuously.
The results that control for work experience
are presented in Table 15, panels A through E.
When work experience is measured more careF E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

fully, the estimated negative effect of having an
AD is somewhat smaller. Overall, people with an
AD earn $2,426 less per year than people with the
same highest degree who have no AD. The earnings gap is smaller for bachelor’s and master’s
degree holders ($2,269 and $2,117, respectively)
and larger for people with doctorates and professional degrees ($6,884 and $7,768, respectively).
Note that gender and race effects remain almost
unchanged compared with Table 14, panels A
through E.
Why does the observed salary gap persist
between people with and without a prior AD
even among the highly educated? Data available
from the NSCG survey are not sufficient to answer
this question. An important caveat of the above
analysis is the lack of ability and school performance measures and data on family characteristics,
such as family income. One hypothesis is that
J A N UA RY / F E B R UA RY

2010

Kolesnikova

because community college students are more
likely to come from families with lower incomes
and education, they are also more likely to attend
lower-performing elementary and secondary
schools. It is possible that they fall far behind
even before entering the postsecondary education
system. The results seem to suggest that this disadvantage affects educational and labor market
outcomes throughout their lives—as a group they
never catch up with their peers.

CONCLUSION
Community colleges play a significant role
in U.S. higher education, enrolling 46 percent of
current U.S. undergraduates. They offer the opportunity to receive a postsecondary education to
many students who would not attend college
otherwise: first-generation college students, students from low-income families, and older students who continue to work as they attend classes
part-time. Attending a community college even
without completing a degree results in economic
payoffs—in particular, annual earnings increase
by 5 to 8 percent for each year of community college education—and better job opportunities.
Today, the number of U.S. undergraduates is at
an all-time high as more people understand the
necessity of higher education in our technologyintensive world. In addition, historically, college
enrollments in general increase during economic
downturns. Community colleges are important
in helping to absorb this increasing number of
students. Currently, community colleges have
additional appeal because tuition and fees at four-

J A N UA RY / F E B R UA RY

2010

year colleges continue to increase while financial
aid and student loans are harder to obtain.
Despite all the benefits of community colleges,
there are downsides as well. The original goal of
community colleges was to prepare students to
transfer to four-year colleges. Associate’s degree
programs were intended to accomplish that goal.
However, only about 29 percent of community
college students transfer to four-year institutions,
and only about 16 percent eventually receive a
bachelor’s degree or higher. Even among those who
start their postsecondary education intending to
receive a bachelor’s degree, only 26 percent accomplish it. They are also much less likely to pursue
postgraduate studies.
In addition, the salary gap persists between
those with a bachelor’s degree or higher and a
prior AD and similar individuals without an AD,
even among the highly educated. This gap remains
even for people of the same gender, race, education, experience level, field of study, and type of
college they attended.
Still, for many students, community colleges
offer the best chance to obtain a college education.
It is important, however, for individuals to know
both the benefits and the disadvantages of attending a community college when making decisions
about higher education.
This paper attempts to present a comprehensive overview of how community colleges improve
the economic mobility of a significant subset of
the U.S. population. A better understanding of
all aspects of this complicated subject should
be an important priority for researchers and
policymakers.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Kolesnikova

REFERENCES
Black, Dan; Kolesnikova, Natalia and Taylor, Lowell. “Earnings Functions When Wages and Prices Vary by
Location.” Journal of Labor Economics, January 2009, 27(1), pp. 21-47.
Hoachlande, Gary; Sikora, Anna C.; Horn, Laura and Carroll, C. Dennis. “Community College Students: Goals,
Academic Preparation, and Outcomes.” NCES Report 2003-164, National Center for Education Statistics,
June 2003; http://nces.ed.gov/pubs2003/2003164.pdf.
Horn, Laura and Nevill, Stephanie. “Profile of Undergraduates in U.S. Postsecondary Education Institutions:
2003-04, With a Special Analysis of Community College Students.” NCES Report 2006-184, National Center
for Education Statistics, June 2006; http://nces.ed.gov/pubs2006/2006184_rev.pdf.
Jacobson, Louis S.; LaLonde, Robert J. and Sullivan, Daniel G. “Estimating the Returns to Community College
Schooling for Displaced Workers.” Journal of Econometrics, March/April 2005, 125(1-2), pp. 271-304.
Kane, Thomas J. and Rouse, Cecilia Elena. “Labor Market Returns to Two- and Four-Year College.” American
Economic Review, June 1995, 85(3), pp. 600-14.
Kane, Thomas J. and Rouse, Cecilia E. “The Community College: Educating Students at the Margin Between
College and Work.” Journal of Economic Perspectives, Winter 1999, 13(1), pp. 63-84.
Kolesnikova, Natalia A. and Shimek, Luke. “Community Colleges: Not So Junior Anymore.” Federal Reserve
Bank of St. Louis The Regional Economist, October 2008, pp. 6-11;
http://stlouisfed.org/publications/pub_assets/pdf/re/2008/d/colleges.pdf.
Kolesnikova, Natalia A. “From Community College to a Bachelor’s Degree and Beyond: How Smooth Is the Road?”
Federal Reserve Bank of St. Louis The Regional Economist, July 2009a, pp. 10-11;
http://stlouisfed.org/publications/pub_assets/pdf/re/2009/c/community_college.pdf.
Kolesnikova, Natalia A. “Community Colleges: A Route of Upward Mobility.” Federal Reserve Bank of St. Louis
Community Development Research Report, March 2009b, pp. 1-28;
http://stlouisfed.org/community_development/assets/pdf/CommunityColleges.pdf.
Leigh, Duane E. and Gill, Andrew M. “Labor Market Returns to Community Colleges: Evidence for Returning
Adults.” Journal of Human Resources, Spring 1997, 32(2), pp. 334-53.
Long, Bridget Terry and Kurlaender, Michal. “Do Community Colleges Provide a Viable Pathway to a Baccalaureate
Degree?” NBER Working Paper No. 14367, National Bureau of Economic Research, September 2008;
http://papers.nber.org/papers/w14367.pdf.
Rouse, Cecilia E. “Do Two-Year Colleges Increase Overall Educational Attainment? Evidence from the States.”
Journal of Policy Analysis and Management, Fall 1998, 17(4), pp. 595-620.
Ruggles, Steven; Sobek, Matthew; Alexander, Trent; Fitch, Catherine; Goeken, Ronald; Hall, Patricia; King, Miriam
and Ronnander, Chad. “2000 Public Use Micro Sample of the U.S. Census,” in Integrated Public Use Microdata
Series. Minneapolis, MN: Minnesota Population Center [producer and distributor], 2004;
http://usa.ipums.org/usa/.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

J A N UA RY / F E B R UA RY

Kolesnikova

52
2010

APPENDIX A
Table A1
Census Bureau–Designated Regions
West

Midwest

Northeast

Pacific

East
North Central

West
North Central

New England

Idaho

Alaska

Wisconsin

North Dakota

Maine

Montana

Washington

Michigan

South Dakota New Hampshire

Mountain

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Wyoming

Oregon

Illinois

Nebraska

Vermont

Nevada

California

Indiana

Kansas

Massachusetts

Utah

Hawaii

Ohio

Mid-Atlantic

South
South Atlantic

West
South Central

New York

Delaware

Kentucky

Oklahoma

Pennsylvania

Maryland

Tennessee

Texas

New Jersey

District of Columbia

Mississippi

Arkansas

Virginia

Alabama

Louisiana

Minnesota

Rhode Island

West Virginia

Colorado

Iowa

Connecticut

North Carolina

Arizona

Missouri

New Mexico

East
South Central

South Carolina
Georgia
Florida

Kolesnikova

APPENDIX B
Category Definitions of Carnegie Foundation Classification of Institutions of
Higher Education
The 1994 Carnegie Classification includes all colleges and universities in the United States that
are degree-granting and accredited by an agency recognized by the U.S. Secretary of Education.
Research Universities I: These institutions offer a full range of baccalaureate programs, are committed to graduate
education through the doctorate, and give high priority to research. They award 50 or more doctoral degrees1
each year. In addition, they receive annually $40 million or more in federal support.2
Research Universities II: These institutions offer a full range of baccalaureate programs, are committed to graduate
education through the doctorate, and give high priority to research. They award 50 or more doctoral degrees1
each year. In addition, they receive annually between $15.5 million and $40 million in federal support.2
Doctoral Universities I: These institutions offer a full range of baccalaureate programs and are committed to graduate
education through the doctorate. They award at least 40 doctoral degrees1 annually in five or more disciplines.3
Doctoral Universities II: These institutions offer a full range of baccalaureate programs and are committed to
graduate education through the doctorate. They award annually at least 10 doctoral degrees—in three or more
disciplines—or 20 or more doctoral degrees in one or more disciplines.3
Master’s (Comprehensive) Universities and Colleges I: These institutions offer a full range of baccalaureate
programs and are committed to graduate education through the master’s degree. They award 40 or more master’s
degrees annually in three or more disciplines.3
Master’s (Comprehensive) Universities and Colleges II: These institutions offer a full range of baccalaureate
programs and are committed to graduate education through the master’s degree. They award 20 or more master’s
degrees annually in one or more disciplines.3
Baccalaureate (Liberal Arts) Colleges I: These institutions are primarily undergraduate colleges with major
emphasis on baccalaureate degree programs. They award 40 percent or more of their baccalaureate degrees
in liberal arts fields4 and are restrictive in admissions.
Baccalaureate Colleges II: These institutions are primarily undergraduate colleges with major emphasis on
baccalaureate degree programs. They award less than 40 percent of their baccalaureate degrees in liberal arts
fields4 or are less restrictive in admissions.
Associate of Arts Colleges: These institutions offer associate of arts certificate or degree programs and, with few
exceptions, offer no baccalaureate degrees.5

Doctoral degrees include Doctor of Education, Doctor of Juridical Science, Doctor of Public Health, and the Ph.D. in any field.

Total federal obligation figures are available from the National Science Foundation’s annual report, called Federal Support to Universities,
Colleges, and Nonprofit Institutions. The years used in averaging total federal obligations are 1989, 1990, and 1991.

Distinct disciplines are determined by the U.S. Department of Education’s Classification of Instructional Programs 4-digit series.

The liberal arts disciplines include English language and literature, foreign languages, letters, liberal and general studies, life sciences,
mathematics, philosophy and religion, physical sciences, psychology, social sciences, the visual and performing arts, area and ethnic studies,
and multi- and interdisciplinary studies. The occupational and technical disciplines include agriculture, allied health, architecture, business
and management, communications, conservation and natural resources, education, engineering, health sciences, home economics, law and
legal studies, library and archival sciences, marketing and distribution, military sciences, protective services, public administration and
services, and theology.

This group includes community, junior, and technical colleges.

SOURCE: This information is from A Classification of Institutions of Higher Education. Princeton, NJ: The Carnegie Foundation for the
Advancement of Teaching Carnegie Foundation, 1994, pp. xix-xxi. Used with permission.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Alt-A: The Forgotten Segment of the
Mortgage Market
Rajdeep Sengupta
This study presents a brief overview of the Alt-A mortgage market with the goal of outlining broad
trends in the different borrower and mortgage characteristics of Alt-A market originations between
2000 and 2006. The paper also documents the default patterns of Alt-A mortgages in terms of the
various borrower and mortgage characteristics over this period. (JEL G1, G21)
Federal Reserve Bank of St. Louis Review, January/February 2010, 92(1), pp. 55-71.

igh default rates on subprime mortgages marked the onset of the current
financial crisis. Not surprisingly, both
academic research and policy studies
have focused their attention on the boom and subsequent collapse of the subprime mortgage market.
However, the high incidence of defaults was not
limited to subprime mortgages only; defaults have
also risen rapidly in the other segments of the
mortgage market—for example, the market for
Alt-A (or Alternative-A) mortgages. But our
knowledge of the Alt-A market is significantly
less than our knowledge of subprime mortgages.
This paper aims to fill this void. This study
presents a brief overview of Alt-A mortgage originations with the goal of outlining broad trends in
the different borrower and mortgage characteristics
of Alt-A originations between 2000 and 2006. The
paper also documents the default patterns of Alt-A
mortgages in terms of the various borrower and
mortgage characteristics over this period. We begin
with a broad overview of the different segments
of the overall U.S. mortgage market and their evolution over this period, with a special emphasis
on the Alt-A mortgage segment.1
Since the 1970s, the principal structural
change in the mortgage market has been the use
of securitization. Prior to this, mortgages were

retained by banks in their portfolios until they
matured or were paid off. Securitization is a
process by which mortgages (typically a large pool
of mortgages) are used as collateral to issue securities, also known as mortgage-backed securities
(MBS). Some mortgage securities are backed
implicitly or explicitly by the U.S. government
and are commonly called agency MBS. Such origination of mortgages and issuance of MBS is dominated by loans to prime borrowers conforming to
underwriting standards set by the governmentsponsored agencies. Non-agency MBS issuance
can be split into three broad categories—jumbo,
Alt-A, and subprime. “Loosely speaking, the
Jumbo asset class includes loans to prime borrowers with an original principal balance larger than
the conforming limits2 imposed on the agencies
1

See Lehnert (2009) and Quigley (2006) for a more detailed overview.

Conforming mortgages satisfy balance limits and are typically
securitized either with some form of explicit government guarantees (Federal Housing Administration/Veterans Administration
[FHA/VA] mortgages securitized by the Government National
Mortgage Association [Ginnie Mae]) or with implicit government
guarantees (conventional mortgages securitized by the Federal
National Mortgage Association [Fannie Mae] and Federal Home
Loan Mortgage Corporation [Freddie Mac]). In contrast, the privatelabel market securitizes nonconforming mortgages, which include
the jumbo prime, subprime, and Alt-A markets (for more details,
see Fabozzi, 2006).

Rajdeep Sengupta is an economist at the Federal Reserve Bank of St. Louis. The author thanks Geetesh Bhardwaj, Dan Thornton, and Dave
Wheelock for comments on a previous draft of this article. Yu Man Tam provided research assistance.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Sengupta

by Congress; the Alt-A asset class involves loans
to borrowers with good credit but includes more
aggressive underwriting than the conforming or
Jumbo classes (i.e., no documentation of income,
high leverage); and the Subprime asset class
involves loans to borrowers with poor credit history.”3 Both agency and non-agency jumbo mortgages constitute the prime mortgage market of
high-credit-quality borrowers, while the non-prime
segment comprises subprime and Alt-A mortgages.
At the outset, it is important to mention that
the guidelines for selecting mortgages into subprime and Alt-A pools vary by arranger of the
MBS. Typically, Alt-A mortgages are underwritten
to borrowers of good credit quality—that is, those
who would otherwise qualify for a prime loan in
terms of their credit history. However, Alt-A borrowers do not satisfy the underwriting rules for
prime loans because they are unwilling or unable
to provide full documentation on their mortgage
application.4 Their inability to provide this information is largely due to the fact that such borrowers are in professions characterized by variable
incomes or are self-employed borrowers operating
cash businesses. On the other hand, subprime
originations are primarily to borrowers with
incomplete or impaired credit histories. Therefore,
while the criterion for selection into a particular
pool is not consistent across lenders, the credit
quality for Alt-A pools is characteristically better
than that for subprime pools.
Historically, the Alt-A market has been the
preserve of highly specialized lenders with
expertise in underwriting such loans. Over the
years, this market has grown significantly and
evolved with an increased level of investor sponsorship. Figure 1 shows the evolution of mortgage
originations by market segment in the United
States between 2001 and 2006.5 A significant
decline in prime mortgage interest rates between
2000 and 2003 aided a refinance boom and the

increase in agency mortgages was a major factor
behind the growth in total mortgage originations
over this period (see Figure 1). However, with
the rise in mortgage interest rates, prime originations declined sharply after 2003. Meanwhile,
the growth of non-prime originations continued
unabated. The growth rates in annual originations
for the agency, subprime, and Alt-A segments from
2001 through 2003 were 95 percent, 94 percent,
and 54 percent, respectively, but annual agency
originations declined by 60 percent from 2003 to
2006. In contrast, the comparable growth rates
between 2003 and 2006 for the subprime and
Alt-A segments were 94 and 340 percent, respectively. The higher levels of originations after 2003
were largely sustained by the growth of the nonprime (both the subprime and Alt-A) segment of
the mortgage market.
This paper uses the loan-level data on securitized Alt-A originations from 1998 through 2007
published by LoanPerformance (LP).6 The data
contain details on individual securitized Alt-A
loans, and a loan is classified as subprime or AltA depending on whether it is securitized in a subprime or Alt-A pool.7 The details include mortgage
attributes of the loan, such as the product type,
the interest rate, the loan purpose (purchase or
refinance), documentation (full-doc or low-doc),
loan-to-value (LTV) ratio, and borrower characteristics such as credit scores (FICO8 at the time
of origination). The next section outlines the
broad trends in the underwriting standards for
Alt-A mortgages in terms of these attributes. We
then outline the performance of Alt-A loans in
terms of the borrower and mortgage characteristics
mentioned previously.
6

For details on the coverage of the LP data and the relation to other
available mortgage databases, see Mayer and Pence (2008). According to Mayer and Pence (2008), LP captures “around 90 percent of
the subprime securitized market from 1999 to 2002 and nearly all
the market from 2003 to 2005.”

Ashcraft and Schuermann (2008).

Generally this documentation is regarding their income. In limited
or no-documentation programs, applicants typically state their
income and assets to the loan officer but are not required to show
detailed proof of that information for the lender’s files. They are
often termed stated income mortgages.

As mentioned earlier, different arrangers use different criteria for
this selection. Therefore, it is possible that what is considered to
be subprime by a particular arranger may be classified as Alt-A by
a different arranger.

Borrower credit score at the time of loan origination is denoted by
FICO (an industry standard developed by the Fair Isaac Corporation)
with a number in the range 300 to 850. The score increases with
the creditworthiness of the borrower.

This figure is updated from Sengupta and Tam (2008).

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

Figure 1
Origination and Issue of Mortgage Loans by Market Segment
$ Billions
3,000

Subprime
Alt-A

2,500

Jumbo
Agency

2,000
1,500
1,000
500
0
2001

2002

2003

2004

2005

2006

2007

NOTE: The figure shows the evolution of mortgage originations by market segment in the United States between 2001 and 2007.
SOURCE: Insider Mortgage Finance Publications, Inc.

SUMMARY: TRENDS FOR ALT-A
MORTGAGES
We begin this section by studying the characteristics of Alt-A mortgages originated between
1998 through 2007. The shares of product types
originated in the Alt-A markets by vintage (year
of origination) are given in Table 1. Table 1 shows
that before 2004, the majority of Alt-A mortgages
were fixed-rate mortgages (FRMs). Interestingly,
the share of FRMs as a proportion of total originations fell by half in a single year, from 2003 to
2004. This decline was accompanied by a rise in
the fraction of loans that were adjustable-rate
mortgages (ARMs).
Table 1 also shows the growth of hybrid-ARM
products over this period.9 We define ARMq as
the hybrid-ARM where the first reset occurs after
q years. Typically, the mortgage rate on an ARM
loan resets once every 6 months or a year into an
indexed rate (like the 6-month LIBOR) plus a
margin. Therefore the ARM1 is just the standard
ARM product that resets after the first year, while
the ARM2, ARM3, and ARM5 categories include
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

the more specialized products, such as the 2/28,
the 3/27, and the 5/25 mortgage products,
respectively.10
Table 2 presents a similar table for subprime
mortgages. The data show a gradual rise in ARM2
and ARM3 products in the subprime market from
around 30 percent of the market in 1998 to more
than 70 percent of the market between 2004 and
2006. However, with the exception of 2004-05,
ARM2 and ARM3 products were never more than
10 percent of the Alt-A market. Among hybridARM Alt-A originations, the ARM5 product has
the largest share of the market, growing from less
9

Hybrid-ARM products are specialized products that include an
initial period over which the repayment schedule on the mortgage
resembles that of an FRM and a subsequent period over which the
mortgage product acts like an ARM. During the fixed-leg of the
hybrid-ARM, the mortgagee pays a lower introductory closing rate
called the teaser rate. The teaser rate remains in effect until the
reset date, after which the repayment schedule on the hybrid-ARM
resembles an ARM. The reset date, market index rate used, and
the margin are decided at the closing date.

Therefore, the 2/28 and the 3/27 products are 30-year mortgages
with teaser rates for two and three years, respectively. The rationale
for adopting the ARMq terminology over the traditional 2/28 or
3/27 is that this terminology is inclusive of mortgage products that
have amortization terms of more than 30 years.

J A N UA RY / F E B R UA RY

2010

Sengupta

Table 1
Evolution of Alt-A Securitized Mortgages (Percent Market Share by Product Type)
Mortgage type
Vintage

FRM

ARM1

ARM2

ARM3

ARM5

Other

Share of total

1998

98.97

0.41

0.09

0.02

0.11

0.41

2.7

1999

93.27

1.50

2.64

0.79

0.94

0.87

1.7

2000

85.04

9.15

1.13

0.94

1.88

1.87

1.5

2001

79.36

6.20

5.09

1.50

5.34

2.52

2.6

2002

75.52

9.98

3.68

1.86

7.33

1.64

4.4

2003

71.21

5.88

4.92

4.38

12.90

0.70

8.3

2004

35.72

21.70

8.03

14.07

20.37

0.11

17.4

2005

38.52

31.57

5.24

6.37

18.24

0.05

27.5

2006

37.01

34.40

1.77

3.05

22.90

0.87

25.5

2007

41.56

22.51

0.18

0.91

33.90

0.94

8.4

Share of total

46.5

24.4

4.1

5.6

18.8

0.6

100

NOTE: The table shows the share (percentage) of Alt-A product types by vintage (year of origination). FRM, fixed-rate mortgages;
ARMq mortgages are defined as the hybrid-ARM where the first reset occurs after q years. Remaining mortgage types are classified
as Other.

Table 2
Evolution of Subprime Securitized Mortgages (Percent Market Share by Product Type)
Mortgage type
Vintage

FRM

ARM1

ARM2

ARM3

ARM5

Other

Share of total

1998

51.33

8.20

1999

38.88

2.26

26.53

4.52

0.25

9.17

2.6

29.34

19.21

0.50

9.81

3.8

2000

32.58

2001

31.70

1.20

43.29

14.78

0.56

7.59

4.1

0.51

48.69

12.44

0.54

6.13

5.1

2002
2003

28.37

0.60

54.84

12.62

1.16

2.42

7.7

33.57

0.45

52.60

11.37

1.20

0.81

12.9

2004

23.81

0.35

59.73

14.64

1.30

0.17

19.5

2005

18.66

0.54

65.48

13.22

1.57

0.53

23.0

2006

19.98

0.82

62.56

10.86

1.35

3.44

18.1

2007

27.59

0.45

50.23

9.92

5.87

5.94

3.4

Share of total

25.7

0.8

56.7

12.7

1.6

2.5

100

NOTE: The table shows the share (percentage) of subprime product types by vintage (year of origination). FRM, fixed-rate mortgages;
ARMq mortgages are defined as the hybrid-ARM where the first reset occurs after q years. Remaining mortgage types are classified
as Other.

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

Figure 2
Percentage Share by Total Alt-A Origination by Purpose
Percent
80

Purchase
Refinance (cash out)

Refinance (no cash out)

60
50
40
30
20
10
0
1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

NOTE: The figure shows the monthly trends of share Alt-A originations by purpose in percentages between 1998 and 2007.

than 1 percent of the total in 1998-99 to 33 percent
of the total by 2007. Further research is needed to
determine the causes behind the sudden switch
from FRMs to ARMs after 2003 and the increase in
share of hybrid ARMs in the 2004 and 2005 vintages. This paper presents data on only Alt-A mortgages, which are then compared with the trends
in subprime originations over the same period.11
It is important to point out that most hybridARM Alt-A originations are ARM5 products that
originated after 2003. Therefore, most reset dates
for surviving mortgages in this pool have not yet
arrived at the time of this writing. In contrast, the
majority of subprime hybrid-ARM originations
were ARM2 and ARM3 products, which are currently past their reset dates. Therefore, unless
these products are refinanced earlier, rate resets
can adversely affect repayment behavior and
increase future delinquency rates on surviving
Alt-A originations.
11

Data on the summary trends on subprime mortgages are not
presented here but are available on request.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Figure 2 shows the monthly trends in share
of Alt-A originations by purpose (purchase or
refinance). Purchases make the largest category of
Alt-A originations, but their proportion fluctuates
over the months in our sample period. At their
peak in June 2000, purchases accounted for 74 percent of Alt-A originations. However, their fraction
drops to 34 percent in February 2003. This movement might be explained by the refinancing behavior of households. From 2000 to 2004, the Federal
Reserve adopted an expansionary monetary policy.
To the extent this translated to lower mortgage
rates on Alt-A products, households would choose
to refinance existing mortgages for lower rates.12
While fluctuations in the proportion of nocash-out refinances and purchases might be
explained in terms of mortgage rates, this pattern
does not hold for cash-out refinances. Perhaps the
12

Indeed, prime mortgage rates fell from 8.29 percent in June 2000
to 5.84 percent in March 2004. Individual mortgage rates on Alt-A
loans are tailored to specific borrower and loan attributes. To the
best of our knowledge, there is no known universal contract rate
for Alt-A mortgages.

J A N UA RY / F E B R UA RY

2010

Sengupta

Figure 3
Percentage Share by Total Alt-A Origination by Occupancy
Percent
100

Owner-Occupied
Non–Owner-Occupied (investor)

Second Home
80
70
60
50
40
30
20
10
0
1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

NOTE: The figure shows the monthly trends of share Alt-A originations by occupancy in percentages between 1998 and 2007.

more remarkable trend in Figure 2 is an increase
in the share of cash-out refinances since 2000.
A similar pattern is observed for subprime mortgages as well. In short, the growth in non-prime
mortgages after 2000 has been fueled largely by
households seeking to extract home equity during
a period of appreciating home prices. With the
decline in home prices and the onset of the mortgage crisis, the Fed lowered rates after the second
quarter of 2007. At the same time, the share of both
cash-out refinances and purchases fell sharply,
while that for no-cash-out refinances increased.
In terms of occupancy, most Alt-A originations
were for owner-occupied properties as shown in
Figure 3. The share of owner-occupied housing
increased from a little over 60 percent at the beginning of our sample period to more than 80 percent
toward the end of the sample period. During the
same time, the share of Alt-A second-home originations fell by half: from nearly 35 percent in 1998
to around 17 percent by the end of 2006. The fraction of non–owner-occupied housing has been
60

J A N UA RY / F E B R UA RY

2010

small throughout the sample period. The broad
pattern of shares in each occupancy category has
been similar to that for subprime mortgages. For
example, owner-occupied houses have accounted
for the significant majority (more than 90 percent)
of subprime originations for most of our sample
period. Consequently, the share of second homes
in the subprime category has been much smaller
than that for Alt-A.
Figure 4 shows a sharp increase in the share
of low-doc loans in post-2004 Alt-A originations.
Barring a few exceptions, the share of low-doc
originations has always ranged between 50 and
60 percent of originations until 2004. To the casual
observer this figure may seem very high. But this
is precisely the rationale behind the creation of
the Alt-A market: borrowers of good credit quality
unwilling or unable to provide full documentation for a prime loan. In any case, the share of lowdoc loans rose from 52 percent in April 2004 to
78 percent by the end of 2006. A similar trend
toward low-doc originations was witnessed for
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

Figure 4
Percentage Share by Total Alt-A Origination by Documentation
Percent
90
80

Full Doc
Low Doc
No Doc

70
60
50
40
30
20
10
0
1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

NOTE: The figure shows the monthly trends of share Alt-A originations by documentation in percentages between 1998 and 2007.

subprime originations, although the increase
there was moderate and gradual compared with
that of Alt-A. It is unclear what prompted this
sharp increase in low-doc originations in the nonprime market segment. More recently, the share
of low-doc mortgages dropped sharply from its
peak of 78 percent at the beginning of 2007 to 50
percent at the end of 2007.
An important measure of underwriting is the
credit quality on the originations as represented
by the credit (FICO) scores of borrowers at the
time of origination. The majority of borrowers
who originate Alt-A mortgages have FICO scores
in excess of 680 (Figure 5). This is a major distinguishing characteristic between the subprime and
Alt-A mortgage pools. The average credit quality
of Alt-A pools is significantly higher than that for
subprime pools. It needs to be mentioned here
that this is not the only distinguishing characteristic; it is often possible to identify a mortgage
that belongs to the subprime pool but has a FICO
score above 700. The reason a mortgage with a
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

high FICO score could be characterized as subprime (and not Alt-A) is possibly because the
mortgage fails to qualify as Alt-A (or even prime)
on one or more criteria other than credit score
such as documentation, lien type, and LTV ratio.13
Figure 5 shows that, for the most part, the
shares of originations with FICO scores in the
621-680 range and the 740 or higher range have
been similar. The share of originations with FICO
scores above 740 fell for a period between October
1998 and December 1999; this was accompanied
by a rise in the share of originations in the 621-680
range. More recently, the onset of defaults in nonprime mortgages tightened lending standards in
this market, leading to a sharp increase in the
percentage of originations with FICO scores in
excess of 740 from around 26 percent in December
2006 to 54 percent in November 2007. To sum up,
Figure 5 shows that, except for the two periods
mentioned above, the share of originations across
13

The choice of a non-prime mortgage between Alt-A and subprime
typically varies with the arranger of the security.

J A N UA RY / F E B R UA RY

2010

Sengupta

Figure 5
Percentage Share by Total Alt-A Origination by Credit Quality (FICO Score)
Percent
60

FICO ⱕ 620
FICO 620-680
FICO 681-740

FICO ⱖ 741
40

0
1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

NOTE: The figure shows the monthly trends of share Alt-A originations by credit quality in percentages between 1998 and 2007.

the various FICO score categories are fairly stable
over the years in our sample period.
At this point, it is important to highlight the
difference between the originations of subprime
and Alt-A loans between 2000 and 2007 with
regard to documentation and credit scores. In both
cases, the share of low-doc loans increased over
the years, as shown in Figure 6. However, in the
case of the subprime market, there is evidence
that average credit scores on originations with
lower documentation increased. This is shown
in Figure 6B as the decline in the proportion of
low-doc originations with FICO scores less than
620. This feature of underwriting suggests that
lenders’ emphasis on FICO score was not only
an adequate indicator of credit risk, but also a
means to adjust for other riskier attributes on the
origination. On the other hand, there does not
appear to be such a trend toward higher FICO
scores for loans with low documentation in the
case of the Alt-A mortgage market. The proportion
of low-doc loans with FICO scores less than 680
remains roughly the same over the years in our
62

J A N UA RY / F E B R UA RY

2010

sample period (Figure 6A).14 This pattern seems
to point to secular deterioration in the underwriting for Alt-A mortgages, unlike that observed for
subprime originations.
A final measure of underwriting on Alt-A
originations in our dataset is the LTV ratio on
the mortgage (Figure 7).15 The majority of Alt-A
originations have LTV ratios that are less than 80
percent, and it is important to note that the LTV
threshold of 80 percent is one of the requirements
on prime mortgages. Figure 7 shows that the share
of originations with LTVs less than or equal to 80
percent has declined over the years in our sample
period. This is accompanied by an increase in the
14

The FICO scores chosen are higher for Alt-A because on average
Alt-A credit scores are higher than subprime. The weighted average
of FICO scores for the Alt-A market is presumably higher, but we
have chosen our cutoff conservatively.

The LTV ratio is calculated as the closing balance/value of the
property, and where available we have used the cumulative loanto-value (CLTV) ratio because it provides a better measure of the
home equity of the borrower. The CLTV ratio is the proportion of
loans (secured by the property) on all liens in relation to the
property’s value.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

Figure 6
Evidence on Underwriting: Documentation and FICO Score
A. Alt-A
Percent
90
80
70
60
50
40
30
20

Low Doc

10
0
2000

Low Doc with FICO < 680
2001

2002

2003

2004

2005

2006

2007

B. Subprime
Percent
70
60
50
40
30
20
Low Doc

10
0
2000

Low Doc with FICO < 620
2001

2002

2003

2004

2005

2006

2007

NOTE: The figure shows the monthly trends of share Alt-A (subprime) originations with low documentation, and among those, with
FICO scores less than 680 (620) in percentages between 2000 and 2007.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Sengupta

Figure 7
Percentage Share by Total Alt-A Origination by LTV Ratios
Percent
90

CLTV
90 ⱖ CLTV ⬎ 80

100 ⱖ CLTV ⬎ 90
CLTV ⬎ 100

70
60
50
40
30
20
10
0
1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

NOTE: The figure shows the monthly trends of share Alt-A originations by loan-to-value ratios in percentages between 1998 and 2007.

share of Alt-A originations with LTV in the (90,
100] range. The share in this category increased
from a low of 2 percent of total originations from
the beginning of our sample period to about 32
percent by the end of 2006. Meanwhile, the share
of originations with LTV in the intermediate range
of (80, 90] has remained fairly stable except for a
period between 1999 and 2001 when this share
increased.
In summary, Alt-A mortgages are typically
originated to borrowers of moderate to high credit
quality with a lack of willingness or ability to provide documentation in support of their loan application. First, most Alt-A originations have FICO
scores above 680. At the same time, the share of
low-doc originations in this market has almost
never been below 50 percent. While this has been
the principal characteristic of Alt-A loans, the
market witnessed a significant relaxing of credit
standards both in terms of a greater share of lowdocumentation loans and high-LTV originations
between 2000 and 2006. Perhaps more signifi64

J A N UA RY / F E B R UA RY

2010

cantly, the share of borrowers using Alt-A products to extract equity in their homes has almost
doubled between 2000 and 2006. In the next section, we study the performance of these mortgages
in terms of the attributes on the originations.

Loan Performance of Alt-A Mortgages
The LP data allow for tracking repayment
behavior on mortgages on a monthly basis. Therefore, we can determine the nature (30-day, 60-day,
90-day, or foreclosure) and timing (month) of the
delinquency event. Following industry conventions, we define a mortgage to be in default (or in
serious delinquency) if it records a 90-day delinquency event at any point in its repayment history.16 The nonparametric default probabilities
presented in this paper are calculated using the
Kaplan-Meier product limit estimator (see the
appendix for details on this methodology).
16

Although we use 90-day delinquencies throughout the paper, the
results for 60-day delinquencies and foreclosures are qualitatively
similar and are available on request.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

Figure 8
Default Rate on Alt-A Originations by Vintage
Percent
40

2000
2001
2002
2003
2004
2005
2006
2007

35
30
25
20
15
10
5
0

10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70

NOTE: The plot shows the Kaplan-Meier default probabilities by loan age for securitized Alt-A mortgages. The graph presents the
default probabilities by years of origination (vintage). Each line shows the performance of originations of the same vintage.

Figure 8 presents the overall performance of
Alt-A mortgages by showing estimated default
probabilities for each vintage (year of origination)
by age of the loan. The broad trends in Figure 8
show that defaults started to rise sharply in 2006
and 2007, primarily for originations after 2003.
To give an example, about 10 percent of mortgages
originated in 2001 were in serious delinquency
after the third calendar year (at the beginning of
2003), whereas the same proportion of defaults
for 2006 originations occurred after the first oneand-a-half calendar years (middle of calendar
year 2007).17
The rise in sharp defaults for later vintages is
best viewed by comparing the default rates on
originations of 2003 and 2004 vintages with that
17

The year of origination is counted as the first year of evaluation of
loan performance. In the interest of clarity, the performance plots
for 1998 and 1999 vintages are omitted from Figure 8.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

of originations of 2005 and 2006 vintages. Table 3
shows the default rates on originations of 2003 and
2004 vintages at the end of the second calendar
year were 2.03 percent and 2.47 percent, respectively. In contrast, originations of 2006 vintage
had a default rate of 16.36 percent by the end of
the second calendar year. To summarize, defaults
rise sharply around 2006 and this is largely concentrated on originations after 2003. Perhaps the
most striking feature of this trend is that a significant proportion of the mortgages defaulted very
early. This is also true for subprime mortgages,
and the literature on subprime has focused on
explaining such high early defaults.
An interesting piece of anecdotal evidence is
revealed in the significantly lower default rates on
2003 vintages. Indeed, 2003 is the best-performing
vintage for Alt-A mortgages and this is true for
subprime mortgages as well. For subprime originations, the anomalous behavior for originations
J A N UA RY / F E B R UA RY

2010

Sengupta

Table 3
Performance of 2003-2006 Alt-A Vintages for the First Three Calendar Years
Default rate (%) per year of origination
Calendar date

2003

2004

2005

2006

End of 2003

0.64

—

End of 2004

2.03

0.63

—

End of 2005

3.55

2.45

0.97

—

End of 2006

—

4.86

4.67

2.66

End of 2007

—

13.84

16.36

End of 2008

—

36.60

The table shows the default rates on Alt-A mortgages originated between 2003 and 2006 at the end of the first three calendar years.

Table 4
90-Day Delinquent Alt-A Mortgages (Percent by Attributes on Origination)
Product type
All

FRM

ARM

Purpose

ARM5

Occupancy

Refinance Refinance
Purchase (cash-out) (no cash-out) Owner

Second
home

Non-owner
(investor)

A. Percent delinquent after first 18 calendar months
2000

3.1

3.5

0.4

0.7

3.3

2.6

3.1

3.3

1.5

2.6

2001

3.1

2.7

1.2

3.9

2.3

2.0

3.0

1.8

3.9

2002

2.7

2.4

1.5

3.4

2.2

1.5

2.8

2.1

2.2

2003

1.3

1.2

1.7

1.0

1.9

1.1

0.7

1.6

1.0

0.7

2004

1.5

1.4

1.5

1.0

1.7

1.2

1.3

1.6

0.9

1.3

2005

2.4

1.9

2.8

3.2

2.9

1.7

2.2

2.4

1.8

2.6

2006

7.9

5.6

9.3

12.9

9.5

5.3

9.0

8.0

6.7

7.8

2007

13.1

8.4

16.6

18.8

16.6

8.9

14.1

13.3

13.9

11.8

B. Percent delinquent after two calendar years
2000

5.0

5.5

1.1

2.2

5.1

4.1

5.1

4.8

2.7

4.4

2001

5.1

5.2

4.3

2.3

6.3

3.9

3.5

4.7

3.1

5.9

2002

4.3

4.5

3.4

2.3

5.4

3.7

2.6

4.3

3.2

3.5

2003

2.0

1.8

2.5

1.5

2.9

1.7

1.1

2.3

1.4

1.2

2004

2.4

2.3

2.6

1.8

2.7

2.1

2.3

2.4

1.5

2.2

2005

4.7

3.3

5.7

5.9

5.5

3.5

4.3

4.2

3.8

4.9

2006

16.4

10.7

19.9

24.2

18.5

12.4

18.8

15.0

14.8

15.6

2007

24.0

15.7

30.3

33.0

28.6

18.2

26.0

22.4

24.0

22.0

The table shows the percentage of originations of a given vintage that are in default within a given time period across various attributes
on the origination, including product types, purpose, and occupancy.

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

Table 5
90-day Delinquent Alt-A Mortgages (Percent by Attributes on Origination)
Documentation
All

Full

Low

Credit score
621-680

CLTV ratio

681-740

740+

<70

[70, 80)

(80,90]

(90,100]

A. Percent delinquent after first 18 calendar months
2000

3.1

2.4

3.6

5.1

2.1

1.2

1.3

2.5

5.1

4.2

2001

3.1

1.9

3.9

5.5

2.0

1.1

1.3

2.4

5.5

5.2

2002

2.7

1.9

3.3

4.9

1.8

0.8

0.9

2.2

4.4

5.1

2003

1.3

0.9

1.6

2.8

0.9

0.4

0.9

2.3

3.1

2004

1.5

1.1

1.7

2.7

1.2

0.5

1.0

2.2

2.7

2005

2.4

1.6

2.8

4.3

2.2

0.9

0.7

1.5

2.8

4.8

2006

7.9

3.6

9.1

12.1

7.8

3.4

1.7

4.8

9.5

14.1

2007

13.1

5.8

14.9

19.3

14.0

6.4

3.0

9.3

18.1

23.6

B. Percent delinquent after two calendar years
2000

5.0

4.1

5.5

7.8

3.6

2.0

2.2

4.1

7.9

6.4

2001

5.1

3.4

6.3

8.7

3.5

1.8

2.2

4.0

8.5

8.7

2002

4.3

3.3

5.1

8.0

3.0

1.2

1.5

3.6

6.9

8.1

2003

2.0

1.5

2.3

4.3

1.5

0.6

1.5

3.6

4.6

2004

2.4

1.9

2.8

4.5

2.0

0.9

1.0

1.7

3.6

4.2

2005

4.7

2.9

5.5

8.2

4.4

1.8

1.4

3.1

6.1

8.7

2006

16.4

7.7

18.7

23.6

16.5

7.7

4.2

11.5

21.7

25.4

2007

24.0

11.5

27.0

34.2

25.6

12.3

7.1

19.2

33.3

37.9

The table shows the percentage of originations of a given vintage that are in default within a given time period across various attributes
on the origination, including documentation, credit score, and CLTV ratio.

of 2003 vintage has been explained in terms of the
high prepayment rates on subprime mortgages.
As many as 83 percent of surviving subprime
hybrid-ARMs that were originated in 2003 were
prepaid by the end of 2007 (see Bhardwaj and
Sengupta, 2009b). This is not surprising for subprime mortgages, given that prepayment is an integral part of the mortgage design for hybrid-ARM
products (see Gorton, 2008, for details). However,
hybrid-ARMs are not a significantly large part of
the Alt-A pool. Therefore, it would be interesting
to explore whether the low default rates on 2003
Alt-A products were also driven by high prepayments. In what follows, we show that this broad
trend of a significant increase in the default rates
on post-2004 originations can be seen across various mortgage attributes such as product type,
purpose, occupancy, and documentation. These
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

trends show a high degree of correlation between
default rates and some origination attributes. Of
course, the standard caveat applies to interpreting
these correlations as causation.
Tables 4 and 5 show the percentage of originations of a given vintage that are in default within
a given time period, by various attributes of the
origination. Panel A presents the default rates for
the first 18 calendar months since the year of
origination, whereas Panel B reports the same for
the first two years since the year of origination.
These choices of time periods are driven by two
reasons. First, we have only the first two years
of data for every origination vintage from 2000
through 2007, allowing for a comparison across
all vintages. Second, as demonstrated earlier, the
crisis in the mortgage markets was characterized
by high early defaults.
J A N UA RY / F E B R UA RY

2010

Sengupta

For most years in our sample period, ARMs
have registered higher default rates than FRMs,
but the difference was much greater for originations between 2005 and 2007 (columns 3 through
5 in Table 4). For 2003, however, the performance
of ARMs is similar to that of FRMs. The lower
default rates on ARMs for earlier vintages might
be explained by the low interest rate environment
during the early part of this decade. However, as
the Federal Reserve tightened monetary policy
after the second half of 2004, the burden of interest
payments on ARMs would have increased significantly. Also, the share of ARMs in total originations for earlier vintages was low compared with
the share for later vintages. Therefore, it is difficult to interpret the default patterns as being reflective of the risk underlying each product type.
Evidently, the default rates on ARM5 products18
are even higher than those on ARM products overall. This is interesting, given that the loan maturity
period under consideration is well before the
reset dates on the ARM5 products. These results
seem to suggest that the defaults on Alt-A products have little to do with interest rate resets on
hybrid-ARM products.
Next, we study the default patterns by purpose
of origination (columns 6 through 8 in Table 4).
Purchase originations show significantly higher
rates of default over the years in our sample
period. This may be attributed to greater adverse
selection problems for first-time buyers than for
refinances, where the borrower is likely to have
had a recorded history of mortgage payments, presumably with the same lender. Here, too, default
rates rise significantly for originations after 2005.
Under occupancy, we find that non–owneroccupied homes have the highest default rates,
followed by second homes, while owner-occupied
homes have the lowest default rates (columns 9
through 11 in Table 4). Anecdotal evidence often
points to the role of investors using non-prime
mortgage products to speculate on residential
property after 2004. This has been claimed as a
proximate cause of the mortgage crisis in the
United States. Of course, this would also explain
18

The choice of ARM5 is motivated by the fact that, among Alt-A
originations with hybrid products, the ARM5 product has the
largest market share.

J A N UA RY / F E B R UA RY

2010

the deterioration of lending standards and the
high early default rates on originations after 2004.
However, the summary data presented above
show little evidence in support of this hypothesis.
The proportion of second homes had been declining over the sample period. Moreover, non–owneroccupied properties were a small fraction of the
loans throughout the sample period.
Not surprisingly, low-doc originations show
a higher rate of default than full-doc loans
(columns 3 and 4 in Table 5). Given the higher
default rates on such loans, even for the earlier
vintages, it is surprising to see the increasing
share of Alt-A originations after 2004 that do not
provide full documentation. To most observers
this would bring into question the wisdom of
originators who increased the proportion of lowdoc loans in their mortgage pools. However, as
noted previously, Alt-A mortgages are originated
in an effort to capture borrowers who have good
credit but are otherwise unable to provide documentation on their loans. Moreover, it is difficult
to interpret the lack of documentation on loans as
the principal cause behind the high default rates
in the Alt-A market. For example, more than half
of the originations in 2003 were loans without
full documentation. However, the difference in
the default rates on full-doc and low-doc loans
for this vintage was less than 1 percentage point
even after three calendar years.19
Next, we turn our attention to default rates in
terms of credit quality as measured by borrower
FICO at the time of origination (columns 5 through
7 in Table 5). A number of observers have pointed
to higher default rates on a given FICO score as
an indication of the poor performance of FICO.
However, one needs to approach this argument
with caution. For instance, if some exogenous
factor were driving defaults in the mortgage market, one is likely to see poor performance for the
said vintages across all FICO score groups. This
is precisely what we observe in the data. A more
relevant test of the effectiveness of FICO would
19

Among originations of 2003 vintage, the default rate after three
calendar years on full-doc loans was 2.82 percent, whereas the
default rate on low-doc loans was 3.79 percent. The comparable
figures for originations of 2006 vintage were 19.03 and 39.18 percent, respectively.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

be a comparison across the different FICO score
groups for a given vintage. The three panels show
the performance of Alt-A by FICO score groups of
621-680, 681-740, and higher than 740.20 Within
each FICO score group, the later vintages (post2004 originations) show higher default rates. More
importantly, across the various score groups, the
default rates for the same vintage are higher as one
moves from a higher FICO score group to a lower
FICO group.
Lastly, we study the effect of LTV on the summary measures of default (columns 8 through 11
in Table 5). Default probabilities increase when
one moves from originations with lower LTV ratios
to those with higher LTV ratios. Note that the loan
performance for the 80 to 90 percent and the 90
to 100 percent LTV ratio categories are somewhat
similar, especially for the early vintages.21 However, the gap in the default rates widens on later
vintages.
In summary, Tables 4 and 5 confirm our a
priori knowledge on underwriting. First, riskier
attributes such as lower documentation, lower
FICO scores, and higher LTV ratios perform poorly.
Second, the differences in default probabilities
between a more-risky attribute and that of a lessrisky attribute increase for originations after 2004.
Third, there is no monotonic trend over the years
in the default rates across these attributes, however. Default rates have typically fallen for 2003
and 2004 originations, but risen sharply for later
vintages. Finally, even for the later vintages, the
defaults have risen across all attributes, irrespective of ex ante risk on the attribute. These summary
results emphasize that one must exert caution
when interpreting the riskier attributes on the
origination as proximate causes of high early
defaults on Alt-A mortgages in 2006 and 2007.

We do not report default rates for the FICO score group less than
620, as their share throughout has been small and declining. In
fact, fewer than 1 percent of post-2004 originations in the Alt-A
market have FICO scores less than 620.

Again, since they form a small share of the total market, the plots
for Alt-A originations with LTV in excess of 100 are not reported
here.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

CONCLUSION
This paper provides a preliminary overview
of Alt-A mortgages that were originated in the
United States from 1998 through 2007. First, the
summary data indicate a shift of Alt-A originations toward a greater share of owner-occupied
properties, adjustable-rate products, and cash-out
refinances. This is accompanied by a deterioration
of underwriting standards for a greater proportion of mortgages with lower documentation and
higher loan-to-value ratios. Serious delinquencies
on Alt-A originations rose sharply in 2006 and
2007, primarily for originations after 2003. Even
for originations of a later vintage, the defaults have
risen across all attributes, irrespective of ex ante
risk on the attribute.
A final comment addresses the following
question: How does one reconcile the secular
deterioration of underwriting for Alt-A mortgages
with the lack of this evidence in the case of subprime mortgages (see Bhardwaj and Sengupta,
2009a)? In their handbook chapter on Alt-A mortgages, Bhattacharya, Berliner, and Liber (2006,
p. 189) remark that “the demarcation between
Alt-A and subprime loans has been blurred. Over
time Alt-A has expanded to include loans with
progressively less documentation and lower borrower credit scores. At the same time, subprime
loans have, on average experienced a slow but
steady rise in average credit scores. A result of
this convergence has been the creation of the
so-called Alt-B sector, where loans using this
nomenclature were securitized in 2004.”

J A N UA RY / F E B R UA RY

2010

Sengupta

REFERENCES
Ashcraft, Adam B. and Schuermann, Til. “Understanding the Securitization of Subprime Mortgage Credit.”
Staff Reports No. 318, Federal Reserve Bank of New York, March 2008;
www.newyorkfed.org/research/staff_reports/sr318.pdf.
Bhardwaj, Geetesh and Sengupta, Rajdeep. “Where’s the Smoking Gun? A Study of Underwriting Standards for
US Subprime Mortgages.” Working Paper 2008-036C, Federal Reserve Bank of St. Louis, October 1, 2009a;
http://research.stlouisfed.org/wp/2008/2008-036.pdf.
Bhardwaj, Geetesh and Sengupta, Rajdeep. “Did Prepayments Sustain the Subprime Market?” Working Paper
2008-039B, Federal Reserve Bank of St. Louis, May 2009b; http://research.stlouisfed.org/wp/2008/2008-039.pdf.
Bhattacharya, Anand K.; Berliner, William S. and Lieber, Jonathan. “Alt-A Mortgages and MBS,” in Fabozzi, Frank,
ed., The Handbook of Mortgage-Backed Securities. Sixth edition. New York: McGraw-Hill, 2006, pp. 187-206.
Fabozzi, Frank J., ed. The Handbook of Mortgage-Backed Securities. Sixth edition. New York: McGraw-Hill, 2006.
Gorton, Gary. “The Panic of 2007.” Manuscript prepared for the Federal Reserve Bank of Kansas City, Jackson
Hole Conference, August 4, 2008; www.kc.frb.org/publicat/sympos/2008/gorton.08.04.08.pdf.
Kaplan, E. and Meier, P. “Nonparametric Estimation from Incomplete Observations.” Journal of the American
Statistical Association, 1958, 53(282), pp. 457-81.
Lehnert, Andreas. “Residential Mortgages,” in Allen Berger, Phillip Molyneux, and John Wilson, eds., Oxford
Handbook of Banking. Oxford, UK: Oxford University Press, 2009.
Mayer, Christopher K. and Pence, Karen. “Subprime Mortgages: What, Where, and to Whom?” Working Paper
No. W14083, National Bureau of Economic Research, June 2008; http://papers.nber.org/papers/w14083.pdf.
Quigley, John M. “Federal Credit and Insurance Programs: Housing.” Federal Reserve Bank of St. Louis Review,
July/August 2006, 88(4), pp. 281-309; http://research.stlouisfed.org/publications/review/06/07/Quigley.pdf.
Sengupta, Rajdeep and Tam, Yu Man. “Mortgage Originations: 2000-2006.” Federal Reserve Bank of St. Louis
Economic Synopses, 2008, No. 18; http://research.stlouisfed.org/publications/es/08/ES0818.pdf.

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Sengupta

APPENDIX
Default rates and the probability of surviving a delinquency are calculated by using the Kaplan and
Meier (1958) product limit estimator. We begin this nonparametric approach to survival and hazard
function estimation by formalizing it in the current context of mortgage repayment behavior.
Following Kaplan and Meier (1958), the delinquency rate D共t兲 at month t (the age of the mortgage
in months) is defined as

D (t ) = 1 − P (T > t ),
where T is the age in months for the delinquency event (60-day, 90-day, or foreclosure) of a randomly
selected mortgage and S共t兲 ⬅ P共T > t兲 is the survivor function or the probability of surviving the delinquency event beyond age t. Let t共1兲 < t共2兲 < … < t共k兲 represent the ordered age in months at the time of the
delinquency event. For all these months, let n共i 兲 be the number of surviving mortgages just prior to
month t共i兲. Surviving mortgages exclude not only the ones that have been delinquent, but also the ones
that have been refinanced prior to age t共i兲. If d共t兲 is the number of mortgages that become delinquent at
age t共i兲, then the Kaplan-Meier estimator of surviving the event of delinquency is defined as
⎛
d ⎞
P̂ (T > t ) = ∏ ik=1 ⎜ 1 − i ⎟ .
ni ⎠
⎝

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

The Relationship Between the Daily and
Policy-Relevant Liquidity Effects
Daniel L. Thornton
The phrase “liquidity effect” was introduced by Milton Friedman (1969) to describe the first of
three effects on interest rates caused by an exogenous change in the money supply. The lack of
empirical support for the liquidity effect using monthly and quarterly monetary and reserve aggregates data led Hamilton (1997) to suggest that more convincing evidence of the liquidity effect
could be obtained with daily data—the daily liquidity effect. This paper investigates the implications of the daily liquidity effect for Friedman’s liquidity effect using a more comprehensive model
of the Federal Reserve’s daily operating procedure than has been previously used in the literature.
The evidence indicates that it is no easier to find convincing evidence of a Friedman liquidity effect
using daily data than it has been with lower-frequency monthly and quarterly data. (JEL E40, E52)
Federal Reserve Bank of St. Louis Review, January/February 2010, 92(1), 73-87.

he phrase “liquidity effect” (LE) was
first used by Milton Friedman (1969)
to describe the first of three effects on
interest rates caused by an exogenous
change in the supply of money.1 Despite its
prominent role in conventional theories of the
monetary policy transmission mechanism, there
has been little evidence of a statistically significant or economically meaningful LE.2 Suggesting
that previous attempts to identify the LE have
been unsuccessful because low-frequency data
mix the effects of policy on economic variables
with the effects of economic variables on policy,
Hamilton (1997) sought to develop a “more con1

The other two effects are the “income” and “price expectation” or
“inflation expectation” effects (e.g., Friedman, 1969; and Gibson,
1970a,b). These effects have roots in classical economics (e.g.,
Humphrey, 1983a,b). Because of the inflation expectation effect,
an exogenous change in money growth eventually leads to higher,
rather than lower, equilibrium nominal interest rates.

The empirical literature on the LE dates back at least to Cagan and
Gandolfi (1969) and Gibson (1970a,b).

vincing measure of the liquidity effect” by estimating the response of the federal funds rate to
exogenous reserve supply shocks using daily
data. This is referred to as the “daily liquidity
effect” (DLE). Thornton (2001a) showed that (i)
Hamilton’s estimates of the DLE were the consequence of a few extreme observations and (ii)
there was no evidence of a DLE using Hamilton’s
model and methodology for his sample period
and for sample periods before or after that period.
Recently, however, Carpenter and Demiralp (2006)
report evidence of a DLE using a more complete
model of the operating procedure of the Trading
Desk of the Federal Reserve Bank of New York
(hereafter, the Desk) than that used by Hamilton.
They also use a reserve supply shock measure
that more adequately reflects reserve supply
shocks that the Desk creates each day in the conduct of open market operations.
Carpenter and Demiralp (2006) and Hamilton
(1997) claim that estimates of the DLE provide

Daniel L. Thornton is a vice president and economic adviser at the Federal Reserve Bank of St. Louis. The author thanks Jim Hamilton and
Sherry Edwards for useful comments and Jonathan Ahlbrecht, John McAdams, Daniel McDonald, and Aaron Albert for helpful research
assistance.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Thornton

evidence of the existence of the Friedman LE. I
argue that evidence of a DLE need not provide evidence of the existence of Friedman’s LE. Specifically, I analyze the relationship between the DLE
and Friedman’s policy-relevant LE. The analysis
shows that because of specific features in the Fed’s
operating procedure, its system of reserve requirements, and other factors, the relationship between
the DLE and Friedman’s LE is neither simple nor
direct. In particular, statistically significant estimates of the former do not necessarily imply the
existence of the latter. In so doing, I estimate the
DLE using (i) Carpenter and Demiralp’s (2006)
reserve shock measure and (ii) a more complete
model of the Fed’s daily operating procedure than
that used by either Hamilton (1997) or Carpenter
and Demiralp. The empirical evidence indicates
that it is no easier to find convincing evidence of
Friedman’s LE using high-frequency daily data
than it has been using monetary and reserve aggregates at monthly or quarterly frequencies.
The remainder of the paper is divided into
three sections. The upcoming section investigates
the relationship between the DLE and Friedman’s
LE using a detailed model of the Desk’s operating procedure. Following the literature, in the
next section I develop estimates of an exponential autoregressive conditional heteroskedasticity
(EGARCH) model of the DLE based on the model.
The empirical model uses Carpenter and
Demiralp’s (2006) reserve supply shock measure.

THE POLICY-RELEVANT AND
DAILY LIQUIDITY EFFECTS
Milton Friedman (1969) termed the first of
three effects of an exogenous change in the supply
of money on nominal interest rates the “liquidity
effect.” Friedman’s LE is relevant for monetary
policy. Consequently, Friedman’s LE is called the
“policy-relevant liquidity effect” (hereafter, LE).
To understand why the DLE need not imply the
existence of the LE, it is important to understand
the mechanism that links the DLE to the LE. In
this regard, it is important to note that the LE stems
from the demand for money; that is,
74

J A N UA RY / F E B R UA RY

2010

(1)

M td = f ( it , y t ),

where M td denotes the demand for money, which,
for purposes of illustrating the relationship
between the DLE and LE, is assumed to be a simple
function of a nominal interest rate, i, and nominal
income, yt. Because individuals tend to economize
their holding of money when interest rates rise,
∂ f /∂ i < 0.
Equilibrium requires that the supply of
money, M ts (which, for simplicity, is assumed to
be exogenously controlled by the Fed), equals
demand; that is,
(2)

M ts = M td .

The LE is the initial effect of an exogenous change
in the money supply on the interest rates and is
given by
(3)

dit dM s = ( ∂f ∂i ) ,
−1

where it is assumed that neither nominal income
nor inflation expectations respond immediately
to the Fed’s actions. Friedman (1969) called equation (3) the “liquidity effect.”
Considerable empirical evidence indicates
that the demand for money is negatively related
to the interest rate and interest inelastic. The interest inelasticity of money demand implies that a
small exogenous change in the supply of money
should cause a relatively large response in interest rates—a relatively large LE. Consequently,
the inability of researchers to find a statistically
significant and economically meaningful LE is
referred to as the “liquidity puzzle.”3
The failure to find the LE using low-frequency
monetary and reserve aggregates has been attributed to several factors, such as the response of
nominal income or inflation expectations to
money supply shocks and the inability to isolate
exogenous monetary shocks. Researchers have
attempted to overcome these problems using,
among other things, structural vector autoregressions (SVARs). The recursive SVAR, or RSVAR,
has been particularly popular in this literature.
SVAR models have been estimated using a vari3

See Strongin (1995).

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Thornton

ety of monetary and reserve aggregates. Pagan
and Robertson (1995) show that it is difficult to
find convincing evidence of the LE with these
models.4
The inability of researchers to find evidence
of the LE using monthly or quarterly data led
Hamilton (1997) to suggest that the failure of the
RSVAR approach likely stemmed from the fact
that changes in Fed policy are frequently due to
information about “current or future values of output, inflation, exchanges rates, or other magnitudes,” so that “the correlation between such a
‘policy innovation’ and the future level of output
of necessity mixes together the effect of policy
on output with the effect of output forecasts on
policy.”5 He suggested that the LE could be more
easily identified by estimating the response of the
funds rate to reserve supply shocks measured at the
daily frequency. Specifically, he estimated reserve
supply shocks from a simple time-series model of
the Treasury’s daily deposits at the Fed. Assuming
that the errors from this model proxy the reserve
supply shocks that the Desk makes in conducting
daily open market operations, Hamilton (1997)
estimated the response of the federal funds rate
to his estimated reserve supply shocks; that is,
he estimated the DLE. He suggested that his estimates of a DLE implied the existence of the LE.

The Relationship Between the PolicyRelevant and Daily Liquidity Effects
The relationship between the DLE and the
LE is a result of the Fed’s imposition of reserve
requirements on some components of money.
This creates a direct link between the demand
for money—the source of LE—and the demand
4

The exception is using a RSVAR with nonborrowed reserves as the
monetary aggregate. Coleman, Gilles, and Labadie (1996) pointed
out, however, that evidence of an LE using nonborrowed reserves
may be a consequence of the Desk’s efforts to offset the effect of
changes in discount window borrowing. Thornton (2001b) confirmed this by showing that the estimated LE using nonborrowed
reserves is a consequence of the interest sensitivity of discount
window borrowing and the Desk’s operating procedure under
either monetary aggregate or funds rate targeting. He shows that
this “liquidity effect” using nonborrowed reserves vanishes in the
early 1980s when borrowing declined dramatically and became
relatively interest-insensitive.
Hamilton (1997), p. 80.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

for reserves—the source of the DLE. This relationship can be illustrated by assuming that the
demand for reserves is given by
(4)

)

Rtd = RR ( M td ,

where R td denotes the demand for reserves and
RR共M td 兲 denotes the Federal Reserve–imposed
system of reserve requirements, which depend
on the demand for money. Equation (4) shows
that the demand for reserves is derived from the
demand for money. Hence, in principle, it is possible to estimate the LE by estimating the response
of interest rates to an exogenous change in the
supply of reserves; that is, by estimating the DLE.
The advantage of using daily data is that measures
of reserve supply shocks at this frequency cannot
be contaminated by the endogenous behavior of
the Fed as Hamilton (1997) noted. Moreover,
since the response will be identical whether the
shock to reserves is due to an error the Desk makes
in conducting daily open market operations or is
a monetary policy–induced exogenous shock to
reserves, there is no identification problem as
there is when higher-frequency monetary and
reserve aggregates are used. It is sufficient to
identify a reserve supply shock from any source.
The strength of this relationship, however,
depends both on the Desk’s daily operating procedure, which has remained essentially the same
since at least the early 1970s, and the Fed’s system
of reserve requirements, which has not.
The Desk’s Operating Procedure. The analysis
begins with a model of the Desk’s operating procedure. Each day the Desk estimates the quantity
of reserves that banks will demand over a maintenance period ending every other Wednesday,
called “settlement Wednesday.”6 The Desk also
estimates the quantity of reserves that will be
supplied if the Desk conducts no open market
6

Until October 1979 the estimate of demand was conditional on the
objective or target for the federal funds rate. From October 1979
to September 1982, the estimate was conditional on the objective
for the growth rate of the M1 monetary aggregate. Beginning in
September 1982, the Fed claimed that the estimate was conditional
on an objective for borrowed reserves; however, Thornton (2006)
provides evidence from Federal Open Market Committee (FOMC)
transcripts suggesting that the real objective was the federal funds
rate. Today the objective is unquestionably the federal funds rate.

J A N UA RY / F E B R UA RY

2010

Thornton

operations that day.7 If the former estimate
exceeds the latter, the operating procedure suggests that the Desk add reserves through an open
market purchase. If the former is smaller than
the latter, the procedure suggests that reserves
be drained through an open market sale.
Specifically, the Desk estimates the demand
for total reserves:
(5)

(

)

E t −1TRtd = E t −1RR f ( it , y t ) + E t −1ERtd ,
TRtd

where
denotes the demand for total reserves,
ERtd denotes depository institutions’ demand for
excess reserves, and Et –1 denotes the expectation
operator conditional on information available
before that day’s open market operation.
The supply of reserves available each day is
given by
(6)

TRts = Bt + BRt + Ft + OMOt ,

where Bt denotes the Fed’s holding of government
debt before that day’s open market operation, BRt
denotes bank borrowing at the discount window,
Ft denotes autonomous factors that affect reserve
supply—currency in circulation, the Treasury’s
balance at the Fed, the float, and so on—and OMOt
denotes the amount of open market purchases or
sales conducted by the Desk that day.8
Each day the Desk estimates the supply of
reserves that will be available if the Desk conducts
no open market operations: OMOt = 0. The Desk
essentially knows the magnitude of Bt, but it must
estimate Ft. The Desk does not estimate borrowing, but rather applies the Federal Open Market
Committee (FOMC)–determined borrowing
assumption, called the initial borrowing assumption (IBAt ).9 The estimate of reserve supply if the
Desk conducts no open market operations is

where Et–1Ft denotes the Desk’s estimate of autonomous factors. The amount of the open market
operations suggested by the Desk’s operating procedure, which I call the operating procedure–
determined open market operation (ODOMOt ),
is given by

A more detailed analysis of the Desk’s operating procedure can be
found in Feinman (1993) and Thornton (2001b, 2007).

Borrowing (and later the initial borrowing assumption) refers to
seasonal plus adjustment borrowing. Extended credit borrowing
is treated separately as one of the autonomous factors affecting
reserve supply.

The initial borrowing assumption was changed relatively infrequently and, most often, when the funds rate target was changed.
Thornton (2006) shows that the initial borrowing assumption was
last mentioned in discussing monetary policy during a conference
call on January 9, 1991. However, it remained part of the Desk’s
formal operating procedure until at least 1996.

J A N UA RY / F E B R UA RY

2010

(

)

OPDOMOt = E t −1RR f ( it , y t ) + E t −1ERtd

(8)

− ( Bt + E t −1Ft + IBAt ) .

)

If OPDOMOt is positive, the procedure directs
the Desk to purchase securities; if it is negative,
the procedure indicates that securities should be
sold.
If the Desk follows its operating procedure
exactly, OMOt = OPDOMOt . The operating procedure is intended only to provide the Desk guidance, however. Judgment is used to conduct each
day’s open market operation. Indeed, over most
of the period examined here, the Desk almost
never followed the operating procedure exactly
(e.g., Thornton, 2007). To allow for this fact, let

OMOt = OPDOMOt + kt ,

(9)

where kt denotes the amount by which actual open
market operation differs from that recommended
by the operating procedure.
Reserve market equilibrium requires that the
demand for reserves equals the supply; that is,

(

)

RR f ( it , y t ) + ERtd = Bt + Ft + BRt + OMOt .

(10)

Substituting equations (8) and (9) into equation
(10) yields
(11)

E t −1TRts = Bt + E t −1Ft + IBAt ,

(7)

(

)

(

) (

RR f ( it , y t ) = E t −1RR f ( it , y t ) − ERtd − E t −1ERtd
− ( Ft − E t −1Ft ) − ( BRt − IBAt ) + kt .

)

The interest rate that equates the reserve market
is the federal funds rate, fft . Thornton (2006) shows
that the FOMC has been targeting the funds rate
to some extent since 1982.10 Consequently, the
10

See Thornton (1988, 2006) for the relevant evidence.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Thornton

Desk’s expectation of reserve demand is conditional on the FOMC’s target for the funds rate.11
Consequently, equation (11) can be rewritten as
(12)

(

)

(

)) (

RR f ( it , y t ) = RR E t −1f fft∗ , y t − ERtd − E t −1ERtd
− ( Ft − E t −1Ft ) − ( BRt − IBAt ) + kt .

)

If the reserve supply shock is given by 共Ft – Et –1Ft兲,
the DLE is given by
(13)

∂fft
1
=
< 0,
∂ ( Ft − E t −1Ft ) RR ′ ( ∂f ∂fft )

where RR′ > 0. Equation (13) shows that the relationship between the DLE and the LE depends on
the Fed’s system of reserve requirements, RR共.兲.

The Role of Reserve Requirements
Several aspects of the Fed’s system of reserve
requirements affect the relationship between the
DLE and the LE. Important among these is the fact
that reserve requirements are not imposed on all
components of money. For example, there are no
reserve requirements on the currency, and the
percentage reserve requirements are different for
various components of money.
Also, reserve requirements have changed
over time, both exogenously and endogenously.
The Fed made two major exogenous reductions
in reserve requirements during the past two
decades—in December 1990 and April 1992.12
In addition, an important endogenous reduction
in effective reserve requirements began in 1994
when banks started “sweeping” their retail transactions deposit accounts to reduce their effective
percentage reserve requirement (e.g., Anderson
11

For a more detailed explanation, see Thornton (2001b).

Effective December 13, 1990, the 3 percent reserve requirement on
non-transaction liabilities was reduced to 1.5 percent for weekly
reporters; effective December 27, 1990, the 1.5 percent reserve
requirement on non-transaction liabilities was reduced to zero for
weekly reporters. The combined effect of these actions reduced
required reserves by an estimated $13.2 billion. Although not
reported here, these changes appear to have had no important effect
on the estimates of the DLE reported in the next text section. There
have been numerous other changes in the Fed’s percentage reserve
requirements over the years; however, these were relatively small
and of little consequence.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

and Rasche, 2001). The result was a significant
reduction in effective reserve requirements and
a significant rise in the number of “nonbound”
banks—banks that satisfy their reserve requirements with vault cash.13 This change has important consequences for the relationship between
the DLE and the LE in that it severs the contemporaneous link between money demand and
reserve demand for nonbound banks.
Importantly, the Fed reintroduced lagged
reserve accounting in July 1998. Beginning with
the maintenance period that began on July 30,
1998, there is a full two–maintenance-period
(four-week) lag in the reserve accounting system.
Reserve requirements for the current maintenance
period now are determined by deposit balances
held during the 14-day period two maintenance
periods before the current one. This system of
lagged reserve accounting severs the contemporaneous link between money demand and reserve
demand for all banks, not simply nonbound banks.
Hence, there is no contemporaneous relationship
between the DLE and the LE after July 1998.
Consequently, statistically significant estimates
of the DLE after this date (e.g., Carpenter and
Demiralp, 2006; and Judson and Klee, 2009) provide no evidence of the LE. The statistically significant negative relationship between the funds
rate and reserve supply shocks merely reflects the
fact that banks have an incentive to economize
on their holdings of non-interest-bearing reserves.
This incentive exists even if the demand for
money does not depend on the interest rate,
because reserve demand is interest sensitive for
reasons other than the interest sensitivity of the
demand for money.
Finally, Thornton (2001a) has noted a two-day
lag in the Fed’s prior reserve accounting system
from March 1984 to July 1998.14 Specifically, a
bank’s maintenance-period reserve requirement
was based on deposit balances held two days
before the end of the maintenance period. The
lack of a contemporaneous relationship between
13

See Anderson and Rasche (2001) for more details on the effects of
retail sweep programs.

From 1968 to March 1984 there was a one–maintenance-period
lag in the Fed’s system of reserve accounting.

J A N UA RY / F E B R UA RY

2010

Thornton

money demand and reserve demand on those
days means that evidence of a DLE on the last two
days of the maintenance period need not imply
anything about the existence of the LE.
Analyses by Clouse and Dow (2002) and
Bartolini, Bertola, and Prati (2002), however, show
that reserve demand may be related to money
demand on the last two days of the maintenance
period if individual banks behave optimally with
respect to the reserve carryover provision.15 These
models do not include the costs of operating such
procedures, however, and these costs could be
large relative to the cost of satisfying a reserve
shortfall at the end of the maintenance period
through the discount window or in the federal
funds market.16 Consequently, it is not clear
whether such intense reserve management—
though technically feasible—is economically
viable.17 In any event, even if banks behave optimally, the relationship between the DLE and the
LE would be affected by the fact that reserve
demand on these days is due to the carryover
provision. Consequently, the extent to which
estimates of the response of the funds rate to a
reserve supply shock on the last two days of the
maintenance period provide evidence of the LE
is uncertain.

requirements applies to all components of the
money supply; for example,
(14)

RR ( M ) = rrM ,

where rr denotes a proportionate reserve requirement, say 0.10. This assumption is crude because
(i) reserve requirements do not apply equally to
all components of the money supply, (ii) rr may
differ for various components of alternative definitions of money, (iii) rr has changed over time
both exogenously and endogenously, and (iv) rr
is effectively zero with the introduction of lagged
reserve accounting in 1998 and during the last
two days of the maintenance period before the
adoption of lagged reserve accounting. Despite
these problems, to maintain comparability with
the previous literature, equation (14) is assumed.
Second, following Hamilton (1997) we assume
that money demand is a linear function of the
federal funds rate; that is,
(15) f ( ff , y t ) = β fft + α y t + ηt ,
where α and β are positive fixed parameters and
ηt denotes an i.i.d. random disturbance with a
mean of zero and a constant variance.18 With these
assumptions, equation (12) can be rewritten as
(16)
⎡ − rr β fft∗ + ( Ft + E t −1Ft )
⎤
⎢
⎥
fft = − (1 rr β ) ⎢ + ( BRt − IBAt ) − ( rr α y t − rr α y t ) ⎥ ,
⎢
⎥
⎢⎣ − ERt − E t −1ERtd + kt − ηt
⎥⎦

ESTIMATING THE DAILY
LIQUIDITY EFFECT
Hamilton (1997) and Carpenter and Demiralp
(2006) estimate the DLE using a model based on
a simpler version of equation (12). Estimating the
DLE requires several additional assumptions.
First, it requires an assumption about the Fed’s
system of reserve requirements. Effectively,
Hamilton (1997) and Carpenter and Demiralp
(2006) assume that the Fed’s system of reserve
15

I thank Jim Hamilton for pointing out this possibility to me.

For example, the one-day cost of paying a 1 percentage point premium on a $100 million dollar reserve shortfall is $2,739.73.

There is also no direct evidence that banks actually implement
such procedures. Indeed, anecdotal evidence from reserve account
managers of two very large New York banks in the late 1990s suggests that these banks did not rely on such procedures to manage
their reserves.

J A N UA RY / F E B R UA RY

2010

(

)

where ~ denotes the Desk’s estimate of the corresponding parameter or variable.
Thornton (2001a) has shown that estimates
of the DLE can give misleading indications about
the LE on days with large idiosyncratic shocks to
the funds rate. In particular, the distortion can
be large on settlement Wednesdays. Hence, special care is taken in estimating the DLE on days
with large idiosyncratic shocks to the funds rate.
18

Equation (15) assumes that the funds rate is a reasonable proxy
for the interest rate in the money demand function. However, this
need not be the case. The literature on monetary demand has
debated whether a long-term or short-term rate should be in the
money demand function and, if it is a short-term rate, which shortterm rate it should be.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Thornton

Finally, Hamilton (1997) and Carpenter and
Demiralp (2006) note that a necessary condition
for obtaining unbiased estimates of the DLE is
that reserve supply shocks be uncorrelated with
shocks to money demand, ηt . However, equation
(16) shows that the measure of reserve supply
shocks that they use (i.e, a measure of Ft – Et–1Ft )
must also be uncorrelated with BRt – IBAt , kt ,
ER td – Et –1ER td, and rrαyt – rrα~ y~t —variables not
included in Hamilton’s (1997) or Carpenter and
Demiralp’s (2006) models.
Following the literature the DLE is estimated
using an EGARCH model based on equation (16).
The EGARCH model is in the class of autoregressive conditional heteroskedastic (ARCH) models
developed by Engle (1982) and was introduced
by Nelson (1991). The specification takes the
general form
(17) fft = X t β + εt , t = 1, 2, …, T ,
where Xt denotes a 1-by-l vector of l regressors
and β denotes the corresponding l-by-1 vector of
coefficients. The errors, εt, are assumed to be
conditionally heteroskedastic. Specifically,
(18)
logσ t2 = ξ + γ

ε
εt −1
+ ψ t −1 + ς logσ t2−1 + Z t δ + ω t ,
σ t −1
σ t −1

where Zt is a 1-by-m vector of observable variables
that determine the evolution of the variance and
δ is a corresponding m-by-1 vector of coefficients.
The coefficient ψ allows for the possibility of
asymmetry in the response of shocks to the funds
rate. Because ARCH models account for heteroskedasticity, they produce estimates of β that
are generally more efficient than ordinary least
squares.19
Figure 1 presents fft and fft* over the period
January 2, 1986, through January 20, 2004. It
shows a number of volatility clusters typical of
ARCH. Some of these are associated with welldefined events, such as the marked increases in
19

However, because the EGARCH specification is not an integral part
of the model, the basic equation was also estimated with ordinary
least squares to determine whether the qualitative conclusions are
affected by using the EGARCH model. The results indicated that the
qualitative conclusions are robust to the use of the EGARCH model.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

volatility associated with the 1987 stock market
crash (bracketed by the first two vertical lines)
and the surprise reduction in reserve requirements in 1990 (bracketed by the third and fourth
vertical lines). There is also a marked decline in
volatility that appears to begin in early 2000
(denoted by the fifth vertical line), which may be
associated with changes in the FOMC’s disclosure
procedures. Moreover, it shows a relatively large
number of volatility spikes—days when the funds
rate changed by a relatively large amount only to
return to essentially its previous-day’s level the
next day. These spikes are often unique to the
funds rate. Some are associated with well-known
events (e.g., settlement Wednesday and the first
and last days of the year or quarter); others are not.
Hamilton (1996) found that a number of
dummy variables were useful in modeling the
behavior of the federal funds rate. Following
Hamilton (1997) and Carpenter and Demiralp
(2006), dummy variables are included for (i)
each of the 10 maintenance-period days (Di ,
i = 1, 2, …,10); (ii) the first and last days of the
month, quarter, and year (bom, eom, boq, eoq, eoy);
(iii) the 15th day of the month (mom); (iv) the day
before and after holidays (bh and dh, respectively);
(v) the day before and after changes in the funds
rate target (btar and atar, respectively); (vi) the
month of December (dec); and (vii) the first and
second week of the maintenance period (w1, w2).20
Dummy variables are also included for the period
of the 1987 stock market crash (d1987) and the
surprise change in reserve requirements (d1990).21
The error made by the staff of the Board of
Governors each day in forecasting Ft is the reserve
supply shock and is denoted miss.22 Separate
20

If the 15th falls on a weekend or a holiday, mom takes on the
value of 1 on the business day closest to the middle of the month.

d1987 takes on the value of 1 from the first day of the stock market
crash, October 19, 1987, through December 31, 1987, and zero elsewhere. d1990 is 1.0 from the first settlement Wednesday affected
by the changes, December 13, 1990, through February 28, 1991,
and zero elsewhere.

The Board staff’s estimate is a proxy because, in reality, the staffs
of the Board and the New York Fed make independent estimates
of the autonomous factors. The Treasury makes an independent
estimate of one of the factors, namely, its balance at the Fed. Exactly
how these estimates are combined each day in conducting open market operations is unclear. See Thornton (2004) for further details.

J A N UA RY / F E B R UA RY

2010

Thornton

Figure 1
The Effective Federal Funds Rate and the FOMC’s Funds Rate Target
(January 2, 1986–January 20, 2004)
Percent
12

1/
2
7/ /86
2/
1/ 86
2
7/ /87
2/
1/ 87
2
7/ /88
2/
1/ 88
2/
7/ 89
2/
1/ 89
2
7/ /90
2/
1/ 90
2/
7/ 91
2
1/ /91
2/
7/ 92
2/
1/ 92
2
7/ /93
2/
1/ 93
2/
7/ 94
2/
1/ 94
2/
7/ 95
2/
1/ 95
2
7/ /96
2
1/ /96
2/
7/ 97
2/
1/ 97
2/
7/ 98
2/
1/ 98
2
7/ /99
2/
1/ 99
2/
7/ 00
2/
1/ 00
2/
7/ 01
2
1/ /01
2
7/ /02
2/
1/ 02
2/
7/ 03
2
1/ /03
2/
04

estimates of the demands for required and excess
reserves are made by the staffs of the Federal
Reserve Bank of New York and the Board of
Governors; however, the Board’s estimates are
used here.
Because of the introduction of sweep accounting in January 1994, initially the model is estimated over sample period January 2, 1986, though
December 31, 1993. Carpenter and Demiralp
(2006) found the DLE to be nonlinear and statistically significant for large shocks (shocks > $1
billion) but not for small shocks (shocks ⱕ $1
billion). Hence, for some specifications, miss is
partitioned into large shocks (miss tlg ) and small
shocks (miss tsm ) using their criterion. Because of
the two-day lag in the Fed’s system of reserve
requirements during this period, settlement days
are partitioned into the last two days of the maintenance period (l2d ) and all other days (nl2d ).23
Also, because the effect of reserve supply shocks
on the funds rate differs on days when the funds
rate target is changed, dummy variables for days
80

J A N UA RY / F E B R UA RY

2010

when the target was changed (dΔfft*) and other
days (dnΔfft*) are included. Finally, the Student
t-distribution, rather that the normal distribution,
is used to account for the thick tails in the distribution of the funds rate.
The results in Table 1 are for three specifications, which differ by the variables included in
the model. The coefficient estimates are presented
in three sections. Table 1A reports the estimates
of β for the parameters that are relevant for evaluating the DLE and the LE. Table 1B reports estimates of coefficients on the dummy variables that
are included to account for various characteristics
of the data. Table 1C reports the estimates of the
variance parameters (equation 18) and the relevant summary statistics.
23

Carpenter and Demiralp (2006) partition miss by each day of the
maintenance period. However, save the last two days of the maintenance period, there is no particular reason to believe that the slope
of the money demand curve should be systematically distinct on
different days of the maintenance period. Consequently, this is
not done here.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Thornton

Table 1A
Estimates of the Reserve Market Model (January 2, 1986–December 31, 1993)
Variable

Specification 1

Specification 2

Specification 3

fft*

0.5555

0.0000

0.5523

0.0000

0.5580

0.0000

Δfft*

0.0003

0.3820

0.0004

0.3238

0.0004

0.3320

0.0109

0.1679

0.0106

0.1763

misst × dnΔfft* × l2d

–0.0327

0.0000

misst ×

–0.0109

0.0000

misst × dnΔfft* × l2d × O

–0.2181

0.0000

misst × dnΔfft* × l2d × NO

–0.0275

0.0001

misst × dnΔfft* × nl2d × O

–0.1195

0.0049

–0.0108

0.0000

misstsm × dΔfft*

0.0104

0.5593

misstsm × dnΔfft* × l2d

–0.0047

0.8203

misstsm × dnΔfft* × nl2d

–0.0083

0.0219

misst ×

dnΔfft*

× nl2d

× nl2d × NO

misstlg × dΔfft*

0.0113

0.1887

misstlg × dnΔfft* × l2d

–0.0323

0.0000

misstlg × dnΔfft* × nl2d

–0.0114

0.0000

BRt – IBAt

0.0243

0.0000

0.0239

0.0000

0.0239

0.0000

errtD

0.0088

0.0000

0.0089

0.0000

0.0088

0.0000

–0.0048

0.0003

–0.0046

0.0005

–0.0047

0.0004

Consistent with the model given by equation
(16), the dependent variable is fft and not fft – fft*,
as in Carpenter and Demiralp (2006), or Δfft , as
in Hamilton (1997). Note that fft – fft* would be
the appropriate dependent variable if and only if
the Desk correctly estimated the interest elasticity
~
of money demand—that is, β = β.24
Specification 1 most closely resembles
Carpenter and Demiralp’s (2006) model. Specifically, misst is partitioned into large and small
misses using their criteria, and the response of
the funds rate is allowed to differ depending on
whether (i) the target changed that day, (ii) the
miss occurred on the last two days of the mainte24

The federal funds rate is very persistent and, hence, close to a unit
root process. Hamilton’s dependent variable is the change in the
funds rate, while Carpenter and Demiralp’s is the spread between
the funds rate and the funds rate, both of which are stationary. The
funds rate is used here because it is consistent with the model given
by equation (16). However, to make sure that the qualitative conclusions reported here are not due solely to the near nonstationarity
of the funds rate, the specifications reported in Table 1 were also
~
estimated imposing the restriction β = β. While the numerical
values of the parameter estimates change, the important qualitative conclusions were the same.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

nance period, or (iii) it occurred on one of the
other eight days. The estimates of the variance
parameters for this specification presented in
Table 1C show that the variance increased significantly during the periods immediately following
the 1987 stock market crash and the 1990 surprise
reduction in reserve requirements. Also, consistent with the thick-tailed distributions, characteristic of interest rates, the estimate of the degrees
of freedom (dof ) parameter is small, 3.77, and
statistically significant, indicating the appropriateness of using the Student t-distribution.
All but a few of the estimates of the coefficients
on the various dummy variables presented in
Table 1B are statistically significant. Not surprisingly, in most cases, the estimated responses are
as one would expect: The funds rate tends to be
higher on settlement Wednesdays, higher at the
end of the quarter, the first and last days of the
month, and so on.
Table 1A reports the estimates relevant for
the DLE and the LE. As expected, reserve supply
shocks that occur on days when the FOMC
J A N UA RY / F E B R UA RY

2010

Thornton

Table 1B
Estimates of the Reserve Market Model (January 2, 1986–December 31, 1993)
Variable

Specification 1

Specification 2

Specification 3

fft–1 × w1

0.4472

0.0000

0.4504

0.0000

0.4447

0.0000

fft–1 × w2

0.4461

0.0000

0.4494

0.0000

0.4436

0.0000

–0.0132

0.2335

–0.0126

0.2555

–0.0130

0.2425

–0.0556

0.0000

–0.0691

0.0000

–0.0683

0.0000

0.0468

0.0000

0.0340

0.0001

0.0342

0.0001

–0.0287

0.0015

–0.0414

0.0000

–0.0413

0.0000

–0.0351

0.0001

–0.0482

0.0000

–0.0482

0.0000

0.0053

0.6869

–0.0085

0.2980

–0.0077

0.3445

–0.0514

0.0001

–0.0649

0.0000

–0.0640

0.0000

0.0542

0.0006

0.0398

0.0006

0.0403

0.0004

–0.0399

0.0224

–0.0537

0.0001

–0.0524

0.0002

D10

0.0817

0.0000

0.0678

0.0000

0.0690

0.0000

eom

0.0871

0.0000

0.0861

0.0000

0.0881

0.0000

bom

0.0572

0.0000

0.0573

0.0000

0.0570

0.0000

eoq

0.2125

0.0032

0.2159

0.0028

0.2000

0.0035

boq

–0.1152

0.0070

–0.1176

0.0056

–0.1202

0.0035

eoy

–0.3804

0.0003

–0.3810

0.0003

–0.3675

0.0004

boy

0.4270

0.0006

0.4301

0.0005

0.4351

0.0005

mom

0.0899

0.0000

0.0904

0.0000

0.0903

0.0000

–0.0169

0.0329

–0.0163

0.0398

–0.0173

0.0297

0.1097

0.0000

0.1094

0.0000

0.1095

0.0000

changed the funds rate target are not statistically
significant, regardless of whether the shocks are
large or small. Also, consistent with Carpenter
and Demiralp (2006), the response of the funds
rate to small shocks on all but the last two days
of the maintenance period is statistically significant and smaller than the response to large shocks.
However, the magnitude of the difference between
the response to large and small shocks is relatively
small. Indeed, the likelihood ratio test statistic for
equality of the response is 0.464, which is not
statistically significant at any reasonable significance level. Contrary to Carpenter and Demiralp’s
(2006) results, there is no evidence of nonlinearity.
Given the absence of nonlinearity, the model
is estimated without partitioning the reserve
supply shocks into large and small shocks. Estimates of this specification are presented as speci82

J A N UA RY / F E B R UA RY

2010

fication 2 of Table 1. Again, there is no statistically
significant response of the funds rate to reserve
supply shocks that occur on days when the target
is changed. Also, consistent with Hamilton (1997)
and Thornton (2001a), the response of the funds
rate on the last two days of the maintenance period
is about three times larger than the response on the
other eight days, and it is statistically significant.
As expected, the coefficients on BRt – IBAt ,
kt , and errtD are all statistically significant. The
coefficient on BRt – IBAt is positive, suggesting
that borrowing above the FOMC’s assumed level
is associated with the funds rate above the target.
The sign of the coefficient is inconsistent with a
supply shock interpretation, but it is consistent
with the evidence that borrowing responds
endogenously to the funds rate (e.g., Thornton,
2001b). The coefficients on kt and errtD have the
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Thornton

Table 1C
Estimates of the Reserve Market Model (January 2, 1986–December 31, 1993)
Variable

Specification 1

Specification 2

Specification 3

Constant

–3.0817

0.0000

–3.0848

0.0000

–3.0208

0.0000

|ε t–1 /σt–1 |

0.7053

0.0000

0.7043

0.0000

0.6821

0.0000

ε t–1 /σt–1

0.0559

0.2237

0.0585

0.2058

0.0638

0.1507

2
log σ t–1

0.5387

0.0000

0.5374

0.0000

0.5466

0.0000

D1 + D2 + D3

1.5364

0.0000

1.5356

0.0000

1.5135

0.0000

btar

0.6902

0.0085

0.6768

0.0086

0.6660

0.0097

1.1983

0.0000

1.2091

0.0000

1.1562

0.0000

eom

0.9886

0.0000

–1.8576

0.0096

–1.6951

0.0161

eoq

2.4000

0.0000

2.4184

0.0000

2.3238

0.0000

eoy

–1.8168

0.0108

0.9883

0.0000

0.9508

0.0000

mom

0.6470

0.0028

0.6558

0.0024

0.6322

0.0033

d1987

0.4993

0.0239

0.4942

0.0251

1.3252

0.0000

d1990

1.3196

0.0000

1.3238

0.0000

0.5740

0.0099

Degrees of freedom

3.7653

0.0000

3.7529

0.0000

3.7440

0.0000

No. of observations
–
R2

1,966

0.9887

0.9885

0.9892

SE
Log likelihood

0.2234

0.2244

0.2180

1477.061

1475.596

1479.130

anticipated signs. The estimated coefficient on kt
suggests that the funds rate tends to be significantly
lower on days when the Desk engages in more
open market operations than the operating procedure suggests. Likewise, if the Desk underestimates the demand for reserves, the funds rate is
somewhat higher.
Equation (16) suggests that the absolute magnitude of the response of the funds rate to miss,
errtD, BRt – IBAt , and kt should be equal; however,
this restriction was not imposed.25 Nevertheless,
it is interesting to note that the estimated coefficients on errtD and miss on days other than the
last two of the maintenance period are similar in
magnitude but opposite in sign as suggested by
25

Given that borrowing is endogenous, it is unlikely that the restriction would hold for borrowing. Also given that the Desk is free to
deviate from the procedure as it sees fit, it seems unlikely that it
would hold for kt as well. Indeed, a test that the absolute values of
the coefficients on miss, errtD, and kt are equal is rejected at the 5
percent significance level or lower.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

equation (16). The likelihood ratio statistic for
the hypothesis that the responses are equal but
opposite in sign is 0.79.
Thornton (2001a) showed that Hamilton’s
(1997) results were sensitive to days with large
shocks to the funds rate. Hence, I investigate the
sensitivity of the estimates of the DLE to unusually
large and idiosyncratic shocks to the funds rate.
This is important because the response of the
funds rate to supply shocks on such days is not
necessarily evidence of an LE. Specifically, miss
is partitioned by days with large and idiosyncratic
shocks to the funds rate: outliers (O) and days
with no outliers (NO). Days with idiosyncratic
shocks to the funds rate are obtained by regressing
the federal funds rate on a constant and the 3month Treasury bill rate over the sample period.
The residuals from this equation are daily changes
in the funds rate that are not associated with
changes in the 3-month T-bill rate: idiosyncratic
shocks to the federal funds rate. Idiosyncratic
J A N UA RY / F E B R UA RY

2010

Thornton

Table 2
Estimates of the Reserve Market Model
(January 3, 1994–December 31, 1996)
Variable
fft*
Δfft*
misstsm × dΔfft*
misst × dnΔfft* × l2d × O

Coefficient

Significance
level

0.769

0.000

–0.000

0.820

0.769

0.000

0.820

misst × dnΔfft* × l2d × NO

–0.008

0.881

misst ×

dnΔfft*

× nl2d × O

–0.011

0.281

misst × dnΔfft* × nl2d × NO –0.004

0.051

BRt – IBAt

0.198

0.000

errtD

0.004

0.006

0.000

0.770

No. of observations
–
R2

754
0.946

0.197

Log likelihood

789.248

shocks to the funds are considered large when
they are more than 80 basis points (roughly two
standard errors [SEs] of the idiosyncratic shocks
to the funds rate).26 There are 62 days when there
were large, idiosyncratic shocks to the funds rate
during the sample period (slightly more than 3
percent of the days), 33 of which occurred on a
settlement Tuesday or Wednesday.
The results are reported in specification 3 of
Table 1. As anticipated, estimates of the DLE are
sensitive to large idiosyncratic shocks to the funds
rate. On days with large idiosyncratic shocks to
the funds rate, the estimated DLE is about 10
times larger than on days without such shocks.
Consistent with the results of Thornton (2001a),
estimates of the DLE appear to be significantly
overestimated on days with large idiosyncratic
shocks to the funds rate. Nevertheless, the estimate on days other than the last two of the maintenance period reported in specification 2 is nearly
26

As a robustness check on the qualitative results, values of 40, 50,
and 60 basis points were also used. The qualitative conclusion about
the coefficient miss on NO days is invariant to the value used.

J A N UA RY / F E B R UA RY

2010

identical to the estimate when there are no outliers in specification 3. Hence, the effect of large,
idiosyncratic shocks to the funds rate is reflected
mostly in estimates on settlement Tuesday and
Wednesday.27

Post-1993 Estimates of the Daily
Liquidity Effect
The introduction of sweep accounts in January
1994 dramatically reduced reserve requirements
for banks over time. Anderson and Rasche (2001)
suggest that by the end of 1999, “the willingness of
bank regulators to permit use of deposit-sweeping
software has made statutory reserve requirements
a ‘voluntary constraint’ for most banks” (p. 71).
To investigate the effect of sweep accounts on
the estimate of the DLE, the model is estimated
over the period from January 3, 1994, through
December 31, 1996. To conserve space, only estimates of the parameters that are relevant for the
LE are reported in Table 2. All estimated coefficients on the various partitions of miss are much
smaller in absolute value than those reported in
Table 1. Moreover, none is statistically significant
at the 5 percent significance level. The estimate
is statistically significant at slightly higher than
the 5 percent significance level when miss is partitioned by nl2d and NO. The estimate is only
about half as large as that for the pre-1994 period.
The smaller estimate is inconsistent with the fact
that sweeps effectively reduce reserve requirements. Other things the same, lower effective
reserve requirements should have resulted in a
larger coefficient estimate. One possible explanation is that the effective elimination of mandatory
reserve requirements for nonbound banks significantly altered the interest sensitivity of reserve
demand independent of money demand. It is
interesting to note that the estimated coefficient
on miss for these days is again equal but opposite
in sign to that of reserve demand shocks.
27

Given the close relationship between the funds rate and the funds
rate target, the model was also estimated using fft – fft* as the
dependent variable. While the coefficient estimates changed somewhat, the qualitative conclusions are not sensitive to whether fft
or fft – fft* is the dependent variable. The quantitative and qualitative results are very sensitive to excluding BRt – IBAt , errtD, and kt,
however. The correlations between miss and BRt – IBAt , errtD, and
kt over this sample period are –0.058, 0.352, and –0.013, respectively.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Thornton

Post-1998 Estimates of the Daily
Liquidity Effect
Finally, the model was estimated over the
period August 3, 1998, through January 30, 2004,
to determine whether such evidence has no implication for the LE. Data on BRt – IBAt , errtD, and kt
are not available over this period, so the estimates
are likely to be biased. The estimate of the DLE
for days other than the last two of the maintenance
period when there were no outliers is small,
–0.007, but statistically significant. This shows
that the demand for reserves is interest sensitive
apart from the interest sensitivity of the demand
for money. Given the interest sensitivity of reserve
demand, caution is necessary in concluding that
there is a statistically significant and economically
relevant LE based on statistically significant estimates of the DLE.

CONCLUSION
The DLE was first estimated by Hamilton
(1997) in an attempt to find evidence of Friedman’s
(1969) policy-relevant LE, which had escaped
detection using lower-frequency (monthly and
quarterly) data. Unfortunately, Hamilton and subsequent researchers did not investigate the linkage between the DLE and the LE. This article fills
this gap in the literature by showing that the DLE
is directly linked to the LE by Federal Reserve–
imposed reserve requirements. The relationship
between the DLE and the LE is then analyzed and
investigated using a more comprehensive model
of the Desk’s operating procedure than has been
used in the literature. The analysis shows that
the relationship between the DLE and the LE
depends on the Desk’s operating procedure, the

Fed’s system of reserve requirements, and other
factors. Importantly, the analysis shows that
there is no relationship between these LEs after
July 1998 when the Fed reinstated lagged reserve
accounting.
Estimates of the DLE before 1994 suggest that
there may have been a statistically significant
policy-relevant LE before 1994. The estimated
DLE is small, however. The estimate suggests that
a $10 billion reserve supply shock generates about
a 20-basis-point change in the funds rate. If one
assumes that the average effective reserve requirement during the sample period is 10 percent, this
would be equivalent to about a $100 billion shock
to the money supply—much larger than any shock
during this sample period.
More problematic is the finding of a statistically significant DLE after July 1998, when the
Fed established lagged reserve accounting. The
existence of a DLE over this period is due to the
fact that banks have an incentive to economize
on their holdings of reserves, independent of the
interest sensitivity of money demand. The fact
that there is a statistically significant DLE during
a period when estimates of the DLE can have no
implication for the LE raises a question of the
extent to which estimates of the DLE have implications for the LE during other periods. It could
be that all estimates of the DLE reflect the interest
sensitivity of reserve demand independent of the
interest sensitivity of money demand. In any
event, the results presented here indicate that it
is no easier to find convincing evidence of a statistically significant and economically important
policy-relevant LE using high-frequency daily data
than it has been using lower-frequency (monthly
and quarterly) data. A resolution of the liquidity
puzzle remains elusive.

REFERENCES
Anderson, Richard G. and Rasche, Robert H. “Retail Sweep Programs and Bank Reserves, 1994-1999.” Federal
Reserve Bank of St. Louis Review, January/February 2001, 83(1), pp. 51-72;
http://research.stlouisfed.org/publications/review/01/0101ra.pdf.
Bartolini, Leonardo; Bertola, Giuseppe and Prati, Alessando. “Day-to-Day Monetary Policy and the Volatility of
the Federal Funds Interest Rate.” Journal of Money, Credit, and Banking, February 2002, 34(1), pp. 137-59.
F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

Thornton

Cagan, Phillip and Gandolfi, Arthur. “The Lag in Monetary Policy as Implied by the Time Pattern of Monetary
Effects on Interest Rates.” American Economic Review, May 1969, 59(2), pp. 277-84.
Carpenter, Seth and Demiralp, Selva. “The Liquidity Effect in the Federal Funds Market: Evidence from Daily
Open Market Operations.” Journal of Money, Credit, and Banking, June 2006, 38(4), pp. 901-20.
Clouse, James A. and Dow, James P. Jr. “A Computational Model of Banks’ Optimal Reserve Management Policy.”
Journal of Economic Dynamics and Control, September 2002, 26(11), pp. 1787-814.
Coleman, Wilbur John II; Gilles, Christian and Labadie, Pamela A. “A Model of the Federal Funds Market.”
Economic Theory, February 1996, 7(2), pp. 337-57.
Engle, Robert F. “Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom
Inflation.” Econometrica, July 1982, 50(4), pp. 987-1007.
Feinman, Joshua N. “Estimating the Open Market Desk’s Daily Reaction Function.” Journal of Money, Credit,
and Banking, May 1993, 25(2), 231-47.
Friedman, M. “Factors Affecting the Level of Interest Rates,” in Proceedings of the 1968 Conference on Saving
and Residential Financing. Chicago: United States Saving and Loan League, 1969, pp. 11-27. Reprinted in
Thomas M. Havrilesky and John T. Boorman, eds., Current Issues in Monetary Theory and Policy. Arlington
Heights, IL: AHM Publishing, 1976, pp. 362-78.
Gibson, William E. “Interest Rates and Monetary Policy.” Journal of Political Economy, May/June 1970a, 78(3),
pp. 431-55.
Gibson, William E. “The Lag in the Effect of Monetary Policy on Income and Interest Rates.” Quarterly Journal
of Economics, May 1970b, 84(2), pp. 288-300.
Hamilton, James D. “The Daily Market for Federal Funds.” Journal of Political Economy, February 1996, 104(1),
pp. 26-56.
Hamilton, James D. “Measuring the Liquidity Effect.” American Economic Review, March 1997, 87(1), pp. 80-97.
Humphrey, Thomas M. “The Early History of the Real/Nominal Interest Rate Relationship.” Federal Reserve Bank
of Richmond Economic Review, May/June 1983a, 69(3), pp. 2-10;
www.richmondfed.org/publications/research/economic_review/1983/pdf/er690301.pdf.
Humphrey, Thomas M. “Can the Central Bank Peg Real Interest Rates? A Survey of Classical and Neoclassical
Opinion.” Federal Reserve Bank of Richmond Economic Review, September/October 1983b, 69(5), pp. 12-21;
http://www.richmondfed.org/publications/research/economic_review/1983/pdf/er690502.pdf.
Judson, Ruth and Klee, Elizabeth. “Whither the Liquidity Effect: The Impact of Federal Reserve Open Market
Operations in Recent Years.” Finance and Economics Discussion Series Working Paper 2009-25, Board of
Governors of the Federal Reserve System; www.federalreserve.gov/pubs/feds/2009/200925/200925pap.pdf.
Nelson, Daniel B. “Conditional Heteroskedasticity in Asset Returns: A New Approach.” Econometrica, March
1991, 59(2), pp. 347-70.
Pagan, Adrian R. and Robertson, John C. “Resolving the Liquidity Effect.” Federal Reserve Bank of St. Louis
Review, May/June 1995, 77(3), pp. 33-54;
http://research.stlouisfed.org/publications/review/95/05/Resolving_May_June1995.pdf.
Strongin, Steven. “The Identification of Monetary Policy Disturbances: Explaining the Liquidity Puzzle.”
Journal of Monetary Economics, June 1995, 35(3), pp. 463-97.
Thornton, Daniel L. “The Borrowed-Reserves Operating Procedure: Theory and Evidence.” Federal Reserve
Bank of St. Louis Review, January/February 1988, 70(1), pp. 30-54;
http://research.stlouisfed.org/publications/review/88/01/Borrowed_Jan_Feb1988.pdf.
86

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Thornton

Thornton, Daniel L. “Identifying the Liquidity Effect at the Daily Frequency.” Federal Reserve Bank of St. Louis
Review, July/August 2001a, 83(4), pp. 59-78;
http://research.stlouisfed.org/publications/review/01/05/59-78Thornton.qxd.pdf.
Thornton, Daniel L. “The Federal Reserve’s Operating Procedure, Nonborrowed Reserves, Borrowed Reserves
and the Liquidity Effect.” Journal of Banking and Finance, September 2001b, 25(9), pp. 1717-39.
Thornton, Daniel L. “Forecasting the Treasury’s Balance at the Fed.” Journal of Forecasting, August 2004, 23(5),
pp. 357-71.
Thornton, Daniel L. “When Did the FOMC Begin Targeting the Federal Funds Rate? What the Verbatim Transcripts
Tell Us.” Journal of Money, Credit, and Banking, December 2006, 38(8), pp. 2039-71.
Thornton, Daniel L. “Open Market Operations and the Federal Funds Rate,” in David G. Mayes and Jan Toporowski,
eds., Open Market Operations and the Financial Markets. Oxford, UK: Routledge, 2007; pp. 178-205.
Reprinted in the Federal Reserve Bank of St. Louis Review, November/December 2007, 89(6), pp. 549-72;
http://research.stlouisfed.org/publications/review/07/11/Thornton.pdf.

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

J A N UA RY / F E B R UA RY

2010

J A N UA RY / F E B R UA RY

2010

F E D E R A L R E S E R V E B A N K O F S T. LO U I S R E V I E W

Full text of Review (Federal Reserve Bank of St. Louis) : January/February 2010, Vol. 92, No. 1

FRASER