The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Using Sibling Data to Estimate the Impact of Neighborhoods on Children's Educational Outcom es Daniel Aaronson Working Papers Series Macroeconomic Issues Research Department Federal Reserve Bank of Chicago November 1996 (W P -96-19) FEDERAL RESERVE BANK OF CHICAGO l i b r a r y NOV 0 5 1996 F D R L R SE V EEA E RE B N O C IC G AK F H AO Using Sibling Data to Estimate the Impact of Neighborhoods on Children’s Educational Outcomes Daniel A aronson daaronson @ frbchi.org Federal Reserve B ank o f C hicago D raft O ctober 1995 Revised O ctober 1996 Abstract Studies that attem pt to measure the im pact o f neighborhoods on children’s outcomes are susceptible to bias because fam ilies choose w here to live. As a result, the effect o f fam ily unobservables, such as the im portance parents place on their children’s welfare, and other unobservables that are com m on to geographically clustered households, may be m istakenly attributed to neighborhood influences. Previous studies that attem pt to correct for this selection bias have used questionable instrum ental variables. This paper introduces an approach based on the observation that the latent factors associated w ith neighborhood choice do not vary across siblings. Therefore, fam ily residential changes provide a source o f neighborhood background variation that is free o f the family-specific heterogeneity biases associated w ith neighborhood selection. U sing a sam ple o f m ultiple-child families w hose kids are separated in age by at least three years, I estim ate fam ily fixed effect equations o f children’s educational outcomes. The fixed effect results suggest that the im pact o f neighborhoods exists even when fam ily-specific unobservables are controlled. This finding is robust to m any changes in estim ation techniques, outcom e measures, neighborhood m easures, variable definitions, and samples. My thanks t Joe A t n i Becky Blank, Greg Duncan, Judy H l e s e n Sandy Jencks, Bruce Meyer, and o loj, elrti, Lauren Sinai for h elpful suggestions. All e r r and omissions are mine. The views expressed i t i paper ros n hs are those of t e author and are not nec s a i y those of th Federal Reserve Bank of Chicago or the Federal h esrl e Reserve System. L Introduction T here is substantial evidence that family background characteristics play an im portant role in th e educational development o f children. H owever, how neighborhoods, schools, and peer g roups affect children remains unsettled. Although a fairly large cross-disciplinary literature on neighborhood effects has emerged over the last fifteen years, th e empirical findings are not robust to data issues, outcom e measures, and estimation techniques . 1 T h e lack o f consistent evidence could stem from a number o f factors. Perhaps survey data cannot adequately represent the com plex nature o f how communities influence children. A nother possibility, often discussed in the literature, is the role o f bias in the estimating equations. Bias may arise because families are not randomly placed in neighborhoods but rather choose their location based on a variety o f factors, including th e im portance they put on their children's developm ent . 2 F or a variety o f reasons, the result is th at neighborhoods are stratified along socioeconom ic lines. This is reflected in the fact th at key family characteristics, such as household income, th e proportion o f single parent families, and average education levels vary substantially across neighborhoods. Studies that ignore the endogeneity o f neighborhood selection risk over- or understating the importance o f neighborhoods for children's outcomes. T he direction o f the bias is related to the w ay the unobservables associated w ith neighborhood selection are correlated with the unobservables associated w ith children's outcomes. I t is generally thought th at this bias is positive, reflecting th e potential o f attributing family characteristics, such as parental competence, tastes fo r education, o r tim e spent w ith their 1Significant neighborhood effects have been found i , among others, Summers and Wolfe (1977), Case and Katz n (1991), Crane (1991), Brooks-Gunn e a . (1993), Duncan (1994), andBoijas (1995). However, other s u i s most tl tde, notably Evans, Oates, and Schwab (1992) and Corcoran e a (1992), have found no evidence t neighborhoods t l hat matter. While Jencks and Mayer (1990) conclude t a no robust evidence of neighborhood e f c s e i t , t e r ht fet xss h i extensive cross-disciplinary summary points to a number ofstudies where neighborhoods seem t matter. o 2 To mini ize the self-selection issues involved in residential location choice, this research has wisely m concentrated on children and t eenagers. A notable exception, which uses quasi random assignment, i the work by s Rosenbaum and Popkin (1991) on Chicago's Gautreaux program. Other econometric problems discussed in the l t r t r are measurement error and the ' e l c i n problem* ieaue rfeto (Manski 1993). The r f e t o problem concerns the p s i i i y t a individuals a f c or are d r c l part of the elcin osblt ht fet iety neighborhood aggregate and therefore i i d f i u t t discern cause from e f c . This problem i p r i u a l t s ifcl o fet s atclry severe in cases l k Case and Katz (1991) where neighborhood variables are aggregates of a small number of ie individuals. In t i paper, neighborhoods const t t roughly 4,000 individuals and therefore should not suffer hs iue from t i endogeneity problem. hs children, to th e neighborhood measures. The bias might be enhanced if th e unobserved heterogeneity th at is com m on to a group o f clustered individuals is correlated w ith m easured neighborhood characteristics. Studies th a t attem pt to correct for this selection bias have used an instrum ental variables approach. H ow ever, this m ethod requires use o f a variable th at is a determ inant o f neighborhood choice but n o t o f th e outcom es o f children. Such an instrum ent is difficult to find. Only three papers th at I am aw are o f attem pt to do so. C ase and K atz's (1991) study o f peer influences surmise th at children are influenced by peers in th eir ow n neighborhood and surrounding neighborhoods but n o t directly by those living in noncontingent communities. This assumption allows them to use neighbors' neighbors as an identifying variable but is susceptible to the criticism that children are likely to be influenced by peers at school w ho do not necessarily live in the same or adjoining neighborhoods. A second paper by Evans, O ates, and Schwab (1992) employs a similar strategy, using a geographically different m easure o f the community-level variable to identify th e selection process. They use m etropolitan area unem ployment, median family income, poverty, and educational attainm ent characteristics as instrum ents, arguing that these variables are likely to be correlated w ith their peer group variable, the log o f the number o f disadvantaged children in a teen's school, b u t do not directly affect their child outcom e measures. To make this assum ption, the authors m ust believe that m etropolitan area characteristics do not affect schooling choice and that parents ta k e their m etropolitan area as given w hen choosing a particular neighborhood o r school. T here are reasons to question b o th assum ptions . 3 Finally, Duncan, Connell and K lebanov (1994) use th e neighborhood th at th e m other lives in after all the children leave the parental hom e as an instrument, under the prem ise th at once children move into their ow n households, parents' residential choice is no longer based on concern about neighborhood influences on their ow n children. H ow ever, inertia in residential choice makes these instrum ent choices suspect. 3 Defending their choice of instrument, EOS note that two-thirds of family moves in a five year period were within the same metropolitan a e . However, t i evidence, which i also borne out in the PSID, suggests that ra hs s families are quite mobile and thus may be s lecting i t metropolitan as well as census t a ta e s e no rc ra. An alternative approach pursued in this paper is based on the observation th at the latent factors associated w ith neighborhood choice are sibling-invariant; households rarely m ove due to the differential ability o f their children. As a result, family residential changes provide a source o f neighborhood background variation within families that can be used to identify neighborhood influences. T he key advantage to using sibling neighborhood background differentials is it offers a natural w ay to eliminate the family-specific heterogeneity biases associated w ith neighborhood selection. Furtherm ore, unlike other studies that use twins o r siblings, such as th e rate o r return to education o r teenage m otherhood debate, the potential endogenous variable is n o t being chosen by the individual. H ow ever, sibling-based fixed effect m odels are no panacea. They may accentuate problem s w ith measurement error and still leave open th e possibility o f omitted variable bias due to unobserved time-varying family characteristics and within-family heterogeneity. U sing a sample from the Panel Study o f Incom e D ynam ics (PSID ) o f multiplechild families w hose kids are separated in age by at least three years, I estim ate family fixed effect equations o f children's educational outcomes. The fixed effect results suggest th a t th e im pact o f neighborhoods exists even w hen family-specific unobservables are controlled. This finding is fairly robust to changes in estimation techniques, outcom e m easures, neighborhood m easures, variable definitions, and samples. T he paper is organized as follows. Section II explains the empirical strategy used in this paper. First, a m odel w here communities m atter to childhood educational opportunities is introduced to show how endogeneity, heterogeneity and functional form assum ptions m ight influence empirical results. Some concerns about the fixed effect estim ates are explained, including th e critical notions o f measurement error and within-family heterogeneity. Section HI describes the data used to develop the sample and the neighborhood characteristics. Section IV discusses the results, including linear probability, logit, instrum ental variables, and fixed effects estimates o f neighborhood influences on the likelihood o f children graduating from high school. Section V outlines a number o f tests o f the robustness o f the findings. The models are rerun using different neighborhood proxies, different outcom e measures, different samples, and different neighborhood variable definitions. M any o f these tests are used to reconcile th e different findings o f this paper and Plotnick and Hoffinan (1995). Plotnick and Hoffman also use PS1D siblings to identify neighborhood effects but conclude th at family fixed effects eliminate th e im pact o f neighborhoods o n post-secondary schooling, teen births, and welfare recipiency. O u r results can be partly reconciled by differences in sample and variable definitions. Som e concluding rem arks are offered in th e last section. n. Empirical S tra te g y M odel T o help clarify these issues, I present a simple variant o f Becker's child quality m odel w ith the additional assum ption th at com munities influence a child's future outcom es . 4 Suppose a parent maximizes a CES production function over her m children's future outcom es (education, say), k t+1, current consum ption, c*, and the quality o f the neighborhood, % (l) u(k +rV ^ , 6 i * v + 1+ 52c? + 53n? 1= 1 Children are indexed by i and tim e by t. wherep<1 Quality o f neighborhood enters the utility function independently and additively to account for th e im portance that households place o n crime, ethnicity, services, housing conditions, and other neighborhood-specific factors. The parent faces a budget constraint o f the form I(n t )= c { + Pn n t - Neighborhood conditions are allow ed to influence the household's income. The parent uses income to purchase consum ption item s and better neighborhood conditions at a price relative to consumption o f Pn. Finally, each child's future education is determ ined by a production function o f the form (2 ) lo g k . t + 1 = P g l o g G + P a l o g a it + P n l o g n t w here Pg,P „ P ne [ 0 ,l) . Family-specific variables are captured in the G term. This w ould include parents' interest in education, and any family background characteristic and ability com ponent th at is constant across siblings. The v ecto r a.^ m easures any variables th at exhibit heterogeneity betw een siblings, m ost 4 For examples of more complicated models with peer e fects, see deBartolome (1990) and Epple and Romano f (1993). Many other l t r t r s have argued that production functions should include c a a t ieaue h r c eristics of the population, including growth theory (Jacobs 1969, Romer 1986, Benabou 1993) and lo a public finance cl (Brueckner and Lee 1989). notably differences in ability, ambition, or, for siblings separated by age, family conditions. w ould also include differences in parental expectations o r treatm ent o f siblings. It A v ector o f neighborhood and peer measures also enters the production function. An important element o f this model is the reason for family moves. W hen a family m oves into a community, it has an expectation about the quality o f the neighborhood. H ow ever, ,there is also uncertainty about th e caliber o f the family-neighborhood m atch since the neighborhood good is, at least in part, an experience good. Therefore, th e actual neighborhood good is a sum o f the random error com ponent that measures uncertainty in m atch quality and the expectation o f the neighborhood prior to th e move. An unexpected negative shock in the match quality param eter may cause households to migrate to a new neighborhood. This uncertainty is used to obtain the variation in sibling background that is needed for the m odel to work. H owever, another source o f neighborhood migration may be due to changes in the family's background. Changes in marital, income, o r employment states may cause families to reevaluate their neighborhood choice. In the empirical w ork, it will be im perative th at these family changes are controlled in o rd er to decom pose the effect o f neighborhood change from other family changes that m ight affect children’s outcomes. T he empirical strategies used to do this are discussed m ore fully below. F o r simplicity, assume neighborhood location does not affect parents' income (dUdo. = 0 ) . 5* Also assume there are tw o siblings that are separated in age by one period. Sibling / lives during period 1 and sibling./ lives during period 2. Maximizing U (ct ,k t + j,n p subject to th e budget constraint and equation ( 2 ) leads to equilibrium ratios o f future outcom es and neighborhood conditions that are a function o f the heterogeneous com ponents between siblings and the functional form assumptions associated w ith p . (3) (4) nl ----- logi L =2 - P d + Pn) % n 2 ' P_(2-p) a. log 1 2 = log- il k .^ 2 - p ( l + Pn) j2 J3 log 1 5 This assumption pertains to the spatial mismatch hypothesis. Most researchers believe that the affect of neighborhoods on a u t s income i n g i i l . Regardless, the assumption does not a f c the main r s l s of the dl' s elgbe fet eut paper. B asing th e estimating equations on within-family differences provides th e empirical advantage o f eliminating the im pact o f any family-constant com ponent, including decisions about neighborhood selection. Selectivity has an im pact only if parents choose neighborhoods based on th e differential ability o f their children (equation 3 ) . 6 I f this selection process exists, th e sign o f th e bias is dependent on th e functional form assumptions. I f p > 0, parents u se a reinforcing strategy in their choice o f neighborhood by taking special effort to live in b etter neighborhoods fo r m ore able children. A s p approaches zero (Cobb-Douglas), ability is independent o f neighborhood choice. Finally, if p < 0, parents adopt a com pensating strategy w here the parents try to place kids w ith lesser ability in better communities. Econom etrically, this preference can be seen m ore clearly by looking at th e error term o f an educational production function w here I rew rite ( 2 ) as: (5) k .r . , . = B - x - ,+ 6 x.~ +B a . « + B n . « + e . A v ' i f , t + l Kx f f * x iff ' a iff * n in iff w here / indexes families. H ere the family and individual factors in the educational production function are separated. The x term s represent family characteristics th at vary over tim e and siblings (xjft) and those that are tim e and sibling invariant (xf). The error term is broken into three com ponents (6 ) e jft = "*"n if t<pf + M ift* (Pf is a sibling invariant error com ponent. I f <p^ is correlated w ith family residential preferences, then first differencing equation ( 6 ) across i will eliminate selectivity concerns. H ow ever, if the selection o f neighborhood is correlated w ith the individual-specific error com ponents, then selection bias remains a problem. In the results presented below, child ability is assum ed to be 6 This d f erential s l c i i ycould be due to other factors relatedto neighborhood choice. For example, suppose a if eetvt parent decides whether to work based on the a i i y of th i c i blt e r h ldren. In t i ca e the budget constraint and the hs s, educational production function includes the parent's decision on how much e f r to put into t e r work and t e r fot hi hi children. (2a) (l-Sjt)Rku = c t + Pn nt (21, l o g k i t + 1 = P g G + P a l o g a i t + P s logsi t k i t + P n I og nt The parent with human c p t l k t chooses whether t work a wage R or to put more e f r and time in o the aia j o t fot t production of the c i d s human c p t l I t i work e f r i used t buy b tter neighborhood goods fo th i hl' aia. f hs fot s o e r er children, then the a t v t e are s b t t t s Because of the endogeneity of S t we would need t worry about ciiis usiue. j, o another simultaneous equation that maps the d f e e t a e f r ofthe parents as a function of the d f e e t a a i i y ifrnil fot ifrnil blt ofthe chi d e . lrn independent of neighborhood choice. This seems to be an innocuous assumption, although I hope to better understand the correlation between the neighborhood good and the error term and its im portance in this system in future work. First differencing equation (5) eliminates th e family-specific unobservable and leaves a reduced form equation o f the general form (7) v' A .k.„ =B A .n.ft+B A .x .« + B A .a .^ + A s .^ 1 iff Kn i lft “ x l lft Ka i lft tft w here A differences across siblings. W ithout differential neighborhood selectivity, then A . n ^ can be thought o f as an element o f the a.t /a jt v ector in equation (4). Equation (7) is th e main estimating equation used in this paper. H ow ever, at least four main com plications rem ain in th e estim ation o f equation (7): unobserved heterogeneity within families, m easurem ent bias, th e discrete nature o f th e outcom e measures used, and complications w ith the sample due to age restrictions placed o n the siblings. The latter tw o estimation problems are discussed first as they are handled by conventional m ethods . 7 * Complications w ith the Fixed Effect E stim ator The first complication is that the outcom e variable used in much o f this analysis (high school graduation) is discrete.* In a linear regression fram ew ork w ith continuous dependent variables, one can handle family fixed effects by applying OLS to the data after taking deviations from group m eans . 9 H owever, the nonlinearity o f discrete choice models excludes this strategy. Furtherm ore, the asymptotic properties o f the logit m odel depend on the num ber o f observations per group increasing. Therefore, as shown by Chamberlain (1980), discrete choice fixed effect equations w ith small numbers o f observations p er group are inconsistent. Instead, h e proposes a logit fixed effect model which is estimated using conditional likelihood functions. A n alternative approach that uses within-group variation employs a specification suggested by M undlak. He 7 A f f h complication arises from the effect of neighborhood specific error components on the standard errors. it This i accounted forusing Huber's formula. s * In the f n l s c i n Irun the fixed e f c estimator on a continuous variable-grades completed—as w l . ia eto, fet el 9 Of course, the linear probability model introduces other well-known problems, including heteroskedasticity and predicted probabilities that are not constrained to the zero-one interval. allows individual effects to enter the probit model by simply specifying separate within-family and across-family variables (8) k^ = +P2xft +P3aift +P4aft +<l> 1nift +<l> 2nft +8ift w here variables w ith b ars represent family averages and <> is the within-family neighborhood fx m easure o f interest. R esults are presented using conditional logit, M undlak probit, and linear probability models. H ow ever, it appears to m ake little difference w hether logit o r linear probability techniques are used in the fixed effect models. Furtherm ore, because coefficients from conditional logit equations are in different units than th e simple logit coefficients, they are not comparable. A s a result, because o f their ease o f comparability and use, linear probability equations are em ployed to conduct much o f the estim ator comparisons. A second concern is sampling restrictions. In order to get meaningful variation in the residential location o f siblings, th e sample includes only individuals w ith a sibling th ree o r m ore years younger o r older than themselves . 10 M ost o f the difference in neighborhood background betw een siblings w ho are close in age is likely to be com posed o f measurement noise. T he further siblings are apart in age, the m ore likely they will experience tru e differences in neighborhood com position and enable m e to identify real differences in background influences. Choosing the age restriction is a b it ad-hoc, w ith sample size considerations balanced against the advantages o f using m ore age-separated siblings. I use a cu to ff o f th ree years. Som e experimentation suggests that using fo u r o r five years does not m ake a significant difference but decreases th e precision o f the estim ates due to the smaller sample sizes. R egardless o f th e cu to ff choice, this sampling restriction com plicates th e com putation o f the fixed effect estim ator since one family fixed effect will com bine siblings th at do not fit the age criterion. Therefore, I construct four fixed effect estim ators to see how assumptions about grouping observations affect the evaluation. The first three estim ators are constructed by pairing siblings th at fit the age criterion for selection into th e sample. F o r example, in a family w ith three kids, aged x , x + 2 , and x + 5 , 1 w ould * 9 1 1 Other important restrictions are: the individual turn 18 by 198S, the individual be in the PSID sample for two 0 years between the ages of 10 and 14, and be in the sample for one year a t rage 18 so that high school graduation fe , can be ascertained. I i i not possible to ascertain grades completed and the individual has data only through age fts 1 , then he i dropped from the sample. 9 s include tw o pairs o f siblings in the sample: the oldest w ith th e third child (5 years apart in age), and the oldest w ith the second child (three years apart in age). This setup might oversample certain individuals and therefore could introduce bias to th e estimates. Therefore, a second estim ator w eights the variables by the number o f times each individual in a sibling pair is in the total sam ple o r ( l / n i + l / 1 1 2 ), w here nj is the number o f tim es individual i is included in any sample pair. I f there is concern that this weighting procedure will not com pensate for the over sampling o f individuals, as a third alternative, I select one pair o f siblings— the oldest and youngest— from each family. Finally, as a fourth option, I estimate an equation th at includes a single fixed effect for each family. This alternative eliminates the multiple sampling problem but is problem atic in my exam ple because it contains groupings o f siblings w ho do not fulfill the age criterion . 11 1 I f 2 inference is similar for all four estimators, I can be more confident in the robustness o f th e results. U nobserved Heterogeneity within Families A m ore serious concern relates to the reasons for neighborhood change and th e individual error com ponents that describe the unobservable differences betw een siblings. In particular, tw o problem s could potentially complicate the interpretation o f the sibling difference estim ator. First, a t som e level, siblings may differ in, among other factors, ability, ambition, o r parental expectations and treatment; these unobserved factors m ay be correlated w ith neighborhood characteristics . 13 However, this is likely to be a serious concern only if parents choose neighborhoods based on these sibling differentials, an unlikely scenario. W hile there appears to be scant evidence on the impact o f siblings' differential ability on neighborhood choice, the research that exists places little significance on differential selection. Altonji and D unn (1995) estim ate the effect o f IQ scores on school choice and find little evidence th at ability m atters to this decision w ithin families. 1 For example, child x and x+2 would be included in the same family fixed e f ct, which i inconsistent with the 1 fe s r s r c i nthat only s b i g threeyears apartbe compared inthe fixed e f c models. etito ilns fet 1 For example, Summers and Wolfe (1977) find t at lower s i led students are more affected by classmates and 2 h kl school quality than t e rmore able p ers. Other omitted differences between s b i g , such as parents expectations hi e ilns or treatment, might have opposing e f c s Plomin and Daniels (1987) review the genetics and psychology fet. l t r t r and find a wide range of components—including the closeness to the mother, the f iendliness of the ieaue r s b i g , the r of s b i g i family decision making, and parental expectations-that might a f c the outcomes ilns ole ilns n fet ofs b i g , parti u a l those with d ffering a i i i s in d s i c ways. ilns clry i blte, itnt T o be safe, it w ould be useful if some m easure o f sibling differences could b e controlled. U nfortunately, the PSED has no test score reports o r other useful childhood m easure. A potential partial solution to th e missing ability m easure is to use w hether th e child w orks during his youth as a m easure o f unobservable ambition or drive. H ow ever it is also possible th at this variable w ould pick up th e availability o f jo b s 13 or itself be a com ponent o f o ther neighborhood influences. I f this latter interpretation is correct, including the 'w hether w orked' variable will bias the neighborhood coefficient downward. O ne characteristic w here sibling differentials might m atter is age. I f parents learn h o w to care for their children over time, it is possible that their younger children will benefit by being placed into b etter neighborhoods than their older siblings. Fortunately, this possibility is easily observed and controlled for in th e analysis by including th e birth order o f the children. A second com plication arises because siblings separated by age may experience different family environm ents due to, say, marital changes, different family income circum stances, o r less m easurable changes in household characteristics. A s m entioned above, this heterogeneity is particularly w orrisom e if the neighborhood variable is simply picking u p changes in family conditions th at precipitate changes in residential location. In o ther w ords, sibling differences in neighborhood conditions may be a function o f changes in unm easured family conditions and n o t changes in community influences. T o provide insight into w hether this issue is im portant empirically, table 1 rep o rts measurable changes in to tal family income, labor income, marital status, and em ploym ent status for residential stayers and m overs, by type o f geographic m ove , 14 to see if residential m oves are correlated w ith changes in family conditions. The last four columns explore family changes corresponding to m oves into better (columns 6-7) and w orse (colum ns 8-9) neighborhoods as crudely m easured by th e poverty rate in the origin and destination neighborhoods. T he sample includes multiple child families in the PSED during 1971-1974 or 1980-1983. T hese tim e periods 1 See Holzer (1991) fora good summary ofthe spatialmismatch lit r t r . 3 eaue 1 Moves are categorized by geographic l v l s a e county, neighborhood, and residence. Neighborhood is the 4 ee: tt, census t a t I the person does not l v in a census t a t then enumeration d s r c ( h ru a equivalent of census rc. f ie rc, itit t e r l t a t ) i used. I the person does not l v i an enumeration d s r c ,then f v d g t zip codes are used. rcs s f ie n itit ie ii w ere chosen so th at it w ould be clear w hen moves occur. B ecause the geocode database is missing addresses fo r 1969, 1975, 1977, and 1978, the level o f a m ove cannot be identified for these years and the year th at follows w ithout making som e ad hoc assum ptions about timing. A s can be seen in th e income, marital status, and employment status variables, longer distance m oves are m ore likely to occur among better-off families. State and county m overs look like stayers w ith regard to income, marital status and employment status. B u t neighborhood and residential m overs are poorer, less attached to the labor market, and less likely to be married. M ost telling fo r this paper is the w ay these variables change in relation to moves. In the row s titled “Change (t-i,t)n, I calculate the difference betw een the year after th e m ove (t) and the tw o years preceding the m ove ( t-l,t-2). Asterisks represent w hether these changes are significantly different from th e changes prior to stayer years. A pound sign represents w hether changes preceding m oves to better neighborhoods a re significantly different from changes preceding m oves to w orse neighborhoods at the five percent level. I also calculate transitions between marital and employment states in the tw o years before a m ove in th e row s directly below the “change” row s. These transitions represent the percentage o f family-year m oves (o r stays) that are prefaced in th e tw o previous years by that particular change. F o r example, 11.4 percent o f state changers experience a divorce (married— >divorced) in the previous tw o years before the state move. The results suggest som e significant change in observable family environm ent preceding moves, although these changes vary by the distance and type o f mobility. T he tw o income categories show fluctuations in years before moves but these changes are generally not significantly different from th e years before stays. There is a decline in total income and labor income preceding moves to neighborhoods w ith higher poverty rates (colum n 6 ) and an increase in total income but not labor incom e in years preceding moves to neighborhoods w ith low er poverty rates (colum n 8 ). M oves into higher poverty neighborhoods are also preceded by a small spike in the variance o f total and labor income, suggesting th at income instability could be a factor in these community changes. H ow ever, the overall picture from these income changes seems to suggest little relationship betw een family changes and moves. O ther w ork using larger samples is consistent w ith this conclusion. T here is m uch m ore activity in marital and employment status changes prior to m oves. Transitions into and o u t o f m arriage are significantly different for m ove years th an stay years am ong every m ove category. Changes preceding moves into high poverty neighborhoods are especially noticeable, w ith strong evidence o f transitions into divorce but much less evidence into marriage. Transition into m arriage is often followed by moves into low er poverty neighborhoods and different counties and states. H ow ever, there also appears to be high levels o f recently divorced households m oving into low er poverty neighborhoods. Em ploym ent status changes play an im portant role in sh o rt distance m oves (residential and neighborhood), particularly by those moving to low er poverty neighborhoods. M oves into low er poverty neighborhoods are often preceded by transitions into retirem ent from full-time employment and into em ploym ent from unemployment o r tem porarily laid off. Transitions into and out o f retirem ent are often related to moves to higher poverty neighborhoods. In sum, the inform ation in table 1 offers evidence that moves are associated w ith changes in family background. H ow ever, the evidence is fairly w eak in tw o ways. First, incom e changes are not highly correlated w ith moves. Second, negative shocks to family com position are as likely to be followed by m oves into better neighborhoods as w orse neighborhoods. A s a result, there does not appear to b e a consistent pattern in th e relationship betw een changes in observable family environment and changes in neighborhood choice. N evertheless, because a large part o f the variation used to identify the neighborhood effect is from moves, th e analysis m ust control as much as possible for these different circum stances. One im portant strategy is to directly control for family mobility. A nother promising feature o f the fixed effect estim ates is th at if the unobserved family change is constant across siblings, then the fixed family effect should eliminate this concern. H owever, if these changes affect siblings in different ways, w hich is possible given the age differences in the sample that I will w o rk with, then some discretion m ust be used in interpreting the results as caused by changes in com m unity influences rather than differing latent family conditions. M easurem ent E rror Since the neighborhood inputs, nj, are imperfect measures of the true effects o f communities, a final concern w ith the fixed effect (and simple linear probability and logit) * neighborhood equations is classical measurement error. In particular, assume th a t n. = n + v . , w here vj is an iid random variable and n* is the tru e community measure. D ifferencing across siblings exaggerates the measurement bias by creating a correlation betw een th e differenced inputs, n, and the differenced idiosyncratic shocks, v. eliminated while the noise remains. A s a result, much o f the tru e variance is The direction o f this measurement error bias is to w ard zero. The size o f the bias would be proportional to th e difference betw een the signal to noise ratio in 9 9 9 the estim ator w ithout fixed effects ( a /(or + o % ) and the fixed effect signal to noise ratio v v n 9 9 9 o ^ v/ ( a ^ v + a *). A com mon solution to this problem is to estim ate fixed effect, instrum ental An variable (FE-IV ) equations. B u t this reintroduces the problem o f finding believable instrum ents that are related to neighborhood differences but not to differences in outcom es betw een siblings that arise from other sources. I tried some specifications using a differenced version o f Evans, Oates, and Schwab’s county-level measures, although this instrum ent is susceptible to th e same criticisms as before. I f this is a classic m easurem ent error story, one neighborhood m easure could be used as an instrument for another, but this relies o n the precarious assum ption that the measurement error is uncorrelated between variables. Fortunately, the results presented in the next section show no evidence that first differencing increases the im portance o f classical measurement error . 13 In fact, m ost o f the fixed effect results are o f the same size o r larger than the simple OLS or logit models w ithout fixed effects. 1 Measurement error in the other right hand side variables could introduce upward bias in the neighborhood 3 estimates if there is a correlation between these measures and the neighborhood variables. For example, if there is measurement error in the change in family income and a correlation between family income changes and neighborhood changes, then the neighborhood measures will be positively biased. In most cases, however, several years of family data are available before and after the moves. Under the assumption that measurement error is iid, one could exploit this fact to construct an IV estimator. m Data T he data used in this pap er are from the Panel Study o f Incom e Dynamics (PSID ) and its accompanying geocode file for 1968-1985. Individuals are included in the sibling sam ple if they have a sibling at least three years apart in age and are in a respondent household fo r a t least tw o years while they are. betw een ages 10 and 14. Furtherm ore, the individual must have at least one year o f d ata after age 18 so that high school graduation can be ascertained. These constraints result in a sample o f 2,178 individuals from 742 families. A problem w ith using the age restriction is th at th e sample is com posed solely o f children from larger families. excluded. This can bias th e results in an unknow n direction w hen fixed effects are In this paper's model, parents invest less p er child in large families, resulting in a positive bias in the neighborhood effects measure. H ow ever, if there are spillovers from large families th a t m ake it easier for children to relate to community externalities, the bias m ight w o rk in the opposite direction. Fixed effect specifications sweeps out this family-specific heterogeneity. N evertheless, to gauge th e im portance o f this nonrandom sampling on the model w ithout fixed effects, I also construct a sample that includes all individuals th at fit the data demands except th e sibling requirements. The all-youth sam ple includes 4,410 people from 1,199 families. Table 2 includes descriptive statistics on th e main variables used in the analysis fo r th e sibling sample and the all youth sample. All statistics are w eighted using the PSDD-constructed probability o f selection into the sample. T he sibling sample appears to be roughly com parable to the all youth sample. Some small differences reflect th e larger family sizes o f the sibling sample. In particular, the sibling sample has low er education levels for the parents and children, low er household income, low er mobility, m ore minorities, and m ore tw o parent families. The N eighborhood M easure The believability o f the neighborhood proxy is key to the measurement o f neighborhood effects. Previous studies have used many different measures, including the fraction o f disadvantaged students in the individual's school (Evans, Oates, and Schwab 1992), the percentage o f families living in th e neighborhood w ith incomes below $10,000 and above $30,000 (B rooks-G unn et al. 1993), the percent o f families on w elfare (Corcoran et al. 1992), the percentage o f female-headed households (C orcoran et al 1992), and racial com position (Summers and W olfe 1977). D uncan (1994) tests many o f these m easures within the same data set and specification. Case and K atz (1991) aggregate household data to derive neighborhood averages o f the num erous outcom es they study. I concentrate the analysis on tw o variables that should pick u p many o f the dimensions hinted at in the above analyses. First, because th e outcom e m easure o f interest is high school graduation, I use a similarly defined variable th at m easures the percentage o f young adults in a census tract w ho w ere aged 16 to 19 in 1980 (16 to 21 in 1970) and w ho had not graduated from high school and w ere not in school. This variable can be thought o f as an extremely rough proxy for peer effects. Second, I use a variable that m easures the percentage o f households below the federal poverty threshold. This variable might be thought o f as a proxy for the effect o f adult neighbors and relative neighborhood conditions on youth achievement. I m ake no claims that these tw o variables will pick up all community-level influences. H ow ever, as much as the various influences are highly correlated, these tw o measures should be representative o f th e size o f the neighborhood effects on educational achievement. In the final section, I also report some preliminary results on three other neighborhood proxies — percentage o f female heads, average family income, and percentage o f population that is w hite -- to see if any patterns em erge when using these different measures. The data for these neighborhood variables com e from tw o sources. Geographic identifiers are reported in the PSID's geocode file. The geocode file is a set o f addresses collected from mailings to respondents. F rom these addresses, identifiers are assigned for various levels o f geographic aggregation. The smallest geographic area classified by the Census bureau is the census tract or block numbering area (BNA), w hich is the basis for the neighborhood measure used in this paper. A census tract is an area of, on average, 4,200 people that local authorities deem to be a 'neighborhood.' B N A s are the equivalent o f census tracts for untracted urban areas. W hen census tracts are unavailable, enum eration districts (ED), the rural equivalent o f census tracts, are used. W hen tracts, BN As, and E D s are unavailable, I employ five digit zip codes, w hich tend to encompass a larger area than the other identifiers. T hese geographic identifiers are m atched to 1970 and 1980 Census identifiers and data o n num erous area dimensions, including family structure, income, employment, race, education, housing, and mobility. I linearly interpolate neighborhood variables for the years 1970 to 1978 and set 1980 to 1985 and 1968 to 1969 values equal to their closest census year. T he effect o f this im putation schem e is examined in section V. T h e main neighborhood measure used is an average o f th e community conditions that the person lived in from ages 10 to 18. This averaging technique implicitly w eights each age equally in th e neighborhood impact estimate. It does n o t pick up any additional effect th at m ay occur from neighborhoods lived in prior to age 10 . 16 A s an alternative com putation, I also explore th e robustness o f th e results to using the age 14 m easure o f neighborhood conditions. This latter m easure is m ore commonly used in the literature but does not describe th e full history o f background influences that the person experiences. Therefore, it may be m ore susceptible to m easurem ent error relative to the averaged variable. T he O utcom e and Control Variables A lthough there are many dimensions in w hich peer and neighborhoods m ight be influential, I concentrate on educational outcom es. M o st o f the analysis focuses on w hether the individual graduated from high school. Section V reports som e findings using college attendance and grades com pleted . 1 7 I chose these education variables because, unlike teenage pregnancy, they do n o t limit the sample to a single gender. O ther variables th at are often studied in this literature, such as crime rates and drug use, are n ot available in the PSID . 1 For evidence ofneighborhood effects on younger children, see Brooks-Gunn et al (1993). No outcome measure 6 currently e i t for younger children in the PSID, which makes i d f i u t to determine neighborhood e f c s on xss t ifcl fet children prior to age 1 . 0 1 Using the PSID education variables might cause an attrition problem because grades completed are not 7 reported until an individual has finished f ll-time schooling. As a r s l , a number of individuals over age 18 u eut leave the sample without ever reporting data on grades completed. The high school graduation variable i coded s as 1 i the individual ever reports completing 12 grades or the individual remains in the sample a t r age 20 but f fe si l reports being a student Individuals who a t i e from the sample before age 20 without reporting grades tl trt completed are excluded. The analysis using college attendance and grades completed includes only individuals who report grades completed a t rage 1 . Grades completed are the greatest number ofgrades reported by age 25, fe 9 unless the individual has not reported a grade by age 25. In t i c s , the f r treported grade completed i used. hs ae is s Covariates in the basic regressions include gender, race, parental education, household income, parents' marital status, the number o f children living in the household, w hether the teenager w orked during her youth, and year b o m and region dummies. The income measure includes all labor income, transfers and asset income, n et o f th e teenager's income. Like the neighborhood variables, the time-vaiying family background variables are averaged over the years from age 10 to 18 th at the youth lives at home. This averaging technique results in m ore 'permanent' measures o f these variables. H owever, if tem porary changes, such as large income fluctuations, marital changes, or residential moves, are im portant, then this averaging might miss im portant factors in the likelihood o f continuing schooling. Therefore, I also include controls for a num ber o f transition variables that measure instability in th e family environment, including the variance o f income during the youth's years in the household, th e percentage o f years that the household moves, detailed marital transitions, and detailed em ployment transitions.1 I also * experiment w ith controls for birth order, whether the individual has an older sibling th at graduated from high school, and w hether the teenager moves into their ow n household by age 18 to see if these m easures o f individual heterogeneity change the estim ate o f com munity influences . *19 IV. Results H o w M uch Variation Is There in the W ithin-Familv M easures? B ecause I elect to use within-family variation rather than across-family variation to identify community influences, I first report some findings in table 3 on th e am ount o f variance that exists within families for four variables: grades com pleted, family income, and the tw o neighborhood variables. The first tw o rows display the m ean and standard deviation for th e entire sample. The third row gives the standard deviation w ithin families. W ithin family standard 1 The employment transitions a e: employment to unemployment, employment to r t r d employment to 1 r eie, temporarily l i o f unemployment t employment, r t r d to employment, and temporarily laid off t a d f, o eie o employment. The marriage t a s t o s a e marriage t divorce, marriage t widow, divorce to marriage, single t r n i i n r: o o o marriage, and widow to marriage. 1 I include the l t e variable only as a t s of the robustness of the results. Because of the endogeneity of the 9 atr et measure, i i probably best not included. ts deviation estim ates adjust fo r degrees o f freedom lost in taking deviations from family m eans . 30 The fourth ro w rep o rts th e fraction o f total variance th at is within-family. T here appears to be plenty o f variation in the education variable, w ith approxim ately 56 percent o f th e to tal variance in grades completed com ing from within-family differences . 31 U nfortunately, th ere is m uch less variation in the neighborhood variables. A bout 7.5 to 13.6 percent o f the to tal variance in the time-averaged neighborhood variables is attributable to withinfamily differences. Interestingly, this fraction o f variance is consistent w ith other time-varying family variables, such as total averaged family money income, w here 10.3 percent o f th e variance is from within-family differences. Therefore, family changes are n o t likely to dom inate changes in neighborhood background characteristics. This finding, combined w ith those from table 1 on incom e and family com position changes before moves, makes it less likely th at family changes will drive the neighborhood change param eters. As a result, w hile I rem ain cautious about th e effect o f family changes, there is reason to be optimistic that neighborhood effects can b e reliably estimated, especially if observable family changes are controlled. W hen th e ag e 14 m easure o f the neighborhood variable is used, the within-family standard deviation rises to 0.057-0.060 (from 0.034-0.036 w ith th e averaged neighborhood m easure), and the fraction o f th e to tal variance due to within-family deviations is 18.2 percent fo r th e poverty rate and 30.6 percent for the dropout rate, approximately tw ice as high as th e averaged m easures . 33 H ow ever, this difference is most likely picking up additional m easurem ent error because o f the shorter and less reliable window it measures. R esults reported in the next section that use these m easures are consistent with this measurement error story. 2 The within family standard deviation i calculated as 0 s 1-1 f I-F where N.~is the value of if the neighborhood variable for individual iin family f N^is the family mean of the neighborhood varia l , I i , be s the number of individuals in the sample, and F i the number of f m l e . s aiis The correlation between s blings in grades completed, high school graduation, and college attendance i i s .approximately . 5 . 8 3-3. 22 The correlation among a l siblings i around . for the two neighborhood measures. Correlation using the age l s 9 14 neighborhood measures are .85 for the poverty r t and . 5 f r the dropout r t . ae 7 o ae 2 1 T he bottom o f table 3 presents m ore information on the am ount o f differentiation that exists betw een siblings using the averaged neighborhood variables. E ach ro w displays the percentage o f sibling pairs w hose average neighborhood background m easure differs by 5, 10, 20, 30, o r 50 percent.2* Approximately 71 (76) percent o f th e sibling pairs live in an average neighborhood during ages 10-18 that is at least 5 percent different ia poverty (dropout) rates. The percentage th at lives in neighborhoods differentiated by at least 10 percent rem ains fairly high at 58 percent fo r the poverty measure and 62 percent for the dropout measure. At 2 0 , 30 and 50 percent differentiation, few er pairs qualify: 35 to 41 percent o f th e pairs g ro w u p in 20 percent different average community environments and 7 to 11 percent g ro w up in 50 percent different average com munity environments. So while the majority o f the variation is clearly across families, a small am ount still exists across siblings. N ow , I tu rn to the estimation. First, I present som e base case linear probability and logit models using high school graduation as the outcom e measure. N ext, I estim ate tw o stage equations th at are similar in spirit to Evans, Oates, and Schw ab’s (EO S from here on) instrumental variables technique for correcting neighborhood selectivity. I then present the main part o f the analysis, the fixed effect equations. All results to this point use high school graduation as the outcom e m easure and poverty and dropout rates as th e neighborhood proxy. T he next section tests th e robustness o f th e findings to changes in th e neighborhood proxy, outcom e measure, and sample. Some further tests to determine the im portance o f d ata variability are also reported in the following section. Single Stage Estim ates Table 4 displays neighborhood coefficients from simple, one-stage ordinary least squares and logit high school graduation equations using the 2,178 children in the sibling sample. Full linear probability and logit regression results for a few selected equations are reported in appendices l a and lb . The appendices include findings from the sibling sam ple and the all youth sample but table 4 reports results only for the sibling sample. 2 This i not an absolute deviation but r ther a r l t v d v a i n Therefore, a poverty r t of 13 percent for one 3 s a eaie eito. ae s b i gversus 10 percent f r another i reported as a 30 percent difference in table 3 iln o s . The top row of table 4 reports the r sults from regressions that allow the neighborhood e variable to enter log-linearly.2 For both the neighborhood poverty and dropout r t s higher 4 ae, values signify a reduced probability of graduating from high school. As an interpretation of the size of these (linear probability) coeffi i n s a 10 percentage point increase in the neighborhood cet, poverty rate would reduce the likelihood of graduating from high school by 2.1 percent. The corresponding impact of a 10 percentage.point increase in the dropout rate i 3.6 percent. These s effects seem f i l large, so I reran the regressions using the all-youth sample of 4,410 children ary that does not require the existence of a 3 year age-separated s b i g This sample produces 35 iln. percent smaller point estimates than the s b i g sample, but t s s of the sample coefficients show iln et that these differences are not s a i t c l y s g i i a t The all-youth sample results are in line with ttsial infcn. the small neighborhood effects findings from previous work. Therefore, i should be kept in mind t that the estimates presented below are representative of the impact of neighborhood, family, and individual characteristics on the educational attainment of children from la r g e f m l e . aiis These regressions include controls for race, gender, household income, parents' marital s a u , whether the father or mother graduated from high school, whether the kid worked, the tts number of kids in the household, and the county unskilled wage r t . Because these variables are ae mostly averages over the c i d s youth from ages 10 to 18, they may not capture important hl' fluctuations in environmental conditions. These fluctuations may be c i i a i changes in rtcl f neighborhood conditions are picking up unobserved heterogeneity in family or individual background rather than the true effects of the community. Therefore, 1 experimented with a variety of such measures to see i omitting them affects the magnitude of the neighborhood f c e f c e t In particular, I t ofiin. ried the percentage of years that the household moved, the variance of family income over the youth’ data, detailed transitions into employment, detailed transitions s into marital s a u , whether the individual moved into her own household by age 18, birth order, tts and whether the individual has an older s b i g that graduated from high school. The results i iln n table 4 include the marital and employment t a s t o s Many of these factors are important rniin. determinants in the probability of graduating from high school, particularly the mobility measures 24 The analyses to follow are similar if the neighborhood measure is specified linearly. 20 and whether the individual moved into her own household. However, none signifi a t y affect the cnl magnitude of the neighborhood c e f c e t While I may not be picking up other important ofiin. factors that might be correlated with the neighborhood variable, I am reassured that adding these factors, especially the indicator variable for moves into own household, do not affect the neighborhopd parameters. The bottom of table 4 reports the neighborhood coefficients when they are allowed to enter nonlinearly. A spline i created at the 25th, 50th, and 75th percentile of the neighborhood s measure and also at only the 90th percentile to determine i there are nonlinear slopes i the f n neighborhood coefficient depending on the "quality" of the neighborhood. In both the poverty rate and dropout rate cases, there does not appear to be much evidence that such a nonlinearity e i t . A notable exception i that the neighborhoods i the bottom decile of dropout rates xss s n (above 28 percent) exhibit stronger effects in the linear probability case. However, no such pattern i detected in the l g t model. Therefore, the remaining analysis ignores any possible s oi spline e f c s 2 f e t .3 A BriefNote on IV Estimates Before presenting the fixed effect estimates, I t ried to replicate E O S ’ two stage estimator s that uses a variety of metropolitan area characteristics to instrument for neighborhood selection. They find that single stage equations show a s g i i a t impact ofneighborhoods on teen birth and infcn high school dropout rates but modeling the selection process using IV eliminates t i entire effect hs Although I use a different data set than EOS, the re u t are quite similar when using the sls neighborhood poverty r t . 4 This i reassuring since t e r peer variable i similar to the poverty ae2 s hi s rate variable used here. However, when the dropout rate i used as the neighborhood measure, the s 23 I also looked at interactions between the neighborhood variables and a number of the family and individual characteristics to see if nonlinearities enter this way. The importance of these nonlinearities is sensitive to the choice of the neighborhood measure. For the poverty variable, only the gender interaction is significant Females are less likely to be affected by high community poverty rates. The dropout rate interactions appear to be more important Income and 'whether worked1 interactions are positive and significant at the one percent level; the number of kids in the household is negative and significant at the one percent level. These results suggest that kids who do not work during their youth and are from lower income households with more children are more susceptible to negative neighborhoods externalities when the youth dropout rate is higher. 24 The results are available upon request. findings are di f r n . Using EOS's instruments, the neighborhood effects are sil of the expected feet tl sign and the point estimate i bigger than the single stage r s l s although they are insignificant at s eut, any conventional level due to a substantial increase in the standard err r Therefore, m y IV o. results suggest that controlling for s l c i i y can eliminate the s gnificant effect of neighborhoods eetvt i on children's high school graduation. But t i conclusion i f i l sensitive and prone to substantial hs s ary increases in imprecision. Fixed Effect Estimates As explained e r i r an alternative approach to correcting the s lectivity b a , or at least ale, e is the family-specific component of location decisions, i to estimate family fixed effect equations. s Table 5 reports these estimates using eight methods.2 In row one oftable 5 the 2,178 individuals 7 , are paired off with siblings that meet the age c i e i n This leaves 1,892 sibling pairs that are rtro. differenced to eliminate the family constant error term. Full results of th s equation are reported i in appendix 2 This procedure has a large impact on the poverty rate measures. The point . estimate of -0.144 corresponds to a seven percent decrease in the likelihood of graduating when the neighborhood poverty rate increases by ten percent. This result i significant at the two s percent l v l On the other hand, the dropout rate coefficient i not affected by the f r t difference ee. s is estimator. The coefficient increases s i h l from -0.061 i the linear probability model to -0.068 lgty n in the f r t difference model. However, a substantial increase in the standard error produces an is increase in the p-value to about the s percent significance l v l ix ee. These findings are robust to adding more controls to account for transitions that may d f e across s b i g . Like the linear probability regressions, I add controls for the variance of ifr ilns money income and labor income, the percentage of years that the household moves, and whether the teenager moved into her own household by age 18. None of these variables, individually or as a group, affects the neighborhood estimates using either neighborhood measure. Even the 'own household' variable, which i highly significant i the regressions, has no indirect effect on the s n magnitude of the neighborhood estimates. 27 The regressions include controls for employment and marital transitions that occur after the older child has left the parents' home. Therefore, these variables should measure changes in employment and family states that are experienced by the younger sibling but not the older sibling in the pair. These findings form the basic r s l s However, as noted i section n, the inferences eut. n would be more convincing i the estimates were consistent across a number of changes in the f specification and assumptions of the model. In the second row, I weight the sample as described in section I using the inverse of the number of times each person i included in a sibling p i . I s ar This change has no effect on either neighborhood measure. In row t r a single fixed effect i h ee, s employed for each family. This lowers the coefficients and standard errors on both measures, but the findings remain similarto the pairwise estimates. A simpler method to a l v a e concern about over sampling i to choose one random pair leit s of s b i g from each family. This i shown i row four, where only the oldest and youngest ilns s n s b ings from the 742 families are used. The point estimate for the dropout variable remains the il same but the poverty rate estimate (standard error) i smaller at -.115 (.067) than the other f r t s is difference estimates, although silremaining almost one standard deviation higher than the simple tl l n a probability estimates. ier All of these estimates use two sources of variation to identify the neighborhood c e f c e t time and differences in resid n i l location. Time influences these results because the ofiin: eta interpolation of the decennial census figures implies that neighborhood conditions will d f e ifr between s b ings of different ages even i they l v in the same neighborhood.2 To see whether il f ie * t i time component i driving the r s l s i rows five and s x I reestimate the unweighted hs s eut, n i, pairwise f r t difference estimator i row one using only those si l n pairs who lived in different is n big neighborhoods at age 14 (row f v ) and who have a d fferent average neighborhood during the r ie i i youth (row s x . The l t e restriction eliminates only those pairs who never moved. i) atr Although the point estimates in the age 14 estimator are very similar i magnitude to the other estimates i n n table 5 the small sample from using the diff r n neighborhood age 14 estimator results in huge , eet standard errors and thus i s n ignificant estimates. The d fferent average neighborhood re t i t o i srcin r esults i an increase i the point estimate ofthe poverty rate and no change i the dropout r t . n n n ae2 22 Secular trends in high school graduation rates are controlled with year turned age 15 dummies. Here, I use time to refer to within-neighborhood changes over time (as opposed to across neighborhood changes due to household moves). In rows 7-9, I return to the l g s i framework. oitc R o w 8 reports unweighted logit estimates using Chamberlain's fixed effect l g t model. Because only pairs of siblings that had oi different educational outcomes contribute to the likelihood function, 441 of the 1,892 pairs are usable. The conditional logit coefficients are in different units and therefore are not comparable to the single stage l g t But the estimates, l k the linear probability estimates in row one, are oi. ie signi i a t and of the expected sign for both neighborhood measures. Other coefficients in the fcn model seem to react similarly in the conditional l g t and linear probability equations. oi F n l y in row 9 l ial, , ogit equations were estimated using the Mundlak formulation where separate within-family and across-family variables are defined. For the poverty rate measure, (the within-family neighborhood coefficient) i -0.872 (0.432) and s dropout r t , ae the findings on i -0.407 (0.357) and s <x j > <2 f > <2 j > $x i 0.415 (0.456). For the s i -0.436 (0.402). Not surprisingly given their s m l r t , s iiaiy are very similar in terms of t-values to the one family fixed effect estimates in row 3 In terms of other si l n or time-varying variables, the within-family coefficient i . big s signi i a t for the female indicator, marital s a u , and whether worked dummy. Family income fcn tts and marital status are highly significant in the across-family point estimates. V. Robustness Checks Using Different Neighborhood Measures H o w sensitive are the results to the choice and computation of the neighborhood proxy? To test t i measurement i s e I reran the models using different proxies and aggregations of the hs su, neighborhood variable, a l of which have appeared in the l t r t r i some form. Tables 6 and 7 l ieaue n report the re u t of th s investigation. sls i In table 6 I examine whether the way the neighborhood measure i calculated has any , s bearing on the findings. In p r i u a , t i table looks a the eff c of the imputation scheme and atclr hs t et the averaging of the neighborhood v riable. Linear probability and unweighted fixed effects a estimates are reported for four variants of the neighborhood measure. F r t i row one, I report is, n the basic results from e r i r tables that use an imputed, time-averaged neighborhood measure. ale As described in section ID, the imputation r f to the l n a interpolation between 1970 and e ers ier 1980 ofthe decennial Census variables, which allows some variation i the neighborhood measure n from time. The time averaging refers to the averaging of variables from ages 10 to 18 for each s b i g In row two, I allow no imputation, setting the neighborhood measure for each year equal iln. to the 1980 Census report for that neighborhood. Therefore, a l si l n differences in l big neighborhood measures w l be from neighborhood moves. The res l of t i change i s il ut hs s mall. The fixed e f estimator even increases for both the poverty and dropout equations. Therefore, f ect biases caused by the current imputation scheme, i anything, dampen the size of the neighborhood f e f c , suggesting that t i data assumption i probably not a problem. fet hs s In rows three and f our, I use the age 14 neighborhood measure instead of the average neighborhood c a a t r s i . I run t i experiment because one-year windows are a common way hrceitc hs to measure the influence of neighborhoods and schools on children. However, t i variable may hs not be a r l a l measure of the true effects of neighborhoods (or any other time-varying eibe covariate) since i i ignores the re t of the individual’ neighborhood history and thus potentially ts s s introduces more measurement noise into the estimation. However, as a methodological and comparative point, i seems to be an useful exercise to compare the results using th s measure t i with the averaged measure. The magnitude of the age 14 estimates i quite different from the averaged variable. In s row three, the census imputation i allowed. s Both the fixed effect and linear probability estimators are i s g i i a t for the poverty measure; the dropout measure shows significant effects ninfcn with the l n a probability estimate but s i h l smaller and i s ier lgty n ignificant results using the fixed e f c equation due to a doubling of the standard e fet rrors. Furthermore, the size of the linear probability coefficients i smaller, although sil s g i i a t R o w four drops the imputation and s tl i n f c n . finds that the fixed eff c estimator i signi i a t a the 10 percent level for the dropout rate but et s fcn t not s g i i a t y d fferent from zero with regard to the poverty r t . I would be comforting to infcnl i ae t find that the r esults using the age 14 measures match the findings from the averaged variable measures. That t i i not the case i not a f t l contradiction. F r t the findings are somewhat hs s s aa is, supportive of the main conclusion of the fixed e f c estimates; correcting for s l c i i y and fet eetvt unobserved family heterogeneity does not completely eliminate the p s i i i y of community osblt influences. While three out of four of the findings are insign f c n a even the 10 percent l v l iiat t ee, the point estimates are in l n with the linear probability c e f c e t , just much le s precisely ie ofiins s estimated. Second, I would expect that the age 14 measure i not as good a proxy of the youth's s f l history of neighborhood background influences and thus i more prone to measurement e r r ul s ro. Therefore, while i would comforting to find that the age 14 measure and the averaged measure t come to exactly the same conclusions, i i not surprising that they do not.1 ts 9 As a second t s , the linear probability and unweighted fixed effect estimators were rerun et for three other neighborhood variables commonly used in the l t r t r : the percentage of ieaue households headed by females, the percentage of the population that i white, and average s household income. The findings are reported in table 7 Column one reports linear probability . estimates using each of the five neighborhood variables. To gauge the r l tive size of these ea coefficients against each other, column two displays derivatives calculated at the mean for each measure. The size of these derivatives i f i l stable across the variables, with the exception s ary being the white composition, which i about the same size as the other variables for white students s but zero for nonwhite students. Column three reports the fixed effect estimates. The three variables not discussed above have point estimates very similar to the linear probability model, but with standard errors roughly three times as l r e In a l three cases, the increase i standard errors res l i insi n f c n ag. l n ut n giiat neighborhood e f c s However, no cases show the dramatic changes i magnitude, much l s fet. n es sign switches, that are reported i EOS. However, the lack of precision of the within-family n estimates does not discount t i p s i i i y hs osblt. Using Different Outcome Measures In table 8 I explore how the fixed effect estimator influences two other education , outcome variables — whether the individual attended college and the number of grades completed by age 25. Two points need to be made about variable definitions and the sample. F r t in order is, to maximize sample size and to avoid problems due to a t i i n I include a l individuals who were trto, l 19 These findings on single age windows versus averaged values is consistent with those reported in An, Ha ve ma n and Wolfe (1992). 26 in the sample and had a grade completion report after age 19, much l k the high school ie graduation equations. However, t i may cause problems, particularly in the grades reported hs variable, since differences i grades reported may be partly due to differences in age. Therefore, a n variable that measures the l s age in the sample used to determine grade completed i included. at s Second, I drop 41 individuals for whom high school graduation was inferred from t heir status as students well into t heir twenties but who never report a grade completed due to a t i i n or the trto end ofthe sampling period. This sampling alteration makes l t l difference to the findings. ite Li rows 1 2 and 4,1 report the neighborhood coefficients for equations that employ high ,, school graduation, college attendance, and number of grades completed as the outcome measure. The regressions were also run for l g t models with identical implications. The findings suggest oi that the effect of neighborhoods on college attendance i much smaller than on high school s graduation, especially when looking at a subsample of siblings who graduated from high school (row 3 . Once again, the neighborhood proxies seem to be acting quite d f e e t y Fixed e f c ) ifrnl. fet estimates using the poverty rate measure suggest that unobserved heterogeneity does not eliminate the e f of communities on educational attainment. In a l three outcome measures, f ect l point estimates (and standard error) increase in magnitude, but remain relatively constant in significance. The dropout rate shows strong support for neighborhood effects in the l inear probability and l g t specific t o s but no evidence of neighborhood effects in the fixed e f c oi ain, fet models, especially with regard to the college attendance and grades completed outcomes. Further analysis using the percentage of households over $30,000 in 1979 dollars as the neighborhood proxy finds some support for the importance of neighborhood conditions on college attendance decisions. Using Different Samples— Separating the Responses Bv Race. Gender, and Income I may also be of i t r s to see how s r t f i g the sample by race and gender affects the t neet taiyn magnitude and significance of the neighborhood estimates. Table 9 reports these r esults using the high school graduation rate as the dependent variable. The f r t rows c a sify the sample by race. is ls The nonwhite sample experiences larger neighborhood influences than the white population as measured i the single stage framework when either the poverty rate or dropout rate i used as the n s neighborhood variable. However, the f r tdifference estimator suggests larger effects in the white is sample for both neighborhood measures. In the nonwhite sample, no s a i t c l y significant effect ttsial i found. s None of these results are s a i t c l y different across groups at conventional ttsial significance l v l . ees When s s e s and brothers are s r t f e i rows three and four, the results are rather itr taiid n surprising. R o w three (four) include a lfemales (males) in the linear probability estimates but only* l s s e (brother) pairs in the fixed effect regressions. The s s e s do not seem to react to poverty itr itr conditions, especially relative to the brothers. However the s s e s have a strong response to itr neighborhood dropout r t s The brothers respond strongly to both neighborhood conditions, ae. although heterogeneity corrections have quite different effects depending on the neighborhood proxy used. With regard to the poverty r t , the point estimates are extremely large (although so ae are the standard errors), while the dropout rate parameters are similar to those found in the aggregate. Again, although some differences a rise between the two groups, these differences are not s a i t c l y s g i i a t ttsial infcn. A Test ofData Variability Given the small variation that I rely on to identify the point estimates, i might be useful t t o redo the analysis using si l n pairs that experience larger neighborhood d f e e t a s 3 There may big i f r n i l .0 be i t rest i sibling pairs with larger neighborhood differences i there i concern that those with ne n f s small differences are especially noisy. However, there i a trade-off as the sample with larger s differences i more susceptible to bias from l t family characteristics that may have caused s a ent large changes in neighborhood location. Furthermore, a p i r , i i not clear whether the roi t s neighborhood influence should be l r e , smaller, or the same size with greater differences in agr sibling neighborhood backgrounds. If families are selecting neighborhoods based on the d f e e t a a i i y of their children, then larger changes in the quality of the neighborhood would ifrnil blt 3aThis issue is essentially one of measurement error. Since, classical measurement error is likely to lead to a larger downward bias in the fixed effect estimates, it is not of great concern in this case. I did run some FE-IV models to formally account for measurement error in the difference neighborhood input, but because of the difficulty in finding a reliable instrument for this model, it is not clear that these techniques will solve any problem. Further, the results are hard to interpret The standard errors increase dramatically with the loss of efficiency overwhelming any information that might allow better estimates of the neighborhood parameters. 28 show bigger effects on the outcome measures i the family i following a reinforcing strategy f s (moving to better neighborhoods to accommodate the more able child) and smaller effects i the f family i following a compensating stra e y Ifthere i no d f e e t a s l ction, then the estimates s tg. s ifrnil ee should be roughly the same magnitude regardless of the size of the neighborhood characteristic difference. Table 10 gives the r sults when the high school graduation model i remn on samples that e s are s r t f e based on the percent difference in the neighborhood characteristics between the taiid sbig. ilns In row zero, the unweighted pairwise estimator from table 5 i reported as the base s case. Rows one through four break down the s b i g sample into those pairs with 5,10, 20, and iln 30 percent differences in t heir neighborhood measures. The results are consistent, although far from conclusive, that parents do not sel c neighborhoods based on the a i i y of their children. et blt There does not appear to be much difference in the point estimates across the categories. For the dropout r t , a Wald te t ofthe equality of coefficients overwhelmingly shows no difference i the ae s n magnitude of the neighborhood impact even when the sample i limited to only those pairs whose s average neighborhood dropout rate i 30 percent d f e e t As for the poverty r t , the larger s ifrn. ae differences, especially the 30 percent l v l have somewhat smaller e f c s but these differences ee, fet, are s a i t c l y indistinguishable from a l the other categories. When the sample excludes the ttsial l largest s b i g d f e e t a s the results are also comparable, suggesting that outliers are not driving iln ifrnil, these f ndings. Therefore, the r i esults seem robust to the neighborhood d f e e t a used. ifrnil Reconciling The Findings With Plotnick and Hoffman Plotnick and Hoffinan (1995) use the same family fixed effect approach and find no evidence that neighborhoods matter. The discrepancy between our r esults i partly due to s differences in variable and sample d f n t o . This i exemplified in the robustness checks of eiiin s tables 6 to 9 Many of t e r specification and variable definition choices are shown in these tables . hi to res l i insigni i a t neighborhood c e f c e t . In p r i u a , I highlight four i s e . ut n fcn ofiins atclr sus F r t Plotnick and Hoffman's neighborhood measures are composed of averages over is, three years (age 16 to 1 ) Results reported i table 6 suggest that shorter time frames can lead 8. n to a reduction i parameter estimates. An, Haveman, and Wolfe (1992) argue that these shorter n 29 windows are consistent with added measurement e r r or, which i l k s i ely to bias estimates downward. Furthermore, Plotnick and Hoffman acknowledge that the use of a 16 to 18 window ignores potential effects a e r i r ages or the accumulation of neighborhood significance over t ale many years. Second, several of the neighborhood measures employed in table 7 some of which , are used in Plotnick and Hoffman’ paper, display i s s n ignificant neighborhood e f c s Therefore, fet. the results are sensitive to the precise neighborhood measure chosen. Third, their education dependent variable i post-secondary schooling, which I show in table 8 to display a much weaker s impact from neighborhood conditions. The stronger effects arise in high school graduation outcomes. Fourth, Plotnick and Hoffman include only s s e p i s which I find in table 9 to itr ar, display smaller neighborhood effects than brothers and brother-sister pairings. Some of these specification, sample, and variable definitions are a b t a y especially the choice of neighborhood rirr, measure. However, the averaging problem seems to be an important measurement issue where more ‘ permanent’ covariates are preferable. Other is u s such as the choice of dependent se, variable and the sampling of s s e s versus brothers, suggests that neighborhoods could matter in itr certain cases. V L Conclusions A well-known complication of estimating the influence of neighborhoods on children's outcomes arises because families are not randomly assigned to neighborhoods but rather choose t heir location based on many factors, including the importance they place on their children’ s welfare. As a r s l , the effects of family unobservables, such as parental competence, taste for eut education, and time spent with their children, and other unobservables that are common to geographically clustered households, may be mistakenly attributed to the neighborhood measures. Previous studies that attempt to correct for t i selection bias have used questionable instrument hs varia l s be. This paper introduces an approach that r l e on the observation that the latent factors eis associated with neighborhood choice do not vary across s b i g . Therefore, family residential ilns changes provide a source of neighborhood background variation within families that i free of s family-specific heterogeneity biases associated with neighborhood se e t o . lcin This approach i s feasible because of the high levels of r s d n i l migration i the United S ates. Using a sample of eieta n t multiple-child PS1D families where the kids are separated i age by at l a t three years, I estimate n es family fixed effect equations of children's educational outcomes. The fixed effect results suggest that the impact of neighborhoods exists even when family-specific unobservables are controlled. In f c , fhmily fixed effect regressions that use the neighborhood poverty rate as the proxy for at community conditions show even larger community ef e t on high school graduation and grades fcs completed compared with the models without fixed e f c s When the neighborhood dropout rate fet. i employed, there appears to be l t l difference i point estimates between the fixed effect high s ite n school graduation equations and the simple l n a probability or l g t r s l s but the effect on ier oi eut, college attendance and grades disappears. Other neighborhood proxies show similar patterns. When s r t f e by race and gender, whites and males are impacted the most by neighborhood taiid conditions i the fixed eff c sp cifications. Therefore, the re u t suggest t a , contrary to n et e sls ht Evans, Oates, and Schwab's findings, corrections for neighborhood selection biases do not necessarily eliminate the potential for signi i a t community e f c s fcn fet. However, the findings are tempered to some degree by large standard errors due to small sample s izes and noise that might a rise i there i not enough variation in the differenced f s neighborhood variables. While attempts to control for family environment are introduced, there i s also the p s i i i y that the empirical models have not adequately isolated latent changes in family osblt background or individual s b i g heterogeneity. This i exemplified by the surprising finding that iln s parameters sometimes increase when moving from the single stage models to the fixed e ffect models. Therefore, in future research, I hope to replicate the r esults on a different sample of the PSID (younger children using grade retention data currently being collected) or a different data s t such as the National Longitudinal Survey of Youth. e, Table 1 Changes in Household Income, Employment Status and Family Composition Preceding a Residential Mov$ (1,2 1971-1974,1980-1983 Stavers d) Family years (3 Families State Movers (2) County Movers (3) 3,705 911 44 39 109 82 32.95 0.57 0.99 32.42 -0.66 •0.58 24.16 0.05 -0.11 26.20 -206 •3.29* Parents' marital status in move year Change (t-1 ,t) Change (t-2,t) Married->divorced Divorced->married 0.737 -0.015 -0.024 0.030 0.014 0.727 -0.046 -0.046 0.114 ••• 0.114 — 0.798 0.055 *** 0.027 ** 0.064 ** 0.110*** Head's employment status in move year Change (t-1 ,t) Change (t-2,t) Employed->unempl. Employed->retired Employed->laid off Unemployed->empl. Retired->emptoyed Laid off->employed 0.800 -0.014 -0.021 0.022 0.018 0.020 0.013 0.022 0.018 0.818 -0.023 0.000 0.023 0.023 0.068 ** 0.023 0.045 0.045 0.844 0.046 ** 0.018 0.037 0.037 0.064 *** 0.055 *** 0.028 0.018 Family money income (4 in move year Change (t-1,t) Change (t-2,t) Head and wife labor income (4 in move year Change (t-1,t) Change (t-2,t) Residence Movers (4) Neighborhood Movers Absolute poverty rate relative to previous neighborhood Hloher 5% Higher Lower 5% Lower (6) (7) (8) (9) m (5) 589 371 387 260 164 143 30.50 -0.80 -0.40 24.44 -0.42* 0.31 25.42 -0.57* 0.30 25.29 -0.67 -0.39 17.09 -059 -0.18 102 94 205 174 147 127 25.51 -1.86 ** •0.59 22.55 •0.64 -0.53 24.96 0.28 1.09 24.97 0.88 1.27 17.42 -0.86* -0.47 18.19 -1.46 ** •0.98 13.62 -1.33 •0.99 16.32 -0.58 •0.26 15.59 -0.02 0.26 0.565 -0.014 -0.051 ** 0.087 *** 0.058 *** 0.553 •0.018 -0.059 ** 0.093 *** 0.054 ** 0.530 -0.019 -0.055 0.091 *** 0.037 0.402 •0.049* -0.069* 0.098 *** 0.020# 0.566 •0.014 -0.063 ** 0.102 *** 0.073 *** 0.551 0.014 -0.020 0.082 0.095 •** 0.696 0.012 ** -0.014 0.034* 0.051 *** 0.044 *** 0.031 *** 0.037 ** 0.026 0.703 0.026 ** -0.008 0.036* 0.054 *** 0.036** 0.034 *** 0.041 ** 0.018 0.677 0.012 -0.030 0.037 0.043 ** 0.037 0.012 # 0.043* 0.012 0.588 0.010 •0.039 0.039 0.039 0.049 ** 0.020 0.059 ** 0.020 0.727 0.049 *** 0.020* 0.034 0.068 *** 0.039 0.054 *** 0.039 0.024 0.721 0.061 * 0.048 *• 0.020 0.088 *** 0.020 0.048 0.020 0.027 Notes: 1) Asterisks represent significance levels from mean tests of the mover groups against the stayer group. *(**,***)«mean of the movers' characteristics is different from the stayers at the 10% (5%,1 %) level. # means that columns (6) or (7) are significantly different from columns (8) or (9) at the 5 percent level. 2) The sample includes households that have a child under age 17 living in the household. Only years 1971-1974 and 1980-1983 are used to avoid difficulties in determining when geographic moves occurred during periods when geocode data is missing (1969,1975,1977,1978). 3) Family year observations are included for all moves where that period's income can be determined. Since the PSID does not report income until the following year, geographic moves during the final year of a household's response are not included. 4) Income is in thousands of 1982-1984 dollars. Table 2 Descriptive Statistics of Main Individual, Neighborhood and Family Variables ( 1 Weighted by PSID Sample Weights Sibling sample_____ Mean Std. Dev. () 2 (D High school graduate College attendance Number of grades completed Nonwhite Female Percent worked during youth Number of kids i household n M o m high school graduate Dad high school graduate Household money income (82-84 $) Parents' married a l years (2 l Percentage of years that family moved between ages 10 and 18 Whether ever moved, ages 10 to 18 Neighborhood Characteristics: Percent households i poverty n Percent youth not employed or i school n Percent white Percent female household heads Average income 0.871 0.425 12.88 0.186 0.494 0.801 3.23 0.631 0.561 40,046 0.740 0.122 0.335 0.494 1.89 0.389 0.500 0.400 1.56 0.483 0.496 23,737 0.439 0.168 0.878 0.445 12.99 0.177 0.493 0.808 2.97 0.678 0.595 40,482 0.722 0.131 0.327 0.497 1.95 0.382 0.500 0.394. 1.60 0.467 0.491 24,452 0.448 0.175 0.485 0.500 0.514 0.500 0.125 0.134 0.852 0.134 40,761 0.097 0.094 0.252 0.088 14,051 0.127 0.131 0.856 0.132 41,309 0.097 0.095 0.240 0.087 14,802 0.359 0.220 0.288 0.110 0.145 0.315 0.243 0.298 0.146 0.284 0.252 0.320 0.169 0.166 0.055 0.114 0.013 0.026 0.120 0.073 0.093 0.028 0.091 0.071 0.110 0.031 0.372 0.228 0.318 0.116 0.159 0.325 0.257 0.291 0.165 0.288 0.257 0.313 0.174 Head experienced at least one transition during ages 10-18 of youth: Married -> divorced 0.152 Mamed -> widowed 0.051 Divorced -> married 0.091 Single -> married 0.012 Widowed -> married 0.021 Employed -> unemployed 0.111 Employed -> retired 0.063 Employed -> temp, l i off ad 0.098 Employed -> disabled 0.022 Unemployed -> employed 0.088 Retired -> employed 0.068 Temp, l i off-> employed ad 0.115 Disabled -> employed 0.029 Number of unique individuals Number of unique families A l youth sample l Mean Std. Dev. () 3 (4) 2,178 742 4,410 1,199 Notes: 1 Sibling sample includes a l individuals with ( ) one sibling that i three years apart i age, ( ) two ) l 1 s n 2 years of data between ages 10 and 14, and one year after age 18 that can distinguish whether the individual graduated from high school. A l youth sample does not require condition ( ) l 1. Variables are averaged for each individual between ages 10 and 18. Family background variables are averaged over the years that the person lived a home. Some 9% of the sample moved out of t their parents' household by age 18. 2) Equals one i the parents stay married while the child i l v n at home between ages 10 and 18. f s iig Table 3 Within-Family Variance i Some Key Variables n - Averaged Variable ( 1 Mean of variable Total standard deviation in sample Standard deviation within families Fraction of variance within families Age 14 Variable (2 Mean of variable Total standard deviation i sample n Standard deviation within families Fraction of variance within families Percentage of sibling pairs whose neighborhood measures are different by: >5% >10% >20% >30% >50% Grades Completed (D 12.52 1.87 1.40 0.560 Neighborhood Poverty Dropout Rate Rate () 2 0) 0.198 0.131 0.036 0.075 0.167 0.093 0.034 0.136 0.199 0.140 0.060 0.182 0.169 0.104 0.057 0.306 0.71 0.58 0.35 0.21 0.07 0.76 0.62 0.41 0.26 0.11 Notes: 1) Neighborhood and income variables are averaged over ages 10 to 18 for each s ibling. Education outcome variables are based on the highest reported grade completed from age 19 to 25. 2) Variable i the measure at age 14 (orthe closest age to 14). s Family Money Income () 4 30,433 19,975 6,403 0.103 Table 4 Effect of Neighborhood Poverty and Dropout Rate on High School Graduation Rates (1 Neighborhood Measures; logfpoverty rate) or logfdropout rate) Dependent variable: 1 if high school graduate (Huber standard errors in parentheses) (2 d) Linear Specification Neighborhood Measure: log(poverty rate)_______________ Linear Probability________ ____________ Logit____________ (3) (2) (4) (6) (5) -0.042 ••• (0.016) -0.034 * (0.019) -0.042 ••• (0.016) -0.543 — (0.155) -1.360 *** (0.352) -0.596 *** (0.164) ________________Neighborhood Measure: log(dropout rate) ________ Linear Probability________ _____________Logit (8) (9) (10) (7) (11) -0.061 **• (0.013) -0.016 (0.013) -0.047 •** (0.012) -0.805 •*• (0.163) -0.784 ** (0.371) (12) -0.749 *** (0.178) Spline Specification (3 Neighborhood measure 25th percentile 50 percentile 75th percentile -0.002 (0.013) 0.001 (0.015) -0.008 (0.016) Sign, level from test of nonlinear coefficients All spline slopes=0 0.079 (0.116) -0.003 (0.020) 90th percentile 0.943 •0.023 * (0.014) 0.004 (0.017) -0.034* (0.018) 0.249 * (0.132) 0.153 (0.109) 0.045 (0.096) 0.867 0.088 0.506 Notes: ‘“ -significant at 1% level “ =significant at 5% level ^significant at 10% level 1) Sample size is 2,178. Regressions control for gender, race, parents' education, parents' marital status, household income, number of siblings, whether the child worked, the year the child turned 15, five marital transition variables, eight employment transition variables, and the wage for unskilled workers in the county of residence. 2) Standard errors corrected for clustering by 1968 neighborhood. 3) Regressions include a spline at the 25th, 50th, and 75th (or 90th) percentiles of the neighborhood measure. The breakpoints for the poverty rate are 8.9,17.5,28.0, and 38.3%. Corresponding breakpoints for the dropout rate are 9.9,16.2,22.6, and 28.0%. -0.053 (0.152) 0.127 (0.118) -0.128 (0.121) -0.045 ** (0.021) 0.009 0.035 -0.078 (0.120) 0.678 0.516 Table 5 Effect of Neighborhood Poverty and Dropout Rates on High School Graduation Fixed Effect Estimates ( 1 Neighborhood Measures: log(poverty rate) or log(dropout rate) Dependent variable: 1 i high school graduate f (Huber standard errors i parentheses) n Estimators (2 Linear Probability Models (0) Base case loa (Dovertv) () 1 loa (droDout) (2) Size () 3 -0.042 * * * (0.016) -0.061 * * * (0.013) 2,178 (1 Unweighted pairwise f r tdifference ) is -0.144 * * (0.060) -0.068 * (0.036) 1,892 (2) Weighted pairwise f r tdifference is -0.146 * * (0.060) -0.068 * (0.037) 1,892 (3) Single family fixed effect -0.129 * * (0.052) -0.045 (0.033) 2,178 (4 Oldest-Youngest pairs ) -0.115 * (0.067) -0.068 * (0.038) 742 (5) Unweighted pairwise f r tdifference is Different age 14 neighborhood variable -0.110 (0.074) -0.066 (0.064) 600 (6 Unweighted pairwise f r tdifference ) is Different average neighborhood variable -0.166 * • * (0.060) -0.069 * (0.042) 1,554 -0.543 • * * (0.155) -0.805 * * * (0.163) 2,178 (8 Conditional l g t pairwise sample ) oi -1.250 "* (0.408) -0.774 * * (0.390) (9 Mundlak fixed effect l g t ) oi -0.872 " (0.432) -0.407 (0.357) Loait Models ( ) Base case 7 441 2,178 Notes: •••^significant at 1 % level **=significant at 5 % level *=significant at 10% level 1) Regressions control for gender, race, parents’education, parents’marital status, household income, number of siblings, whether the child worked, the year the child turned 15, the wage for unskilled workers i the county of residence, the variance of money income while the n youth lives at home, f v parent marital transition variables, and eight head employment ie transition variables. These transition variables are l s e i table 2 Only transitions that are itd n . experienced by the younger child ( e after the older child has l f the household) are coded i. et as 1 i the difference estimators. n 2) See text for explanation of different estimators. Table 6 The Effect of the Neighborhood Imputation and Averaging on the High School Graduation Results ( 1 Neighborhood Measures: log(poverty rate) or log(dropout rate) Dependent variable: 1 i high school graduate f (Huber standard errors i parentheses) n log(poverty r te)_________ a Unweighted Fixed Linear Effect Probabilitv () 2 (D _____ log(dropout rate) Unweighted Fixed Linear Effect Probabilitv () 4 () 3 Imputed, time averaged (3,5 -0.042 * * * (0.016) -0.144 ~ (0.060) -0.061 * * * (0.013) -0.068 * (0.036) No imputation, time avg. (3,6 -0.044 (0.017) -0.142 ** (0.064) -0.055 * * * (0.014) -0.081 ** (0.041) imputed, age 14 (4,5 -0.024 (0.015) -0.042 (0.037) -0.048 *** (0.012) -0.037 (0.025) No imputation, age 14 (4,6 -0.030 * * (0.015) -0.020 (0.041) -0.040 *** (0.011) -0.053 * (0.028) Notes: ***=significant at 1% level t "=significant a 5% level *=significant at 10% level 1) See notes to tables 4 and 5 for list of control variables. 2) Instruments are county poverty rate, unemployment rate, average household income, and percentage of adults who did not graduate from high school. 3) Neighborhood variables are averaged overages 10 to 18. 4) Neighborhood variables set to age 14 (or closest age to 14) measure. 5) Neighborhood measures are imputed between Census years (1969 and 1979) and held constant at 1969 (and 1979) values before1969 (after 1979) 6) No imputations are calculated. All neighborhood measures are from the 1980 Census reports. Table 7 Linear Probability and Fixed Effect Estimates of Neighborhood Impact on High School Graduation Using Different Neighborhood Proxies ( 1 Dependent variable: 1 i high school graduate f (Huber standard errors i parentheses) n Neiahborhood Measure (2 Linear Probability (D Derivative at Mean () 2 Unweighted Fixed Effect (3) ( ) Poverty rate 1 -0.042 *** (0.016) -0.0021 -0.144 ~ (0.060) ( ) Dropout rate 2 -0.061 *** (0.013) -0.0036 -0.068 * (0.036) ( ) Percent white population 3 0.012 (0.012) ( ) Percent households 4 that are female headed ( ) Average income 5 0.0002 -0.001 (0.029) -0.060 *** (0.021) -0.0029 -0.089 (0.066) 0.073 ** (0.038) 0.0021 0.051 (0.100) Notes: ***=significant at 1% level **=significant at 5% level *=significant at 10% level 1) See notes to tables 4 and 5 for ls of controls. it 2) Neighborhood measures are entered into the high school graduation equations one at a time. Table 8 The Effect of Neighborhoods on Different Educational Outcome Measures ( 1 Neighborhood Measures: log(poverty rate) or log(dropout rate) (Huber standard errors i parentheses) n Outcome Measure Unweighted Mean of Outcome () 1 log(poverty rate) Unweighted Linear Probability Fixed Effects () 2 () 3 log(dropout rate) Linear Unweighted Probability Fixed Effects () 4 () 5 ( ) High school graduation 1 0.803 -0.038 ** (0.017) -0.165 *" (0.060) -0.056 *** (0.013) -0.064 (0.040) ( ) College attendance 2 0.351 -0.037 * (0.020) -0.082 (0.050) -0.081 (0.017) -0.012 (0.039) (3 College attendance ) conditional on h s graduation .. 0.437 -0.027 (0.022) -0.058 (0.067) -0.068 *** (0.018) 0.033 (0.048) ( ) Grades completed (2 4 12.52 -0.180 * * (0.080) -0.519 * * * (0.199) -0.362 *** (0.068) -0.082 (0.148) Sample Size (3 Rows (1,2,4) Row ( ) 3 2,137 1,716 1,822 1,206 2,137 1,716 Notes: ***=significant at 1% level **=significant at 5 % level *=significant at 10% level 1 See tables 4 and 5 for l s of control variables. Regressions also include the maximum age ) it used to determine an individual's educational outcome measure. 2) Maximum number of grades reported from age 19 to 25. I no grades have been reported by f age 25 ( e the indiviudal i sil a student), then the f r tgrade report after age 25 i used. i. s tl is s 3) The sample includes only those siblings where grades completed are easily determined. This eliminates 41 individuals who were assumed to be high school graduates i previous n tables because they were sil students i their early 20s when they a t i e from the sample. tl n trtd This assumption makes no difference to the results. Row 3 includes only those siblings who graduated from high school. 1,822 1,206 Table 9 Effects of Neighborhoods on High School Graduation, By Race, Gender, and Income (Huber standard errors in parentheses) Neighborhood Measures: log(poverty rate) or log(dropout rate) Dependent variable: 1 if high school graduate [sample size in brackets] Sample log(poverty rate)__________ Unweighted Linear fixed effect orobabilitv (2) (1) log(dropout rate) Linear Unweighted Drobabilitv fixed effect (3) (4) White -0.013 (0.017) -0.204 *** (0.073) -0.031 *** (0.012) -0.077 ** (0.037) Nonwhite -0.055 * (0.031) -0.092 (0.086) -0.097 *** (0.032) Female (4 -0.016 (0.019) -0.030 (0.083) Male (5 -0.057 " (0.023) -0.218 ** (0.087) Sample size f2 (5) 985 769 -0.040 (0.072) 1,193 1,123 -0.077 *** (0.015) -0.086 * (0.051) 1,104 495 -0.039 ** (0.017) -0.083 (0.074) 1,074 449 Notes: ***=significant at 1% level **=significant at 5% level *=significant at 10% level 1) Instruments are county poverty rate, unemployment rate, average household income, and percentage of adults who did not graduate from high school. 2) Sample size for linear probability and IV models. 3) Sample size of sibling pairs for pairwise first difference model. See text for explanantion. 4) Fixed effect sample includes only those sibling pairs that are sisters. 5) Fixed effect sample includes only those sibling pairs that are brothers. Sample Dairs 13 (6) Table 10 The Importance of Sibling Differences i Neighborhood Background n on High School Graduation Rates Unweighted F r t Difference Equations is Neighborhood Measures: log(poverty rate) or log(dropout rate) Dependent variable: 1 i high school graduate f (Huber standard errors i parentheses) n [Sample size i brackets] n Row Neighborhood Differential ( 1 Poverty Rate (D Dropout Rate () 2 ( ) * 0% 0 -0.144 * * (0.060) [1,892] -0.068 (0.036) [1,892] ( ) > 5% 1 -0.151 * * (0.059) [1,350] -0.070 (0.037) [1,445] ( ) > 10% 2 -0.148 * * (0.059) [1,101] -0.070 (0.037) [1.173] ( ) > 20% 3 -0.126 * * (0.058) [657] -0.070 (0.036) [779] ( ) > 30% 4 -0.108 * (0.062) [398] -0.072 (0.038) [488] Significance level from Wald test of difference between row 0 and row 4 0.89 0.98 Notes: ***=significant at 1% level **=significant at 5 % level *=sign‘f c n at 10% level riat 1 Neighborhood d f e e t a refers to the percentage difference between s ) ifrnil ibling pairs i n th i average neighborhood characteristic. For example, the >10% category includes er only those sibling pairs whose average poverty rate (or dropout rate) d f e s by more ifr than ten percent. Appendix 1a Linear Probability High School Graduation Regressions Dependent variable: 1 i high school graduate f (Huber standard errors i parentheses) n Neighborhood measure: log(poverty rate) Al youth l samDle Sibling sample (3) (D (2) Intercept Neighborhood variable Whether female Log(household income) Whether nonwhite Parents' married Dad high school grad. M o m high school grad. No. kids in household Whether kid worked County unskilled wage Variance of income Parents divorced while youth was aged 10-18 Percentage of years moved, aged 10-14 Own household by 18 Adusted R-squared Sample size 0.317 (0.298) -0.042*** (0.016) 0.060 * * * (0.015) 0.056 * * (0.028) 0.086 *** (0.028) -0.012 (0.033) 0.063 * * * (0.023) 0.084 * * * (0.023) -0.031 * * * (0.007) 0.068 * * * (0.020) 0.015 (0.010) -0.020 (0.015) -0.083 * * (0.036) 0.133 2,178 0.470 * (0.287) -0.041 ** (0.016) 0.079 *** (0.015) 0.046 * (0.027) 0.058 ** (0.028) -0.020 (0.031) 0.059 *** (0.023) 0.077*** (0.021) -0.027 *** (0.007) 0.053 *** (0.020) 0.015 (0.010) -0.015 (0.015) -0.082 ** (0.035) -0.202 *** (0.047) -0.225 *** (0.037) 0.172 2,178 0.210 (0.199) -0.024 ** (0.011) 0.061 *** (0.011) 0.067 * ** (0.018) 0.049 ** (0.020) -0.016 (0.021) 0.060 *** (0.017) 0.070 *** (0.015) -0.020 *** (0.004) 0.043 *** (0.014) 0.004 (0.007) -0.004 (0.006) -0.042 (0.026) -0.152 *** (0.034) -0.195 * * * (0.024) 0.140 4,410 Neighborhood measure: log(dropout rate) All youth Sibling sample sample (4 ) () 5 (6) 0.320 (0.279) -0.061 * * (0.013) 0.062 * * * (0.015) 0.060 ** (0.027) 0.079 * * * (0.027) -0.010 (0.032) 0.057 * * (0.024) 0.081 * * * (0.023) -0.031 *** (0.007) 0.069 * * * (0.020) 0.017 * (0.010) -0.021 (0.016) -0.080 ** (0.035) 0.139 2,178 0.424 (0.273) -0.049 *** (0.012) 0.081 *** (0.015) 0.052 * * (0.026) 0.049 * * (0.026) -0.019 (0.031) 0.055 * * (0.023) 0.076 * * * (0.021) -0.027 *** (0.007) 0.056 * * * (0.020) 0.017 * (0.010) -0.016 (0.016) -0.080 * * (0.035) -0.194 * * * (0.047) -0.218 * * * (0.036) 0.175 2,178 Notes: ***=significant at 1% level **=significant at 5% level *=significant at 10% level 1) All regressions include 8 employment transition variables, 4 other m arital status transitions, and region and age 15 dummies. 0.198 (0.187) -0.034*** (0.008) 0.061 *** (0.011) 0.070 *** (0.018) 0.046 ** (0.019) -0.016 (0.021) 0.057 * * * (0.017) 0.068 *** (0.016) -0.020 * * * (0.004) 0.043 *** (0.014) 0.005 (0.007) -0.006 (0.006) -0.042 (0.026) -0.148 *** (0.034) -0.191 *** (0.024) 0.142 4,410 Appendix 1b Logit High School Graduation Regressions Dependent variable: 1 i high school graduate f (Huber standard errors i parentheses) n Neighborhood measure: iog(poverty rate) All youth samDle Sibling sample O) (D () 2 Intercept * Neighborhood variable Whether female Log(household income) Whether nonwhite Parents' married Dad high school grad. M o m high school grad. No. kids i household n Whether kid worked County unskilled wage Variance of income Parents divorced while youth was aged 10-18 Percentage of years moved, aged 10-14 Own household by 18 Log likelihood Sample size -1.470 (2.858) -0.543 * * * (0.133) 0.463 * * * (0.124) 0.403 * * * (0.157) 0.765 * * * (0.185) -0.093 (0.181) 0.663 * * (0.177) 0.705 *** (0.144) -0.221 * * * (0.039) 0.437 * * * (0.133) 0.144 * * (0.068) 0.019 (0.274) -0.562 * * * (0.195) -867.3 2,178 -0.309 (2.912) -0.586 *** (0.158) 0.659 *** (0.133) 0.343 (0.280) 0.556 * * (0.222) -0.153 (0.239) 0.630 * * * (0.213) 0.690 *** (0.174) -0.195 *** (0.045) 0.348 * * * (0.138) 0.145 * (0.087) 0.058 (0.264) -0.593 * * * (0.223) -1.329 * * * (0.282) -1.323 * * * (0.202) -827.8 2,178 -2.581 (1.859) -0.348*** (0.116) 0.525 *** (0.094) 0.511 *** (0.176) 0.458 *** (0.162) -0.132 (0.159) 0.636 *** (0.157) 0.581 *** (0.123) -0.145 *** (0.028) 0.292 *** (0.098) 0.028 (0.064) 0.227 (0.248) -0.320 * (0.167) -0.964 * * * (0.211) -1.169 * * * (0.134) -1,678.2 4,410 Neighborhood measure: iog(dropout rate) Al youth l Sibling sample samDle () 4 (5) () 6 -1.611 (2.778) -0.805 *** (0.163) 0.495 *** (0.125) 0.483 * (0.277) 0.572 * * * (0.197) -0.105 (0.244) 0.604 * * * (0.212) 0.667 *** (0.178) -0.215 *** (0.045) 0.438 * * * (0.133) 0.174 ** (0.087) 0.045 (0.252) -0.553 * * (0.218) -854.9 2,178 Notes: ***=significant at 1% level 1) All regressions include 8 employment transition variables, 4 other marital status transitions, and region and age 15 dummies. -1.058 (2.889) -0.726 * * * (0.161) 0.682 * * * (0.133) 0.449 (0.289) 0.340 * (0.200) -0.172 (0.248) 0.587 * * * (0.212) 0.662*** (0.173) -0.190 * * * (0.045) 0.358 * * * (0.137) 0.172 ** (0.088) 0.074 (0.265) -0.597 * * * (0.227) -1.228 * * * (0.287) -1.244 * * * (0.197) -821.0 2,178 -2.959 * (1.706) -0.473 (0.103) 0.530 (0.094) 0.579 (0.170) 0.361 (0.149) -0.137 (0.160) 0.604 (0.157) 0.561 (0.124) * -0.143 * * (0.027) 0.283 (0.097) 0.046 (0.063) 0.197 (0.235) -0.334 (0.166) -0.919 (0.210) * -1.127 * * (0.134) -1,669.3 4,410 Appendix 2 Unweighted Pairwise Fi s Difference High School Graduation Regressions rt Neighborhood Measures: log(poverty rate) or log(dropout rate) Dependent variable: 1 i high school graduate f (Huber standard errors i parentheses) n logfpoverty rate)_____ (2) (D Intercept Neighborhood Var. Whether female Log(hh income) Parents' married Number of kids in hh Whether kid ever worked during youth County unskilled wage -0.057 * * (0.028) -0.144 * * (0.060) 0.072 * * * (0.022) 0.003 (0.053) 0.004 (0.063) -0.037 * * (0.019) 0.070 * * * (0.025) 0.020 (0.018) Own household by 18 Adusted R-squared 0.058 -0.053 ** (0.027) -0.142 ** (0.060) 0.087 *** (0.021) -0.003 (0.051) 0.011 (0.065) -0.029 (0.018) 0.048 ** (0.024) 0.016 (0.018) -0.217 * * * (0.048) 0.084 _____ log(dropout rate) (3) (4) -0.052 *** (0.028) -0.068 * (0.037) 0.070 *** (0.021) 0.018 (0.055) 0.001 (0.063) -0.036 * (0.019) 0.066 ** (0.026) 0.022 (0.019) 0.054 Notes: ***=significant at 1% level **=significant at 5 % level *=significant at 10% level 1) A l regressions also include sibling differences i parents' marital and employment l n status dummies, variance of household income, whether the child participated i n housework, and region and year turned age 15 dummies. -0.057 (0.027) -0.068 (0.036) 0.085 (0.021) 0.012 (0.054) 0.009 (0.065) -0.029 (0.019) 0.044 (0.024) 0.019 (0.018) -0.218 (0.049) 0.080 Bibliography A lto n ji, J o sep h and T hom as D u n n. 1 9 9 5 . "The E ffe c ts o f S c h o o l and F am ily C haracteristics o n th e R etu rn to E du cation ." m im eo. An, Chong Bung, Robert Haveman and Barbara Wolfe. 1992. "The Window Problem' in Studies of Children's Attainments: A Methodological Exploration.” mimeo, National Bureau ofEconomic Research. Benabou, Roland. 1993. "Workings of a City: Location, Education, and Production." J o u r n a l o f E c o n o m ic s , p 619-652. . Q u a r te r ly Boijas, George. 1995. "Ethnicity, Neighborhoods, and Human Capital Externalities." A m p 365-390. . e ric a n E c o n o m ic R e v ie w , Brooks-Gunn, Jeanne, Greg Duncan, Pamela Klebanov and Naomi Sealand. 1993. "Do Neighborhoods Influence Child and Adolescent Development?" A m e ric a n J o u r n a l o f S o c io lo g y . p 353-395. . Brueckner, Jan and Kangoh Lee. 1989. " Club Theory and a Peer Group Effect." R e g io n a l p 399-420. . S c ie n c e a n d U rb a n E c o n o m ic s , Case, Anne and Lawrence Katz. 1991. "The Company You Keep: The Effects of Family and Neighborhood on Disadvantaged Youth." mimeo, National Bureau ofEconomic Research. Chamberlain, Gary. 1980. "Analysis of Covariance with Qualitative Data." R e v ie w S tu d ie s , p 225-238. . o f E c o n o m ic Corcoran, Mary, Roger Gordon, Deborah Laren and Gary Solon. 1992. "The Association between Men's Economic Status and Their Family and Community Origins." J o u r n a l o f H u m a n R e s o u rc e s , p 575-601. . Crane, Jonathan. 1991. "The Epidemic Theory of Ghettos and Neighborhood Effects on Dropping Out and Teenage Childbearing." A m e ric a n J o u r n a l o f S o c io lo g y , p 1226-1259. . deBartolome, Charles. 1990. "Equilibrium and Inefficiency i a Community Model With Peer n Group Effects." J o u r n a l o f P o lit ic a l E c o n o m y , p 110-133. . Duncan, Greg. 1994. ‘ Tamilies and Neighborhoods as Sources of Disadvantage in the Schooling Decisions of White and Black Adolescents.” A m e ric a n J o u r n a l o f E d u c a tio n , p 20-53. . Duncan, Greg and James Connell and Pamela Klebanov. 1994. “Selection Bias in the Estimation of the Effects ofNeighborhood Conditions on Children’ Development.” mimeo. s E p p le, D e n n is and R ichard R om an o. 1 9 9 3 . "C om petition B e tw e e n P riv a te and P u b lic S c h o o ls, V o u ch ers and P ee r G roup E ffects." m im eo, Septem ber. Evans, William and Wallace Oates and Robert Schwab. 1992. "Measuring Peer Group Effects: A Study of Teenage Behavior." J o u r n a l o f P o lit ic a l E c o n o m y , p 966-991. . Holzer, Harry. 1991. "The Spatial Mismatch Hypothesis: What Has the Evidence Shown?" U r b a n S tu d ie s , p 105-122. . Jacobs, Jane. 1969. T h e E c o n o m y o f C it ie s . N e w York: Vintage Press. Jencks, Christopher and Mayer, Susan. 1990. "The Social Consequences of Growing Up in a Poor Neighborhood: A Review." in Lynn and McGeary (eds) I n n e r C it y P o v e r ty in th e U n ite d S ta te s . Washington, DC: National Academy Press. Manski, Charles. 1993. "Identification of Endogeneous Social Effects: The Reflection Problem?" p 531-542. . R e v ie w o f E c o n o m ic S tu d ie s , Plomin, Robert and Denise Daniels. 1987. "Why are Children in the Same Family So Different From One Another?" B e h a v io r a l a n d B r a in S c ie n c e s , p 1-60. . Plotnick, Robert and Saul Hoffinan., 1995. "Fixed Effect Estimates of Neighborhood Effects.” mimeo. Romer, Paul. 1986. "Increasing Returns and Long-Run Growth." J o u r n a l p. 1002-1037. o f P o lit ic a l E c o n o m y . Rosenbaum, James and Susan Popkin. 1991. "Employment and Earnings of Low-Income Blacks W h o Move to Middle-Class Suburbs." in Jencks and Peterson (eds) T h e U r b a n U n d e r c la s s . Washington, D.C.: The Brookings I s i u i n nttto. Summers, Anita and Barbara Wolfe. 1977. "Do Schools Make a Difference?" p 639-652. . E c o n o m ic R e v ie w , A m e ric a n