The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
REVIEW FEDERAL RESERVE BANK OF ST. LOUIS FIRST QUARTER 2022 VOLUME 104 | NUMBER 1 Economic Activity during the COVID-19 Pandemic: A Model with “Acquired Immunity” Juan Esteban Carranza, Juan David Martin, and Álvaro José Riascos Sectoral Impacts of Trade Wars Wan-Jung Cheng and Ping Wang The Impact of Juvenile Conviction on Human Capital and Labor Market Outcomes Limor Golan, Rong Hai, and Hayley Wabiszewski Further Evidence on Greenspan’s Conundrum Cletus C. Coughlin and Daniel L. Thornton REVIEW Volume 104 • Number 1 President and CEO James Bullard Director of Research 1 Carlos Garriga Economic Activity during the COVID-19 Pandemic: A Model with “Acquired Immunity” Deputy Director of Research B. Ravikumar Review Editors-in-Chief Michael T. Owyang Juan M. Sánchez Juan Esteban Carranza, Juan David Martin, and Álvaro José Riascos 17 Special Policy Advisor Sectoral Impacts of Trade Wars David C. Wheelock Wan-Jung Cheng and Ping Wang Economists David Andolfatto Subhayu Bandyopadhyay Serdar Birinci Yu-Ting Chiang YiLi Chien Riccardo DiCecio William Dupor Maximiliano Dvorkin Miguel Faria-e-Castro Charles S. Gascon Victoria Gregory Nathan Jefferson Kevin L. Kliesen Julian Kozlowski Fernando Leibovici Oksana Leukhina Fernando M. Martin Michael W. McCracken Amanda M. Michaud Alexander Monge-Naranjo Christopher J. Neely Serdar Ozkan Paulina Restrepo-Echavarría Hannah Rubinton Ana Maria Santacreu Guillaume Vandenbroucke Christian M. Zimmermann 41 The Impact of Juvenile Conviction on Human Capital and Labor Market Outcomes Limor Golan, Rong Hai, and Hayley Wabiszewski 70 Further Evidence on Greenspan’s Conundrum Cletus C. Coughlin and Daniel L. Thornton Managing Editor Lydia H. Johnson Contributing Editors George E. Fortier Jennifer M. Ives Designer | Production Coordinator Donna M. Stiller i Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Review Review is published four times per year by the Research Division of the Federal Reserve Bank of St. Louis. Full online access is available to all, free of charge. Online Access to Current and Past Issues The current issue and past issues dating back to 1967 may be accessed through our Research Division website: http://research.stlouisfed.org/publications/review. All nonproprietary and nonconfidential data and programs for the articles written by Federal Reserve Bank of St. Louis staff and published in Review also are available to our readers on this website. Review articles published before 1967 may be accessed through our digital archive, FRASER: http://fraser.stlouisfed.org/publication/?pid=820. Review is indexed in Fed in Print, the catalog of Federal Reserve publications (http://www.fedinprint.org/), and in IDEAS/RePEc, the free online bibliography hosted by the Research Division (http://ideas.repec.org/). Subscriptions and Alerts The Review is no longer printed or mailed to subscribers. Our last printed issue was the first quarter of 2020. Our monthly email newsletter keeps you informed when new issues of Review, Economic Synopses, Regional Economist, and other publications become available; it also alerts you to new or enhanced data and information services provided by the St. Louis Fed. Subscribe to the newsletter here: http://research.stlouisfed.org/newsletter-subscribe.html. Authorship and Disclaimer The majority of research published in Review is authored by economists on staff at the Federal Reserve Bank of St. Louis. Visiting scholars and others affiliated with the St. Louis Fed or the Federal Reserve System occasionally provide content as well. Review does not accept unsolicited manuscripts for publication. The views expressed in Review are those of the individual authors and do not necessarily reflect official positions of the Federal Reserve Bank of St. Louis, the Federal Reserve System, or the Board of Governors. Copyright and Permissions Articles may be reprinted, reproduced, republished, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. In these cases, there is no need to request written permission or approval. Please send a copy of any reprinted or republished materials to Review, Research Division of the Federal Reserve Bank of St. Louis, P.O. Box 442, St. Louis, MO 63166-0442; STLS.Research.Publications@stls.frb.org. Please note that any abstracts, synopses, translations, or other derivative work based on content published in Review may be made only with prior written permission of the Federal Reserve Bank of St. Louis. Please contact the Review editor at the above address to request this permission. Economic Data General economic data can be obtained through FRED®, our free database with over 800,000 national, international, and regional data series, including data for our own Eighth Federal Reserve District. You may access FRED through our website: https://fred.stlouisfed.org. © 2022, Federal Reserve Bank of St. Louis. ISSN 0014-9187 ii Economic Activity during the COVID-19 Pandemic: A Model with “Acquired Immunity” Juan Esteban Carranza, Juan David Martin, and Álvaro José Riascos We calibrate a macroeconomic model with epidemiological restrictions using Colombian data. The key feature of our model is that a portion of the population is immune and cannot transmit the virus, which improves substantially the fit of the model to the observed contagion and economic activity data. The model implies that during 2020, government restrictions and the endogenous changes in individual behavior saved around 15,000 lives and decreased consumption by about 4.7 percent. The results suggest that most of this effect was the result of government policies. (JEL E1, I1, H0) Federal Reserve Bank of St. Louis Review, First Quarter 2022, 104(1), pp. 1-16. https://doi.org/10.20955/r.104.1-16 1 INTRODUCTION In this article we formulate and calibrate a dynamic macroeconomic model in which optimizing agents respond to the risk of contagion and restrictions imposed by the government during the recent public health crisis. Our model is similar to the model by Eichenbaum, Rebelo, and Trabandt (2020; henceforth ERT), except for the inclusion of a modified epidemiological model that incorporates the possibility of exogenous immunity to contagion. We calibrate the model with Colombian data and use it to simulate counterfactual policies. The original ERT model was the first of a wave of new articles using variations of a simple susceptible-infected-recovered (SIR) epidemiological model to account for the endogenous risk of contagion faced by economic agents. Other articles with similar approaches include Atkeson (2020); Alvarez, Argente, and Lippi (2020); Acemoglu et al. (2020); and Berger, Herkenhoff, and Mongey (2020). In these models, both contagion and economic activity are the results of a dynamic programming problem in which agents maximize their intertemporal utility, accounting for the risk of contagion over the course of an epidemic. In the SIR model and its variations (based on seminal work by Kermack and McKendrick, 1927), an epidemic runs its course as it infects individuals who then become immune. The epidemic ends Juan Esteban Carranza is the Director of Economic Studies at the Banco de la República; Juan David Martin is a researcher at the Banco de la República; and Álvaro José Riascos is a professor at the University of los Andes and Director of Quantil. © 2022, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. 1 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos when the population reaches “herd immunity,” which occurs when enough people are immune and the virus cannot spread anymore. The main drawback of these models is the difficulty they have in fitting the observed COVID-19 contagion data. In particular, the standard model predicts very high numbers of both infections and deaths compared with the relatively low numbers observed in the data. Our main contribution to understanding the pandemic is the modification of the SIR model to allow for the presence of individuals who are unaffected by the virus and who become immune over time at an exogenous rate. As our results show, the presence of this “immune” population helps to fit the model to the observed data. In particular, the model replicates well the rapid decline of observed deaths after the infection of a relatively low portion of the population during the first wave of the pandemic. In contrast to ERT, we calibrate the model to match measures of both the pandemic and economic activity. We follow ERT in modeling government restrictions as a consumption tax, which induces consumers to cut back in their consumption activities, but we actually calibrate the parameterization of this tax. The calibrated model predicts that almost all infections in Colombia will have already occurred by December 2020 and that the economy will be back on its long-term path by mid-2021. Our simulations suggest that government restrictions decreased yearly 2020 consumption by around 3 percent and saved around 10,000 lives. Without government restrictions, the economy would have still faced a 1 percent contraction, generated by consumers cutting back on consumption and labor to avoid contagion. The model can be easily extended to accommodate successive contagion waves by allowing agents who become immune to become susceptible again. In the model, new waves can be triggered by making immunity disappear after an exogenous number of weeks or by letting immune agents become susceptible every period at an exogenous rate. This type of modeling would capture the possibility that, for example, either variations of the original virus appear or antibodies acquired from a mild infection disappear. Because the reasons individuals may or may not be immune are not well understood, we focus on the modeling of one wave of the pandemic and leave extensions for further research. The article is organized as follows: Section 2 describes the model and its calibration. Section 3 contains the baseline results and counterfactual simulations. Section 4 concludes. 2 THE MACRO-SIOD MODEL 2.1 Description of the Model The Model of the Pandemic with Endogenous Contagion. As we indicated above, we follow closely ERT and formulate a model with infinitely lived consumers who choose to allocate time into labor and consumption to maximize their lifetime expected utility. Their choices determine both the level of observed economic activity and the probability of contagion. There are four types of agents in the model, depending on their exposure to the infection. These include susceptible (S), infected (I), survivor (O), and dead (D)—SIOD. A survivor agent can in turn be immune (M) or recovered (R). Notice that, in contrast to the standard SIR model embedded in ERT, we add the additional immune type (M), which is a type of patient that develops no symptoms and is not infectious. 2 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos In the model, agents become immune over time at an exogenous rate, which we calibrate.1 Once they become immune, the immune agents behave similarly to recovered agents. The difference between an immune agent and a susceptible agent who becomes infected and then recovers is that the immune agent is never contagious and therefore never propagates the pandemic. The existence of this type of “immune” agent is consistent with the increasing evidence of preexisting immunity to the SARS-CoV-2 virus among a significant part of the population (Doshi, 2020). More broadly, our definition of immunity is consistent with other biological mechanisms that are not well understood yet, the details of which fall beyond the scope of this article. For example, this “immunity” is equivalent to situations in which individuals become infected but transmit the virus at variable rates (Adam, 2020). More specifically, if an individual does not transmit the virus after infection, they are “immune” according to our definition. Denote Tt as the flow of newly infected individuals at time t, which depends on the probability that the stock of susceptible agents St becomes infected while interacting with the stock of infected agents It during consumption activities, work, or other activities, denoted π1, π2, and π3, respectively; that is, (1) Tt = St (1− π m ) It (π 1CtSCtI + π 2 N tS N tI + π 3 ) , where C tj and Ntj correspond to the consumption and work hours of j-type agents with j = S,I. The stock of infected agents evolves over time depending on (1) as follows: (2) It+1 = It (1− π dt − π r ) +Tt , where πdt and πr are the probabilities of death and recovery, conditional on I. The probability of death changes over time depending on the capacity restrictions of the health care system, denoted as ξ in the following equation: (3) π dt = π d +1{It >ξ }κ It2 , where the probability increases in a quadratic way when the number of infected individuals surpasses the capacity ξ, while κ is a parameter to be calibrated. The main innovation of our work is the inclusion of type M agents who become immune over time without being infected and whose stock evolves, as follows: (4) Mt+1 = Mt + π m St , where πm is the exogenous probability of becoming immune. Notice that we assume that agents become immune over time at a constant rate, probably as a result of their exposure to the virus through interactions with infected agents. This simplifying assumption recognizes the fact that this immunity is not yet well understood. Notice that setting M0 = 0 and πm = 0 yields the standard SIR model, which we can also calibrate as a particular case in our model.2 To complete the description of the epidemiological model, the following equations describe the evolution of the stock of agent types D, R, and S: 3 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos (5) Dt+1 = Dt + π dt It , Rt+1 = Rt + π r It , and St+1 = St (1− π m ) −Tt , where we assume that the initial stock of susceptible agents is the initial population S0 = Pop0 and that the initial stock of infected agents is nil; that is, I0 = є > 0, a small portion of the population, which is calibrated. The Economic Model. The economic problem of agents is the maximization of their lifetime utility through consumption and work decisions. Their choices determine the total number of hours devoted to consumption Ct and work Nt , which in turn determines the transition across agent types described above and the endogenous probability of contagion faced by a susceptible agent; that is, (6) τ t = π 1 ( ctS )( It CtI ) + π 2 (ntS )( It N tI ) + π 3 ( It ). On the supply side, we assume there is a continuum of competitive, identical firms that use only labor and maximize period-by-period profits; that is, Πt = AN t − wt N t , (7) where A is a productivity parameter and wt is the competitive wage. In this closed economy, in equilibrium, total production must be equal to total consumption. At each time t, agents are identified by their infection status j ∈ S,I,O. The problem of each agent type j at t = t0 is given by (8) ∞ max U t0 = ∑ β j c,n t−t0 t=t0 u(ctj ,ntj ), where u ( ctj ,ntj ) = lnctj − θ j 2 (nt ) 2 and θ is a parameter to be calibrated. The agent faces a budget constraint given by (9) ctj = φ j wt n j − µt ctj + Γ t , where wt is the hourly wage, μt is an exogenous consumption tax rate, and Γt is a government transfer that does not depend on the type of agent and that balances government finances. The parameter ϕ tj is a measure of the work ability of agent j due to infection, with the assumption that ϕS = ϕO = 1 and ϕI ≤ 1. Thus, the probability of contagion τt affects the economic decisions of each type of agent through their expected lifetime utility at period t; that is, 4 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos (10) ( ) S I O ⎤, U tS =u ( ctS ,ntS ) + β ⎡⎣(1− π m ) (1− τ t )U t+1 + τ tU t+1 + π mU t+1 ⎦ I O ⎤⎦ , and U tI =u ( ctI ,ntI ) + β ⎡⎣(1− π r − π dt )U t+1 + π rU t+1 O U tO =u ( ctO ,ntO ) + βU t+1 . The tax rate μt plays an important role in the model because it absorbs all the restrictions imposed by the government. Policies such as quarantines and limits to the gathering of people are rationalized in the model as a consumption tax. In our application, we will identify this parameter using the period-by-period measure of economic activity. In equilibrium, it must hold that all agents maximize their expected lifetime utility, firms maximize profits, the government balances its budget, and both labor and goods markets clear. We specify the equilibrium conditions in the appendix. 2.2 Calibration Table 1 We calibrate the model using weekly Colombian data matching the observed path of the COVID-19 Calibrated Parameters pandemic throughout 2020. For any parametrizaParameter Value tion of the model, it is solved using a backward induction algorithm that involves (i) finding the Calibrated with minimum distance algorithm optimal sequence of working hours for every type π1 1.5403 × 10 –6 of agent along 250 weeks and then, similarly to π2 1.0014 × 10 –5 ERT, (ii) computing the rest of the equilibrium π3 0.9703 sequences using the first-order conditions of the πd 0.0061 model. πr 0.7141 We follow closely the criteria in ERT in choosπm 0.0185 ing the parameters of the economic model, which we show in Table 1. We then calibrate the parameμ 0.3783 ters π = {π1,π2,π3,πd ,πr ,πm}, which determine the Calibrated a la ERT dynamics of the pandemic. Moreover, we also A 19.2308 calibrate the maximum percentage of the populaθ 0.0015 tion that gets infected, after which there is “herd ϕ1 0.8000 immunity.” β 0.9992 As pointed out, another difference between our model and ERT is the treatment of the tax rate, κ 1.5000 which, as we pointed out, reflects the restrictions є 2.3897 × 10 –5 imposed by the government to contain the panNOTE: ERT, Eichenbaum, Rebelo, and Trabandt (2020). demic. We set μt = μ for t =1 until t =19, which SOURCE: Authors’ calculations. corresponds to August 31, 2020, when the national lockdown imposed by the government was officially ended. This policy had been in place since March 23 (t = –4). For t ≥ 20, we set μt = 0.9μt –1 so that restrictions decreased gradually toward zero. As indicated above, we need to calibrate the parameters π and μ. To identify π, we match the number of deaths predicted by the model with the observed deaths reported by the Colombian 5 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos health care authorities. We focus on the deaths to avoid problems associated with the underreporting of detected cases. To identify μ, we match the weekly consumption predicted by the model to a proxy of weekly economic activity that we observe in almost real time. Specifically, we use electricity consumption, which historically roughly matches the economic cycle, with the understanding that electricity is both an input used in any type of consumption activity and difficult to substitute in the short run. The model is calibrated over 31 weeks, starting during the second week of April 2020, which is eight weeks after the presumed beginning of the pandemic in Colombia. As occurs almost anywhere, there is very substantial underreporting of contagion, as most infections are asymptomatic. On the other hand, the Colombian health care system collects death data with accuracy. Therefore, instead of using reported cases, as in ERT, we calibrate the model matching the predicted and observed weekly deaths. Our proxy of consumption is the gap between observed weekly electricity consumption and a simulated trend estimated from historic data. The electricity consumption is reported daily by the national electricity system operator and aggregated into weekly data. Our calibration algorithm minimizes the following loss function: (11) t=31 ( Ω = ∑ ⎡ω Dt − D̂t ⎢ t=1 ⎣ ) 2 ( ) 2 + (1− ω ) Et − Êt ⎤ , ⎥⎦ where Dt and D̂t are the observed and predicted weekly deaths, respectively. On the other hand, Et and Êt are, respectively, the realized and predicted weekly electricity consumption gaps with respect to a scenario without the pandemic. The initial time period t =1 corresponds to the thirteenth week of 2020. As mentioned previously, we use electricity consumption as a proxy to measure economic activity. However, to compute Et we need a measure of the average electricity consumption that would have been observed in a scenario without a pandemic. We do so by projecting the trend of electricity consumption implied by the data observed up to March 2020.3 As for the weighting scalar, ω, we use a backtesting approach to choose the value that minimizes the average out-of-sample prediction error.4 3 RESULTS 3.1 Baseline In Figures 1 and 2 we show the simulated and observed measures of the pandemic and the economic activity, respectively. The data correspond to the weekly number of COVID-19-related deaths and the weekly gap in electricity consumption. We calibrate two models. On the one hand, we calibrate our model with immune agents as specified above. On the other hand, we also calibrate a model with no immune individuals (Mt = 0, ∀t ), which is equivalent to the standard macro-SIRD model a la ERT—that is, susceptible (S), infected (I), recovered (R), and dead (D). The contrast between both calibrations illustrates the contribution of our approach. As shown in Figures 1 and 2, the macro-SIOD model is able to predict well the pattern of both variables. Relative to the observed deaths, the model predicts a later peak at a level of around 2,000 weekly deaths, which is slightly lower than the observed peak deaths. The number of observed deaths experienced a slight acceleration during September 2020 that the model cannot replicate. 6 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos Figure 1 Observed and Simulated Deaths Deaths (thousands) Nov-18 SIRD model 7.5 SIOD model Observed 5.0 2.5 0.0 2020−07 2021−01 2021−07 2022−01 Week SOURCE: Authors’ calculations based on data from the Instituto Nacional de Salud (INS). Figure 2 Observed and Simulated Consumption Gap Gap (percent) 0 –5 Nov-18 –10 SIRD model SIOD model Observed –15 2020−07 2021−01 2021−07 Week SOURCE: Authors’ calculations based on data from the Instituto Nacional de Salud (INS). 7 2022−01 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos On the other hand, the model predicts well the collapse of economic activity observed during the strict quarantine that was in place during April 2020. The observed recovery afterward is much bumpier than the prediction of the model. In any case, the model predicts well the rapid recovery of economic activity, as measured by the electricity consumption gap. The macro-SIOD model predicts that the COVID-19 pandemic will have been mostly over by the beginning of 2021. This prediction is relatively optimistic compared with standard epidemiological models that have been used to forecast the progress of the pandemic in Colombia. It should also be said that these standard models have consistently predicted much higher deaths than observed. On the other hand, the macro-SIOD model predicts that economic activity will have converged to its long-run path by mid-2021. The model predicts that consumption will have fallen 4.5 percent below its long-run level in 2020 and have recovered almost fully in 2021. We should note that our calibration is based on electricity consumption, which is an imperfect measure of consumption and of economic activity in general. In particular, there are sectors of the economy that might be permanently affected, such as tourism and entertainment activities, that are not intensive in electricity use. It should not be a surprise, then, that the model predicts a full recovery, whereas a portion of the economy most probably will be underperforming for a long while. In contrast to our macro-SIOD model, the calibrated macro-SIRD model has more difficulties matching the data. As shown in Figure 1, this model predicts a much later and higher peak in weekly deaths than what we observe in the data. The model predicts a total of almost 200,000 deaths, whereas the preferred macro-SIOD model predicts no more than 36,000 deaths. Standard epidemiological models, such as SIRD, commonly predict a much higher number of deaths than are observed. As shown in Figure 2, the macro-SIRD model also has difficulties predicting the path of economic activity. The model replicates well the initial dip of consumption, but it then shows a second dip in early 2021. In this model, the second dip is a result of the relaxation of government restrictions, which increases substantially the risk of contagion and which in turn induces consumers to reduce consumption and labor. The contrast between the calibrated SIRD and SIOD models suggests that the addition of immune/non-contagious agents allows the model to replicate the data much better than the standard model. The standard macro-SIRD model is unable to generate simultaneously reasonable predictions for both the number of deaths and economic activity, using our standard calibration methodology. Because our model assumes that immune agents stay immune forever, it predicts that the pandemic ends after the first wave of contagion. It should be noted, though, that our model can be easily extended to allow immune agents to become susceptible again by calibrating an additional exogenous probability of immune agents reverting to susceptibility. Moreover, vaccinations with full or limited immunity can also be incorporated into the model. Since at this point it is not clear how immunity works, even with vaccines, we abstract from the problem and focus on a one-wave pandemic. In the analysis that follows we use the preferred macro-SIOD model to evaluate counterfactual scenarios. We focus on the counterfactual behavior of the economy during 2020 and avoid using the model to predict the future, which will be affected by precisely how immunity evolves over time. 8 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos 3.2 Counterfactual Analysis: The Impact of Government Restrictions and Individual Choices We use the calibrated macro-SIOD model to evaluate the impact of imposed government measures and individual self-regulation on the pandemic and economic activity. Specifically, we simulate the model assuming that there are no government restrictions and that individuals do not account for the contagion risk when consuming or working. We call this last assumption “suboptimal consumption.” We perform three counterfactual simulations under combinations of these counterfactual assumptions. More specifically, each counterfactual simulation can be described as follows: 1. No government restrictions plus suboptimal consumption, μt = 0, and consumers ignore the additional risk of contagion associated with ct and nt : This simulation assesses the joint effect of government restrictions and individual behavior. It provides an upper bound on deaths relative to the baseline model. 2. No government restrictions plus optimal consumption decisions, μt = 0: In other words, there are no limits to consumption activities, and therefore individuals freely maximize their welfare, accounting for the risk of contagion. 3. Observed government restrictions plus suboptimal consumption: Individuals ignore the additional risk of contagion associated with ct and nt . We describe each simulation below. A common feature of the simulations is that they all predict the pandemic will have been over by early 2021. As pointed out above, this is an optimistic forecast that is very robust in the model. We should also reiterate that our consumption calculations are based on the use of electricity, which is not a precise measure of total consumption. In particular, activities intensive in personal interactions, such as dining at restaurants or meeting at entertainment venues, are less intensive in electricity use than are the production and consumption of, for example, manufactured goods. Therefore, the demand for electricity has shown a faster recovery than has the overall economy, and our model predicts a fast recovery as well, along with a full convergence to the long-run equilibrium path by the middle of 2021. The Impact of Both Government Restrictions and Individual Behavior. We first simulate the model assuming that there are no restrictions, setting μt = 0, ∀t , and assuming that consumers perceive the probability of contagion in (6) as not related to consumption and labor activities; that is, individuals believe π1 = π2 = 0 in (6). In this fully unrestricted model, individuals behave as if the pandemic follows the standard epidemiological model that assumes an exogenous probability of contagion. Nevertheless, the actual probability of contagion in (1) is still affected by the behavior of individuals. Under our assumption, individuals are irrational in the sense that they believe the contagion risk is exogenous and given solely by π3. Therefore, in this model the effect of the pandemic on economic activity is very low and driven only by the number of individuals who die, which is a small share of total population. The results of the simulation are shown in Figures 3 and 4. The predicted number of deaths in this simulation is 50,506, which is almost 42 percent higher than in the baseline. Recall that this number of deaths would have been the result if the government had imposed no restrictions and if 9 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos Figure 3 Simulated Deaths: Baseline vs. No Restrictions and Suboptimal Consumption Deaths (thousands) 3 Nov−18 2 1 Baseline No restrictions + suboptimal consumption 0 2020−07 2021−01 2021−07 2022−01 Week SOURCE: Model's predictions. Figure 4 Simulated Consumption Gap: Baseline vs. No Restrictions and Suboptimal Consumption Gap (percent) 0 Nov−18 –5 –10 Baseline No restrictions + suboptimal consumption –15 2020−07 2021−01 2021−07 2022−01 Week SOURCE: Authors’ calculations based on data from the Colombian Electricity Independent System Operator (XM). 10 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos individuals had not changed their behavior endogenously. In that sense, this figure is the upper bound in the number of deaths according to the calibrated model. Notice that the consumption path is almost constant, which is a reflection of the fact that behavior is not affected by the pandemic in this simulation.5 The baseline consumption in 2020 is 4.7 percent lower than this unrestricted consumption level. This figure is a rough estimate of the economic cost of the pandemic in the model. The Role of Government Restrictions. We now simulate the model assuming that there are no restrictions but that individuals fully account for the effects of their behavior on the risk of contagion. In other words, we set μt = 0, ∀t and keep the remaining parameters of the model as in the baseline simulation. The results of this simulation are shown in Figures 5 and 6. As shown in Figure 5 and as expected, the imposed quarantine did have a substantial effect on the number of deaths. Without the restrictions, the model predicts a total of 45,654 deaths by the end of the pandemic. The model implies that the excess deaths would have occurred mostly around the peak. The restrictions delayed the peak for several weeks and decreased its level by around 1,000 deaths per week. As shown in Figure 6, the model without government restrictions shows a much smaller dip in consumption than observed, which coincides in time with the predicted peak in deaths. In the model, this decrease in consumption is a result of consumers’ efforts to avoid contagion. In other words, the restrictions had an immediate and substantial effect on economic activity. The Role of Individual Behavior. Finally, we isolate the impact of individual efforts to self-regulate their behavior on both contagion and economic activity. We fix government restrictions as in the baseline calibration but set the perceived probability of contagion in (6) equal to zero; that is, π1 = π2 = 0. As explained above, in this model, individuals believe that contagion risk is given by π3. Because contagion is perceived to be unaffected by behavior, the simulated dip in consumption is entirely a result of government restrictions. We show the results of this simulation in Figures 7 and 8: There are around 2,800 more deaths in this scenario than in the baseline simulation. Compared with the results of the previous simulation, the model suggests that individual self-regulation had less of an effect on deaths than did government restrictions. The effect on economic activity is an increase of around 1.4 percent relative to the 2020 baseline value. In other words, the change in individual behavior explains a relatively small portion of the decrease in economic activity. We show a summary of the results of these simulations in Table 2, including consumption and deaths in 2020 and 2021 for the baseline and each simulation. As pointed out above, without government restrictions and agents’ behavior, the total number of deaths would have been 50,506, which is 42 percent higher than in the baseline simulation. Consumption in 2020 would have been 4.7 percent higher than in the baseline simulation and would have grown less than 1 percent in 2021. These figures are a measure of the total cost of the pandemic, in terms of both deaths and economic activity. Without government restrictions, agents’ behavior would have resulted in 45,458 deaths. Therefore, government restrictions saved 10,060 lives, which is 28 percent of baseline deaths. In this scenario, consumption would have been 3 percent higher in 2020 and then would have fully recovered in 2021. 11 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos Figure 5 Simulated Deaths: Baseline vs. No Restrictions and Fully Optimal Consumption Deaths (thousands) Baseline Nov−18 No restrictions 2 1 0 2020−07 2021−01 2021−07 2022−01 Week SOURCE: Model's predictions. Figure 6 Simulated Consumption Gap: Baseline vs. No Restrictions and Fully Optimal Consumption Gap (percent) 0 Nov−18 –5 –10 Baseline No restrictions –15 2020−07 2021−01 2021−07 2022−01 Week SOURCE: Authors’ calculations based on data from the Colombian Electricity Independent System Operator (XM). 12 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos Figure 7 Simulated Deaths: Baseline vs. Suboptimal Consumption with Restrictions Deaths (thousands) 2.0 Nov−18 Baseline Suboptimal consumption 1.5 1.0 0.5 0.0 2020−07 2021−01 2021−07 2022−01 Week SOURCE: Model's predictions. Figure 8 Simulated Consumption Gap: Baseline vs. Suboptimal Consumption with Restrictions Gap (percent) 0 Nov−18 –5 –10 Baseline Suboptimal consumption –15 2020−07 2021−01 2021−07 2022−01 Week SOURCE: Authors’ calculations based on data from the Colombian Electricity Independent System Operator (XM). 13 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos The role of individual behavior is more limited. Without the rational response of individuals to the risk of contagion, the number of deaths would have been 38,458, which is 2,864 more than in the baseline simulation. Therefore, individual behavior saved 8 percent of deaths, relative to the baseline. In this scenario, 2020 consumption would have been only 1.4 percent higher than in the baseline simulation and then would have almost fully recovered in 2021. These results imply that the combination of policy and individual behavior saved around 42 percent of baseline deaths, with an economic cost of around 4.7 percent of consumption in 2020. The simulations imply that government restrictions had a bigger impact than the endogenous changes in individual behavior. 4 CONCLUSION Table 2 Summary Results of Simulated Scenarios Baseline Deaths up to Dec 2020 35,311 Deaths up to Jun 2021 35,594 Deaths total 35,594 Consumption 2020 (trillions COP) 979.6 Consumption 2021 (trillions COP) 1,027.3 No restrictions plus suboptimal consumption Deaths up to Dec 2020 50,323 Deaths up to Jun 2021 50,506 Deaths total 50,506 Consumption 2020 (trillions COP) 1,027.9 Consumption 2021 (trillions COP) 1,027.9 No government restrictions Deaths up to Dec 2020 45,458 Deaths up to Jun 2021 45,654 Deaths total 45,654 Consumption 2020 (trillions COP) 1,009.6 Consumption 2021 (trillions COP) 1,027.9 We have calibrated a model of ecoSuboptimal consumption nomic behavior during the COVID-19 pandemic, as it applies to the Colombian 38,172 Deaths up to Dec 2020 economy. Our model incorporates an 38,458 Deaths up to Jun 2021 “immune” type of agent that better explains Deaths total 38,458 the data than do standard epidemiological Consumption 2020 (trillions COP) 993.2 models. In our model, the pandemic falls Consumption 2021 (trillions COP) 1,027.3 rapidly and disappears during early 2021. Consumption falls substantially during NOTE: COP, Colombian peso. 2020 but recovers fully by mid-2021. It SOURCE: Authors’ calculations. also implies that government restrictions and consumers’ self-regulation helped to avert around 15,000 deaths, or around 42 percent of baseline deaths. Government restrictions account for more than 67 percent of this effect. The model focuses on the first wave of the pandemic and is able to reproduce only this first wave. However, a shortcoming of the model is the assumption that immune agents stay immune forever. To generate additional waves of contagion, the model can be extended to allow for limited immunity so that immune agents become susceptible again after a period of time, or with some probability. Immunity can also be modeled to incorporate vaccinations. Given the uncertainty that surrounds the evolution of immunity to this virus over time, we leave these issues for future research. n 14 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos APPENDIX In this appendix, we show the full set of equilibrium conditions of the model. The definition of the variables and their transitions over time are defined in the body of the article. The additional equilibrium conditions of the model imply that all agents maximize their expected lifetime utility, firms maximize profits, the government balances its budget, and both labor and goods markets clear. First-order conditions for c and n: For j = S (susceptible agents), u1 ctS , ntS 1 t btS t 1 m 1 I t CtI = 0 and u2 ctS , ntS wt btS t 1 m 1 I t N tI = 0. First-order conditions for c and n: For j = I (infected agents), u1 ( ctI ,ntI ) = λbtI (1+ µt ) and u2 ( ctI ,ntI ) = −φ I wt λbtI . First-order conditions for c and n: For j = O (surviving agents), u1 ( ctO ,ntO ) = λbtO (1+ µt ) and u2 ( ctO ,ntO ) = −wt λbtO . The first-order condition for τt is I S β (1− π m )(U t+1 −U t+1 ) − λτt = 0. The government budget constraint implies that the government balances its budget; that is, µt ( St ctS + It ctI + Ot ctO ) = Γ t ( St + It + Ot ) . Finally, market-clearing conditions guarantee that both labor and goods markets clear; that is, St CtS + It CtI + Ot CtO St N tS + It N tI I φ + Ot N tO 15 = AN t = Nt . Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Carranza, Martin, Riascos NOTES 1 As pointed out by a referee, immunity arises independently of agents’ decisions. If immunity were a consequence of interactions, it would be equivalent to infection and the model would revert to the standard SIRD model: susceptible (S), infected (I), recovered (R), and dead (D). 2 An alternative specification would simply assume an exogenous number of immune agents. Our current specification recognizes the possibility that agents can move in and out of immunity and would easily accommodate vaccinations with limited effectiveness. 3 More specifically, we estimate such a trend fitting an autoregressive integrated moving average (ARIMA) model with both monthly fixed effects and a time trend and then project the implied, predicted mean starting from March 2020. 4 Our backtesting approach uses a moving window of 10 weeks starting from the third week of March 2020. After standardizing scales—that is, dividing each series by its own sample standard deviation—we find ω = 0.55. 5 It is worth noting that the slight drop in consumption of this counterfactual is due to the lower productivity faced by infected workers, which directly affects consumption in the equilibrium. REFERENCES Acemoglu, D.; Chernozhukov, V.; Werning, I. and Whinston, M.D. “Optimal Targeted Lockdowns in a Multi-Group SIR Model.” NBER Working Paper 27102, National Bureau of Economic Research, 2020; https://doi.org/10.3386/w27102. Adam, D. “A Guide to R—the Pandemic’s Misunderstood Metric.” Nature, 2020, 583(7816) pp. 346-48; https://doi.org/10.1038/d41586-020-02009-w. Alvarez, F.E.; Argente, D. and Lippi, F. “A Simple Planning Problem for COVID-19 Lockdown.” NBER Working Paper 26981, National Bureau of Economic Research, 2020; https://doi.org/10.3386/w26981. Atkeson, A. “What Will be the Economic Impact of COVID-19 in the U.S.? Rough Estimates of Disease Scenarios.” Staff Report 595, Federal Reserve Bank of Minneapolis, 2020; https://doi.org/10.21034/sr.595. Berger, D.W.; Herkenhoff, K.F. and Mongey, S. “An SEIR Infectious Disease Model with Testing and Conditional Quarantine.” Staff Report 597, Federal Reserve Bank of Minneapolis, 2020; https://doi.org/10.21034/sr.597. Doshi, P. “COVID-19: Do Many People Have Pre-Existing Immunity?” BMJ, 2020, 370; https://doi.org/10.1136/bmj.m3563. Eichenbaum, M.S.; Rebelo, S. and Trabandt, M. “The Macroeconomics of Epidemics.” NBER Working Paper 26882, National Bureau of Economic Research, 2020; https://doi.org/10.3386/w26882. Kermack, W.O. and McKendrick, A.G. “A Contribution to the Mathematical Theory of Epidemics.” Royal Society, 1927, 115(772) pp. 700-21; https://doi.org/10.1098/rspa.1927.0118. 16 Sectoral Impacts of Trade Wars Wan-Jung Cheng and Ping Wang In recent years, we have witnessed rising trade protectionism with broad ranges of tariffs imposed on intermediate products. In this article, we develop an accounting framework to evaluate the sectoral impacts of the current U.S.-China trade war. We find that U.S. final demand and intermediate demand for goods produced by China decline significantly, with the largest losses occurring in the Electronic and ICT (information and communications technology) industry and the Electrical industry. We obtain sizable deadweight losses for the United States, particularly in the Electronic and ICT; Electrical; and Furniture industries. We also find that, with a leakage rate of 20 percent, total losses to U.S. consumers and importers are $3.3 billion, about 0.05 percent of gross U.S. output, whereas the full leakage losses are $10.7 billion, or 0.16 percent of gross U.S. output, which is twice as much as the annual welfare gains from the North America Free Trade Agreement. (JEL D20, F10, O50) Federal Reserve Bank of St. Louis Review, First Quarter 2022, 104(1), pp. 17-40. https://doi.org/10.20955/r.104.17-40 1 INTRODUCTION Not long after the worldwide Great Recession, protectionism began to rise, from the battled renegotiations of the North America Free Trade Agreement (NAFTA), to the recently escalated U.S.-China trade war, to the ongoing Japan-Korea trade war. Rising protectionism concerns some economists, particularly those who view free trade as beneficial to both developed countries (hereafter, the North) and developing countries (hereafter, the South), by advancing world productivity and enhancing global consumer welfare. A particular concern is that recent trade protectionism has included broad ranges of tariffs imposed on intermediate products (for example, in the United States, nearly 90 percent of intermediate imports from China faced increased tariffs in 2018, as computed by Bown, 2019). Such tariffs violate the so-called Diamond and Mirrlees (1971) principle of optimal taxation: Taxing intermediate goods creates much larger economic distortions and is more harmful to economic development. Wan-Jung Cheng is an assistant research fellow at Academia Sinica. Ping Wang is the Seigle Family Professor at Washington University in St. Louis, a senior fellow at the Federal Reserve Bank of St. Louis, and a research associate at the National Bureau of Economic Research. We are grateful for useful suggestions from Ching-mu Chen, Tain-Jy Chen, Wen-Tai Hsu, Shin-Kun Peng, and Ray Riezman. © 2022, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. 17 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang In this article, we provide an overview of various key findings in the literature on gains from trade and review the literature on trade wars. We then develop an accounting framework to evaluate the sectoral impacts of the current U.S.-China trade war. Using the international input-output linkage between the two countries and isoelastic demands, we compute how much each U.S. sector’s demand for goods produced in China declines when the United States raises its import tariffs as it has under the ongoing trade war. The resulting sectoral deadweight losses and full leakage losses are also computed, where the latter considers an extreme case with no tariff revenues redistributed back to consumers or importers. We find that in response to the trade war, U.S. final demand for Chinese goods drops by $39 billion and intermediate demand by $13 billion. Among others, the Electronic and Information and Communications Technology (ICT) industry and the Electrical industry suffer the largest losses, with their demands lowered by $23 billion and $9.5 billion, respectively. U.S. aggregate deadweight losses are about $1.5 billion, with the Electronic and ICT; Electrical; Metal Products; and Furniture industries suffering the greatest total losses. With a leakage rate of 20 percent, total losses to U.S. consumers and importers are $3.3 billion, about 0.05 percent of gross output and two-thirds as much as the annual welfare gains from NAFTA. The full leakage losses are $10.7 billion, or 0.16 percent of gross U.S. output, which is twice as much as the annual welfare gains from NAFTA. 2 BACKGROUND AND LITERATURE REVIEW As argued by Kindleberger (1989), the Smoot-Hawley Tariff Act of 1930 passed by the U.S. Congress led to tariff wars and defensive trade blocs, with a peak sector-weighted average tariff of 24.4 percent for the United States, 29.4 percent for France, and 47.7 percent for the United Kingdom. Germany, Italy, Japan, and the Soviet Union had explicit autarkic aims and militaristic motivations behind their defensive trade blocs, and historians have argued that tariff wars might have subsequently triggered WWII. Thus, when the world returned to peace, many major industrialized countries worked hard on international cooperation. As a result, the General Agreement on Tariffs and Trade (GATT) was signed by 23 nations in Geneva on October 30, 1947, and took effect on January 1, 1948. Since then, there have been several more rounds of negotiations on international cooperation, including the most crucial Kennedy Round (1964-67), Tokyo Round (1973-79), and Uruguay Round (1986-94), each featuring a major tariff cut of over 30 percent. As a result, the average tariff for major GATT participants reduced from about 22 percent in 1947 all the way down to about 5 percent in 1995. Especially, toward the end of the Uruguay Round agreements, the World Trade Organization was signed by 123 nations in Marrakesh on April 14, 1994, and established on January 1, 1995, replacing the previous workhorse, GATT. By 2000, the average tariffs in North America and the European Union (EU) were 4.0 percent and 4.2 percent, respectively, whereas in Asia the average was 9.0 percent (in China the average was 16 percent). The relatively high average tariff in Asia should not be surprising, because many economies there are less developed and several are centrally planned to lean toward protecting domestic, less-competitive firms. The prevalence of trade protection is particularly high in the Agricultural; Food and Beverage; and Light industries (often exceeding 20 percent, and even 40 percent in the Agricultural industry). 18 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang The over-half-a-century history of international cooperation is supported by a vast literature on gains from trade. For the sake of brevity, we provide an overview of only the more recent literature pioneered by Arkolakis, Costinot, and Rodriguez-Clare (2012, hereafter ACR). For comparison, we only discuss articles that measure gains from trade by contrasting the welfare under current trade costs with the welfare in an autarky world without international trade (that is, infinite trade costs). Based on a general accounting framework consistent with many static trade models, ACR and their followers found modest gains from trade, usually below 2 percent, compared with autarky. Such gains from trade, however, rise significantly in dynamic frameworks, particularly when technologies are allowed to advance over time (for example, Hsieh, Klenow, and Nath, 2019, find they could be as high as 12.2 percent; Perla, Tonetti, and Waugh, 2021, as high as 13.3 percent; and Bloom et al., 2013, as high as 16.3 percent). In short, in dynamic models with technological improvements, one expects sizable gains from trade because the reduction in trade barriers can promote better technology, raise world productivity, and enhance global consumer welfare. The sizable gains from trade justify the long-devoted effort toward international trade agreements and tariff reductions. The effort toward trade liberalization has unfortunately reversed lately. We have seen Brexit, the battled renegotiations of NAFTA, the recently escalated U.S.-China trade war, the ongoing JapanKorea trade war, and the possible U.S.-EU trade war—each started by an advanced high-income country. More than three decades ago, Kennan and Riezman (1988) showed that when two countries are engaged in a trade war, the more-advanced larger economy can more easily manipulate international prices and “win” over the less-developed smaller economy. If both countries are of comparable size and at similar development stages, both countries lose. The authors illustrate the battle by using an Edgeworth box, where the cigar-shaped area in the middle represents the set of relative country sizes in which both countries lose a trade war. In this pivotal article, trade protection is imposed strictly on final goods. In practice, this is not the case. In view of the current U.S.China trade war, would it still be true that the United States would win, whereas China would lose? We begin by illustrating the trade war between two economies that are comparable in size. Consider a hypothetical trade war between the United States and the EU, a case where the relative country sizes lie in the cigar-shaped area. Ossa (2014) found that a large increase in tariffs—of over 50 percent—would only result in a 2 percent welfare loss in the United States and a 2.6 percent loss in the EU. Next, we turn to the relatively thin literature that focuses on the U.S.-China trade war. This ongoing trade war consists of a series of announcements between the two governments over three years, starting from the Trump administration’s memorandum on March 22, 2018, that spelled out the intention to impose a 25 percent tariff on over $50 billion of imports from China. Bown (2019) provides a comprehensive review of the detailed content of this trade war. Focusing on the major waves of the trade war in 2018, Amiti, Redding, and Weinstein (2019) find almost complete tariff pass-through that raised consumer prices of importables almost one-for-one (see a graphical illustration in their Figure 4). They find that rising consumer prices lowered consumer welfare, which they estimate as having deadweight losses of $8.2 billion (in 2018 prices). This welfare loss is comparable to the welfare gains of NAFTA estimated by Caliendo and Parro (2015). Yet, Amiti, Redding, and Weinstein (2019) also estimate an additional cost of $14 billion to U.S. importers and consumers as a result of tariff revenues transferred to the government. Moreover, they find that exporters also suffered due to retaliatory tariffs by China, with lost exports estimated at about 19 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang $28 billion. Summing both imports and exports together, Amiti, Redding, and Weinstein (2019) estimate that about $183 billion of trade was redirected. Notably, such reshuffling may induce severe increases in business costs and resource misallocation that have not been counted in the deadweight loss measures mentioned. Using a general equilibrium model, Fajgelbaum et al. (2020) compute U.S. losses as a result of rising consumer prices as $51 billion for importers and consumers and $7.2 billion in aggregate real income by the end of 2018 (based on an aggregate equivalent variation measure frequently used by trade theorists). They further identify that the industrial regions of the Midwest and Northeast suffered less (more domestic protection with less retaliation), whereas the rural regions of the Midwest and Mountain West suffered more. Turning the focus to firms, Amiti, Kong, and Weinstein (2020) conduct an event-study analysis and find the various waves of the U.S.-China trade war over 2018-19 caused the stock prices of U.S. firms that trade with China to drop by 2.6 percentage points and those of other U.S. firms to drop by 3.4 percentage points, a total drop of 6.0 percentage points. Lower returns to U.S. publicly listed firms thus induced a 1.9-percentage-point reduction in those firms’ business investments, which hurts future productivity. In an independent work in this Review, Santacreu and Peake (2020) estimate responses across states and find that those states more exposed to trade experienced lower increases or even decreases in output growth and employment growth between 2018 and 2019. Finally, one may wonder what happened to the retaliating party—China. While there are few systematic studies, we may learn from a more general framework of the North (which includes the United States) and the South (which includes China), developed by Chen et al. (2020). They consider a global supply chain along which the North produces intermediate goods with more advanced technologies and the South with inferior technologies. Thus, trade in intermediate goods that embody different technologies may act as a mechanism to transfer technology from the North to the South. They find that when the South uses higher-end intermediate goods more intensively along the supply chain in response to a tariff war, its loss may be mitigated by rising average productivity. 3 AN ACCOUNTING FRAMEWORK We now establish an accounting framework to assess sectoral impacts of the trade war. Similar to the argument made by ACR, the advantage of adopting an accounting framework is that the results are less sensitive to model specifics. That is, the framework is consistent with a larger set of models. We base our analysis on three fundamental structures: (i) There is an international input-output linkage between the source country, the destination country, and the rest of the world (ROW), constituting the global value chain. (ii) Sectoral demands are isoelastic, where both the demand shifters and the price elasticities are sector specific. (iii) The increase in import tariffs in the United States during the trade war is completely passed through to U.S. domestic prices. While the first two are standard assumptions in the trade literature, the third is rooted in the empirical findings of Amiti, Redding, and Weinstein (2019) elaborated on below. 20 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Consider a Leontief framework that features a global (domestic and international) input-output matrix. Denoting X as gross output, Z as intermediate demand, F as total final demand, and A = [aij] as the input technology matrix, according to the Leontief framework we have X Zi F AX F . (1) Under this Leontief framework, production of one unit of the jth good uses aij units of the ith good as inputs. The Leontief input-output framework is based on a Leontief fixed-proportion technology. The input coefficients are allowed to vary over time, but how such coefficients evolve dynamically is not modeled explicitly. Denote I as the identity matrix and let 1 I A , and by rearranging (1) we know that final demand F and global supply X satisfy the following relationship: X F. Ω is the Leontief inverse matrix: k i s j , where ωk(i)s( j) captures the backward linkage effect of a change in the final demand for goods in sector j sourced from country s on the change in the output of sector i in country k. We first provide a simplified example for illustration and then construct the general framework. Suppose there are three countries, the United States, China, and ROW, denoted as {U,C,R}, and two sectors, denoted as {1,2}. To describe the international flow of goods, we denote the source country as country s and the destination country as country d. Let Xkj denote the gross output of sector j in country k and Fsjd country d’s final demand for sector-j goods sourced from country s. The Leontief framework thus implies that (2) FUU1 FUC1 FUR1 XU1 U C R X FU 2 FU 2 FU 2 U 2 U X C1 FC1 FCC1 FCR1 , U FC 2 FCC2 FCR2 XC2 X R1 FRU1 FRC1 FRR1 U C R X R 2 FR 2 FR 2 FR 2 U U where, for example, FC1 and FC2 are final U.S. demand for China’s goods in sectors 1 and 2, respectively. Moreover, we have 21 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang (3) FUU1 FUC1 FUR1 X U 1 U C R X F F F 2 2 2 U U U U2 U C R X C 1 FC1 FC1 FC1 . U F FCC2 FCR2 X C 2 C2 X R1 FRU1 FRC1 FRR1 U C R X R 2 FR 2 FR 2 FR 2 Note that k i s j captures the effect of the international input-output linkage. For example, ωU(i)C( j) captures the effect of a change in the final demand for Chinese goods of sector j (from ΔFCjU , ΔFCjC , ΔFCjR ) on the change in the output of U.S. sector i (ΔXUi ). We further suppose that the rise in the import tariff on Chinese goods imposed by the United States only affects FCjU . That is, U.S. final demand (including final consumption and investment) for domestic goods and imports from ROW are not affected, and China’s and ROW’s final demands for goods from any country are not affected. The latter assumption is natural given that we focus on the tariff increases imposed by the United States and do not consider retaliatory tariffs by China. The former assumption restricts the United States from substituting between Chinese goods and other goods.1 Given these assumptions, (3) becomes (4) 0 X U 1 0 X U2 FCU1 X C 1 k i s j U . X FC 2 C 2 X R1 0 0 X R 2 To estimate ΔFCjU we need to specify the sectoral demand structure for imports, and we adopt the isoelastic demand assumption that is commonly used in the trade literature. Let qsjd denote the quantity of sector-j goods imported from country s to country d. By properly choosing the base year, let the price psjd be normalized by psjd = 1 + τ sjd , where τ sjd is country d’s tariff rate imposed on imports in sector j from source country s. Given isoelastic demand, U.S. demand for sector-j goods from China in logs before the trade war is given by ln qsjd ln D j sjd ln 1 sjd , where σsjd < 0 is the import demand elasticity, which is estimated by subtracting the fixed effects for the destination country from the fixed effects for the source country. The demand scaling parameter Dj is thus by definition specified as Dj = qsjd /(1 + τ sjd )σ . In the World Input-Output Database (WIOD), Fsjd is expressed in terms of value. Thus, writing the demand relationship in terms of values, we have d sj (5) ln Fsjd ln p j ln qsjd ln p j ln D j sjd ln 1 sjd ln V j sjd ln 1 sjd , 22 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang where Vj ≡ pj Dj . Let the prime symbol (ʹ ) indicate the quantity or price after the trade war. Analogously, the value of demand after the trade war is (6) ln Fsjd ln p j qsjd ln V j sjd ln 1 sjd , where (7) p jqsjd p j Fsjd V j p j D j sjd pj 1 d 1 sjd sj sjd p j V j , pj where the last equality is calculated by applying (5). Empirical evidence in Amiti, Redding, and Weinstein (2019), based on monthly data from January 2017 to December 2018, suggests that “the Trump administration’s tariff changes have been almost entirely passed through into domestic prices” (p. 197). Their findings are consistent with more recent studies by Fajgelbaum et al. (2020) and Cavallo et al. (2021). Thus, based on all of these findings, we assume complete pass-through of tariffs in our analysis.2 That is, we have p j (8) pj 1 sjd 1 sjd . Taking the difference between (5) and (6) and applying (7) and (8), we obtain F d 1 d sj sj , ln d 1 sjd ln Fsj 1 sjd or, equivalently, (9) 1 Fsjd Fsjd sjd 1 1 sjd 1 sjd . Accordingly, the impact of the trade war on country d’s final demand for country s’s goods in sector j in response to country d’s tariff hike is measured by (10) sjd d Fsj 1 1 d sj 1 sjd 1 Fsjd . Let τ denote the average tariff in destination country d, and consider a trade war that causes the average tariff to increase by Δτ. The different tariff rates in different sectors are reflected by country d’s tariff coverage rate imposed on imports in sector j from source country s, which is denoted by TCRsjd and can be computed from the data.3 The sector-specific tariff increase is thus measured by Δτ sjd /(1 + τ sjd ) = (TCRsjd ∙ Δτ)/(1 + τ). The import demand elasticity σsjd can be measured by the trade elasticity, which captures the impact of trade costs such as tariffs on the quantity of country d’s demand for sector-j goods. From (4), the impact on sectoral output can be computed as 23 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang (11) j 1,2 U 1C j FCjU X U 1 U F j 1,2 U 2 C j Cj X U 2 U F j 1,2 C 1C j X C1 Cj . U X C 2 j 1,2 C 2 C j FCj X U R1 j 1,2 R 1C j FCj X R 2 U FCj j 1,2 R 2 C j ∆Xkj derived by (11) measures how much the output of sector j in country k would be affected through the international input-output linkage when the United States raises its import tariff on Chinese goods by Δτ on average. Some comments are in order. As mentioned above, this impact measure is based on a general accounting framework that is consistent with a wide class of models. Because trade elasticities could be measured in the short or the long run, the potential dynamic effects could be partially accounted for. The only limitation is that in (3), we assumed that, in response to the trade war, (i) U.S. final U U demand for its own and ROW’s goods in any sector (ΔFUj and ΔFRj ) remain unchanged and (ii) China’s and ROW’s final demands for any country’s goods in any sector remain unchanged. While the latter is not our concern, because we are not characterizing U.S. exporter responses, the former prohibits us from estimating redirection of U.S. demand to other sources. Think of this scenario as being under the framework of Alvarez and Lucas (2007), with buyers searching globally for the cheapest goods of similar quality. Imports from China are associated with a vector of observed minimum prices, but what would the next-highest augmented trade price be? We cannot know unless we have empirically observed what happens after the trade war when redirection starts. The reshuffling of the global value chain requires deep structures of international demand and pricing strategies, which is beyond the scope of this article. Now we are ready to extend to the general framework. Suppose there are K countries and J sectors. The U.S. tariff increase imposed on Chinese imports by Δτ on average would impact the output in country k’s sector i through the international input-output linkage as follows: (12) d J 1 sj X ki kK ,iJ j 1 k (i )C ( j ) 1 d sj 1 sjd U 1 FCj . kK ,iJ Note that this impact measure is the general form of (11). To compute welfare losses, we extend the framework in Amiti, Redding, and Weinstein (2019). They note that almost complete pass-through of tariffs means that the source country’s supply of exports is close to perfectly elastic, which is illustrated in Figure 2 in their article. Following their analysis, we compute the sector-specific deadweight losses in destination country d under a trade war it initiates toward source country s as follows: (13) DWLdsj 1 sjd qsjd , 2 24 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang where qsjd is the quantity of sector-j goods imported from country s to country d, and the unobservable quantity changes ∆qsjd will be estimated. Recall that from isoelastic demand, sjd d q D j psjd sj qsjd d we have inverse demand psj Dj , 1 sjd , and hence in value terms, 1 d sj j d D j psjd After the trade war, we have qsj 1 d 1 d sj sj 1 D q vsjd psjd qsjd sjd . . Thus, vsjd D j sjd qsjd 1 1 qd d sj vsj d qsj 1 sjd 1 sjd , and (14) qsjd d d d d vsj vsj vsj vsj 1 d qsj 1 1 sjd 1 . Therefore, from (14) and with the price prior to the trade war normalized by psjd = 1 + τ sjd , we can rewrite the quantity changes as follows: vsjd q j 1 1 d vsj sjd 1 sjd vsjd 1 1 d qj vsj 1 sjd vsjd , d 1 sj sjd where νsjd is the value of sector-j goods imported from country s to country d. Denote Zsjd as country d’s intermediate demand for sector-j goods from country s. The total value of sector-j goods imported from country s to country d is the sum of intermediate and final demand, whereby ΔZsjd can be computed by (15) sjd d Z sj 1 1 d sj 1 sjd 1 Z sjd . Thus, the deadweight losses defined in (13) become (16) vsjd 1 DWLdsj sjd 1 1 d 2 vsj 25 sjd 1 sjd vsjd d 1 sj Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang with νsjd = Fsjd + Zsjd and ∆νsjd = ∆Fsjd + ∆Zsjd. Note that the implicit assumptions in Amiti, Redding, and Weinstein (2019) include a partial equilibrium, perfect competition, U.S. tariff increases treated as an exogenous shock, and no retaliation by China. The increase in tariff revenue due to the trade war is computed as TRevsjd d sj (17) qsjd qsjd 1 d qsj 1 d sj d sj vsjd vsjd d qsj sjd 1 sjd vsjd , 1 sjd where the last equality is calculated by applying (14) and psjd = 1 + τsjd . As mentioned by Amiti, Redding, and Weinstein (2019), the cost of the U.S. tariff increase is twofold. First, the higher prices caused by the trade war lead to distortion of domestic consumption and production decisions, constituting the deadweight losses captured by (13). The increase in tariffs widens the wedge between the prices charged by foreign producers and the prices paid by domestic consumers and producers and leads to further distortion of final demand and intermediate demand. Second, complete tariff pass-through suggests that higher prices resulting from increased tariffs were almost entirely borne by U.S. consumers and importers. Incremental tariff revenue is thus a transfer from domestic consumers to the government. If the government does not use the tariff revenue to generate social welfare, then the welfare loss from the tariff war to the economy equals the sum of deadweight losses (13) and the incremental tariff payments (17). Formally, we define a leakage rate λ as the fraction of tariff revenues not redistributed back to consumers or importers but used for other purposes that do not benefit consumers or importers. In our quantitative analysis below, we will link 1 – λ with the fiscal multiplier. Total losses are thus given by d TLoss DWLdsj TRevsjd . sj Thus, deadweight losses, DWL, are a special case with λ = 0 and full leakage losses, FLoss, are a special case with λ = 1. Total losses, TLoss, are in between the lower-bound measure of DWL and the upper-bound measure of FLoss. 4 QUANTITATIVE ANALYSIS We are now prepared to compute the impact measures for the ongoing U.S.-China trade war, where the destination country d is the United States and the source country s is China. The primary data source is the World Input-Output Tables (WIOT) from the 2016 release of the WIOD (see the discussion in Timmer et al., 2015, concerning the 2013 release). We then adopt the trade elasticities from Caliendo and Parro (2015), using the estimates with 1 percent trimming, and obtain the U.S. tariff coverage rates by sector and the average tariff increase of ∆τ = 9.3 percent from Bown (2019).4 Because sectoral classifications in these three sources are based on different systems, we make additional adjustments. The final list of sectors (industries) is provided in Table 1, and the sectoral trade elasticities and tariff coverage rates are provided in Table 2. 26 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Table 1 Classifications of Sectors Sectors WIOD sectors Caliendo-Parro (2015) sectors Bown-Crowley (2016) sectors Primary A01-03 Agriculture Hides and Skins; Animal Products; Vegetable Products Food and Tobacco C10-12 Food Prepared Food Textile C13-15 Textile Textiles and Clothing; Footwear Wood C16 Wood Wood Paper C17 Paper Wood Printing and Media C18 Paper Wood Petroleum C19 Petroleum Fuel Chemicals C20 Chemicals Chemicals Pharmaceutical C21 Chemicals Chemicals Plastic and Rubber C22 Plastic Plastics and Rubber Non-metallic Mineral C23 Minerals Mineral Products; Stone and Glass Basic Metals C24 Basic Metals Metals Metal Products C25 Metal Products Metals Electronic and ICT C26 Communication; Office; Medical Electronics and Electrical Machinery Electrical Equipment C27 Electrical Electronics and Electrical Machinery Machinery C28 Machinery Machinery Motor Vehicles C29 Auto (Motor Vehicles) Transportation Equipment Other Transport Furniture Repair and Installation C30 Other Transport Transportation Equipment C31-32 Other Miscellaneous C33 Machinery Machinery NOTE: For details, refer to Table A1 in Caliendo and Parro (2015) and Table Appendix B in Bown and Crowley (2016). CaliendoParro (2015) sectors are based on International Standard Industrial Classification Revision 3; Bown-Crowley (2016) sectors are based on Harmonized System sections. The Miscellaneous sector in Bown (2019) contains Optical, Precision, and Medical Products industry, while the Other Manufacturing industry includes furniture. Since the share of the former is expected to be small, we assign it to map with WIOD sectors C31-32 (Manufacture of Furniture; Other Manufacturing). In the analysis below, we use primarily the classification of the 2016 WIOD release, which classifies sectors according to International Standard Industrial Classification Revision 4. Specifically, it separates the Furniture industry from the Other Wood Products industry; the Printing industry from the Paper industry; the Pharmaceutical industry from the Chemical industry; the Machinery Repair and Installation industry from the Machinery industry; and the Motor Vehicles industry from the Other Transport industry. Whenever a measure for a sub-industry industry is not available, we apply the measure for the broader industry to all sub-industries. Two points are noteworthy. First, in the Wood; Paper; Petroleum; and Electrical industries, trade elasticities are high and hence U.S. demand for products from these industries is more sensitive to a tariff war. Second, in more than half of the industries under study, tariff coverage rates exceed 50 percent. In the Food and Tobacco; Transport (Motor Vehicles and Other Transport); Metal 27 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Table 2 Sectoral Trade Elasticities and Tariff Coverage Rates Sector Trade elasticity Tariff coverage rate (%) Primary –9.11 77.99† Food and Tobacco –2.62 94.22 Textile –8.1 Wood –11.5 69.92 Paper –16.52 69.92 Printing and Media –16.52 69.92 Petroleum –64.85 84.63 Chemicals –3.13 55.76 Pharmaceutical –3.13 55.76 Plastic and Rubber –1.67 70.42 Non-metallic Mineral –2.41 47.65§ Basic Metals –3.28 85.56 Metal Products –6.99 85.56 Electronic and ICT –8.54* 45.11 Electrical Equipment 9.84‡ –12.91 45.11 Machinery –1.45 51.49 Motor Vehicles –1.84 91.86 Other Transport –0.39 91.86 Furniture –3.98 66.26 Repair and Installation –1.45 51.49 NOTE: *Average of –3.95, –12.95, and –8.71. †Average of 100, 78.88, and 55.10. ‡Average of 12.66 and 7.01. §Average of 44.11 and 51.19. See Table 1 for the corresponding sectors for the averages. (Basic Metal and Metal Products); and Petroleum industries, tariff coverage rates are especially high—above 80 percent each. Thus, U.S. demand for products from these industries is more likely to be discouraged. We use WIOT data and make the base year 2014, the last year for which data are available. We first compute U.S. final demand for sector-j goods from China as a ratio of total U.S. final demand in sector j, defined as (18) U FDRCj FCjU U kK1 Fkj . Similarly, we also compute U.S. intermediate demand for sector-j goods from China as a ratio of total U.S. intermediate demand in sector j, defined as (19) U IDRCj U Z Cj U kK1 Z kj 28 . Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Table 3 China’s Contribution to U.S. Demand FDRUCj (%) IDRUCj (%) DRUCj (%) A01-03: Primary 0.4 0.1 0.2 C10-12: Food and Tobacco 0.8 0.2 0.6 C13-15: Textile 29.5 4.9 25.0 C16: Wood 7.2 1.9 3.1 C17: Paper 2.4 1.4 1.6 C18: Printing and Media 0.9 0.4 0.8 C19: Petroleum 0.2 0.6 0.3 C20: Chemicals 1.1 3.8 2.6 C21: Pharmaceutical 2.1 0.8 1.4 C22: Plastic and Rubber 8.4 2.7 4.5 C23: Non-metallic Mineral 9.8 3.4 5.2 C24: Basic Metals 26.0 1.1 2.0 C25: Metal Products 10.9 2.4 4.2 C26: Electronic and ICT 26.1 13.6 22.4 C27: Electrical Equipment 26.0 11.7 20.4 C28: Machinery 6.7 7.3 6.9 C29: Motor Vehicles 0.9 3.6 1.6 C30: Other Transport 1.5 1.1 1.4 C31-32: Furniture 11.6 2.6 10.4 C33: Repair and Installation 0.0 0.0 0.0 Total 7.1 2.7 5.1 WIOD sector Finally, we define the aggregate demand ratio of U.S. imports from China as follows: (20) U DRCj U FCjU ZCj U U kK1 Fkj Z kj , which captures overall U.S. dependence on Chinese imports. We report FDR, IDR, and DR in Table 3. The results suggest that in the Textile; Electronic and ICT, Electrical Equipment; Basic Metals; Furniture; and Metal Product industries, U.S. final demand is highly China dependent, with each industry’s demand ratio exceeding 10 percent. U.S. intermediate demand is highly China dependent only for the Electronic and ICT industry and Electrical Equipment industry, which is not surprising, because Chinese products for these industries are more heavily downstream in the global value chain. By aggregation, overall U.S. dependence on Chinese imports is particularly high in the Textile; Electronic and ICT; Electrical Equipment; and Furniture industries. Using the formula derived in Section 3, we can compute the measures of sectoral impact on output, {∆Xkj}k = {U,C,R}; deadweight losses, DWLUCj ; and tariff revenues, TRevUCj , as follows. There are 29 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang three countries (the United States, China, and ROW) and 21 WIOD sectors (as listed in Table 1), including a “Remainder” sector (not shown) consisting of all non-agriculture and non-manufacturing industries. Following Bown (2019), we set τ = 3.1 percent and τʹ = 12.4 percent. In the WIOT, the data are observed as X = Zi + F, where X, Z, and F are gross output, intermediate demand, and final demand, respectively.5 Specifically, the intermediate demand matrix is expressed as ZU (1)U (1) ZU (21)U (1) Z C (1)U (1) Z Z C (21)U (1) Z R (1)U (1) Z R (21)U (1) ZU (1)U (21) ZU (21)U (21) ZU (1)C (1) ZU (1)C (21) ZU (21)C (1) ZU (21)C (21) Z C (1)U (21) Z C (21)U (21) Z R (1)U (21) Z C (1)C (1) Z C (21)C (1) Z R (1) C (1) Z R (21)U (21) Z R (21)C (1) Z R (21) C (21) Z C (1)C (21) Z C (21)C (21) Z R (1) C (21) ZU (1) R (1) ZU (1) R (21) ZU (21) R (1) ZU (21) R (21) Z C (1) R (1) Z C (1) R (21) , Z C (21) R (1) Z C (21) R (21) Z R (1) R (1) Z R (1) R (21) Z R (21) R (1) Z R (21) R (21) where z k(i)s( j) denotes the demand for country k’s sector-i goods from sector j in country s. The zk i s j , and therefore ak i s j Leontief input technology matrix can be thus calculated as A X s j we know that X = AX + F = ΩF, where Ω = [ωk(i)s( j)]. Given this matrix and the tariff coverage rates, the explicit form of the sectoral impact on output through the international input-output linkage (12) can be expressed as (21) U X ki 20 U , C , R , i 1 20, j 1k i C j FCj for k where (22) U TCRCj FCjU 1 1 U 1 Cj 1 FCjU . The explicit form of the deadweight losses (13) for the United States is thus given by (23) DWLUCj 1 U U TCRCj qCj for j 1 20, 2 wherein we are particularly interested in three measures: • Measure 1 (our benchmark, as in Amiti, Redding, and Weinstein, 2019), considers both final demand and intermediate demand: U FCjU ZCj U qCj 1 1 U FCjU Z Cj 30 U 1 CjU FCjU ZCj , 1 U Cj Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang where U.S. intermediate demand for sector-j goods from China ZUCj is calculated as U Z Cj i211 zC j U i and (24) U Z Cj U TCRCj 1 1 U 1 Cj U 1 Z Cj . • Measure 2 considers only final demand: FCjU U qCj 1 1 U FCj U Cj 1 CjU FCjU . 1 • Measure 3 considers only private consumption demand: U FconsCj U qCj 1 1 U FconsCj U 1 CjU FconsCj , 1 U Cj where note that in the WIOT the variables CONS_h and CONS_np denote the final consumption expenditure by households and by non-profit organizations, respectively. Accordingly, we define private final consumption as U U U FconsCj CONS _ hCj CONS _ npCj , and thus U FconsCj FCjU U U CONS _ hCj CONS _ npCj FCjU . Finally, the incremental U.S. tariff revenue from the tariff increase is U U (25) TRev 1 TCRCj Cj FCjU FCjU U Z Cj U Z Cj U Cj 1 CjU U FCjU Z Cj 1 for j 1 20. As discussed above, total losses (TLossUCj ) are measured by DWLUCj + λ ∙ TRevUCj , depending crucially on the leakage rate λ. In the absence of a precise measure of λ, we measure it as 1 minus the fiscal multiplier. Owyang, Ramey, and Zubairy (2013) estimate that the U.S. fiscal multipliers range from 0.7 to 0.9. We take their findings and assume 1 – λ = 0.8. We then report the total losses for each sector as the sum of our benchmark measure of deadweight losses and λ times the incremental tariff revenue. We summarize these results in Tables 4 to 7 (measured in US$ millions). 31 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Table 4 Sectoral Impacts on Chinese Output WIOD sector ΔXC j (US$ millions) ΔF UCj + ΔZUCj (US$ millions) A01-03: Primary –3,216 –303 C10-12: Food and Tobacco –2,980 –599 C13-15: Textile –7,181 –3,175 C16: Wood –2,335 –774 C17: Paper –1,534 –1,385 –348 –93 C19: Petroleum –3,084 –1,207 C20: Chemicals –5,979 –1,290 –650 –243 C22: Plastic and Rubber –2,855 –278 C23: Non-metallic Mineral –2,429 –174 C24: Basic Metals –8,939 –973 C18: Printing and Media C21: Pharmaceutical C25: Metal Products –4,525 –3,892 C26: Electronic and ICT –32,849 –23,103 C27: Electrical Equipment –11,730 –9,520 C28: Machinery –2,773 –478 C29: Motor Vehicles –1,377 –688 –98 139 –3,692 –3,504 0 0 –98,574 –51,540 C30: Other Transport C31-32: Furniture C33: Repair and Installation Total NOTE: ΔXC j , derived according to (21), measures the impacts on the sectoral output in China through the international input- U output linkage when the United States raises import tariffs against China. ΔF U Cj + ΔZ Cj , derived according to (22) and (24), is the estimated change in U.S. import demand for Chinese goods due to the U.S. tariff war against China. Two measures of the sectoral impacts on Chinese output are reported in Table 4. We classify sectoral impacts into four groups by the magnitude of the impact. The groups for the first measure (the changes in the sum of U.S. final demand and intermediate demand for Chinese goods) are as follows: (i) The high-impact group (U.S. demand lowered by more than $5 billion) includes the Electronic and ICT industry and the Electrical industry, whose U.S. demands are lowered by $23 and $9.5 billion, respectively. (ii) The sizable-impact group (U.S. demand lowered by more than $1 billion but less than $5 billion) includes the Metal Products; Furniture; Textiles; Paper; Chemicals; and Petroleum industries, ordered by size from high to low. (iii) The moderate-impact group (U.S. demand lowered by more than $500 million but less than $1 billion) includes the Basic Metals; Wood; Motor Vehicles; and Food industries, ordered by size from high to low. 32 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Table 5 Sectoral Impacts on U.S. Output WIOD sector ΔXUj (US$ millions) ΔXUj /XUj (‰) A01-03: Primary –46.41 –0.095 C10-12: Food and Tobacco –22.45 –0.023 C13-15: Textile –4.89 –0.052 C16: Wood –9.44 –0.096 C17: Paper –22.60 –0.117 –2.70 –0.032 C19: Petroleum –37.69 –0.046 C20: Chemicals –97.24 –0.163 C21: Pharmaceutical –15.34 –0.072 C22: Plastic and Rubber –17.95 –0.078 –7.28 –0.063 –36.28 –0.129 C18: Printing and Media C23: Non-metallic Mineral C24: Basic Metals C25: Metal Products –27.19 –0.072 –137.27 –0.355 C27: Electrical Equipment –22.08 –0.176 C28: Machinery –31.31 –0.077 C29: Motor Vehicles –10.82 –0.018 C30: Other Transport –12.33 –0.036 C31-32: Furniture –10.61 –0.044 –0.25 –0.007 –572.13 –0.085 C26: Electronic and ICT C33: Repair and Installation Total NOTE: ΔXUj , derived according to (21), measures the impacts on the sectoral output in the United States through the inter national input-output linkage when the United States raises import tariffs against China. XUj is directly obtained from WIOT and corresponds to sectoral gross output in the base year. (iv) The low-impact group (U.S. demand lowered by less than $500 million) includes all other industries: Machinery; Primary; Plastic and Rubber; Pharmaceutical; Mineral; Printing and Media; Other Transport; and Repair and Installation. The groups for the second measure (the changes in sectoral output in China resulting from the changes in U.S. demand for Chinese goods, through international input-output linkages) are as follows: (i) The high-impact group (Chinese sectoral output lowered by more than $5 billion) includes the Electronic and ICT; Electrical; Basic Metals; Textiles; and Chemicals industries, whose Chinese output is lowered by $33 billion, $12 billion, $9 billion, $7 billion, and $6 billion, respectively. (ii) The sizable-impact group (Chinese sectoral output lowered by more than $3 billion but less than $5 billion) includes the Metal Products; Furniture; Primary; and Petroleum industries, ordered by size from high to low. 33 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Table 6 U.S. Deadweight Losses WIOD sector Measure 1 (US$ millions) Measure 2 (US$ millions) Measure 3 (US$ millions) A01-03: Primary 11.62 5.12 5.12 C10-12: Food and Tobacco 39.51 35.17 34.80 C13-15: Textile 16.00 15.43 15.41 C16: Wood 26.02 13.08 11.05 C17: Paper 45.30 15.20 14.60 C18: Printing and Media 3.04 2.45 2.09 C19: Petroleum 46.12 20.62 19.81 C20: Chemicals 46.56 8.76 7.74 C21: Pharmaceutical C22: Plastic and Rubber C23: Non-metallic Mineral C24: Basic Metals 8.77 5.75 5.55 21.37 12.43 12.24 6.26 3.23 2.97 52.14 23.77 19.30 C25: Metal Products 169.38 93.42 54.27 C26: Electronic and ICT 522.47 428.28 92.18 C27: Electrical Equipment 206.16 159.65 73.83 C28: Machinery 35.12 24.22 2.08 C29: Motor Vehicles 60.08 23.93 11.15 C30: Other Transport C31-32: Furniture C33: Repair and Installation Total 3.56 2.65 1.83 135.94 131.61 119.28 0.00 0.00 0.00 1,455.42 1,024.77 505.29 NOTE: Measure 1 of the DWL is our benchmark, which considers both final demand and intermediate demand. Measure 2 of the DWL considers only final demand. Measure 3 of the DWL considers only private final consumption demand. (iii) The moderate-impact group (Chinese sectoral output lowered by more than $1 billion but less than $3 billion) includes the Food; Plastic and Rubber; Machinery; Mineral; Wood; Paper; and Motor Vehicles industries, ordered by size from high to low. (iv) The low-impact group (Chinese sectoral output lowered by less than $1 billion) includes all other industries: Pharmaceuticals; Printing and Media; Other Transport; and Repair and Installation. Notably, the aggregate impact on Chinese output is –$98.5 billion, while the measured changes in U.S. demand for Chinese goods are –$39 billion for final demand and –$13 billion for intermediate demand. These findings suggest that, as U.S. import demand for Chinese goods is depressed by higher tariffs, the resulting impact on Chinese output is amplified when the international input-output linkage is considered. Similarly, we can classify the U.S. sectoral deadweight losses, as reported in Table 6, into the four groups by size: 34 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Table 7 U.S. Deadweight Losses, Tariff Revenue, and Total Losses DWLUCj (F & Z) (US$ millions) TRevUCj (US$ millions) TLossUCj (US$ millions) FLossUCj (US$ millions) A01-03: Primary 11.62 27.09 17.04 38.71 WIOD sector C10-12: Food and Tobacco 39.51 331.15 105.74 370.67 C13-15: Textile 16.00 431.60 102.32 447.60 C16: Wood 26.02 51.02 36.22 77.04 C17: Paper 45.30 51.88 55.68 97.18 C18: Printing and Media 3.04 3.48 3.74 6.52 C19: Petroleum 46.12 0.79 46.28 46.91 C20: Chemicals 46.56 561.64 158.89 608.21 C21: Pharmaceutical C22: Plastic and Rubber C23: Non-metallic Mineral C24: Basic Metals 8.77 105.74 29.92 114.51 21.37 394.79 100.33 416.17 6.26 117.05 29.67 123.30 52.14 378.19 127.78 430.33 C25: Metal Products 169.38 497.33 268.85 666.72 C26: Electronic and ICT 522.47 2,575.14 1,037.50 3,097.61 C27: Electrical Equipment 206.16 612.25 328.61 818.42 C28: Machinery 35.12 1,033.87 241.89 1,068.99 C29: Motor Vehicles 60.08 761.87 212.45 821.94 C30: Other Transport C31-32: Furniture C33: Repair and Installation Total 3.56 225.38 48.64 228.94 135.94 1,044.79 344.90 1,180.73 0.00 0.00 0.00 0.00 1,455.42 9,205.06 3,296.43 10,660.48 NOTE: The deadweight losses are measured in dollar terms (numbers are reported in million US$). DWLUCj reports our benchmark measures of the deadweight loss. TRevU Cj , derived from (3), estimates the incremental tariff revenue due to the U.S. import tariff rise. Total losses TLossUCj are the sum of the benchmark measure of deadweight losses and λ times the incremental tariff revenue. The full leakage losses FLossUCj are the sum of DWLUCj and TRevU Cj . (i) The high-impact group (sectoral deadweight losses of more than $100 million) includes the Electronic and ICT; Electrical; Metal Products; and Furniture industries, whose benchmark deadweight losses are $522 million, $206 million, $169 million, and $136 million, respectively. (ii) The sizable-impact group (sectoral deadweight losses of more than $40 million but less than $100 million) includes the Motor Vehicles; Basic Metals; Chemicals; Petroleum; and Paper industries, ordered by the size from high to low. (iii) The moderate-impact group (sectoral deadweight losses of more than $20 million but less than $40 million) includes the Food; Machinery; Wood; and Plastic and Rubber industries, ordered by the size from high to low. 35 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang (iv) The low-impact group (sectoral deadweight losses of less than $20 million) includes all other industries: Textiles; Primary; Pharmaceutical; Mineral; Other Transport; Printing and Media; and Repair and Installation. It is worth noting that, as reported in Table 5, the trade war impacts the output of the U.S. Electronic and ICT industry and Chemicals industry the most, $137 million and $97 million, respectively. These industries also bear the highest measured deadweight losses (Electronic and ICT in the high-impact group and Chemicals in the sizable-impact group). Likewise, most industries also sort into comparable groups in the sectoral output impact measure and the deadweight losses measure. However, it is interesting to note that although the Electrical; Metal Products; and Furniture industries only face moderate-to-low sectoral output impacts ($27 million, $22 million, and $11 million, respectively), they fall into the high-impact group for deadweight losses—because of differences in their demand responses and supply responses under heterogeneous tariffs and differences in their positions in the global value chain upstream and downstream. The Electrical industry has high trade elasticity (highly responsive demand), whereas the Metal Products industry faces high tariff coverage (high TCR); both are highly China dependent, particularly for their downstream products (high FDR). Similarly, the impact on the output of the Motor Vehicles industry is low ($11 million), but that industry encounters sizable deadweight losses. Conversely, the impact on the output of the Primary industry is sizable ($46 million), but the industry is classified in the low- impact group for deadweight losses. Aggregate deadweight losses are $1.5 billion in our benchmark measure, which considers both intermediate demand and final demand. When only considering final demand or private final consumption demand, the aggregate deadweight losses are $1 billion and $0.5 billion, respectively. Finally, as reported in Table 7, the measured total losses, which sum the benchmark deadweight losses and the estimated incremental tariff revenue that is not redistributed back to consumers or importers, are an aggregate $3.3 billion, around 0.05 percent of gross U.S. output. Our results can also be compared with the estimates in Amiti, Redding, and Weinstein (2019), where for 2018 the yearly cumulative deadweight losses are estimated (at 2018 prices) to be $8.2 billion and the tariff revenue is estimated to be $15.6 billion, for total losses of $23.8 billion. The main takeaways of the quantitative analysis are summarized as follows. First, because of imbalanced trade, a trade war’s sectoral impacts on Chinese sectors are generally larger than those on U.S. sectors, with Chinese output dropping by $98.5 billion and U.S. demand for Chinese goods dropping by $39 billion for final demand and by $13 billion for intermediate demand. Second, the Electronic and ICT industry and the Electrical industry encounter the largest drops in demand for Chinese goods, with their total demands dropping by $23 billion and $9.5 billion, respectively. Third, based on our benchmark measure, the aggregate U.S. deadweight losses are $1.5 billion. Among all sectors, the Electronic and ICT; Electrical; Metal Products; and Furniture industries suffer the greatest losses, $522 million, $206 million, $169 million, and $136 million, respectively. Fourth, with a leakage rate of 20 percent, total losses to U.S. consumers and importers are $3.3 billion, about 0.05 percent of gross output and two-thirds as much as the annual welfare gains from NAFTA. The full leakage losses are $10.7 billion, or 0.16 percent of gross output, which is twice as much as the annual welfare gains from NAFTA. 36 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang 5 CONCLUSIONS In this article, we have established an accounting framework that assesses the sectoral impacts of the U.S.-China trade war with regards to output, demand, deadweight losses, total losses, and full leakage losses. We have found nonnegligible detrimental effects for the United States, with total losses and full leakage losses about two-thirds and twice as much as the annual welfare gains from NAFTA, respectively. While the takeaway message is clear—initiating a trade war is costly to the domestic economy—it would also be valuable to evaluate the detrimental consequences of such a global conflict by using a broader scope. It would be particularly rewarding to build a deep structural model to quantify how the global value chain is reshuffled during a trade war, which may shed light on the longer-run impacts on international demands, pricing strategies, and hence welfare. n APPENDIX Table A1 World Input-Output Table for Two Industries and Three Countries, in Monetary Units Country 1 Country 1 Country 2 Country 3 Country 2 Country 3 Ind 1 Ind 2 Ind 1 Ind 2 Ind 1 Ind 2 FD1 FD2 FD3 Total Ind 1 2 3 1 2 0 1 2 1 0 12 Ind 2 0 1 0 1 2 0 2 0 0 6 Ind 1 1 1 2 0 1 1 0 2 0 8 Ind 2 3 0 0 2 0 0 0 3 0 8 Ind 1 1 0 1 0 3 0 0 0 3 8 Ind 2 0 0 0 1 1 2 0 1 1 6 VA 5 1 4 2 1 2 GO 12 6 8 8 8 6 NOTE: Ind, industry; FD, final demand, where FD j denotes the final demand by country j; VA, value added; and GO, gross output. Table A1 is a simple example of the data in the WIOT. Transforming the WIOT into the expression of the Leontief framework, we have X Zi F AX F , where X is gross output, Z intermediate demand, i a column vector of 1, and A the input technology matrix. The gross output matrix is 37 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang X c1 d 1 X c1 d 2 X c 2 d 1 X X c 2 d 2 X c 3 d 1 X c 3 d 2 12 6 8 , 8 8 6 the intermediate demand is zc1 d 1c1 d 1 zc1 d 2 c1 d 1 zc 2 d 1c1 d 1 Z zc 2 d 2 c1 d 1 z c 3 d 1c1 d 1 z c 3 d 2 c1 d 1 zc1 d 1c1 d 2 zc1 d 1c 2 d 1 zc1 d 1c 2 d 2 zc1 d 1c 3 d 1 zc1 d 2 c1 d 2 zc1 d 2 c 2 d 1 zc1 d 2 c 2 d 2 zc1 d 2 c 3 d 1 zc 2 d 1c1 d 2 zc 2 d 1c 2 d 1 zc 2 d 1c 2 d 2 zc 2 d 1c 3 d 1 zc 2 d 2c1 d 2 zc 2 d 2 c 2 d 1 zc 2 d 2c 2 d 2 zc 2 d 2 c 3 d 1 zc 3 d 1c1 d 2 zc 3 d 1c 2 d 1 zc 3 d 1c 2 d 2 zc 3 d 1c 3 d 1 zc 3 d 2 c1 d 2 zc 3 d 2c 2 d 1 zc 3 d 2c 2 d 2 zc 3 d 2c 3 d 1 zc1 d 1c 3 d 2 2 zc1 d 2c 3 d 2 0 zc 2 d 1c 3 d 2 1 zc 2 d 2 c 3 d 2 3 zc 3 d 1c 3 d 2 1 0 zc 3 d 2 c 3 d 2 3 1 1 0 1 0 2 0 2 1 0 2 0 2 1 0 1 0 1 , 0 0 1 0 3 0 0 0 1 1 2 and the final demand is Fcc11d 1 Fcc12d 1 Fcc13d 1 c1 c2 c3 Fc1d 2 Fc1d 2 Fc1d 2 c1 Fc 2 d 1 Fcc22d 1 Fcc23d 1 F F c1 F c 2 F c 3 c 2d 2 c 2d 2 c 2d 2 c1 c2 Fc 3d 1 Fc 3d 1 Fcc33d 1 c1 c2 c3 Fc 3d 2 Fc 3d 2 Fc 3d 2 2 1 0 3 2 0 0 2 0 2 0 2 . 0 3 0 3 0 0 3 3 0 1 1 2 zk i s j is thus ak i s j The Leontief input technology matrix specified as A X s j Ac1 d 1c1 d 1 Ac1 d 2 c1 d 1 Ac 2 d 1c1 d 1 A Ac 2 d 2 c1 d 1 A c 3 d 1c1 d 1 A c 3 d 2 c1 d 1 Ac1 d 1c1 d 2 Ac1 d 1c 2 d 1 Ac1 d 1c 2 d 2 Ac1 d 1c3 d 1 Ac1 d 2c1 d 2 Ac1 d 2 c 2 d 1 Ac1 d 2 c 2 d 2 Ac1 d 2 c 3 d1 Ac 2 d 1c1 d 2 Ac 2 d 1c 2 d 1 Ac 2 d 1c 2 d 2 Ac 2 d 1c 3 d1 Ac 2 d 2c1 d 2 Ac 2 d 2 c 2 d 1 Ac 2 d 2 c 2 d 2 Ac 2 d 2 c 3 d 1 Ac 3 d 1c1 d 2 Ac 3 d 1c 2 d 1 Ac 3 d 1c 2 d 2 Ac 3 d 1c 3 d 1 Ac 3 d 2 c1 d 2 Ac 3 d 2 c 2 d 1 Ac 3 d 2 c 2 d 2 Ac 3 d 2 c 3 d 1 38 2 12 Ac1 d 1c 3 d 2 0 Ac1 d 2 c 3 d 2 1 Ac 2 d1c 3 d 2 12 Ac 2 d 2c 3 d 2 3 Ac 3 d 1c 3 d 2 12 1 Ac 3 d 2c 3 d 2 12 0 3 6 1 6 1 6 1 8 0 0 0 1 8 0 0 0 2 8 2 8 1 8 0 2 8 0 1 8 0 2 8 1 8 0 3 8 1 8 1 6 0 1 6 . 0 0 2 6 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang NOTES 1 This assumption is possibly in effect with a quality ladder model, where sector-j goods produced by different countries are of different quality, consumers and firms treat goods of different quality as different varieties, and both the preference and investment technology are Leontief. 2 It is possible that complete pass-through may not be hold up universally in other circumstances. 3 The tariff coverage rate, taken from Bown (2019), is defined as the share of U.S. imports from China affected by U.S. special protection. 4 Caliendo and Parro (2015) estimate trade elasticities using NAFTA data. We use their estimates for the value of σ UCj . It should be noted that it implies the implicit assumption that σsjd = σsʹjd for all s ≠ sʹ. 5 We provide a simple example in the appendix to illustrate the data structure of the WIOT and how to transform the data to the Leontief framework. REFERENCES Alvarez, Fernando and Lucas, Robert E. “General Equilibrium Analysis of the Eaton-Kortum Model of International Trade.” Journal of Monetary Economics, 2007, 54(6), pp. 1726-68; https://doi.org/10.1016/j.jmoneco.2006.07.006. Amiti, Mary; Kong, Sang Hoon and Weinstein, David E. “The Effect of the U.S.-China Trade War on U.S. Investment.” NBER Working Paper 27114, National Bureau of Economic Research, 2020; https://doi.org/10.3386/w27114. Amiti, Mary; Redding, Stephen J. and Weinstein, David E. “The Impact of the 2018 Tariffs on Prices and Welfare.” Journal of Economic Perspectives, 2019, 33(4), pp. 187-210; https://doi.org/10.1257/jep.33.4.187. Arkolakis, Costas; Costinot, Arnaud and Rodríguez-Clare, Andrés. “New Trade Models, Same Old Gains?” American Economic Review, 2012, 102(1), pp. 94-130; https://doi.org/10.1257/aer.102.1.94. Bloom, Nicholas; Romer, Paul M.; Terry, Stephen J. and Van Reenen, John. “A Trapped-Factors Model of Innovation.” American Economic Review, 2013, 103(3), pp. 208-13; http://dx.doi.Org/10.1257/aer.103.3.208. Bown, Chad P. “The 2018 US-China Trade Conflict After Forty Years of Special Protection.” CEPR Discussion Paper No. 13695, Centre for Economic Policy Research, 2019; https://cepr.org/active/publications/discussion_papers/dp.php?dpno=13695. Bown, Chad P. and Crowley, Meredith A. “The Empirical Landscape of Trade Policy,” in Kyle Bagwell and Robert W. Staiger, eds., Handbook of Commercial Policy. Volume 1A. Elsevier, North Holland, 2016, pp. 3-108; https://www.sciencedirect.com/science/article/pii/S2214312216300151. Caliendo, Lorenzo and Parro, Fernando. “Estimates of the Trade and Welfare Effects of NAFTA.” Review of Economic Studies, 2015, 82(1), pp. 1-44; https://doi.org/10.1093/restud/rdu035. Cavallo, Alberto; Gopinath, Gita; Neiman, Brent and Tang, Jenny. “Tariff Pass-Through at the Border and at the Store: Evidence from US Trade Policy.” American Economic Review: Insights, 2021, 3(1), pp. 19-34; https://doi.org/10.1257/aeri.20190536. Chen, Ching-Mu; Cheng, Wan-Jung; Peng, Shin-Kun; Riezman, Raymond G. and Wang, Ping. “Trade Wars and Technology Choice.” NBER Working Paper 26468, National Bureau of Economic Research, 2020; https://doi.org/10.3386/w26468. Diamond, Peter A. and Mirrlees, James A. “Optimal Taxation and Public Production I: Production Efficiency.” American Economic Review, 1971, 61(1), pp. 8-27; https://www.jstor.org/stable/1910538. Fajgelbaum, Pablo D.; Goldberg, Pinelopi K.; Kennedy, Patrick J. and Khandelwal Amit K. “The Return to Protectionism.” Quarterly Journal of Economics, 2020, 135(1), pp. 1-55; https://doi.org/10.1093/qje/qjz036. Hsieh, Chang-Tai; Klenow, Peter J. and Nath, Ishan B. “A Global View of Creative Destruction.” NBER Working Paper No. w26461, 2019; https://doi.org/10.3386/w26461. Kennan, John and Riezman, Raymond G. “Do Big Countries Win Tariff Wars?” International Economic Review, 1988, 29(1), pp. 81-85; https://doi.org/10.2307/2526808. Kindleberger, Charles P. “Commercial Policy Between the Wars,” in Peter Mathias, and Sidney Pollard, eds., The Cambridge Economic History of Europe from the Decline of the Roman Empire. Chapter 2. Cambridge: Cambridge University Press, 1989; https://doi.org/10.1017/CHOL9780521225045.003. 39 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Cheng and Wang Ossa, Ralph. “Trade Wars and Trade Talks with Data.” American Economic Review, 2014, 104(12), pp. 4104-46; https://doi.org/10.1257/aer.104.12.4104. Owyang, Michael T.; Ramey, Valerie A. and Zubairy, Sarah. “Are Government Spending Multipliers Greater during Periods of Slack? Evidence from Twentieth-Century Historical Data.” American Economic Review, 2013, 103(3), pp. 129-134; https://doi.org/10.1257/aer.103.3.129. Perla, Jesse; Tonetti, Christopher and Waugh, Michael E. “Equilibrium Technology Diffusion, Trade, and Growth.” American Economic Review, January 2021, 111(1), pp. 73-128; https://doi.org/10.1257/aer.20151645. Santacreu, Ana Maria and Peake, Makenzie. “The Economic Effects of the 2018 U.S. Trade Policy: A State-Level Analysis.” Federal Reserve Bank of St. Louis Review, Fourth Quarter 2020, 102(4), pp. 385-412; https://doi.org/10.20955/r.102.385-412. Timmer, Marcel P.; Dietzenbacher, Eric; Los, Bart; Stehrer, Robert and de Vries, Gaaitzen J. “An Illustrated User Guide to the World Input-Output Database: the Case of Global Automotive Production.” Review of International Economics, 2015, 23(3), pp. 575-605; https://doi.org/10.1111/roie.12178. 40 The Impact of Juvenile Conviction on Human Capital and Labor Market Outcomes Limor Golan, Rong Hai, and Hayley Wabiszewski This article documents the long-term relationship among juvenile conviction, occupational choices, employment, wages, and recidivism. Using data from the National Longitudinal Survey of Youth 1997 (NLSY97), we document that youth convicted at or before age 17 have lower a full-time employment rate and lower wage growth rate even after 10 years in the labor market. Merging the NLSY97 with occupational characteristics data from the Occupational Information Network (O*NET), we show that youth with a juvenile conviction are less likely to be employed in occupations that have a high on-the-job training requirement and that these occupations have higher wages and wage growth. Accumulated occupation-specific work experience, general experience, and education are important for explaining the gaps in wage and recidivism between youth with and without a juvenile conviction. Our results highlight the important role of occupational choices as a human capital investment vehicle through which juvenile crimes have a longterm impact on wages and recidivism. (JEL K42, I24, J2, J3) Federal Reserve Bank of St. Louis Review, First Quarter 2022, 104(1), pp. 41-69. https://doi.org/10.20955/r.104.41-69 1 INTRODUCTION In this article, we document the empirical relationship among juvenile conviction, education, adult labor market occupational choices, employment, wages, and recidivism. Although several studies have shown that juvenile adjudication is associated with lower formal educational attainment and an increased likelihood of dropping out of high school, no existing study examines human capital accumulation through on-the-job (OTJ) training. Our data are from the National Longitudinal Survey of Youth 1997 (NLSY97) and the Occupa tional Information Network (O*NET). NLSY97 is a longitudinal survey that follows the lives of a sample of American youth born between 1980-84. It provides detailed information on each individual’s convictions and incarcerations over time as well as the age and date of the first time the individual had an interaction with the correctional system. It also collects detailed information on each individual’s history of employment, occupations, and wages. Finally, it has information on each Limor Golan is a professor at Washington University in St. Louis and a research fellow at the Federal Reserve Bank of St. Louis. Rong Hai is an assistant professor at the University of Miami. Hayley Wabiszewski is a PhD candidate at Washington University in St. Louis. © 2022, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. 41 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski individual’s education, age, gender, race, and measures of cognitive ability. Also, the O*NET data survey provides detailed information on job requirements for and the characteristics of each occupation. Matching the O*NET job requirement data to the NLSY97 data, we are able to analyze the impact of a juvenile conviction on occupational choices, which have long-term consequences on wages. Using the NLSY97 and O*NET data, we first document that youth with juvenile adjudications have worse educational outcomes. We then show that juvenile convictions are associated with a lower full-time employment rate, even after controlling for ability, education, and general work experience. We also find that individuals who had a juvenile adjudication are less likely to be employed in occupations with high-OTJ-training requirements. We also show that the wage gap between youth with and without a juvenile conviction can be explained by the differences in accumulated occupation-specific work experience, general work experience, education, and ability. We analyze wage growth over a 10-year period of employment and find that a juvenile conviction still reduces the wage growth rates even after controlling for education and occupation-specific work experience. Finally, we document a juvenile conviction to be a strong predictor of the likelihood of adult incarceration. We do find race and gender differences both in the effects of a juvenile conviction and in the above outcomes. For women, having a juvenile conviction does not have a statistically significant effect on the overall employment probability, but it reduces the probability of full-time employment. For men, having a juvenile conviction reduces both the overall employment probability and the full-time employment probability. In addition, male Black workers are less likely to be employed in all specifications; this finding is consistent with findings in the literature (see Ritter and Taylor, 2011). Moreover, Black workers are less likely to be employed in occupations with high training requirements, even after controlling for test scores, education, and experience. This finding is consistent with the findings in Golan, James, and Sanders (2019). Among women, however, once we control for test scores, the coefficient on the Black race dummy becomes statistically insignificant. This finding is true for wages as well. These findings are consistent with the differences in labor force participation of and selection into the labor market for Black and White women (see Neal, 2004). For Hispanic men and women, the negative effects on outcomes either lose statistical significance or become positive once we account for the differences in test scores. While our results are suggestive regarding the effects of a juvenile conviction on education, employment, occupational choices, and recidivism, it highlights the rich dynamic relationship among youth crime and labor market choices and outcomes (including occupational choices). We argue that a juvenile conviction reduces the probability of future employment in occupations with high-OTJ-training requirements and that this is an important channel through which youth crime interacts with labor market outcomes. Specifically, this channel helps to generate a long-term impact of youth crime on labor market outcomes and it also acts as a cost that affects a youth’s decision to commit crime ex ante in a forward-looking model with crimes. Our article relates to three strands of the literature. First, our article contributes to the literature on juvenile crime and human capital investment. These existing studies primarily focus on schooling as the measure of human capital investment and find that (i) juvenile arrest/adjudication reduces schooling and (ii) school enrollment reduces future crimes. Regarding (i), Kirk and Sampson (2013) and Aizer and Doyle (2015) both find that juvenile arrest, adjudication, or incarceration reduces the probability of high school graduation. Kirk and Sampson (2013) further show that juvenile arrest 42 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski reduces the likelihood of four-year college enrollment conditional on high school graduation. Litwok (2015) supports this result, finding that automatic expungement of juvenile conviction records—unconditionally—increases the probability of college attendance and graduation. Evidence for (ii) in Lochner (2004) shows that high school graduates are less likely than high school dropouts to be incarcerated in their 20s. Similarly, Merlo and Wolpin (2015) find that attending school at age 16 reduces the probability of committing a crime at age 19. Second, our article relates to the literature on juvenile crime and labor market outcomes (see Western, Kling, and Weiman, 2001, for a survey).1 Litwok (2015) shows that automatic expungement of juvenile criminal records increases an individual’s average income in their late 20s. Imai and Krishna (2004) estimate a dynamic discrete choice model of criminal behavior where forward- looking youth make decisions about whether to commit a crime. The authors show that policies that reduce future labor market punishment for committing a crime lead youth to commit more crime ex ante. Nagin and Waldfogel (1995) look at the impact of conviction at ages 17 and 18 on labor market outcomes at age 19 of young British offenders and find mixed results. They find that conviction status decreases job stability, via more weeks unemployed, a decrease in job duration, and an increase in the number of jobs ever held, but increases weekly earning. Western and Beckett (1999) analyze youth incarceration between the ages of 15 and 22 and its impact on future employment using the NLSY79, finding a long-lasting decrease in employment that does not decay with time. Using NLSY97 data, Apel and Sweeten (2010) find that youth incarceration has a persistent negative impact on formal employment, driven mostly by an increased probability and duration of labor force non-participation. They find that incarceration reduces annual income and that this income gap widens over time. Third, our article also relates to the literature that investigates the relationship between juvenile crime and future recidivism. This literature is vast, especially in criminology. Nagin and Paternoster (1991), Nagin and Land (1993), and Nagin, Farrington, and Moffitt (1995) evaluate the change in criminal behavior over the life cycle and find that participating in crime early in the life cycle increases the likelihood of participating in crime in the future as social and professional relationships deteriorate. Paternoster, Brame, and Farrington (2001) find some evidence that variation in the propensity to commit crimes as an adult can be attributed to differences in individual criminal behavior established during adolescence as opposed to processes that occur during adulthood. Several studies in economics also evaluate this relationship. Levitt (1998) shows deterrence is empirically more important than incapacitation in reducing crime, particularly in the case of property crimes. Aizer and Doyle (2015) find that individuals on the margin of juvenile incarceration who are incarcerated are significantly more likely to recidivate as adults, especially for serious crimes, relative to those who are not incarcerated. Indeed, Bayer, Hjalmarsson, and Pozen (2009) explore the peer effects of juvenile incarceration on juvenile recidivism and find that there are significant peer effects that increase the probability of recidivism for crimes in which an individual already has experience. 2 DATA AND BASIC ANALYSIS 2.1 Data The data are compiled from NLSY97 and O*NET. NLSY97 is a longitudinal survey that follows the lives of a sample of American youth born between 1980-84. 43 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski NLSY97 collects information on each individual’s criminal behavior, arrests, convictions, and incarceration in each survey round; it also has information regarding the age and date of the first time the individual interacted with the correctional system. We construct an indicator variable of juvenile convictions that equals 1 if and only if the individual was convicted at least once before age 18. NLSY97 also asks individuals to report their monthly incarceration status. We define an individual as incarcerated in the corresponding year if the individual was incarcerated at least one month of the year. We obtain a complete weekly history for each individual on their employment status and the number of jobs worked. We also have complete weekly information on the occupation, hours worked, and hourly wage for every job the individual worked, as well as the job starting date and job ending date, over the period 1997 to 2013. We aggregate the weekly information into annual information. We define an individual as employed if their average weekly hours worked in the reference year is more than or equal to 10 hours; we define an individual as full-time employed if their average weekly hours worked is more than or equal to 30 hours. The number of years an individual has worked can be observed from the start and end dates of each job. An individual’s main job in the reference year is characterized by the occupation the individual worked in for the most hours in the year. We focus on the hourly wage of the individual’s main job. In addition to hourly wage levels, we also look at wage growth over 10 years; this variable is constructed using the percentage change in an individual’s hourly wage in their 11th working year (when their number of years worked is 10) versus their 1st working year (when their number of years worked is 0). All wage data are in 2000 dollars. NLSY97 also provides us information on an individual’s age, race, education, year of graduation, and Armed Services Vocational Aptitude Battery (ASVAB) test score. We classify educational attainment into five categories: high school dropout (i.e., no high school diploma and no General Educational Development certificate [GED]), GED, high school graduate, associate college degree, and four-year college degree or higher. We merge NLSY97 data with O*NET data based on each employed individual’s occupation. The O*NET data include detailed job requirements and characteristics for each occupation. Specifi cally, O*NET asks questions regarding the amount of OTJ training required to perform the job. OTJ training includes apprenticeships, internships, and other supervised experiences. For each occupation, we obtain an OTJ-training intensity variable that documents the percentage of the jobs in that occupation that require more than one month of OTJ training, ranging from 0 percent to 100 percent. We merge this variable with each individual’s occupation in the NLSY97 data. Average OTJ-training intensity increases with a worker’s education level. In particular, in our final sample, average OTJtraining intensity is 69 percent among college graduates with a four-year degree or higher, indicating that these college graduates are employed in occupations where 69 percent of the jobs require at least one month of OTJ training. The average OTJ-training intensities are 65 percent, 58 percent, 57 percent, and 54 percent for individuals with an associate degree, with a high school diploma, with a GED, or that are a high school dropout, respectively. We say an individual is employed in a high-OTJ-training occupation if the percentage of jobs in that occupation that require one month of OTJ training is higher than the sample median level (i.e., 62 percent in our sample). Examples of low-OTJ-training occupations include parking lot attendants, dining room and cafeteria attendants, and bartender helpers. Examples of high-OTJ-training occupations include mechanics, installers, repairers, electricians, and first-line supervisors/managers of construction and production 44 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 1 Key Variables by Gender Men Women Juvenile conviction 0.079 0.038 Less than a high school diploma/GED 0.088 0.076 GED 0.120 0.087 High school graduate 0.552 0.498 Associate college degree 0.052 0.069 Four-year college degree or higher 0.188 0.270 Employment 0.869 0.786 Full-time employment (among employed workers) 0.668 0.589 Employed in high-OTJ-training occupation 0.555 0.431 Years worked 4.663 4.160 Years worked in high-OTJ-training occupation 2.657 1.873 Hourly wage ($) 13.059 11.488 Change in wage over 10 years worked (%) 88.504 78.227 Incarceration 0.027 0.005 Race = Black 0.153 0.155 Race = Hispanic 0.124 0.116 25.575 25.837 0.470 0.484 Age ASVAB score SOURCE: NLSY97. workers. High-OTJ-training occupations allow for more human capital accumulation post school and have higher wage growth. The original NLSY97 data have 8,984 respondents. We drop the observations with no information on the highest degree of education or the year when the individual left school. We also drop observations with missing juvenile conviction, ASVAB, or race information. We only keep observations starting from the year the individual enters the labor market (after obtaining the highest degree of education). 2.2 Summary Statistics In this section, we discuss summary statistics of our data. Because men and women have very different patterns of human capital accumulation and crime behaviors over the life cycle, we present summary statistics as well as analysis results for men and women separately. Table 1 presents the mean values of key variables used in our analysis. In our sample, the average juvenile conviction rate is 7.9 percent for men and 3.8 percent for woman. In terms of educational outcomes for men, 8.8 percent are high school dropouts, 12 percent have a GED, 55.2 percent are high school graduates, 5.2 percent have an associate college degree, and the remaining 18.8 percent have a four-year college degree or higher. In terms of educational outcomes for women, 7.6 percent 45 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Figure 1 Educational Outcomes for Those with and without a Juvenile Adjudication A. Males B. Females 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 No conviction 0 Juvenile conviction <HSG HSG CLG No conviction GED SCL Juvenile conviction NOTE: <HSG, less than a high school diploma and no GED; HSG, high school graduate; CLG, four-year college degree or higher; and SCL, associate college degree. SOURCE: NLSY97. are high school dropouts, 8.7 percent have a GED, 49.8 percent are high school graduates, 6.9 percent have an associate college degree, and the remaining 27 percent have a four-year college degree or higher. Figure 1 plots the distribution of educational outcomes by juvenile conviction status for men and women separately. Compared with men without a juvenile conviction, men with a juvenile conviction are more likely to drop out of high school or have a GED and are less likely to have a four-year college degree or higher. Similar patterns hold for women. Next, we discuss the outcome variables on labor market employment, occupation, and wages. For men, the employment rate is 86.9 percent and the full-time employment rate is slightly lower, 66.8 percent; the average number of years worked in our sample period is 4.66 years; and 55.5 percent worked in high-OTJ-training occupations, with an average of 2.66 years worked in those occupations. For women, the employment rate is 78.6 percent and the full-time employment rate is 58.9 percent; the average number of years worked in our sample period is 4.16 years; and only 43.1 percent worked in high-OTJ-training occupations, with an average of 1.87 years worked in those occupations. Figure 2 plots the employment rate by age for individuals with and without a juvenile conviction for men and women seperately. There is a large employment gap among men based on their juvenile conviction status, and the gap increases over age as the employment patterns for men with and without juvenile convictions diverge over time. The employment gap also exists for women, but the size of the gap is much smaller and the time trend is less clear than they are for men. For employed men, the average hourly wage is $13.06 (in 2000 dollars) and the average wage growth rate over 10 years worked is 88.5 percent. For employed women, the average hourly wage is $11.49 and the average wage growth rate over 10 years worked is 78 percent. Figure 3 plots the average hourly wages by age based on juvenile conviction status for men and women separately. Similar to employment patterns seen in Figure 2, Figure 3A shows that men have a clear widening of the wage gap based on conviction status. The gap starts out relatively small at age 19: Men with a juvenile conviction earn an average hourly wage of $7.85 compared with $8.63 for those without 46 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Figure 2 Employment Rates for Those with and without a Juvenile Adjudication A. Males B. Females Probability of employment 1.0 Probability of employment 1.0 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 Juvenile conviction = 0 Juvenile conviction = 1 0.4 19 21 23 25 Age 27 Juvenile conviction = 0 Juvenile conviction = 1 0.4 29 19 21 23 25 Age 27 29 SOURCE: NLSY97. Figure 3 Wages for Those with and without a Juvenile Adjudication A. Males B. Females Hourly wage 20 Hourly wage 20 15 15 10 10 Juvenile conviction = 0 Juvenile conviction = 1 5 19 21 23 25 Age 27 Juvenile conviction = 0 Juvenile conviction = 1 5 19 29 SOURCE: NLSY97. 47 21 23 25 Age 27 29 19 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 2 Key Variables by Occupation Categories Low OTJT High OTJT Hourly wage ($) 10.071 14.161 Change in wage over 10 years worked (%) 59.709 101.298 Less than a high school diploma/GED 0.091 0.042 GED 0.116 0.078 High school diploma 0.585 0.473 Associate college degree 0.050 0.074 Four-year college degree or higher 0.158 0.333 ASVAB score 0.443 0.543 NOTE: OTJT, on-the-job training. Figure 4 Adulthood Incarceration Rates for Those with and without a Juvenile Adjudication A. Males B. Females Probability of incarceration 0.20 Probability of incarceration 0.20 Juvenile conviction = 0 Juvenile conviction = 1 0.16 Juvenile conviction = 0 Juvenile conviction = 1 0.16 0.12 0.12 0.08 0.08 0.04 0.04 0 0 19 21 23 25 Age 27 29 19 21 23 25 Age 27 SOURCE: NLSY97. a juvenile conviction, for a wage gap of $0.78. By age 30, the gap increases over five-fold, to $3.91, with average hourly wages of $12.30 and $16.21, respectively. The hourly wage gap for women has a similar trend: It increases from $0.39 at age 19 to $3.27 by age 30. Table 2 reports the average worker characteristics for high- and low-OTJ-training occupations separately. Both hourly wages and the wage growth rate are lower in low-OTJ-training occupations than in high-OTJ-training occupations. Average hourly wages are $10.07 in low-OTJ-training occupations and $14.16 in high-OTJ-training occupations. The average 10-year wage growth rates 48 29 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski are 60 percent and 101 percent in low- and high-OTJ-training occupations, respectively. Regarding the education levels of workers in low-OTJ-training occupations, 9.1 percent are high school dropouts, 11.6 percent have a GED, 58.5 percent are high school graduates, 5 percent have an associate college degree, and the remaining 15.8 percent have a four-year college degree or higher. Of the workers in high-OTJ-training occupations, 4.2 percent are high school dropouts, 7.8 percent have a GED, 47.3 percent are high school graduates, 7.4 percent have an associate college degree, and the remaining 33.3 percent have a four-year college degree or higher. The average ASVAB score is 0.443 in low-OTJ-training occupations and 0.543 in high-OTJ-training occupations. Finally, we discuss the adulthood crime outcomes measured by an incarceration indicator variable. As shown in Table 1, the average incarceration rates are 2.7 percent and 0.5 percent for men and women, respectively, and the average age of men and women is about 26 over our sample period. Figure 4 presents the probability of incarceration by juvenile conviction status by age for men and women separately. Individuals with a juvenile conviction are more likely to be incarcerated during adulthood than individuals without a juvenile conviction. This is suggestive evidence that a juvenile conviction predicts adult recidivism. The effects of a juvenile conviction on adulthood incarceration exist both for men and women, but the magnitude is larger for men. 3 CONCEPTUAL FRAMEWORK AND EMPIRICAL STRATEGY 3.1 Conceptual Framework Our conceptual framework is a dynamic model of human capital accumulation and crime behaviors. Heterogeneous individuals are forward-looking and make decisions on schooling, employment, occupational choices, and crime behaviors. Firms are also forward-looking and make decisions on occupation-specific job offers that differ in wages and OTJ-training requirements based on observed worker characteristics (also see the framework in Gayle and Golan, 2012). High-OTJtraining occupations incur higher training costs presently but also have higher future productivity growth if the employment relationship continues. Human capital investment in such a framework takes the form of education and post-school occupation-specific OTJ-training investment. Finally, the model allows for past choices as well as the returns to accumulated human capital and criminal capital to affect current choices (Merlo and Wolpin, 2015, and Mancino, Navarro, and Rivers, 2016). In this model, a juvenile conviction can have a long-lasting impact on a youth’s human capital accumulation, labor market outcomes, and future criminal activity through three potential channels. The first channel is through changing an individual’s schooling by increasing the psychic costs of schooling and reducing college admission probabilities. A juvenile conviction affects a youth’s psychic cost of schooling because interactions with the juvenile justice system can disrupt the youth’s schooling activities (Kirk and Sampson, 2013; Aizer and Doyle, 2015; and Litwok, 2015) and harm the youth’s mental health (Kashani et al., 1980, and Forrest et al., 2000). Youth with a record may find it hard to be re-enrolled if schooling is disrupted. College admission probabilities may also be reduced because schools may be unwilling to admit juveniles with a conviction record. The second channel is through changing a youth’s post-school human capital accumulation, as measured by accumulated work experience in occupations with different OTJ-training requirements. On one hand, youth with a juvenile record may have less incentive to invest in post-school human capital accumulation because juvenile corrections encourage the accumulation of “criminal 49 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski capital” (Bayer, Hjalmarsson, and Pozem, 2009) and reduce the psychic cost of future incarceration. On the other hand, a potential employer with asymmetric information may be less likely to offer them jobs and less likely to offer them jobs with higher OTJ-training requirements, because the employer anticipates these individuals have higher probabilities of quitting the job and committing crimes in the future. The last and third channel through which a juvenile conviction has a long-lasting effect is the dynamic interaction between human capital investment and crime behaviors over time. As noted with the previous two channels, individuals with a juvenile record are likely to have lower human capital investment and hence lower future wages and wage growth. Lower wages and wage growth increase the likelihood of future crime activities by reducing the opportunity cost of going to jail and leaving the labor force. As a result, youth with a juvenile record have a higher probability of recidivism, which, in turn, reduces their ex ante incentive to invest in human capital and decreases firms’ ex ante willingness to offer them good jobs. The state dependence between past behaviors and current choices further reinforces such interactions. These dynamic interactions between human capital investment and crime behavior exacerbate the negative impact of a juvenile conviction over time. Another factor that affects the correlations between a juvenile record and the outcomes discussed above is unobserved heterogeneity that affects both the likelihood of having a criminal record early on and the likelihood of high educational attainment, stability in the labor market, selection into high OTJ-training occupations, and high wages. However, in the empirical analysis below we are unable to quantify separately unobserved heterogeneity and state dependence. 3.2 Empirical Strategy The preceding figures show considerable differences in educational attainment, employment, wages, and adult incarceration between individuals with and without a juvenile conviction. These differences are representative of an agglomeration of observable and unobservable differences between the two groups aside from juvenile conviction status. To begin sorting out the effects of the differences in the compositions of the two groups and the effect of juvenile conviction status, we next present regression results for outcome variables of interest, including educational outcomes, employment, occupational choices, wages, and adult incarceration. As evidenced by Figures 1 to 4, age is an important determinant of these outcomes. There are also important differences in the racial makeup of each group. Other important omitted variables that are correlated with both juvenile conviction and the outcome variables of interest include educational attainment, individual ability or skills, and job market experience. For example, if individuals without juvenile adjudication have better skills and more education on average, both of which make them more employable, then the coefficient on juvenile adjudication will be biased downward. It is therefore important to control for these confounding variables when analyzing the effect of juvenile conviction status on the relevant outcome variables. In this section we use the following regression model to conduct our analysis: yi,t = β ∙ JuvenileConvictioni + Xi,t γ + єi,t , where yi,t is the outcome variables of interests, including educational outcomes, employment, occupational choices, wages, and adulthood incarceration. JuvenileConviction is an indicator variable that equals 1 if the youth had been convicted at least once before age 18. Xi,t is a vector that includes 50 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 3 Table 4 The Effects of a Juvenile Conviction on Male Educational Outcomes: Less Than a High School Diploma or GED The Effects of a Juvenile Conviction on Male Educational Outcomes: Four-Year College Degree or Higher (1) (2) (1) (2) Juvenile conviction 0.069*** (0.0124) 0.033*** (0.0086) Juvenile conviction –0.360*** (0.0634) –0.272*** (0.0557) Race = Black 0.047*** (0.0090) –0.009 (0.0069) Race = Black –0.217*** (0.0278) 0.013 (0.0286) Race = Hispanic 0.049*** (0.0094) 0.005 (0.0070) Race = Hispanic –0.165*** (0.0291) –0.033 (0.0297) ASVAB score –0.246*** (0.0512) ASVAB score 0.742*** (0.1446) ASVAB score squared 0.101* (0.0532) ASVAB score squared 0.012 (0.1316) Mean value 0.091 0.091 Mean value 0.239 0.239 Observations 2,646 2,646 Observations 2,646 2,646 Pseudo R 2 0.040 0.181 Pseudo R 2 0.047 0.241 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. individual variables such as race, education, and work experience, and єi,t is an error term. We use the logit model when the dependent variable is a dummy variable (including employment, occupational choices, and educational outcomes), and we report the marginal effects associated with each regression variable. We perform OLS analysis when the outcome variables are continuous variables such as log wages and wage growth. 4 RESULTS 4.1 Juvenile Conviction and Educational Outcomes In this section, we investigate the effects of a juvenile conviction on educational outcomes. Educational outcomes are central to accumulation of human capital after individuals complete their education because it is a strong determinant of labor market attachment, occupational sorting, and earnings growth. Consistent with the existing literature, we also find that a juvenile conviction increases the probability of dropping out of high school and decreases the probability of having a four-year college degree or higher. Starting with the results for men, Column 1 of Table 3 shows that men with a juvenile conviction are 6.9 percentage points more likely to drop out of high school. Race also affects the probability of dropping out of high school, with Black men 4.7 percentage points and Hispanic men 4.9 percentage points more likely to drop out. Once we control for the ability measures (Column 2), the effect of a juvenile conviction on men is reduced to 3.3 percentage points and the effects of the two race dummies lose statistical significance. Column 1 of Table 4 shows that a juvenile conviction reduces the 51 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 5 Table 6 The Effects of a Juvenile Conviction on Female Educational Outcomes: Less Than a High School Diploma or GED The Effects of a Juvenile Conviction on Female Educational Outcomes: Four-Year College Degree or Higher (1) (2) (1) (2) Juvenile conviction 0.064*** (0.0148) 0.014** (0.0066) Juvenile conviction –0.418*** (0.0929) –0.341*** (0.0931) Race = Black 0.029*** (0.0084) –0.014*** (0.0043) Race = Black –0.213*** (0.0261) 0.020 (0.0298) Race = Hispanic 0.031*** (0.0090) –0.005 (0.0036) Race = Hispanic –0.248*** (0.0308) –0.100*** (0.0345) ASVAB score –0.146*** (0.0384) ASVAB score 1.136*** (0.1754) ASVAB score squared 0.047 (0.0320) ASVAB score squared –0.166 (0.1601) Mean value 0.067 0.067 Mean value 0.324 0.324 Observations 2,781 2,781 Observations 2,781 2,781 Pseudo R 2 0.023 0.271 Pseudo R 2 0.039 0.203 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. probability of men obtaining a four-year college degree or higher by 36 percentage points. Once we control for the ability measures (Column 2), a juvenile conviction leads to a 27.2- percentage- point reduction in the probability of men graduating with a four-year college degree or higher and both race dummies lose statistical significance. Tables 5 and 6 repeat the analysis for women. After controlling for the ability measures, for women, a juvenile conviction increases the probability of dropping out of high school by 1.4 percentage points (Column 2 of Table 5) and decreases the probability of graduating with a four-year college degree or higher by 34.1 percentage points (Column 2 of Table 6). Compared with the existing studies, Kirk and Sampson (2013) find a larger effect of an arrest on high school dropout probability, which is likely due to a difference in how dropout is defined: In the Kirk and Sampson study, high school dropouts include individuals who went on to obtain a GED. 4.2 Juvenile Conviction and Labor Market Employment Section 4.1 shows that individuals with a juvenile conviction have worse educational outcomes and hence may have worse labor market outcomes. In this section, we show that a juvenile record has a negative impact on labor market employment even after controlling for education and ability. Employment and attachment to the labor market are important to understanding the amount of human capital workers accumulate and their wage growth over time, all important determinants of the likelihood of recidivism and future incarceration. We find large and statistically significant negative impacts of a juvenile conviction on both employment (extensive margin) and full-time employment (intensive margin) for men. In contrast, we find that for women, a juvenile conviction does not have a statistically significant impact on 52 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 7 The Effects of a Juvenile Conviction on Adult Male Labor Market Employment (1) (2) (3) (4) (5) (6) Juvenile conviction –0.078*** (0.0120) –0.066*** (0.0117) –0.035*** (0.0106) –0.021*** (0.0075) –0.046*** (0.0135) –0.030*** (0.0100) Race = Black –0.109*** (0.0081) –0.070*** (0.0091) –0.064*** (0.0086) –0.030*** (0.0054) –0.082*** (0.0110) –0.040*** (0.0072) Race = Hispanic –0.015 (0.0107) 0.009 (0.0114) 0.011 (0.0105) 0.007 (0.0064) 0.015 (0.0136) 0.011 (0.0087) Age 0.004*** (0.0007) 0.004*** (0.0008) 0.002** (0.0007) –0.023*** (0.0013) 0.002** (0.0009) –0.032*** (0.0018) ASVAB score 0.387*** (0.0592) 0.329*** (0.0582) 0.094*** (0.0325) 0.422*** (0.0771) 0.115*** (0.0441) ASVAB score squared –0.281*** (0.0609) –0.316*** (0.0598) –0.105*** (0.0340) –0.412*** (0.0809) –0.133*** (0.0468) GED 0.017 (0.0132) 0.061*** (0.0087) 0.022 (0.0166) 0.084*** (0.0116) High school diploma 0.064*** (0.0121) 0.038*** (0.0070) 0.081*** (0.0152) 0.051*** (0.0092) Associate college degree 0.117*** (0.0252) 0.214*** (0.0205) 0.148*** (0.0316) 0.293*** (0.0272) Four-year college degree or higher 0.199*** (0.0191) 0.297*** (0.0141) Years worked 0.049*** (0.0023) 0.066*** (0.0030) Years worked squared –0.001*** (0.0002) –0.002*** (0.0002) Mean value 0.846 0.846 0.846 0.852 0.824 0.829 Observations 26,610 26,610 26,610 25,757 22,407 21,596 0.036 0.057 0.084 0.207 0.057 0.189 Pseudo R 2 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. employment (extensive margin), but it does have a statistically significant negative impact on fulltime employment probability conditional on employment (intensive margin). Table 7 examines the marginal effects of a juvenile conviction on male labor market employment using a logit regression model. Controlling for age and two race dummies, a juvenile conviction reduces the probability of employment by 7.8 percentage points for men (Column 1). Columns 2, 3, and 4 sequentially add controls for ASVAB scores, educational attainment, and labor market experience. With the test scores included (Column 2), the magnitudes of the negative effects on men become smaller: A juvenile conviction now reduces the probablity of employment by 6.6 percentage points and also reduces the coefficient on the race dummy for Black workers, but it remains statistically significant. These findings suggest a role for a selection effect of innate skills on the probability of unemployment. Adding educational attainment has a large effect on the probability of employment for men and reduces it further to 3.5 percentage points, which can be due to different traits 53 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 8 The Effects of a Juvenile Conviction on Adult Male Labor Market Full-Time Employment (Among Employed Workers) (1) (2) (3) (4) (5) (6) Juvenile conviction –0.107*** (0.0211) –0.104*** (0.0211) –0.076*** (0.0217) –0.067*** (0.0224) –0.080*** (0.0233) –0.069*** (0.0246) Race = Black –0.098*** (0.0129) –0.086*** (0.0142) –0.087*** (0.0142) –0.067*** (0.0145) –0.093*** (0.0159) –0.066*** (0.0163) Race = Hispanic –0.019 (0.0136) –0.011 (0.0143) –0.009 (0.0142) –0.003 (0.0143) –0.007 (0.0159) 0.001 (0.0161) Age 0.031*** (0.0011) 0.031*** (0.0011) 0.029*** (0.0012) –0.010*** (0.0030) 0.029*** (0.0013) –0.020*** (0.0038) ASVAB score 0.103 (0.0782) 0.075 (0.0802) –0.055 (0.0804) 0.203** (0.0940) 0.028 (0.0947) ASVAB score squared –0.062 (0.0761) –0.102 (0.0786) –0.004 (0.0787) –0.268*** (0.0972) –0.130 (0.0975) GED –0.004 (0.0240) 0.052** (0.0246) –0.005 (0.0254) 0.068** (0.0266) High school diploma 0.056*** (0.0208) 0.050** (0.0208) 0.061*** (0.0220) 0.053** (0.0221) Associate college degree 0.108*** (0.0310) 0.265*** (0.0342) 0.118*** (0.0326) 0.325*** (0.0384) Four-year college degree or higher 0.137*** (0.0263) 0.349*** (0.0316) Years worked 0.074*** (0.0044) 0.086*** (0.0056) Years worked squared –0.002*** (0.0003) –0.002*** (0.0003) Mean value 0.652 0.652 0.652 0.655 0.624 0.626 Observations 22,487 22,487 22,487 21,908 18,424 17,883 0.057 0.057 0.062 0.087 0.051 0.082 Pseudo R 2 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. and unobserved skills of workers who have higher educational attainment and also due to differences in the labor market conditions for high- and low-skilled workers. As seen in Column 4, which is the table’s most exhaustive specification and includes labor market experience in the regression, a juvenile conviction reduces the probability of employment for men by 2.1 percentage points. This is not surprising, because workers who are attached to the labor market and have more experience accumulate more human capital and are more likely to continue to be employed. Another interesting finding is that as we move from Column 1 to Column 4, the negative impact associated with the Black race dummy becomes smaller in magnitude, moving from –10.9 percentage points to –3.0 percentage points. Columns 5 and 6 focus on men without a four-year college degree or higher. This subpopulation has low skills and may be more vulnerable to the negative impact of a juvenile conviction. As expected, we find a juvenile conviction has a larger negative impact on employment 54 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 9 The Effects of a Juvenile Conviction on Adult Female Labor Market Employment (1) (2) (3) (4) (5) (6) Juvenile conviction –0.068** (0.0303) –0.037 (0.0301) 0.009 (0.0284) 0.011 (0.0195) 0.009 (0.0358) 0.012 (0.0256) Race = Black –0.079*** (0.0127) 0.014 (0.0144) –0.002 (0.0134) 0.006 (0.0095) –0.009 (0.0174) –0.001 (0.0127) Race = Hispanic –0.022 (0.0142) 0.038** (0.0151) 0.036** (0.0147) 0.025** (0.0106) 0.048** (0.0190) 0.033** (0.0142) Age –0.001 (0.0010) –0.003*** (0.0010) –0.007*** (0.0010) –0.040*** (0.0018) –0.007*** (0.0013) –0.055*** (0.0026) ASVAB score 0.613*** (0.0829) 0.378*** (0.0802) 0.088 (0.0576) 0.513*** (0.1081) 0.094 (0.0801) ASVAB score squared –0.313*** (0.0815) –0.272*** (0.0798) –0.052 (0.0614) –0.404*** (0.1137) –0.063 (0.0901) GED 0.107*** (0.0226) 0.137*** (0.0169) 0.129*** (0.0275) 0.182*** (0.0215) High school diploma 0.163*** (0.0194) 0.100*** (0.0130) 0.198*** (0.0235) 0.124*** (0.0162) Associate college degree 0.223*** (0.0331) 0.338*** (0.0334) 0.272*** (0.0404) 0.454*** (0.0455) Four-year college degree or higher 0.338*** (0.0252) 0.471*** (0.0199) Years worked 0.064*** (0.0038) 0.094*** (0.0055) Years worked squared –0.001*** (0.0003) –0.002*** (0.0004) Mean value 0.770 0.770 0.770 0.775 0.728 0.732 Observations 26,949 26,949 26,949 26,273 20,948 20,345 0.006 0.049 0.085 0.178 0.053 0.166 Pseudo R 2 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. in this population. In the table’s most exhaustive specification (Column 6), a juvenile conviction reduces the probability of employment by 3 percentage points for men without a four-year college degree or higher. In all specifications, the Black coefficient remains significant. One possible explanation for this finding is that the type of labor experience Black workers receive is different from that of White workers. We further explore this possibility below; however, this finding is consistent with labor market discrimination in hiring and differences in opportunities available to Black workers, as found in Golan, James, and Sanders (2019). Table 8 reports the marginal effects of a juvenile conviction on the probability of full-time employment for employed men. Typically, full-time employment is associated with higher accumulation of human capital and stronger attachment to the labor market. While qualitatively the results are similar to the ones in Table 7, the negative effect of a juvenile conviction is larger and not reduced by test scores, education, or experience to the same extent employment is. As seen in 55 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 10 The Effects of a Juvenile Conviction on Adult Female Labor Market Full-Time Employment (Among Employed Workers) (1) (2) (3) (4) (5) (6) Juvenile conviction –0.121*** (0.0310) –0.103*** (0.0301) –0.064** (0.0314) –0.064** (0.0310) –0.076** (0.0345) –0.082** (0.0347) Race = Black –0.073*** (0.0143) –0.024 (0.0161) –0.033** (0.0160) –0.021 (0.0158) –0.040** (0.0189) –0.027 (0.0186) Race = Hispanic –0.009 (0.0160) 0.024 (0.0170) 0.031* (0.0174) 0.035** (0.0172) 0.043** (0.0197) 0.047** (0.0196) Age 0.032*** (0.0013) 0.031*** (0.0013) 0.026*** (0.0014) 0.000 (0.0029) 0.027*** (0.0016) –0.008** (0.0037) ASVAB score 0.367*** (0.0932) 0.219** (0.0958) 0.121 (0.0971) 0.292** (0.1149) 0.144 (0.1171) ASVAB score squared –0.197** (0.0884) –0.185** (0.0896) –0.119 (0.0908) –0.272** (0.1168) –0.156 (0.1185) GED 0.077** (0.0338) 0.117*** (0.0341) 0.080** (0.0351) 0.132*** (0.0361) High school diploma 0.142*** (0.0277) 0.124*** (0.0287) 0.147*** (0.0289) 0.122*** (0.0306) Associate college degree 0.194*** (0.0374) 0.303*** (0.0404) 0.203*** (0.0391) 0.347*** (0.0444) Four-year college degree or higher 0.269*** (0.0314) 0.404*** (0.0345) Years worked 0.059*** (0.0050) 0.070*** (0.0065) Years worked squared –0.002*** (0.0004) –0.002*** (0.0004) Mean value 0.576 0.576 0.576 0.577 0.520 0.521 Observations 20,729 20,729 20,729 20,341 15,221 14,885 0.046 0.051 0.061 0.076 0.043 0.061 Pseudo R 2 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Column 4 of Table 8, the table’s most exhaustive specification, conditional on ability, education, and labor market experience, having a juvenile conviction reduces the probability of full-time employment for these men by 6.7 percentage points. This estimate is statistically significant at the 1 percent level. Columns 5 and 6 focus on men in this group without a four-year college degree or higher. As seen in Column 6, having a juvenile conviction reduces the probability of full-time employment by 6.9 percentage points for employed men without a four-year college degree or higher. Tables 9 and 10 present the analogous estimates for women. As seen in Columns 2 to 6, once we control for test scores, there is no statistically significant relationship between a juvenile conviction and employment for these women. This result is robust to excluding women with a four-year college degree or higher from the sample (Columns 5 and 6). Hence, the observed employment gap between women with and without a juvenile conviction shown in Figure 2B is primarily explained by the differences in ability and education between the two groups. 56 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Our results suggest a large and statistically significant negative impact of a juvenile conviction on the probability of full-time employment for employed woman (Table 10). This result is robust across all specifications and is particularly larger for the less-educated of this group. Under the table’s most exhaustive specification (Column 4), a juvenile conviction reduces the probability of full-time employment by 6.4 percentage points for all employed woman and by 8.2 percentage points for employed woman who do not have a four-year college degree or higher (Column 6). This result is significant because workers who do not work full time are less attached to the labor market and are less likely to invest in human capital ex ante and are more likely to commit crimes, as discussed in our conceptual framework. We will discuss recidivism in Section 4.5. Compared with the literature, our estimated direct effects of a juvenile record on male employment (after controlling for its impact on education) are in line with the estimated long-term effect of an adult crime record on employment from Prescott and Starr (2020). Specifically, Prescott and Starr (2020) estimate that expunging adult criminal records leads to a 6.7-percentage-point increase in the employment rate of all offenders three years later. We do not find a statistically significant impact of a juvenile conviction on female employment. Possible explanations include that (i) a juvenile conviction has a smaller direct impact than an adult crime record on female employment and (ii) the results in Prescott and Starr (2020) are primarily driven by male offenders. Lastly, our results show that a juvenile conviction has large and statistically significant negative effects on fulltime employment both for employed men and employed women. 4.3 Juvenile Conviction and OTJ Training Investment In Sections 4.1 and 4.2 we investigated the effects of a juvenile conviction on education and employment, and our results are broadly in line with the findings of the literature. We found that workers with more past working experience have a higher probability of employment after controlling for juvenile conviction, ability measures, and education. This finding is partly due to workers with more experience acquiring more human capital while working. To further explore the relationship between training experience and juvenile conviction, we investigate the effects of a juvenile conviction on an individual’s post-school human capital investment as characterized by occupation-specific employment. To the best of our knowledge, our article is the first to investigate such a relationship. Our findings suggest that a juvenile conviction reduces the probability of working in a high-OTJ-training occupation. As shown in the next section, high-OTJ-training occupations have higher wage levels and faster future wage growth. Hence this channel is an important mechanism through which a juvenile conviction can affect wages. Table 11 reports the marginal effects of a juvenile conviction on the probability of working in a high-OTJ-training occupation for men. As seen in Column 1, controlling for the race dummies and age, a man with a juvenile conviction is 13.8 percentage points less likely to work in a high-OTJtraining occupation. Once we control for ASVAB scores and education, the direct impact of juvenile conviction on this probability becomes –8.7 percentage points (Column 3). In the table’s most exhaustive specification (Column 4), a juvenile conviction reduces the probability of working in a high-OTJ-training occupation by 4.9 percentage points for employed men, after controlling for both general work experience and work experience in high-OTJ-training occupations (measured as accumulated years worked in these occupations) as well as education. Notice that both high education levels and experience in high-OTJ-training occupations are correlated with the increased 57 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 11 The Effects of a Juvenile Conviction on Adult Male Occupation Choices: Employed in a High-OTJ-Training Occupation (1) (2) (3) (4) (5) (6) Juvenile conviction –0.138*** (0.0263) –0.126*** (0.0271) –0.087*** (0.0272) –0.049** (0.0201) –0.092*** (0.0285) –0.056** (0.0220) Race = Black –0.188*** (0.0189) –0.137*** (0.0206) –0.138*** (0.0206) –0.069*** (0.0156) –0.159*** (0.0230) –0.087*** (0.0181) Race = Hispanic –0.061*** (0.0196) –0.026 (0.0206) –0.021 (0.0202) –0.007 (0.0144) –0.023 (0.0217) –0.013 (0.0162) Age 0.024*** (0.0014) 0.023*** (0.0014) 0.019*** (0.0014) 0.002 (0.0039) 0.020*** (0.0016) 0.004 (0.0050) ASVAB score 0.511*** (0.1033) 0.483*** (0.1064) 0.373*** (0.0805) 0.466*** (0.1210) 0.378*** (0.0933) ASVAB score squared –0.328*** (0.1018) –0.423*** (0.1053) –0.353*** (0.0797) –0.409*** (0.1262) –0.359*** (0.0960) GED 0.018 (0.0341) 0.019 (0.0259) 0.019 (0.0348) 0.021 (0.0280) High school diploma 0.068** (0.0293) 0.028 (0.0213) 0.071** (0.0300) 0.032 (0.0226) Associate college degree 0.166*** (0.0470) 0.138*** (0.0388) 0.166*** (0.0479) 0.146*** (0.0441) Four-year college degree or higher 0.243*** (0.0369) 0.207*** (0.0356) Years worked –0.032*** (0.0074) –0.017* (0.0085) Years worked squared –0.002*** (0.0006) –0.003*** (0.0006) Years worked in high-OTJtraining occupations 0.149*** (0.0045) 0.143*** (0.0046) Mean value 0.521 0.521 0.521 0.549 0.477 0.501 Observations 23,912 23,912 23,912 19,114 19,859 15,292 Pseudo R 2 0.040 0.049 0.060 0.234 0.042 0.216 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. probability of men working in a high-OTJ-training occupation. Therefore, the 4.9-percentage- point reduction reported in Column 4 is only the direct contemporaneous effects of a juvenile conviction. The long-term overall effect of a juvenile conviction on OTJ-training occupational choices is larger due to its accumulated effects through education and past work experience. As in our previous tables, Columns 5 and 6 focus on low-skill male workers who do not have a four-year college degree or higher. We find that a juvenile conviction directly reduces the chances of these men working in a high-OTJ-training occupation by 5.6 percentage points. As expected, the negative impact of a juvenile conviction is larger among this disadvantaged population. 58 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 12 The Effects of a Juvenile Conviction on Adult Female Occupation Choices: Employed in a High-OTJ-Training Occupation (1) (2) (3) (4) (5) (6) Juvenile conviction –0.165*** (0.0449) –0.128*** (0.0470) –0.084* (0.0438) –0.039 (0.0336) –0.108** (0.0457) –0.074** (0.0353) Race = Black –0.100*** (0.0180) 0.006 (0.0192) –0.002 (0.0193) 0.004 (0.0165) –0.022 (0.0211) –0.014 (0.0180) Race = Hispanic –0.016 (0.0203) 0.059*** (0.0215) 0.072*** (0.0219) 0.043** (0.0181) 0.067*** (0.0223) 0.041** (0.0185) Age 0.022*** (0.0015) 0.020*** (0.0015) 0.013*** (0.0016) 0.001 (0.0036) 0.014*** (0.0017) 0.001 (0.0048) ASVAB score 0.700*** (0.1124) 0.518*** (0.1152) 0.199** (0.1003) 0.523*** (0.1241) 0.222** (0.1095) ASVAB score squared –0.315*** (0.1062) –0.317*** (0.1079) –0.115 (0.0929) –0.351*** (0.1251) –0.153 (0.1094) GED 0.139*** (0.0425) 0.133*** (0.0412) 0.126*** (0.0394) 0.128*** (0.0387) High school diploma 0.172*** (0.0353) 0.133*** (0.0347) 0.157*** (0.0328) 0.115*** (0.0317) Associate college degree 0.289*** (0.0458) 0.262*** (0.0467) 0.261*** (0.0428) 0.254*** (0.0474) Four-year college degree or higher 0.368*** (0.0393) 0.316*** (0.0431) Years worked –0.058*** (0.0077) –0.020** (0.0084) Years worked squared 0.000 (0.0007) –0.001* (0.0007) Years worked in high-OTJtraining occupations 0.162*** (0.0055) 0.129*** (0.0053) Mean value 0.413 0.413 0.413 0.445 0.343 0.367 Observations 22,762 22,762 22,762 17,424 17,149 12,239 Pseudo R 2 0.028 0.054 0.073 0.224 0.042 0.185 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. As with employment, the probability of Black male workers being employed in a high-OTCtraining occupation is substantially lower than that of White male workers, close to 19 percentage points (see Column 1 of Table 11). For Hispanic male workers it is 6.1 percentage points less than that of White male workers. Controlling for test scores reduces the coefficient on the Black indicator variable to 13.7 percentage points (Column 2), and it is not further reduced once education is included (see Column 3). However, experience in high-OTC-training occupations reduces the coefficient on the Black indicator variable to 8.7 percentage points (see Column 6). For Hispanic male workers, the race effect on the probability of employment in a high-OTC-training occupation becomes small and statistically insignificant once we account for ASVAB scores. The fact that Black 59 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski individuals are less likely to be employed in high-OTJ-training occupations is consistent with Golan, James, and Sanders (2019), who find evidence for discrimination in the assignment and promotion of Black workers to occupations with demand for complex and non-routine tasks in the NLSY79. It is likely that these occupations are also occupations with higher training requirements, although this variable is not directly analyzed in their article. We conduct the same analysis for women in Table 12. Looking at Column 1 of Table 12, women with a juvenile conviction are 16.5 percentage points less likely to be employed in a high-OTCtraining occupation than women without a juvenile conviction. After further controlling for test scores and education, the reduction in the probability of women with a juvenile conviction being employed in a high-OTC-training occupation is 8.4 percentage points (and significant at the 10 percent level; Column 3). Controlling for general work experience and experience in high-OTCtraining occupations, the coefficient in front of the juvenile conviction variable becomes smaller in magnitude (–3.9 percentage points) and loses statistical significance (Column 4). However, a juvenile conviction may still affect a women’s occupational choices through its indirect impact via education. Furthermore, the selection of women into high-OTC-training occupations may be different from that of men due to the effects of fertility and marriage on occupational choices and the glass ceiling women face.2 These issues, however, are beyond the scope of this article. It is interesting to note that, once we control for test scores, the coefficient on the Black indicator variable becomes small and statistically insignificant, while the coefficient on the Hispanic indicator variable becomes positive and statistically significant. In Columns 5 and 6 of Table 12, we focus on women without a four-year college degree or higher, who are relatively disadvantaged in the labor market. As seen in Column 6, a juvenile conviction reduces their probability of working in a high-OTJ-training occupation by 7.4 percentage points. Moreover, we find large state dependence in occupational choices, as work experience in high-OTJ-training occupations increases the probability of working in such an occupation in the future. This finding implies that the overall life-cycle effect of a juvenile conviction on an individual’s occupational choices is likely to be bigger and more persistent. 4.4 Juvenile Conviction and Wages So far, our results establish that a juvenile conviction reduces a youth’s educational achievement, employment, and probability of working in a high-OTJ-training occupation. In this section, we document that the work experience in high-OTJ-training occupations is associated with higher wages and wage growth. Higher wages and income imply a higher opportunity cost of devoting time to crime and a higher opportunity cost of time spent in jail. The observed wage gap by juvenile conviction status (see Figure 3) is primarily due to the accumulated effects of a juvenile conviction on employment and occupational choices over time. We first examine the effect of a juvenile conviction on wage levels. Table 13 presents regression analysis of log hourly wages for employed men. Starting with the estimate in Column 1, we find a statistically significant negative impact of a juvenile conviction on employed men’s wages. Columns 2 and 3 repeat the initial estimate, now adding main effects for ability and education sequentially; the negative effects remain statistically significant but smaller in magnitude. Column 4 includes years worked and years worked squared; the coefficient in front of the juvenile conviction variable becomes slightly smaller in magnitude but remains negative and statistically significant. However, 60 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 13 The Effects of a Juvenile Conviction on Adult Male Log Hourly Wages (1) (2) (3) (4) (5) (6) (7) (8) Juvenile conviction –0.147*** (0.0295) –0.128*** (0.0289) –0.077*** (0.0287) –0.068** (0.0299) –0.027 (0.0285) –0.087*** (0.0286) –0.076** (0.0301) –0.038 (0.0285) Race = Black –0.215*** (0.0201) –0.147*** (0.0212) –0.147*** (0.0206) –0.129*** (0.0195) –0.098*** (0.0211) –0.161*** (0.0220) –0.135*** (0.0207) –0.105*** (0.0224) Race = Hispanic –0.075*** (0.0196) –0.027 (0.0203) –0.018 (0.0192) –0.020 (0.0187) –0.022 (0.0192) –0.023 (0.0201) –0.024 (0.0195) –0.023 (0.0198) Age 0.047*** (0.0013) 0.045*** (0.0013) 0.038*** (0.0013) 0.007* (0.0041) 0.016*** (0.0049) 0.036*** (0.0014) –0.006 (0.0047) –0.005 (0.0057) ASVAB score 0.320*** (0.1062) 0.298*** (0.1041) 0.197* (0.1043) 0.042 (0.1088) 0.468*** (0.1161) 0.325*** (0.1165) 0.202* (0.1220) ASVAB score squared –0.048 (0.1103) –0.207* (0.1068) –0.131 (0.1066) –0.000 (0.1104) –0.433*** (0.1284) –0.316** (0.1278) –0.233* (0.1332) GED 0.012 (0.0303) 0.051* (0.0311) 0.032 (0.0353) 0.011 (0.0303) 0.067** (0.0313) 0.062* (0.0352) High school diploma 0.063** (0.0260) 0.056** (0.0260) 0.043 (0.0295) 0.064** (0.0260) 0.055** (0.0258) 0.048* (0.0287) Associate college degree 0.216*** (0.0451) 0.338*** (0.0482) 0.235*** (0.0491) 0.222*** (0.0451) 0.387*** (0.0494) 0.328*** (0.0510) Four-year college degree or higher 0.328*** (0.0345) 0.494*** (0.0408) 0.365*** (0.0475) Years worked 0.055*** (0.0056) 0.032*** (0.0066) 0.066*** (0.0062) 0.053*** (0.0072) Years worked squared –0.002*** (0.0003) –0.002*** (0.0003) –0.001*** (0.0003) –0.002*** (0.0004) Years worked in highOTJ-training occupations 0.026*** (0.0037) 0.025*** (0.0039) Employed at high-OTJ occupation 0.177*** (0.0135) 0.179*** (0.0137) Constant 1.276*** (0.0327) 1.168*** (0.0398) 1.283*** (0.0444) 1.845*** (0.0873) 1.647*** (0.0985) 1.330*** (0.0459) 2.104*** (0.0971) 2.035*** (0.1128) Mean value 2.384 2.384 2.384 2.385 2.404 2.310 2.310 2.324 Observations 22,009 22,009 22,009 21,456 17,770 18,097 17,579 14,147 R2 0.143 0.163 0.197 0.218 0.278 0.126 0.157 0.233 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. as we further add controls for experience in high-OTJ-training occupations and current employment status in those occupations, the direct effect of a juvenile conviction on wages is no longer significant (Column 5). This evidence suggests that the wage effects of a juvenile conviction mainly come from indirect effects through accumulated work experience in high-OTJ-training occupations and current employment status in a high-OTJ-training occupation. However, it can also reflect selection based on unobserved characteristics and traits of the individuals (in the data). In other words, it could be that unobserved skills and traits make individuals more likely to accumulate 61 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 14 The Effects of a Juvenile Conviction on Adult Female Log Hourly Wages (1) (2) (3) (4) (5) (6) (7) (8) Juvenile conviction –0.211*** (0.0577) –0.148*** (0.0531) –0.075 (0.0488) –0.071 (0.0496) –0.045 (0.0551) –0.104** (0.0510) –0.105** (0.0519) –0.079 (0.0590) Race = Black –0.134*** (0.0199) 0.019 (0.0196) 0.009 (0.0184) 0.024 (0.0183) 0.022 (0.0202) 0.009 (0.0203) 0.028 (0.0200) 0.025 (0.0220) Race = Hispanic –0.045** (0.0201) 0.064*** (0.0207) 0.092*** (0.0196) 0.098*** (0.0196) 0.065*** (0.0211) 0.107*** (0.0219) 0.111*** (0.0217) 0.083*** (0.0237) Age 0.049*** (0.0016) 0.044*** (0.0015) 0.031*** (0.0016) 0.009** (0.0039) 0.025*** (0.0047) 0.027*** (0.0017) –0.005 (0.0050) 0.010 (0.0067) ASVAB score 0.690*** (0.1185) 0.433*** (0.1153) 0.334*** (0.1157) 0.201 (0.1297) 0.503*** (0.1374) 0.366*** (0.1361) 0.261 (0.1600) ASVAB score squared –0.109 (0.1224) –0.193* (0.1168) –0.128 (0.1172) –0.054 (0.1294) –0.287* (0.1522) –0.190 (0.1496) –0.152 (0.1743) GED 0.134*** (0.0422) 0.167*** (0.0426) 0.081* (0.0491) 0.140*** (0.0422) 0.185*** (0.0430) 0.114** (0.0510) High school diploma 0.177*** (0.0315) 0.161*** (0.0316) 0.105*** (0.0374) 0.178*** (0.0315) 0.152*** (0.0319) 0.112*** (0.0380) Associate college degree 0.365*** (0.0477) 0.456*** (0.0510) 0.293*** (0.0593) 0.378*** (0.0478) 0.507*** (0.0532) 0.375*** (0.0655) Four-year college degree or higher 0.568*** (0.0372) 0.680*** (0.0408) 0.476*** (0.0504) Years worked 0.041*** (0.0055) 0.019*** (0.0065) 0.045*** (0.0072) 0.024*** (0.0089) Years worked squared –0.001*** (0.0004) –0.002*** (0.0004) –0.000 (0.0004) –0.001 (0.0005) Years worked in highOTJ-training occupations 0.018*** (0.0045) 0.018*** (0.0052) Employed at high-OTJ occupation 0.182*** (0.0140) 0.170*** (0.0161) Constant 1.046*** (0.0377) 0.818*** (0.0454) 1.011*** (0.0493) 1.438*** (0.0869) 1.164*** (0.0992) 1.097*** (0.0527) 1.742*** (0.1073) 1.474*** (0.1362) Mean value 2.256 2.256 2.256 2.257 2.289 2.118 2.118 2.133 Observations 20,343 20,343 20,343 19,979 15,796 14,960 14,640 10,898 R2 0.131 0.201 0.278 0.288 0.332 0.124 0.144 0.191 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. experience and human capital and as a result these individuals earn higher wages. This can be seen from the statistically positive coefficients in front of both experience in high-OTJ-training occupations and current employment in a high-OTJ-training occupation. Our findings are similar when we focus on less-skilled men who do not have a four-year college degree or higher (Columns 6 to 8). Table 14 repeats the analysis for employed women. The estimates in Columns 1 and 2 suggest that a juvenile conviction has a negative impact on their wages. However, as we introduce dummies for the education categories, the coefficient of a juvenile conviction loses statistical significance. 62 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 15 The Effects of a Juvenile Conviction on Adult Male 10-Year Percent Changes of Hourly Wages (1) (2) (3) (4) (5) (6) Juvenile conviction –40.730*** (10.282) –37.674*** (10.109) –34.127*** (10.548) –25.814*** (9.739) –34.445*** (10.636) –26.056*** (9.797) Race = Black –24.034** (9.744) –9.385 (9.996) –9.239 (10.159) –1.803 (10.513) –14.393 (10.062) –5.950 (10.366) Race = Hispanic –15.135* (8.808) –4.010 (8.795) –3.841 (8.840) 3.996 (9.118) –7.347 (8.446) 1.019 (8.535) –4.650 (2.853) –5.751** (2.871) –7.541** (3.559) –11.230*** (4.163) –7.507** (3.570) –11.375*** (4.186) ASVAB score 147.023*** (53.283) 145.579*** (55.591) 108.913** (54.104) 118.132** (58.625) 82.551 (56.990) ASVAB score squared –84.646 (55.010) –89.760 (56.049) –64.766 (55.505) –59.674 (61.199) –36.798 (60.867) GED –13.364 (22.696) –17.409 (26.311) –12.520 (22.601) –16.249 (26.206) High school diploma 4.140 (20.847) –7.065 (23.769) 4.631 (20.789) –6.539 (23.718) Associate college degree 23.102 (32.977) 27.678 (36.484) 23.050 (32.928) 28.152 (36.407) Four-year college degree or higher 21.057 (32.260) 30.340 (36.103) Age Years worked in high-OTJtraining occupations Constant Mean value Observations R 2 3.509*** (1.260) 3.741*** (1.277) 232.643*** (84.279) 218.839*** (83.752) 269.608*** (103.512) 367.359*** (119.797) 273.385*** (103.867) 374.795*** (120.575) 86.017 86.017 86.017 83.420 84.965 82.010 952 952 952 792 907 751 0.014 0.035 0.038 0.054 0.040 0.056 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. This finding suggests that many of the negative effects of a juvenile conviction found earlier can be explained by the effects of a juvenile conviction on educational outcomes. The coefficient of the juvenile conviction variable changes little after further controlling for general work experience (Column 4). In Column 5, the estimates suggest that both experience in high-OTJ-training occupations and current employment status in a high-OTJ-training occupation increase wages, which are important mechanisms through which a juvenile conviction can affect wages. A juvenile conviction has a negative wage impact for women without a four-year college degree or higher, but the significance disappears once we control for experience in high-OTJ-training occupations and current employment status in a high-OTJ-training occupation. An interesting contrast between the hourly wages of men and women is the role of race. Black men earn statistically significant lower wages than White men in all specifications, and Hispanic men generally do not earn significantly different wages from White men. On the other hand, Black 63 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 16 The Effects of a Juvenile Conviction on Adult Female 10-Year Percent Changes of Hourly Wages (1) (2) (3) (4) (5) (6) Juvenile conviction –63.858*** (18.998) –60.016*** (17.093) –63.292*** (23.935) –66.399** (29.042) –62.997*** (24.199) –63.148** (29.415) Race = Black –43.461*** (11.727) –29.748*** (10.675) –28.558*** (10.067) –29.628** (12.523) –28.283*** (10.016) –28.379** (12.651) Race = Hispanic –22.030* (12.725) –11.244 (11.590) –11.970 (12.243) –20.432 (14.853) –10.136 (12.484) –18.533 (15.312) 9.419* (5.259) 6.076 (4.223) 1.554 (5.178) 0.806 (7.747) 1.825 (5.177) 0.994 (7.836) ASVAB score –123.139 (161.425) –165.783 (169.474) –221.081 (213.330) –122.124 (193.447) –163.900 (250.781) ASVAB score squared 221.676 (199.200) 246.790 (201.926) 280.683 (249.045) 199.242 (233.822) 214.115 (295.863) GED 21.079 (19.314) –5.849 (28.303) 19.165 (19.412) –8.951 (27.916) High school diploma 31.952* (16.418) 13.768 (26.013) 30.303* (16.406) 10.782 (25.711) Associate college degree 140.774 (103.715) 151.562 (114.438) 139.333 (102.905) 149.241 (111.290) Four-year college degree or higher 38.241 (35.914) 29.132 (55.290) Age Years worked in high-OTJtraining occupations Constant Mean value Observations R 2 2.852 (3.098) 4.351 (3.455) –188.667 (149.058) –98.178 (121.317) 12.388 (161.566) 56.164 (229.763) –1.428 (162.523) 38.667 (234.264) 71.923 71.923 71.923 73.477 68.712 69.181 709 709 709 543 657 498 0.021 0.045 0.064 0.074 0.060 0.072 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. women generally do not earn significantly different wages from White women, while Hispanic women generally earn statistically significantly higher wages than White women.3 Next we examine the effect of a juvenile conviction on the growth rate of wages in Tables 15 and 16 for men and women, respectively. Starting with the estimates in Column 1 of Table 15, for men, a juvenile conviction reduces the 10-year growth rate of wages by 40.7 percentage points and both the Black race dummy and the Hispanic race dummy have negative coefficients on wage growth. Once we control for the ability measures, the coefficients on the race dummies become statistically insignificant. The negative effects of a juvenile conviction on wage growth remain large and significant across all specifications. As seen in the table’s most exhaustive specification (Column 4), a juvenile conviction reduces wage growth by 25.8 percentage points for men. The results for less- educated male workers are reported in Columns 5 and 6. The negative effects of a juvenile conviction are slightly larger in magnitude among this disadvantaged group. For them, as seen in Column 6, 64 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski a juvenile conviction leads to a 26.1-percentage-point reduction in wage growth. We also note that experience in high-OTJ-training occupations leads to higher wage growth (Columns 4 and 6). Turning to the estimates in Table 16 for women, we find a juvenile conviction reduces wage growth by a larger magnitude for them than for men. In this table’s most exhaustive specification (Column 4), we find that a juvenile conviction reduces wage growth by 66.4 percentage points for women. Compared with the average 10-year wage growth for women, this estimate is very large. Column 6 shows that a juvenile conviction reduces wage growth by 63.1 percentage points for less- educated female workers. The coefficients on experience in high-OTJ-training occupations are insignificant (Columns 4 and 6), while the coefficients on the Black race dummy remain negative and significant. Our findings here suggest an interesting gender difference in wage growth dynamics. 4.5 Juvenile Conviction and Recidivism In Sections 4.1 to 4.4 we discussed the effects of a juvenile conviction on human capital investment and labor market outcomes. In this section, we investigate the relationship between a juvenile conviction and recidivism during adulthood. We find that individuals with a juvenile conviction are more likely to commit crimes in adulthood. As discussed in the conceptual framework, a higher probability of future recidivism reduces an individual’s incentive to invest in human capital ex ante, which may help explain our estimated negative effects of a juvenile conviction on human capital investment. Table 17 reports the estimates of the marginal effects of a juvenline conviction on the probability of adult incarceration using a logit model for men. It shows a positive and statistically significant impact for all specifications. As shown in Column 1, among men, a juvenile conviction increases the probability of adult incarceration by 3.5 percentage points, after controlling for race and age. As we sequentially add the ability and education measures, the estimated effects become 2.7 percentage points and 1.4 percentage points, respectively. In the table’s most exhaustive specification (Column 4), where measures of work experience are added, the probability of adult male incarceration increases by 0.9 percentage points. We also investigate the effects for less-educated men and find a much larger effect (1.4-percentage-point increase). It is also worth noting that once we control for the work-experience measures, the coefficient of the Black race dummy is not statistically significant (Columns 4 and 6), suggesting that employment dynamics hold the key for understanding the racial differences in crime behavior. Table 18 reports the results of the same analysis for women. The effect of a juvenile conviction on the probability of adult incarceration becomes smaller once we control for test scores; adding education further reduces the effect on incarceration by half (Column 3). Focusing on Column 4, adding work experience and experience in high-OTJ-training occupations, the effect of juvenile conviction on the probability of adult incarceration becomes small and statistically insignificant; in addition, we find a statistically significant negative coefficient in front of the variable for experience in high-OTJ-training occupations. Columns 5 and 6 report the analysis for women without a four-year college degree or higher and show a statistically significant positive impact of a juvenile conviction on the incarceration probability of women with lower education. Overall the results confirm the relationship between juvenile conviction and future incarceration, highlighting the importance of education and work experience in reducing this probability. Our analysis is suggestive, but these relationships can be also driven by unobserved traits and skills of individuals who have higher educational attainment and work experience, especially in high-OTJ-training occupations. 65 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 17 The Effects of a Juvenile Conviction on Adult Male Incarceration (1) (2) (3) (4) (5) (6) Juvenile conviction 0.0351*** (0.00401) 0.0274*** (0.00353) 0.0139*** (0.00290) 0.0094*** (0.00200) 0.0218*** (0.00450) 0.0142*** (0.00297) Race = Black 0.0200*** (0.00336) 0.0095*** (0.00350) 0.0066** (0.00269) 0.0022 (0.00182) 0.0103** (0.00422) 0.0031 (0.00278) Race = Hispanic 0.0015 (0.00407) –0.0045 (0.00369) –0.0032 (0.00272) –0.0017 (0.00181) –0.0054 (0.00437) –0.0030 (0.00280) Age 0.0000 (0.00026) 0.0002 (0.00022) 0.0002 (0.00016) 0.0006* (0.00032) 0.0003 (0.00026) 0.0009* (0.00050) ASVAB score 0.0103 (0.01708) 0.0156 (0.01270) 0.0046 (0.00952) 0.0281 (0.02107) 0.0099 (0.01530) ASVAB score squared –0.0533*** (0.01945) –0.0334** (0.01465) –0.0140 (0.00975) –0.0587** (0.02537) –0.0261 (0.01658) GED 0.0031 (0.00262) –0.0006 (0.00199) 0.0050 (0.00414) –0.0009 (0.00303) High school diploma –0.0131*** (0.00277) –0.0080*** (0.00189) –0.0207*** (0.00415) –0.0120*** (0.00272) Associate college degree –0.0318*** (0.00797) –0.0219*** (0.00612) –0.0502*** (0.01217) –0.0332*** (0.00893) Four-year college degree or higher –0.0359*** (0.00551) –0.0229*** (0.00407) Years worked –0.0010* (0.00056) –0.0017* (0.00088) Years worked squared –0.0000 (0.00004) –0.0000 (0.00006) Years worked in high-OTJtraining occupations –0.0002 (0.00039) –0.0004 (0.00060) Mean value 0.030 0.030 0.030 0.017 0.035 0.021 Observations 26,878 26,878 26,878 19,981 22,633 16,076 Pseudo R 2 0.059 0.085 0.132 0.129 0.103 0.106 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. Aizer and Doyle (2015) also produce large recidivism estimates, finding that juvenile incarceration increases the probability of being incarcerated as an adult by age 25 by 23 percentage points. They go on to show that individuals who experience juvenile incarceration are more likely to recidivate for serious crimes including homicide, violence, and drug offenses. This finding suggests there is a behavioral change due to juvenile incarceration, which may be watered down in our sample since we include all convicted juveniles, of which only about 23 percent are incarcerated in the Aizer and Doyles (2015) sample. 66 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Table 18 The Effects of a Juvenile Conviction on Adult Female Incarceration (1) (2) (3) (4) (5) (6) Juvenile conviction 0.0098*** (0.00201) 0.0077*** (0.00204) 0.0035** (0.00158) 0.0017 (0.00103) 0.0086*** (0.00327) 0.0042** (0.00213) Race = Black 0.0012 (0.00154) –0.0010 (0.00160) –0.0003 (0.00090) –0.0015* (0.00088) –0.0006 (0.00219) –0.0037* (0.00190) Race = Hispanic 0.0008 (0.00189) –0.0007 (0.00175) –0.0002 (0.00099) –0.0006 (0.00076) –0.0005 (0.00242) –0.0016 (0.00189) Age 0.0000 (0.00013) 0.0001 (0.00011) 0.0001 (0.00008) 0.0002 (0.00012) 0.0002 (0.00018) 0.0005 (0.00031) ASVAB score –0.0062 (0.00775) –0.0002 (0.00414) –0.0005 (0.00405) 0.0015 (0.01023) 0.0020 (0.01078) ASVAB score squared –0.0027 (0.00784) –0.0007 (0.00432) –0.0011 (0.00455) –0.0046 (0.01108) –0.0074 (0.01342) GED –0.0013 (0.00105) –0.0007 (0.00093) –0.0032 (0.00247) –0.0019 (0.00238) High school diploma –0.0028** (0.00128) –0.0013 (0.00092) –0.0068*** (0.00253) –0.0032* (0.00190) Associate college degree –0.0033* (0.00177) –0.0016 (0.00140) –0.0079** (0.00376) –0.0042 (0.00334) Four-year college degree or higher –0.0100*** (0.00185) –0.0055*** (0.00163) Years worked –0.0001 (0.00017) –0.0005 (0.00044) Years worked squared 0.0000 (0.00001) 0.0000 (0.00003) Years worked in high-OTJtraining occupations –0.0003* (0.00015) –0.0008* (0.00043) Mean value 0.006 0.006 0.006 0.003 0.007 0.005 Observations 27,139 27,139 27,139 18,572 21,109 13,222 Pseudo R 2 0.042 0.062 0.101 0.104 0.062 0.067 NOTE: Marginal effects; standard errors are in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. 5 CONCLUSION In this article, we show that juvenile conviction has a long-term impact on human capital accumulation, wages, and recidivism. Specifically, we find that individuals with a juvenile conviction have lower education levels, lower employment rates, and are less likely to work in occupations with high-OTJ-training requirements. Juvenile conviction reduces wages mainly through its negative impact on education and work experience (including both general experience and occupation- specific work experience). Regarding the effect on recidivism, we find that a juvenile conviction is associated with a higher probability of incarceration in adulthood. Finally, all these effects are more pronounced among individuals without a four-year college degree or higher. Our results highlight 67 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski the rich dynamics and interplay between educational choices, occupational choices, employment, wages, and recidivism when analyzing the effects of a juvenile conviction. While we find negative effects of a juvenile conviction on long-term labor market outcomes as well as on the probability of adult incarceration, even after controlling for measures of ability and education, it is possible that some of the effects are due to unobserved traits and skills differences between individuals with a juvenile conviction and those without. To address this issue, including the interactions of juvenile convictions, individuals’ traits and skills, and employers’ discrimination, future analysis using a structural model to investigate these dynamic mechanisms and evaluate alternative policies is a fruitful direction. n NOTES 1 Much of the existing literature on crime and labor market focuses on the relationship between adult conviction or incarceration and labor market outcomes; see Prescott and Starr (2020), for example. 2 See, for example, Gayle and Golan (2012) for a discussion on occupational sorting and discrimination. 3 There are differences in the patterns of labor market attachment and labor supply of Black and White women, which is discussed in the literature. REFERENCES Aizer, Anna and Doyle, Joseph J., Jr. “Juvenile Incarceration, Human Capital, and Future Crime: Evidence from Randomly Assigned Judges.” Quarterly Journal of Economics, May 2015, 130, pp. 759-803; https://doi.org/10.1093/qje/qjv003. Apel, Robert and Sweeten, Gary. “The Impact of Incarceration on Employment During the Transition to Adulthood.” Social Problems, 2010, 57, pp. 448-79; https://doi.org/10.1525/sp.2010.57.3.448. Bayer, Patrick; Hjalmarsson, Randi and Pozen, David. “Building Criminal Capital Behind Bars: Peer Effects in Juvenile Corrections.” Quarterly Journal of Economics, 2009, 124, pp. 105-47; https://doi.org/10.1162/qjec.2009.124.1.105. Forrest, Christopher; Tambor, Ellen; Riley, A.; Ensminger, Margaret and Starfield, B. “The Health Profile of Incarcerated Male Youth.” Pediatrics, 2000, 105, pp. 286-91. Gayle, George-Levi and Golan, Limor. “Estimating a Dynamic Adverse-Selection Model: Labour-Force Experience and the Changing Gender Earnings Gap 1968-1997.” Review of Economic Studies, 2012, 79, pp. 227-67; https://doi.org/10.1093/restud/rdr019. Golan, Limor; James, Jonathan and Sanders, Carl. “What Explains the Racial Gaps in Task Assignment and Pay Over the Life-Cycle?” Society for Economic Dynamics, 2019. Imai, Susumu and Krishna, Kala. “Employment, Deterrence, and Crime in a Dynamic Model.” International Economic Review, 2004, 45, pp. 845-72; https://doi.org/10.1111/j.0020-6598.2004.00289.x. Kashani, Javad H.; Manning, George W.; McKnew, Donald H.; Cytryn, Leon; Simonds, John F. and Wooderson, Phil C. “Depression Among Incarcerated Delinquents.” Psychiatry Research, 1980, 3, pp. 185-91; https://doi.org/10.1016/0165-1781(80)90035-9. Kirk, David S. and Sampson, Robert J. “Juvenile Arrest and Collateral Educational Damage in the Transition to Adulthood.” Sociology of Education, 2013, 88, pp. 32-62; https://doi.org/10.1177/0038040712448862. Levitt, Steven D. “Why Do Increased Arrest Rates Appear to Reduce Crime: Deterrence, Incapacitation, or Measurement Error?” Economic Inquiry, 1998, 36, pp. 353-72; https://doi.org/10.1111/j.1465-7295.1998.tb01720.x. Litwok, Daniel. “Have You Ever Been Convicted of a Crime? The Effects of Juvenile Expungement on Crime, Educational, and Labor Market Outcomes” in Essays on the Economics of Juvenile Crime and Education, PhD dissertation. Michigan State University, 2015. Lochner, Lance. “Education, Work, and Crime: A Human Capital Approach.” International Economic Review, 2004, 45, pp. 811-43; https://doi.org/10.1111/j.0020-6598.2004.00288.x. 68 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Golan, Hai, Wabiszewski Mancino, Maria Antonella; Navarro, Salvador and Rivers, David A. “Separating State Dependence, Experience, and Heterogeneity in a Model of Youth Crime and Education.” Economics of Education Review, 2016, 54, pp. 274-305; https://doi.org/10.1016/j.econedurev.2016.07.005. Merlo, Antonio and Wolpin, Kenneth I. “The Transition from School to Jail: Youth Crime and High School Completion Among Black Males.” European Economic Review, 2015, 79, pp. 234-51; https://doi.org/10.1016/j.euroecorev.2015.07.015. Nagin, Daniel S.; Farrington, David P. and Moffitt, Terrie E. “Life-Course Trajectories of Different Types of Offenders.” Criminology, 1995, 33, pp. 111-39; https://doi.org/10.1111/j.1745-9125.1995.tb01173.x. Nagin, Daniel S. and Land, Kenneth C. “Age, Criminal Careers, and Population Heterogeneity: Specification and Estimation of a Nonparametric, Mixed Poisson Model.” Criminology, 1993, 31, pp. 327-62; https://doi.org/10.1111/j.1745-9125.1993.tb01133.x. Nagin, Daniel S. and Paternoster, Raymond. “On the Relationship of Past to Future Participation in Delinquency.” Criminology, 1991, 29, pp. 163-89; https://doi.org/10.1111/j.1745-9125.1991.tb01063.x. Nagin, Daniel S. and Waldfogel, Joel. “The Effects of Criminality and Conviction on the Labor Market Status of Young British Offenders.” International Review of Law and Economics, 1995, 15, pp. 109-26; https://doi.org/10.1016/0144-8188(94)00004-E. Neal, Derek. “The Measured Black-White Wage Gap Among Women Is Too Small.” Journal of Political Economy, 2004, 112, pp. S1-28; https://doi.org/10.1086/379940. Paternoster, Raymond; Brame, Robert and Farrington, David P. “On the Relationship Between Adolescent and Adult Conviction Frequencies.” Journal of Quantitative Criminology, 2001, 17, pp. 201-25; https://doi.org/10.1023/A:1011007016387. Prescott, J.J. and Starr, Sonja B. “Expungement of Criminal Convictions: An Empirical Study.” Harvard Law Review, June 2020; https://doi.org/10.2139/ssrn.3353620. Ritter, Joseph A. and Taylor, Lowell J. “Racial Disparity in Unemployment.” Review of Economics and Statistics, 2011, 93, pp. 30-42; https://doi.org/10.1162/REST_a_00063. Western, Bruce and Beckett, Katherine. “How Unregulated Is the U.S. Labor Market? The Penal System as a Labor Market Institution.” American Journal of Sociology, 1999, 104, pp. 1030-60; https://doi.org/10.1086/210135. Western, Bruce; Kling, Jeffrey and Weiman, David. “The Labor Market Consequences of Incarceration.” Working Paper 829, Princeton University, Department of Economics, Industrial Relations Section, 2001. 69 Further Evidence on Greenspan’s Conundrum Cletus C. Coughlin and Daniel L. Thornton During his February 2005 congressional testimony, Alan Greenspan identified what he termed a conundrum. Despite the fact that the Federal Open Market Committee (FOMC) had increased the federal funds rate 150 basis points since June 2004, the 10-year Treasury yield remained essentially unchanged. Greenspan considered several explanations for his observation but rejected each. Thornton (2018) showed that the relationship between the 10-year Treasury yield and the federal funds rate changed in the late 1980s, many years prior to Greenspan’s observation. Moreover, he showed that the relationship changed because the FOMC began using the federal funds rate as its policy instrument. The federal funds rate moved only when the FOMC changed its target for it, while, in contrast, the 10-year Treasury yield continued to respond to news as before. As a consequence of this change in the FOMC’s operating procedure, the correlation between changes in the funds rate and the 10-year Treasury yield declined—effectively to zero. There is no obvious reason that the U.S. experience should be unique. Hence, we explore the experiences of two other countries that implemented a policy of targeting a short-term rate. We find that, as in the United States, the correlation between the policy rate and the long-term sovereign bond yield declined effectively to zero for both the Bank of England and the Reserve Bank of New Zealand after they began using a short-term rate as their policy instrument. (JEL E43, E52, E58) Federal Reserve Bank of St. Louis Review, First Quarter 2022, 104(1), pp. 70-77. https://doi.org/10.20955/r.104.70-77 1 INTRODUCTION During his February 2005 congressional testimony, Alan Greenspan noted that despite the fact that the Federal Open Market Committee (FOMC) had increased the federal funds rate 150 basis points since June 2004, the 10-year Treasury yield remained essentially unchanged. He posited several possible explanations for what he believed was the aberrant behavior of long-term Treasury yields. Rejecting each in turn, he called it a conundrum. Not surprisingly, Greenspan’s observation and ruminations stimulated much research. Several researchers (Backus and Wright, 2007; Kim and Wright, 2005; Rosenberg, 2007; Rudebusch and Cletus C. Coughlin is an emeritus economist of the Federal Reserve Bank of St. Louis and a senior research fellow at the Sinquefield Center for Applied Economic Research at St. Louis University. Daniel L. Thornton, a former vice president and economic advisor at the Federal Reserve Bank of St. Louis, is president of D.L. Thornton Economics, LLC. © 2022, Federal Reserve Bank of St. Louis. The views expressed in this article are those of the author(s) and do not necessarily reflect the views of the Federal Reserve System, the Board of Governors, or the regional Federal Reserve Banks. Articles may be reprinted, reproduced, published, distributed, displayed, and transmitted in their entirety if copyright notice, author name(s), and full citation are included. Abstracts, synopses, and other derivative works may be made only with prior written permission of the Federal Reserve Bank of St. Louis. 70 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Coughlin and Thornton Wu, 2007; and Smith and Taylor, 2009) investigated possible changes in the 10-year yield. Each of these articles generated declining estimates of the 10-year Treasury term premium; however, none were able to explain why the term premium declined. Thus, the apparent aberrant behavior of the 10-year Treasury yield remained a conundrum. Thornton (2018) took a different approach. Rather than assuming the conundrum began at about the time Greenspan observed it, he investigated when it began. He found that the relationship between the 10-year Treasury yield and the federal funds rate changed in the late 1980s, with the most likely date being May 1988. Based on previous research, he hypothesized that the change in behavior occurred when the FOMC began using the federal funds rate as its policy instrument. Once the FOMC began this practice, the federal funds rate moved only when the FOMC changed its target for it. In contrast, the 10-year Treasury yield continued to respond to news as before. The correlation between the federal funds rate and the 10-year Treasury yield declined to zero. This is because the FOMC changed its target for the funds rate infrequently. Thornton (2018) called this the funds-rate-targeting hypothesis (FRTH). This research is motivated by the fact that if the FRTH is correct, other central banks should have had a substantial decline in the correlation between their policy rate and sovereign long-term bond yield when they began using a short-term rate as their policy instrument. Simply stated, the experiences of other countries adopting interest rate targeting should be similar to that of the United States. This article examines the experiences of the Bank of England (BOE) and of the Reserve Bank of New Zealand (RBNZ). We use these countries because they implemented a policy of targeting a short-term rate as their policy rate, and we have sufficient data to see whether their experiences are comparable with that of the United States.1 The remainder of the article is in four sections. Because the FRTH is not the only possible explanation for the disconnect between the federal funds rate and the 10-year Treasury yield, Section 2 is used to summarize alternative hypotheses. To provide a foundation for our analysis, Section 3 reviews Thornton’s (2018) methodology, analysis, and findings. Section 4 investigates the impact of the BOE and the RBNZ, respectively, adopting a policy-rate-targeting regime. Section 5 presents the summary and conclusions. 2 COMPETING HYPOTHESES Thornton (2018) found that once the FOMC began using the funds rate as its policy instrument, the federal funds rate moved only when the FOMC changed its target for it. In contrast, the 10-year Treasury yield continued to respond to news as before. As a consequence, the correlation between changes in the funds rate and changes in the 10-year Treasury yield declined to zero. However, the FRTH is not the only possible explanation for the disconnect between the federal funds rate and the 10-year Treasury yield. For example, Goodfriend (1993) suggested market participants believed the FOMC would not permit inflation to accelerate. If Goodfriend is correct about Fed credibility, long-term Treasury rates would not move with increases in the federal funds rate when rate increases were prompted by inflation scares. Thornton (2018) termed this the inflation expectation hypothesis (IEH). Another explanation for Greenspan’s conundrum requires the behavior of the federal funds rate to become more predictable. In this case, the 10-year Treasury yield would change in advance 71 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Coughlin and Thornton Bank of Canada On February 22, 1996, the Bank of Canada (BOC) announced that it would implement monetary policy by targeting its overnight rate and set its bank rate at the top of its operating band for the overnight rate. The BOC bank rate is the minimum rate of interest the BC charges financial institutions on overnight loans. The overnight rate is the rate at which major financial institutions borrow and lend overnight funds among themselves. The overnight rate is analogous to the federal funds rate; the bank rate is analogous to the Fed’s primary credit rate. The BOC made those choices to make its policy intentions clearer. However, Borio (1997, p. 25) suggests the BOC began using the overnight rate as its policy instrument in June 1994 when it announced “an explicit 50 basis point operating band, communicated and validated by the offer to enter into repurchase transactions at those rates.” Unfortunately, data on the overnight rate are only available beginning January 1, 1996. Consequently, it is impossible to determine whether there was a dramatic change in the relationship between the overnight rate and the Canadian 10-year government bond yield after the BOC began using the overnight rate as its policy instrument. of the FOMC’s action, not when the FOMC changed its federal funds rate target. This hypothesis can be termed the policy predictability hypothesis (PPH). Thornton (2018) performed a battery of tests on the competing hypotheses. These tests effectively ruled out the IEH and the PPH while providing strong support for the FRTH. Let’s take a closer look at the methodology, analyses, and findings that support the FRTH. 3 THORNTON’S (2018) METHODOLOGY AND ANALYSIS Thornton (2018) used both statistical evidence and documentary evidence to examine the FRTH. Let’s begin with the key statistical evidence. To determine when the break in the relationship between changes in the federal funds rate and changes in the 10-year Treasury yield occurred, Thornton (2018) estimated the following simple regression: (1) ΔT10t = α + βΔFFt + εt , where ΔT10 and ΔFF denote the change in the 10-year Treasury yield and in the federal funds rate, respectively; α and β denote constant parameters; and ε denotes a random error with a zero mean and a constant variance. This equation was estimated using a rolling regression with a window of 33 months over the period January 1983 through March 2007. The window size was determined by the number of months from July 2004, the month of the first increase in the FOMC’s funds rate target, to March 2007. The starting date was chosen because Thornton (1988 and 2006) found that the FOMC began paying more attention to the federal funds rate in late 1982. The end date was chosen so the results would not be affected by the 2007-09 Financial Crisis. Thornton (2018) found that – estimates of R 2, which had been fluctuating around 25 percent, dropped to zero in the mid-1990s. While the preceding analysis illustrates the decline in the correlation between the policy rate and the long-term Treasury yield, it does not date precisely when the change occurred. To date the time of the change more accurately, Thornton (2018) used the Andrews (1993) supremum test to identify the most likely date of the change. The test indicated May 1988 as the most likely date. 72 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Coughlin and Thornton Believing the change may have occurred because the FOMC started using the federal funds rate as its policy instrument at about that time, Thornton (2018) turned to the verbatim transcripts of FOMC meetings and found documentary evidence supporting his conjecture. Poole, Rasche, and Thornton (2002) also support the conjecture. They looked at the “Credit Markets” column of the Wall Street Journal published at least two days before the FOMC changed the federal funds rate target and found that May 9, 1988, was the first time market participants were aware the policy action had been taken. Hence, the change in the relationship between the 10-year Treasury yield and the overnight federal funds rate was more likely not due to the “aberrant behavior” of the 10-year yield as Greenspan and others assumed, but rather due to a change in the behavior of the federal funds rate. 4 TESTING THE FRTH FOR ENGLAND AND NEW ZEALAND If the FRTH is correct, the relationship between changes in a central bank’s key interest rate and changes in long-term yields should have weakened substantially when the BOE and the RBNZ each began using those key rates as their policy instrument. Fortunately for our analysis, the timing of when these central banks made the change is well documented, as it is critical for establishing a causal relationship between the change in the implementation of monetary policy and the change in the relationship between changes in the policy rate and changes in the long-term yield. Because these banks are targeting other interest rates, not the federal funds rate, a more accurate term for the hypothesis under investigation is the interest-rate-targeting hypothesis (IRTH). 4.1 The BOE Like the Federal Reserve, the BOE targeted monetary aggregates until the late 1970s. Finding that monetary aggregates were increasingly less reliably connected to output and inflation, the BOE shifted its emphasis to a broad range of economic indicators.2 The BOE’s key rate is the bank lending rate. As with the Fed, the BOE increased the emphasis on the bank lending rate in conducting monetary policy over time. However, during the 1980s the BOE had an exchange rate target that constrained monetary policy. The exchange rate further constrained monetary policy in 1990 when the United Kingdom entered the European Exchange Rate Mechanism. In 1992, the BOE noted that economic conditions in Europe had created tension between setting the interest rate to maintain the exchange rate and setting it as required for the domestic economy. The United Kingdom withdrew from the European Exchange Rate Mechanism in September 1992, and the BOE began using the bank lending rate to implement monetary policy. The IRTH suggests the BOE’s adoption of the bank lending rate as its policy instrument should have produced a marked change in the relationship between changes in the 10-year gilt yield and changes in the policy rate around late 1992. Figure 1 shows for the United Kingdom the estimates – of R 2 from a 50-month rolling regression of changes in the BOE’s policy rate and changes in the 10-year yield government bond yield for the period January 1972 through June 2007.3 The data are plotted on the first month in the sample; the vertical line denotes October 1992. Similar to the Fed, when the BOE began using the bank lending rate as its policy instrument, the relatively strong and statistically significant relationship between changes in the 10-year yield and changes in the BOE’s policy rate declined sharply and virtually vanished in late 1992. The fact that the correlation declined 73 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Coughlin and Thornton Figure 1 50-Month Adjusted R-Squared Estimates from a Rolling Regression of the Change in the 10-Year Gilt Yield on the Change in the BOE’s Policy Rate, January 1972 to June 2007 0.40 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0 –0.05 Feb-72 Feb-74 Feb-76 Feb-78 Feb-80 Feb-82 Feb-84 Feb-86 Feb-88 Feb-90 Feb-92 Feb-94 Feb-96 Feb-98 Feb-00 Feb-02 NOTE: The data are plotted on the first month in the sample; the vertical line denotes October 1992. SOURCE: BOE and authors’ calculations. to zero is a consequence of the fact that, like the FOMC, the BOE changed its target infrequently. This dating was confirmed by the Andrews (1993) break point test, which found October 1992 as the most likely date of the break in the relationship between the BOE’s policy rate and the 10-year gilt yield. 4.2 The RBNZ Until the mid-to-late 1990s, the RBNZ used an eclectic approach to implementing monetary policy (Huxford and Reddell, 1996). In March 1997, the RBNZ proposed implementing policy by targeting the overnight cash rate; however, the policy was not implemented until March 1999. Again, if the IRTH is correct, there should be a marked change in the relationship between the cash rate and the 10-year government bond yield at about that time. – Figure 2 shows for New Zealand the estimates of R 2 from a 50-month rolling regression of the change in the overnight cash rate and the change in the 10-year government bond yield for the period January 1986 through May 2012.4 The data are plotted on the first month in the sample, and the vertical line denotes March 1999. There is a relatively weak and variable relationship between changes in the cash rate and changes in the 10-year yield prior to March 1999. This date was confirmed by the Andrews (1993) test, which determined March 1999 as the most likely date of the – change. However, consistent with the IRTH, the estimate of R 2 dropped to zero a few months before 74 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Coughlin and Thornton Figure 2 50-Month Adjusted R-Squared Estimates from a Rolling Regression of the Change in the New Zealand 10-Year Government Bond Yield on the Change in the RBNZ Cash Rate, January 1986 to May 2012 0.45 0.40 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0 –0.05 Feb-86 Sep-87 Apr-89 Nov-90 Jun-92 Jan-94 Aug-95 Mar-97 Oct-98 May-00 Dec-01 Jul-03 Feb-05 Sep-06 Apr-08 NOTE: The data are plotted on the first month in the sample; the vertical line denotes March 1999. SOURCE: RBNZ and authors’ calculations. March 1999. It began to increase in mid-July 2003, peaking at over 40 percent in December 2004 before declining dramatically and returning to essentially zero by March 2008. – This dramatic rise and fall in the estimate of R 2 is entirely due to five observations from September 2008 through January 2009 and to the fact that ordinary least squares is very sensitive to outliers. Figure 3 shows the change in the cash rate and the change in the 10-year yield from March 1999 through May 2012. The two rates moved independently except for the five noted observations, when the rates moved together. When the equation is estimated over the period March 1999 through May 2012, the relationship is weak; the estimate of β is 0.13 with a t-statistic – of 1.25 and R 2 of 0.015. The relationship is even weaker when the five observations are deleted. The – estimates of β and R 2 are –0.05 and –0.003, respectively. Hence, as was the case for the Fed and the BOE, the correlation fell to zero—and for the same reason: It occurred after the RBNZ began using the cash rate as its policy instrument. The dramatic change occurred just as the IRTH predicted. Just as with the Federal Reserve and the BOE, the weak but statistically significant relationship between the policy rate and the 10-year government bond yield vanished when the RBNZ began using its policy rate—the cash rate—as its policy instrument. 75 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Coughlin and Thornton Figure 3 New Zealand 10-Year Government Bond Yield and RBNZ Cash Rate, March 1999 to May 2012 1.00 Cash rate 10-Year yield 0.50 0 –0.50 –1.00 –1.50 Mar-99 Jun-00 Sep-01 Dec-02 Mar-04 Jun-05 Sep-06 Dec-07 Mar-09 Jun-10 Sep-11 SOURCE: RBNZ. 5 CONCLUSION Thornton (2018) examined the explanatory power of the funds-rate-targeting hypothesis to explain what became commonly known as Greenspan’s conundrum. He demonstrated that the breakdown in the correlation between changes in the federal funds rate and changes in 10-year Treasury yields was due entirely to the FOMC’s adoption of the federal funds rate as its policy instrument. We extend his line of reasoning by exploring the impacts of the adoption of interest rate targeting by the Bank of England and the Reserve Bank of New Zealand. In each case, the adoption of interest rate targeting is found to be closely related in time to a substantial breakdown in the relationship between the targeted interest rate and the long-term sovereign bond yield. Moreover, the date of the adoption of interest rate targeting varies across countries, so the date of the breakdown of the relationship between the policy rate and the long-term sovereign bond yield associated with interest rate targeting varies across countries. Hence, Thornton’s explanation for Greenspan’s conundrum is not limited to the United States, but rather has general applicability. n 76 Federal Reserve Bank of St. Louis REVIEW . First Quarter 2022 Coughlin and Thornton NOTES 1 See the boxed insert regarding the Bank of Canada’s use of a short-term rate as its policy rate. 2 Bank of England (1995). 3 A 50-month window was used because the sample period was larger and could more easily accommodate a longer window. However, the results were not sensitive to other sizes of the window, such as 30 and 40 months. 4 Once again, the use of 30- and 40-month rolling regressions produced similar results. REFERENCES Andrews, D.W.K. “Tests for Parameter Instability and Structural Change with Unknown Change Point.” Econometrica, 1993, 61, pp. 821-56; https://doi.org/10.2307/2951764. Backus, D.K. and Wright, J.H. “Cracking the Conundrum.” Brookings Papers on Economic Activity, 2007, 2007(1), pp. 293-329; https://doi.org/10.1353/eca.2007.0008. Bank of England. “Money Market Operations Since September 1992.” Quarterly Bulletin, February 1995, pp. 13-14. Borio, C.E.V. “The Implementation of Monetary Policy in Industrial Countries: A Survey.” Bank for International Settlements BIS Economic Papers, August 1997, No. 47; https://www.bis.org/publ/econ47.htm. Goodfriend, M. “Interest Rate Policy and the Inflation Scare Problem: 1979-1992.” Federal Reserve Bank of Richmond Economic Quarterly, 1993, 79, pp. 1-23; https://www.richmondfed.org/-/media/richmondfedorg/publications/research/ economic_quarterly/1993/winter/pdf/goodfriend.pdf. Huxford, J. and Reddell, M. “Implementing Monetary Policy in New Zealand.” Reserve Bank of New Zealand Bulletin, 1996, 59, pp. 309-22. Kim, D.H. and Wright, J.H. “An Arbitrage-Free Three-Factor Term Structure Model and Recent Behavior of Long-Term Yields and Distant-Horizon Forward Rates.” Finance and Economics Discussion Series Working Paper No. 2005-33, Board of Governors of the Federal Reserve System, 2005; https://doi.org/10.2139/ssrn.813267. Poole, W.; Rasche, R.H. and Thornton, D.L. “Market Anticipations of Monetary Policy Actions.” Federal Reserve Bank of St. Louis Review, July/August 2002, 84(4), pp. 65-93; https://doi.org/10.20955/r.84.65-94. Rosenburg, J. “Interpreting the Decline in Long-Term Interest Rates.” Internal memo, Board of Governors of the Federal Reserve System, 2007. Rudebusch, G.D. and Wu, T. “Accounting for a Shift in Term Structure Behavior with No-Arbitrage and Micro-Finance Models.” Journal of Money, Credit, and Banking, 2007, 39(2), pp. 395-422. Smith, J.M. and Taylor, J.B. “The Term Structure of Policy Rules.” Journal of Monetary Economic, 2009, 56, pp. 907-17; https://doi.org/10.1016/j.jmoneco.2009.09.004. Thornton, D.L. “The Borrowed-Reserves Operating Procedure: Theory and Evidence.” Federal Reserve Bank of St. Louis Review, January/February 1988, 70(1), pp. 30-54; https://doi.org/10.20955/r.70.30-54.phh. Thornton, D.L. “When Did the FOMC Begin Targeting the Federal Funds Rate? What the Verbatim Transcripts Tell Us.” Journal of Money, Credit, and Banking, 2006, 38(8), pp. 2039-71; https://doi.org/10.1353/mcb.2007.0010. Thornton, D.L. “Greenspan's Conundrum and the Fed's Ability to Affect Long-Term Yields.” Journal of Money, Credit, and Banking, 2018, 50(2-3), pp. 513-43; https://doi.org/10.1111/jmcb.12469. 77