1-s2.0-s0927537113000602-main.pdf

14
Causal effects on employment after rst birth A dynamic treatment approach Bernd Fitzenberger a, , Katrin Sommerfeld b , Susanne Steffes c a University of Freiburg, IFS, IZA, ROA, ZEW, Germany b University of Freiburg, Germany c Centre for European Economic Research, ZEW, Germany HIGHLIGHTS The causal effect of rst childbirth is estimated by a dynamic treatment approach. The causal effect of childbirth on female labor supply is large and persistent. It decreases over the rst ve years as the child grows, but does not level off. The employment loss is particularly pronounced for medium-skilled mothers. We nd a signicant reduction in the employment loss over time. abstract article info Article history: Received 31 October 2012 Received in revised form 17 May 2013 Accepted 20 May 2013 Available online 2 June 2013 JEL classication: C14 J13 J22 Keywords: Female labor supply Maternity leave Dynamic treatment effect Inverse probability weighting The effects of childbirth on future labor market outcomes are a key issue for policy discussion. This paper im- plements a dynamic treatment approach to estimate the effect of having the rst child now versus later on future employment for the case of Germany, a country with a long maternity leave coverage. Effect heteroge- neity is assessed by estimating ex post outcome regressions. Based on SOEP data, we provide estimates at a monthly frequency. The results show that there are very strong negative employment effects which are caus- ally due to childbirth. Although the employment loss is reduced over the rst ve years following childbirth, it does not level off to zero. We nd a signicant reduction in the employment loss over time. © 2013 Elsevier B.V. All rights reserved. 1. Introduction Childbirth is associated with a strong reduction in female employ- ment right after birth, which in a life-cycle perspective has adverse ef- fects on later labor market outcomes (OECD, 2007; Schönberg and Ludsteck, 2012). Decisions about career objectives and about starting a family have to be taken jointly usually between the age of 20 and 40 years. Empirical estimates of the effects of motherhood on subsequent employment have to address the selection into mother- hood. Our paper contributes to this literature, rst, by suggesting and implementing a dynamic treatment approach to estimate the ef- fect of having the rst child now versus later on future employment. Second, we study the case of Germany, a country with long maternity leave coverage (Schönberg and Ludsteck, 2012). A large literature estimates the causal effects of birth on subse- quent employment of mothers by means of different identication strategies. Earlier quasi-experimental studies for the U.S. estimate the labor supply effects in response to having an additional child (Bronars and Grogger, 1994; Angrist and Evans, 1998). Cristia (2008) estimates the effect of a rst child exploiting the quasi-experimental variation in fertility among women who seek help to become preg- nant. There also exists a large literature on the negative effects of having children on mothers' wages. As the main reason for large wage losses, Waldfogel (1998) identies the long employment breaks after childbirth associated with employer changes. Ejrnaes and Kunze Labour Economics 25 (2013) 4962 Corresponding author. E-mail addresses: bernd.[email protected] (B. Fitzenberger), [email protected] (K. Sommerfeld), [email protected] (S. Steffes). 0927-5371/$ see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.labeco.2013.05.003 Contents lists available at ScienceDirect Labour Economics journal homepage: www.elsevier.com/locate/labeco

Upload: kenny-ng

Post on 29-Sep-2015

216 views

Category:

Documents


2 download

TRANSCRIPT

  • ir

    St

    imated by a dynamic treatment approach.labor supply is large and persistent.he child grows, but does not level off.nounced for medium-skilled mothers.ployment loss over time.

    C14

    Labour Economics 25 (2013) 4962

    Contents lists available at ScienceDirect

    Labour Eco

    e lsJ22

    Keywords:Female labor supplyMaternity leaveDynamic treatment effectInverse probability weighting

    1. Introduction

    Childbirth is associated with a strong reduction in female employ-ment right after birth, which in a life-cycle perspective has adverse ef-fects on later labor market outcomes (OECD, 2007; Schnberg andLudsteck, 2012). Decisions about career objectives and about startinga family have to be taken jointly usually between the age of 20

    and implementing a dynamic treatment approach to estimate the ef-fect of having the rst child now versus later on future employment.Second, we study the case of Germany, a country with long maternityleave coverage (Schnberg and Ludsteck, 2012).

    A large literature estimates the causal effects of birth on subse-quent employment of mothers by means of different identicationstrategies. Earlier quasi-experimental studies for the U.S. estimateJ13 2013 Elsevier B.V. All rights reserved.and 40 years. Empirical estimates of the efsubsequent employment have to address thhood. Our paper contributes to this literat

    Corresponding author.E-mail addresses: [email protected]

    [email protected] (K. Sommerfeld

    0927-5371/$ see front matter 2013 Elsevier B.V. Allhttp://dx.doi.org/10.1016/j.labeco.2013.05.003monthly frequency. The results show that there are very strong negative employment effects which are caus-ally due to childbirth. Although the employment loss is reduced over the rst ve years following childbirth,it does not level off to zero. We nd a signicant reduction in the employment loss over time.JEL classication: The causal effect of rst childbirth is est The causal effect of childbirth on female It decreases over the rst ve years as t The employment loss is particularly pro We nd a signicant reduction in the em

    a r t i c l e i n f o

    Article history:Received 31 October 2012Received in revised form 17 May 2013Accepted 20 May 2013Available online 2 June 2013a b s t r a c t

    The effects of childbirth on future labor market outcomes are a key issue for policy discussion. This paper im-plements a dynamic treatment approach to estimate the effect of having the rst child now versus later onfuture employment for the case of Germany, a country with a long maternity leave coverage. Effect heteroge-neity is assessed by estimating ex post outcome regressions. Based on SOEP data, we provide estimates at aH I G H L I G H T SCausal effects on employment after rst b

    Bernd Fitzenberger a,, Katrin Sommerfeld b, Susannea University of Freiburg, IFS, IZA, ROA, ZEW, Germanyb University of Freiburg, Germanyc Centre for European Economic Research, ZEW, Germany

    j ourna l homepage: www.fects of motherhood one selection into mother-ure, rst, by suggesting

    rg.de (B. Fitzenberger),), [email protected] (S. Steffes).

    rights reserved.th A dynamic treatment approach

    effes c

    nomics

    ev ie r .com/ locate / labecothe labor supply effects in response to having an additional child(Bronars and Grogger, 1994; Angrist and Evans, 1998). Cristia (2008)estimates the effect of a rst child exploiting the quasi-experimentalvariation in fertility among women who seek help to become preg-nant. There also exists a large literature on the negative effects ofhaving children on mothers' wages. As the main reason for largewage losses, Waldfogel (1998) identies the long employment breaksafter childbirth associated with employer changes. Ejrnaes and Kunze

  • and 1992 led to a successive rise in the duration of employment inter-ruptions.1 Schnberg and Ludsteck (2012) estimate causal effects byapplying a regression discontinuity design. The advantage of this ap-proach is the random assignment of the treatment, as couples did not

    50 B. Fitzenberger et al. / Labour Economics 25 (2013) 4962(2013) and Schnberg and Ludsteck (2012) use German policy re-forms to investigate the causal effect of births on female wages in Ger-many. In an approach, which is more closely related to our study,Simonsen and Skipper (2006) match nonparametrically mothers andnon-mothers for Denmark based on observable exogenous covariates.The authors account for the effects ofmotherhood on some observablecharacteristics and distinguish between the direct effect holding thesecharacteristics constant and the indirect effects operating through thechanges in these characteristics. Most of the wage costs of mother-hood are found to operate through the indirect effects.

    While the aforementioned studies provide evidence on the negativeeffect of childbirth versus not having a child on subsequent labor mar-ket outcomes, these studies do not address the timing of birth. In con-trast, Miller (2011), Wilde et al. (2010), and Troske and Voicu (2010,2012) investigate the effects of timing of rst birth on labor supplyand wages of mothers in the U.S.

    Troske and Voicu (2011, 2012) analyze the effect of timing andspacing of births on employment by estimating a multinomial probitmodel for different employment states and for the fertility decision.Identication is mainly driven by including the number of childrenof the mother's siblings, based on a tight specication of the timingof the decision process. Furthermore, the authors restrict the analysisto married women who have a rst child after marriage. The estimat-ed model contrasts the employment outcome of mothers with theemployment outcome of childless women. Troske and Voicu (2012)nd that delaying the rst birth leads to higher levels of labor marketinvolvement before the birth of the rst child and reduces the nega-tive employment effects after birth.

    Miller (2011) and Wilde et al. (2010) estimate the effect of higherage at rst birth on subsequent labor market outcomes, using womenwho have a child at a later age as the control group for women whohave a child at an earlier age (see also Herr, 2011, for a relatedstudy). Miller (2011) instruments age at rst birth by naturally occur-ring fertility shocks. Wilde et al. (2010) criticize the instruments usedby the former study, especially with regard to the lack of exogeneity inthe case of time-varying instruments and to measurement errorproblems. Instead, their instruments involve characteristics of parentsor events having occurred at an early age. All these studies nd thatdelaying rst birth increases employment and wages after birth. Thehigher career costs of childbirth may explain the greater tendency ofhigh-skilled women to delay birth or to avoid having a child alto-gether. Correspondingly, Miller (2011) nds that delaying birthshows a stronger positive effect for high-skilled women compared toless-skilled women.

    Why is the case of Germany interesting? In light of the long mater-nity leave coverage of 36 months of job-protected leave, there hasbeen growing concern in Germany that mothers do not return totheir job (quickly) and that their careers are actually negatively affect-ed by such long employment interruptions. Over time, social normsseem to have changed, now favoring employment of mothers withyoung children, and policy changes are asked for which allow for areconciliation of employment and family tasks. Firms take effort tobalance work and family roles, fathers have increased their involve-ment in informal childcare, and the availability of formal childcarefor children below the age of three has increased strongly. Policymakers implemented a policy reform in 2001 providing nancial in-centives for an earlier return-to-job after childbirth and fosteringpart time work when the child is still young. Another reform wasimplemented in 2007 with the intent to foster female labor supplyas well as fertility.

    Until now little is known about the magnitude of longitudinal em-ployment losses after rst childbirth in Germany and their changesover a longer time period. Most of the existing studies for Germanyanalyze the impact of policy changes in maternity leave coverage.Ondrich et al. (2003) and Schnberg and Ludsteck (2012) nd that in-

    creases of the maximum duration of maternity leave between 1986know about the reform at the time of conception. However, one maybe concerned about the external validity of these results in light of thepossibility that individuals adjust their behavior over time when aninstitutional change is associated with changes in social norms (seeFortin, 2009, for the importance of gender role attitudes).

    The timing of rst birth can have a strong impact on future careeroutcomes (Miller, 2011; Wilde et al., 2010). Thus, women of a certainage decide whether to have a child now or wait with the expectednegative career effects varying by age. This raises two questions tobe addressed by our study: 1) Does the causal impact of rst birthdiffer with the age of the mother at childbirth? 2) Do we nd differ-ences in the impact of rst birth with respect to personal characteris-tics, especially regarding career-related decisions before childbirth?Are these differences in accordancewith human capital theory? Clear-ly, the employment behavior in the rst years after childbirth dependson personal circumstances and institutions (e.g. regarding the avail-ability of childcare, see Kreyenfeld and Hank, 2000), but also onpre-birth employment and education. Furthermore, the fertility deci-sion is based on the expected employment chances after childbirth.Women with higher human capital or higher labor force attachmenthave the most to lose from an employment break in terms of humancapital and, thus, career opportunities (Miller, 2011).Wewould there-fore expect older and more educated rst-time mothers to catch upmore quickly to the control group than younger or less educatedrst-time mothers.

    Our paper implements a dynamic treatment approach, which doesnot require a complete modeling of the fertility and labor marketdecisions. We estimate the causal effect of rst childbirth on femaleemployment over the course of ve years after birth. Our empiricalanalysis uses data on rst childbirth between the age of 24 and 33from the German Socio-Economic Panel Study (SOEP) for the timeperiod 1991 to 2008, as the SOEP data provide not only detailedmonthly information on employment, but also a large set of controlvariables.

    Our estimation approach has been used before to estimate the ef-fect of active labor market policies (Sianesi, 2004, 2008; Fitzenbergeret al., 2013) in a situation, where nontreated individuals today maybe treated in the future. We address the selection into childbirthby contrasting rst-time mothers with highly comparable womenwho display a similar propensity to have a rst child soon. We assumethat individuals decide whether to have a child now or wait. Afterconditioning on labor market history and personal characteristics,the exact timing of birth is assumed random, which is the basis ofour identication strategy. We cannot estimate the causal effect ofhaving a child versus never having a child or of having a child at ahigher age versus at a lower age (delay childbirth). We align treatedand control individuals dynamically by age, and we match treatedand controls using inverse probability weighting (IPW) with weightsnormalized to sum up to one (Busso et al., 2009). Finally, we employex-post outcome regressions (Abadie and Imbens, 2011) to analyze theheterogeneity of the estimated employment effects. We investigate towhat extent the estimated effect changes over time, possibly in re-sponse to changing social norms towards employment of mothers.

    Our results show a sharp decline in employment around child-birth. The average treatment effect on the treated slowly convergesto zero, implying that employment rates recover as the child growsolder. However, we also show that within the observation period of

    1 A similar result is found by Lalive and Zweimller (2009) for Austria and by Baker

    and Milligan (2008) for Canada.

  • ve years, the estimated average treatment effects on the treated donot level off. We nd a signicant reduction in the employment lossover time.

    The paper proceeds as follows. The institutional background in

    Euro plus 3140 Euro for each additional child. Furthermore, familiesreceive a monthly child allowance (Kindergeld) or a tax exemption

    3. Econometric approach

    3.1. Dynamic treatment approach

    Our goal is to estimate the average treatment effect for the treated(ATT) on employment based on discrete time data. The treatment is

    51B. Fitzenberger et al. / Labour Economics 25 (2013) 4962for each child. Moreover, at exactly the same date as the reformdiscussed so far (January 2001), there was another political reformwhich facilitated part time work, regardless of having children ornot. The latter reform established a legal claim for part time workand regulated xed term contracts (Teilzeit- und Befristungsgesetz).

    2 The post-birth break is mandatory whereas for the pre-birth period, women canGermany is described in Section 2. Section 3 develops the econometricapproach and Section 4 describes the data used. The empirical resultsare discussed in Section 5. Section 6 concludes. An additional onlineappendix (available at www.empiwifo.uni-freiburg.de/discussion-papers-1) includes further detailed information and results.

    2. Institutional background

    Maternity leave coverage in Germany, which provides employ-ment protection after childbirth, was extended to three years in theearly 1990s, which is very generous in international comparison(Schnberg and Ludsteck, 2012). However, there is concern that thelong maternity leave coverage is one of the reasons for the low em-ployment rates of German mothers (Schnberg and Ludsteck, 2012).As a result, the German government changed the rules that regulateincome support during maternity leave in 2001 and in 2007 in orderto increase the incentive for mothers to return to their jobs sooner.

    Two different laws govern maternity leave coverage (Schnbergand Ludsteck, 2012), both of which concern the job protection andthe regulation of nancial benets, but apply to different time pe-riods. The Maternity Protection Law (Mutterschutzgesetz) requiresa pregnant woman to temporarily stop working six weeks before andeight weeks after birth in order to protect her and the baby's health.2

    During this period, the mother receives her full net labor incomewhich is nanced partly by the employer and partly by the healthinsurance.

    The Parental Leave Law (Elternzeitgesetz) regulates the job pro-tection period and the income assistance during the parental leave pe-riod, which is typically takenmainly by themother. The job protectioninvolves the right to return to a comparable job at the previous em-ployer, but this does not necessarily have to be exactly at the sameworkplace as before birth. The job protection period is three years ofwhich up to twelve months may be delayed until the child reachesthe age of eight. Effects of the extensions of maternity leave coveragefrom the 1970s to the early 1990s on children outcomes are analyzedby Dustmann and Schnberg (2012). According to the Parental LeaveLaw, parental benets are paid only if the parent on leave doesnot work more than the allowed amount of hours, lives in the samehousehold and predominantly cares for the child by him- or herself.Before 2001, a parent on parental leave had been allowed to work atmost 19 hours a week. In January 2001, this upper boundwas extend-ed to 30 hours. In addition, the parent on leave if eligible was giventhe right to choose between receiving 300 Euro per month during therst 24 months or 450 Euro per month during the rst 12 months.This should provide incentives for a return-to-job after 12 or24 months. The maternity benet was means-tested, i.e. it was paidto families with an annual net income of less than 30,000 Euro intwo-parent households and 23,000 Euro in single-parent households.Six months after birth these boundaries decreased to 16,500 (13,500)apply for exemption.rst childbirth at a certain age against the alternative of waiting.The alternative of waiting, i.e. of not having a rst childbirth now, en-tails both the possibility of never having a child and the possibility ofhaving a child at a later age. This treatment effect has been suggestedby Sianesi (2004, 2008) in the context of estimating the effects ofactive labor market programs. Thus, for a certain age, we estimatethe ATT of having a rst child at this age versus not having a child atthis age. Since the observed and unobserved characteristics of treatedand untreated women may change with age, the approach does notallow us to estimate the causal effect of delaying childbirth. Thegroup of nontreated women consists of both women who neverhave a child andwomenwho have a child at a later age. This denitionof the control group avoids conditioning on future outcomes. Treatedand nontreated women at a certain age did not have a child beforethe considered age. Furthermore, we match treated and nontreatedwomen by observable characteristics twelve months before birth.

    The treatment effect we estimate is an example for the dynamictreatment approach applied in the context of program evaluation ofactive labor market policies by Sianesi (2004, 2008) for Sweden orby Fitzenberger et al. (2013) for Germany. In these two countries, un-employed individuals who have not been treated at a certain point intime may receive treatment later. The exit from unemployment toemployment and the entry into program participation are two com-peting risks, and unobserved characteristics are typically correlatedwith the chances of nding a job (quickly) as well as with programparticipation. Correspondingly, Sianesi (2004, 2008) and Fredrikssonand Johansson (2008) argue that excluding future participants fromthe control group would lead to biased estimates of the treatment ef-fect due to selection effects.

    The key identication assumption for our analysis is a dynamicconditional independence assumption (a sequential randomizationassumption). It states that conditional on the variables controlledfor until a certain time period the assignment to treatment in thistime period is random, i.e. independent of the potential outcomes.Our application of the dynamic treatment approach is comparable tothe training literature example insofar as the control group remainschildless up to the considered point in time (similar to unemployedindividuals remaining jobless and/or without training). Therefore,in our application, for women giving birth for the rst time at a certainage, we do not exclude the alternative of giving rst birth at a later age.This corresponds to the dynamic nature of fertility decisions, whichare taken jointly with career decisions. As Ciliberto et al. (2012,p. 10) put it [] employees make their fertility choices many timesin their lives []. Using solely a control group of women who donot give rst birth until a much later age or who will never have achild would bias the control group towards women with a low pro-pensity of having a child. This is due to selection into motherhood.3

    This bias is likely to be correlated with labor market outcomes(e.g. women with a strong unobserved career orientation are morelikely to exhibit a higher labor market attachment and lower fertilityrates).4 The endogeneity problem of fertility and employment alsoarises due to potential labor market shocks and shocks related to per-sonal circumstances (relationships, household characteristics), with

    3 A number of studies use non-mothers for comparison and this way include selec-tion into motherhood in the estimated effect (Ejrnaes and Kunze, 2013; Fernandez-Kranz et al., 2013; Simonsen and Skipper, 2006).

    4 The underlying reason for this type of endogeneity is that both fertility and careerare controlled by women (at least in part, see Troske and Voicu, 2010). For different ap-proaches to address endogeneity in this context, see Miller (2011), Herr (2009, 2011),

    Gustafsson et al. (1996), and Wilde et al. (2010).

  • differential effects on fertility and labor market outcomes. It is plausi-ble that positive labor market shocks are negatively correlated withfuture fertility or that personal shocks, which reduce the fertility inthe future, are also correlated with labor market outcomes.5

    zero).7 Thus, treated and non-treated females are likely to becomeless comparable at a higher age.

    3.2. Inverse probability weighting (IPW)

    To control the selection of treated women, we estimate a sequenceof quarterly propensity scores. The size of the treatment sample is not

    52 B. Fitzenberger et al. / Labour Economics 25 (2013) 4962Our approach assumes that women giving birth to their rst childare comparable before the gestation lag, i.e. at the time before preg-nancy, to women who do not give birth at a certain point in time,i.e. at a certain age. This approach assumes that women do not knowthe exact timing of rst birth before the gestation period. But theymay know the probability of having a rst birth now versus laterand they may act upon the determinants of this probability (Abbringand van den Berg, 2003). Assuming a no-anticipation condition withrespect to the precise date of pregnancy before the gestation period al-lows us to match treated and nontreated women at this date.

    The treatment group in our analysis consists of women who havetheir rst child between the age of 24 and 33 andwho are not disabledor retired. The control group consists of women who did not have achild at the age of the treated mother at the time of birth. We imposean exact alignment of the age in months. We measure both the treat-ment effect and the age of the women at a monthly frequency. For thepurpose of presenting our results, we pool the month-specic treat-ment effects into groups (see Section 5). For example, there shouldbe little difference between a woman who has her rst child at26 years and 11 months and awomanwho has her rst child at exact-ly 27 years. Our analysis estimates an average counterfactual outcomefor each treatedwoman based on the individual-specic group of indi-viduals not treated yet.

    We assume that the rich set of covariates allows us to capturesystematic differences in the propensity to have a rst child at a cer-tain age. Specically, we assume that given the duration of childless-ness and given the covariates, having a rst birth within the next yearis random, i.e. the dynamic conditional independence assumptionholds (DCIA as discussed in Fitzenberger et al., 2013). This random-ness in the exact month of birth is given by nature, as conceptionsare not perfectly controllable events (Hotz and Miller, 1988, p. 91).Furthermore, we assume that the timing within the year considered,i.e. the month of birth, is unrelated to the selection into rst birth.

    The DCIA can bemotivated as follows for the considered age range.Our data allows us to control for a large number of characteristicsregarding education, labor market experience, relationships, and atti-tudes. Note that our data are particularly rich in this regard comparedto administrative data used in the literature (Simonsen and Skipper,2006; Ejrnaes and Kunze, 2013), thus making the DCIA particularlyplausible in our application. One year before birth, the treatedwomen are not likely to differ systematically from those womenwho stay childless until shortly after the birth of the child. The exacttiming of birth cannot be planned with certainty and may dependupon random circumstances not reected in long-run labor marketchoices. It is rather implausible that women plan the exact month ofrst birth more than a year ahead.6 At the same time, women differin their probability to have a child within the next year and this is like-ly to be reected by the characteristics that are being controlled for.We argue that the DCIA is particularly plausible during the age rangeof 24 to 33 years. At age 24, women have realized the major parts oftheir pre-employment human capital investments and most of themhave started their labor market career. In contrast, in their late 30s itis likely that rst time mothers and childless women at the same agebecome less comparable. Because the fecundity starts to drop in the30s and giving birth becomes more risky at a higher age, mostwomen in their late 30s may either be determined to have a child ordecide to stay childless (i.e. the probability to have a child is close to

    5 An extreme example may be a work accident of a woman which prevents her bothfrom working and from having a child in the future.

    6 Cristia (2008) provides a detailed and informative discussion of the inherent ran-

    domness in becoming pregnant even when the women intends to have a child.sufcient for a monthly frequency, and we think that the selectioninto childbirth does not change strongly from month to month. Theprobability of having the rst child within the next year is modeledas a function of human capital and employment history. As sequentiallabor market decisions are correlated (Troske and Voicu, 2010), con-trolling for past labor market career is crucial for successful matching.Moreover, we control for status of the relationship with the partnerand for the self reported importance of having a family.8 We includecovariates of the partner, such as income and education, to controlfor the partner.

    Under the unconfoundedness of the treatment and perfect overlapin the propensity score, Busso et al. (2009) nd that in small sampleswith unknown propensity score, a modied inverse probabilityweighting estimator (IPW) performs best in comparison to variousmatching estimators. This result stands in contrast to the conclusionsobtained by the Monte Carlo study in Frlich (2004). The crucial mod-ication of the IPW estimator involves the normalization of weightsfor the nontreated women.

    Busso et al. (2009, Eq. (7)) suggest to estimate the ATT as follows:

    ^BDM ni1 TiYini1 Ti

    nj1 1Tj

    W^ jYj

    nj1 1Tj

    W^ j1

    with weights W^ j p^ Xj

    = 1p^ Xj

    :

    Ti, Tj denote the treatment dummy variables for individuals i, j(treated and non-treated), respectively, and p^ Xj

    denotes the esti-

    mated propensity score as a function of covariates Xj. The applicationof the weights Wj leads to a reweighting of the nontreated womenaccording to the odds-ratio of having a child within the next year.

    For our application, we have to account for the fact that the groupof eligible comparison women changes by month of age. Correspond-ingly, the alignment between treated and nontreated observationschanges as well by month of age. Thus, we modify the estimator inEq. (1) to

    ^ ni1 Ti Yi

    nj1 1Tj

    W^ i;jYj;age i

    nj1 1Tj

    W^ i;j

    8