It was rumored that on the day President Clinton signed the Personal Responsibility and Work Opportunity Reconciliation Act (PRWORA) of 1996, the Statue of Liberty looked slightly greener than usual.  Someone even said that Lady Liberty threw up.  While the hyperbole associated with this story is obvious, so is the moral; the immigrant provisions of PRWORA represent a backlash against immigrants that is at odds with the fabled inscription of the Statue of Liberty and the perception of the United States as a place hospitable to immigrants.  Federal welfare reform legislation barred future (those arriving after passage of the law) legal immigrants from receiving cash assistance under the Temporary Assistance to Needy Families (TANF) program, as well as most other federal means-tested benefits (e.g., food stamps), for five years, and left it up to states’ discretion whether current legal immigrants would be eligible for such assistance.  Notably, every state but Alabama maintained the TANF eligibility of current legal immigrants and 19 states made state funds available to cover TANF costs associated with future legal immigrants during the five-year period for which they were barred from receiving federal TANF benefits (Zimmerman and Tumlin 1999).  The states’ response to PRWORA, as well as subsequent federal legislation restoring some lost benefits to immigrants, makes clear that our nation does not share a common vision related to the treatment of immigrants.  Therefore, how the welfare reform has affected immigrants is an important area of study.
    Why were immigrants singled out for special treatment?  One reason was that the number of immigrants to the country had been growing rapidly prior to passage of the law, as was the foreign born share of the population.  Naturally, policy makers were concerned about the effect of this large influx of immigrants.  Immigrants may have contributed to the stagnating wages of less skilled workers and worsened wage inequality, which was also growing during this period.  The increase in immigration prior to PRWORA was truly large.  According to the Immigration and Naturalization Service (INS), approximately 13.5 million legal immigrants came to the United States between 1981 and 1996 and perhaps as many as 5 million illegal immigrants also entered the country during this period.  The only other period in which such a large number of immigrants entered the country was between 1900 and 1920.  In sum, the growth in immigration and the potentially adverse effects of that immigration may have created some anti-immigrant sentiment that became manifest in the federal welfare reform law.
    A second reason to treat immigrants differently was that it was seen as a way to address two problems associated with immigration that were directly related to the Aid to Families with Dependent Children (AFDC) program.  Not only was there a large number of immigrants entering the U.S. prior to PRWORA, but the immigrants that were entering were less educated (relative to natives) than previous immigrants reflecting the increase in the number of immigrants from Latin America and Asia (INS 1996).  Moreover, recent immigrants were more likely to use AFDC and other welfare programs than were earlier immigrants and immigrants had higher rates of participation in welfare programs than natives (Borjas 1995, Borjas and Hilton 1996).   This increasingly greater use of public funds erodes the fiscal benefits associated with immigration and led to concerns that immigrants, particularly newer immigrants, were a fiscal drag—i.e., consuming more than they were producing (Borjas and Hilton 1996, Smith and Edmonston 1997).  There was also a debate over whether immigrants were attracted to states such as California and New York that have generous cash assistance benefits (Borjas 1999).  Therefore, barring immigrants from cash assistance and other federal safety net program was seen as a way to save money and increase the fiscal benefits of immigration, and to reduce “adverse” immigration, that is immigration by future welfare recipients.
    The immigrant provisions of PRWORA were not universally supported and future debates over this issue will benefit from evidence of the effect of welfare reform on immigrants.  In this paper, we provide such evidence by investigating the effect of PRWORA on the employment and marriage rates of three groups of low-educated women: foreign-born citizens, foreign-born non-citizens and natives.  The objectives of welfare reform were to reduce dependency on public assistance by encouraging women to work and to change behaviors that create dependency (e.g., non-marital births).  Therefore, our analyses of employment and marriage are central to the evaluation of the success of welfare reform.  In addition, the immigrant provisions of welfare reform grew out of perceptions about immigrant use of cash assistance and were intended to save money by barring some immigrants from receiving benefits.  Our investigation of the effect of welfare reform on foreign-born women, both citizens and non-citizens, will contribute to and expand knowledge about differences between foreign- and native-born women’s use of public assistance and their response to welfare reform.  Importantly, we investigate whether the behavioral response to welfare reform differed by recency of immigration.  And our analysis of the effect of welfare reform on foreign-born non-citizens, a group which consists partly of newly barred legal immigrants, will provide evidence of the cost savings associated with the immigrant provisions of welfare reform.  Finally, because some states created programs to insure that all legal immigrants remained eligible for TANF and others did not, we can compare the response of foreign-born non-citizens (women) between these states to investigate whether there was a “chilling” effect of the immigrant provisions of federal welfare reform legislation.
 In this paper, we also examine the effect of federal welfare reform on immigrants in California and the New York metropolitan region.  These two areas merit special attention for at least two reasons.  First, California and New York State have the largest number of foreign-born residents and most of the foreign-born residents in New York reside in the New York City area.  Second, each state has responded differently to the federal welfare reform law.  California has made cash assistance benefits equal to those obtainable through TANF available to all legal immigrants including those arriving after passage of PRWORA.  New York did not take this step.  Moreover, in spite of its anti-immigration reputation, California has provided more generous benefits to legal immigrants than all other states including New York (Zimmerman and Tumlin 1999).  Thus, a comparison of the effect of welfare reform in these two states provides a natural experiment to evaluate the hypothesis that PRWORA has had a “chilling” effect on immigrant’s use of social services.

PRWORA’s Incentives: Immigrants and Natives
    The federal welfare reform law does not distinguish among citizens by nativity and the time limited benefits and other provisions of PRWORA apply to all citizens.  However, the behavioral response (e.g., changes in employment and marital decisions) to welfare reform may differ between foreign- and native-born citizens.  For example, Borjas and Hilton (1996) show that that foreign-born women have a greater number of spells of welfare receipt and longer average spell lengths than do native-born women.  As a result, welfare reform will have a larger effect on foreign-born citizens (women) than it does on native-born citizens (women).  The reason for this is that the reduction in lifetime benefits associated with PRWORA’s time limits represents a larger change in policy for foreign-born women who expect to receive benefits for a longer period.  Other considerations, however, suggest that foreign-born citizens (women) may have a smaller behavioral response.  Foreign-born women may have inferior labor market opportunities compared to native-born women, say because of language barriers, discrimination and local demand conditions.  In this case, provisions of PRWORA such as the work requirements will not be as utility reducing for foreign-born women relative to native-born women and foreign-born women will be less likely to exit welfare in response to welfare reform (Besley and Coate 1992, 1995).  In general, the behavioral response to welfare reform will differ depending on the underlying cause of welfare participation, which may differ by women’s nativity status.  We examine this issue explicitly in the empirical analysis and in doing so investigate whether there is heterogeneity in the behavioral response of immigrant citizens vis-a-vis natives by recency of immigration.
    A unique aspect of PRWORA is the distinction it makes among foreign-born non-citizens.  Specifically, it creates two eligibility classes among foreign-born non-citizens who entered the country legally.  Legal immigrants who arrived in the U.S. prior to August 22, 1996 were eligible for AFDC and remained eligible for TANF.  Legal immigrants arriving in the U.S. after August 22, 1996 are ineligible for TANF for five years unless they live in one of the 19 states that made state funds available to maintain the eligibility of this group.  Thus, all legal immigrants are affected by welfare reform, but some newly arrived immigrants face the more draconian policy of being denied access to government cash assistance.
    In our empirical analysis, we examine the effect of welfare reform on foreign-born non-citizens.  This group consists of legal and illegal immigrants, although the data do not allow us to distinguish between the two.  Illegal immigrants were never eligible for cash assistance and welfare reform had no effect on their behavior.  Thus, the behavioral response that we are attempting to measure in this analysis is that associated with legal immigrants.  The data allow us to identify new immigrants (i.e., post 1996) and therefore we can test whether new immigrants had a larger behavioral response consistent with the more draconian policy change that they faced.  Finally, since some states maintained eligibility of newly arrived immigrants, we can compare the behavioral response of new immigrants in these states to that of new immigrants in states that did not maintain eligibility to test whether there was “chilling” effect of PRWORA.  There have been many reports that the immigrant provisions of PRWORA have discouraged participation among eligible immigrants who are confused or frightened by the immigrant provisions of the law (Fix and Passel 1999, Zimmerman and Fix 1998).  No difference in the behavioral responses of foreign-born non-citizens in these two groups of states would be evidence consistent with a “chilling” effect since actual eligibility differences did not matter as much as passage of the law.

California and New York
    Our empirical analysis pays special attention to two states: California and New York, or more precisely the New York metropolitan region.  New York and California have the largest number of foreign-born residents and the highest proportion of residents who are foreign born.  Nearly 50 percent of the foreign born population in the U.S. resides in these two states.   In 1996, 25.1 percent of California’s population was foreign born and in New York, 17.7 percent of the population was foreign born.  The two states foreign-born residents differ with respect to their legal status: 16 percent of New York’s foreign-born population was undocumented whereas the similar figure for California was 25 percent.  Obviously, the effect of welfare reform on immigrants in these two states will largely determine the effect nationally and thus these two states are central to our analysis.
 New York and California are also interesting because of the different administrative responses to welfare reform.  The devolution of welfare policy to the states and the immigrant provisions of the legislation resulted in significant state variation in policy.  Zimmerman and Tumlin (199) present a thorough description of this variation and conclude that California provides the most generous benefits to legal immigrants and does the most to offset any adverse consequences of PRWORA on legal immigrants.  For example, California extended TANF, SSI and Medicaid benefits to legal immigrants barred from receiving federal dollars.  In contrast, New York and New Jersey, the two states that are relevant to the New York metropolitan region, as we have defined it, provided no such benefits.  Interestingly, California’s generosity comes in spite of it being the most aggressive state in denying benefits to undocumented immigrants.  A comparison of the effect of welfare reform in these two areas, particularly with respect to the group of foreign-born non-citizens where state differences in policy are greatest, will provide important evidence that can be used to evaluate hypotheses related to the “chilling” effects of welfare.  The impact of welfare reform on immigrant non-citizens should be smaller in California, with its more generous eligibility policies, than in New York.  A rejection of this hypothesis would suggest significant “chilling” effects of the law since it is evidence that actual eligibility does not matter as much as the passage of the law.

Previous Research
    To our knowledge, there has been one previous study of the effect of welfare reform on immigrants.  Fix and Passell (1999) compare changes in welfare participation between 1994 and 1997 among citizens and non-citizens.  They found that non-citizen participation in welfare, defined as participation in AFDC,/TANF, SSI and General Assistance, declined significantly more than did citizen participation.  These authors concluded that most of the decline (relative to citizens) is due to a “chilling” effect since very few of the non-citizens were ineligible for benefits during the period studied.
    The major weakness with this study is it’s simple before and after design.  Changes in welfare participation rates may be driven by a variety of factors, for example macroeconomic changes, and this study does not control for these factors.  Specifically, citizens and non-citizens may face different economic conditions or may have different responses to equivalent economic changes.  Indeed, the spatial concentration of non-citizens strongly suggests that citizens and non-citizens face different labor market conditions.  Therefore, it is not clear whether welfare reform is the cause of the relative (to citizens) decline in welfare participation.  A second weakness of the study is that the data extend only through 1997 and federal welfare reform had not been widely implemented by that time.
    In contrast to this study, our analysis controls for several factors that may affect decisions about work and welfare participation.  Most importantly, we include explicit and implicit controls for macroeconomic changes in the economy.  We also use data that extend through 1999 by which time federal welfare reform had been completed.  Finally, we do not examine welfare participation, but rather the determinants of welfare participation: employment and marital status. While the outcomes are linked, changes in the caseload are not the converse of changes in employment or marital status.  It is interesting from a policy point of view to investigate how much of the change in the caseload can possibly be explained by changes in employment or marriage.
 
Research Strategy
    To estimate the effect of welfare reform on employment and marriage, we use a quasi-experimental research design commonly referred to as a difference-in-differences (DD) analysis.  The DD procedure compares the change over time in employment and marriage of a group affected by welfare reform, the target group, to the change over time in employment and marriage of a similar group that is unaffected by welfare reform, the comparison group.  The underlying logic of the DD methodology is illustrated in Table 1, which refers to employment, but a similar logic is applied to the analysis of marital outcomes.

In Table 1, the difference B-A measures the change in employment of the target treatment group, the group affected by welfare reform, before and after welfare reform.  This difference may be due to welfare reform and other factors that change over time.  The difference D-C measures the change in employment before and after welfare reform of the comparison group, a group similar to the target group, but unaffected by welfare reform.  Changes in the employment of the comparison group are due to other factors since this group unaffected by welfare reform.  Thus, the difference-in-differences, (B-A)-(D-C), measures effect of welfare reform on the target group.  The effect of other factors have been eliminated by subtracting the before and after change of the comparison group from the before and after change of the target group.  Obviously, a critical assumption of the DD procedure is that changes in employment caused by other factors are the same for the target and comparison groups.
The DD analysis can also be cast in a regression framework.  For example, the regression specification that corresponds to

Table 1 is as follows: Click here to view the regression.
(1)  where Empikt is an indicator of whether woman ‘i’ with marital status ‘k’ in year ‘t’ is employed.  The variable Treatk in equation (1) denotes membership in our target group and therefore equals one if woman ‘i’ is unmarried with 12 or fewer years of education, and zero otherwise.  Reformt is an indicator equal to one if it is an observation taken after welfare reform.  The key parameter in equation (1) is ?3, which is the DD estimate corresponding to (B-A)-(D-C).
As written, the DD estimate obtained from a regression using the equation (1) specification would equal the estimate obtained by the subtractions in Table 1.  This simple specification of the regression model generates no advantage over the differences in means in Table 1.  A more complex specification of the regression model that includes controls for personal characteristics, unmeasured state effects and unmeasured time effects does have some advantage.  Notably, DD estimates obtained from such a model will be more precise and there can be a more refined specification of time effects vis-a-vis policy effects.  In practice, we estimate a model similar to equation (2) below. Click here to view equation 2.
(2)
Equation (2) includes controls for state effects (?j), year effects (?t), personal characteristics (Xijkt) such as age, race and recency of immigration (for foreign born), and state level variables (Zjt) such as the unemployment rate.  Equation (2) also reflects a more general specification of welfare reform by differentiating between AFDC waivers and TANF.  This is particularly important because of the special immigrant provisions of TANF.  Coefficients ?4 and ?5 measure the effect of AFDC waivers (?4) and TANF (?5) on the employment of the target group holding constant unmeasured time-varying factors correlated with welfare reform.
    One strength of the difference-in-differences analysis is that it controls for time variation in outcomes that is unrelated to welfare reform, for example due to macroeconomic changes.  Clearly, this statement is correct only if the comparison group is appropriate.  Assuming this to be the case, it implies that it is not as crucial in a DD analysis to control for macroeconomic activity as it is in other approaches.  To be cautious, however, our regression model includes a control for macroeconomic activity, specifically the state unemployment rate in the month of the survey.  We also interact the unemployment rates with the dummy variable indicating membership in the target group.
    As noted, a crucial aspect of the DD analysis is the validity of the comparison group.  For the employment analysis, we define the target group to be unmarried women with 12 or fewer years of education.  These women are a reasonable target group since many of them are at risk of welfare receipt.  Education and marital status are strong correlates of welfare participation and a large portion of the AFDC/TANF caseload consists of low-educated, unmarried women.  Indeed, approximately 80 percent of the caseload in the early 1990s consisted of women with 12 or fewer years of education and an equally large proportion were unmarried (Kaushal and Kaestner 2000).  The comparison group corresponding to this target group is married women with 12 or fewer years of education.  Eligibility for AFDC/TANF is largely determined by family composition and there is a large difference in welfare participation rates between low-educated married women and low-educated unmarried women.  While some low-educated married women are surely at risk of welfare participation, the majority is not.   CPS data show that approximately 20 percent of the welfare caseload is married.  The inclusion of women at risk of welfare receipt in the comparison group will bias our estimates toward zero with the size of the bias depending on the difference in the proportion of at risk women in the target and comparison groups.  For example, if 60 percent of the women in the target group are at risk of welfare receipt and 20 percent of the women in the comparison group are at risk of welfare receipt, the DD estimate will be 40 percent of the true estimate.
    In order to gauge the sensitivity of estimates to our choice of comparison group, we estimate equation (2) using an alternative comparison group.  Our second comparison group is unmarried women with 13 to 15 years of education.  The problem with using this group is that a significant proportion of this group, for example those who are unmarried with children, are at risk of welfare receipt.  Thus, the DD estimates using this group are likely to be severely downward biased.  It is also the case that marital status is a more important determinant of employment than education (see Table 1), and thus it is more likely that married women with similar education levels will be a more appropriate comparison group than the unmarried women with more education.  We do not include unmarried women with 16 (BA) or more years of education in the alternative comparison group because it is unlikely that their labor market experiences will be similar to that of women with 12 or fewer years of education.  In sum, we believe that the married women comparison group is superior, but we present estimates using the alternative group for the reader’s benefit since the validity of the comparison group is non-verifiable.
    A difference-in-differences approach is also used to examine the effect of welfare reform on marital status.  For this analysis, we define the target group to be all women with 12 or fewer years of education and the comparison group to be all women with 13 to 15 years of education.  Obviously, since the outcome is marital status, we cannot stratify on the basis of this variable.  Therefore, we are relegated to using education to classify women into those most and least likely to be affected by welfare reform.  This will undoubtedly increase the bias of the DD estimate, but there are few feasible alternatives.

Data
    The data used in the study is the Current Population Survey – Outgoing Rotation Group File for the years 1994 to 1999.  The Outgoing Rotation Group File is a 25 percent sample from each monthly Current Population Survey (CPS) file and contains information on nativity status, citizen status, and recency of immigration, as well as information on important welfare related outcomes such as employment and marital status.  Importantly, the Outgoing Rotation Group File (ORG) does not include information on fertility or welfare receipt and so we do not examine the effect of welfare reform on these outcomes.  This ORG does, however, have information on other individual characteristics such as age, education level, ethnicity, race, and place of residence.  We can use these in the regression analysis to increase the precision of the DD estimates.
One of the most important characteristics of the ORG data for the purposes of this study is the relatively large number of observations in these data.  The March CPS file contains more comprehensive information about respondent’s (e.g., welfare receipt) than the ORG, but it does not have sufficient number of observations to carry out analyses separately by nativity and citizen status.  This stratification is essential to the goals of this study.  Similarly, the March CPS does not have sufficient observations to do state- or metropolitan-region level analyses, even in the largest states such as California and New York.  In contrast, the Outgoing Rotation Group File has an adequate sample size for these purposes.
    We focus on three samples of women: native-born citizens, foreign-born citizens and foreign-born non-citizens.  In all cases, we limit the sample to women between the ages of 18 and 54 years of age because very few women over age 54 are at risk of welfare receipt.  We exclude younger women because we use education to stratify the sample and there is little variation in education among women below age 18.  We also exclude from the analysis women with more than 15 years of education.  These women are not likely to be at risk of welfare receipt and their labor market and marital experiences are not likely to be comparable to those of the low-educated women in our target groups.  For native-born citizens, we select a 25 percent random sample from the ORG to reduce the computational burden associated with the large number of observations for this demographic group.  The 25 percent sample provides sufficient sample sizes.
    Information about state level policies related to welfare reform was merged to the individual level data.  Policy variables are measured as of the date they became effective.  We used a variety of data sources to define the policy variables including the 1999 CEA report, information reported in Schoeni and Blank (2000), and data collected by the National Governors Association.  We use two broad categories of reform: AFDC waivers and TANF.
    Finally, for the analysis of the New York Metropolitan area, we selected observations from New York City and several surrounding areas: Bergen-Passaic PMSA, Jersey City PMSA, Nassau-Suffolk PMSA, New York PMSA, Newark PMSA, and Orange County PMSA.  These areas were selected because they all have relatively high concentrations of immigrants and are part of a common metropolitan area and labor market.  For California, we selected all observations from that state.  For both New York and California, we did not take a 25 percent random sample of the native-born sample, but instead used all sample observations.
    Table 1 provides descriptive information about the employment rates of the samples and the sample sizes of our target and comparison groups for the U.S. as a whole.  Among the foreign born groups, the sample sizes are not large, particularly for foreign-born citizens, but are sufficient to detect reasonable sized effects.  For example, an estimate of the effect welfare reform on employment is simply the difference in mean employment before and after welfare reform, or
(3). Click here to view this.

The variance of the effect is simply the variance of the difference in mean employment before and after welfare reform:
(4) .Click here to view this.

    In equation (4), ?2 is the variance of the binomial outcome of employment, which we assume is equal to 0.24 (implying mean employment of 0.60).  Equation (4) also assumes that there are an equal number of observations before and after welfare reform and that the variance of employment is equal in each period.  To detect a significant effect, the estimate has to be 1.96 times larger than its standard error, or
 (7) Click here to view this.

    The necessary sample size depends on the size of the true effect.  So, to detect a true effect of 0.05, or a five percentage point change in the employment rate, the required sample size is 1475 (2N in equation 7).  The preceding calculation is meant to be more illustrative than definitive, for example this sample size calculation ignores type II errors and is based on a tw0-tailed (1.96) test, but it clearly demonstrates that the sample sizes in Table 1 are sufficient.
    The second point to note about Table 1 is the relatively similar employment rates of our target and comparison groups, particularly when marital status and education are used to define target and comparison groups.  For example, the employment rate of unmarried women without a high school degree is 0.436; the similar figure for married women without a high school degree is 0.499.  Table 1 also indicates that education is a stronger correlate of employment than is marriage.  The similarity of means between the target and comparison groups is important because it provides initial evidence that the two groups have similar labor market experiences and is consistent with the assumption underlying the DD analysis that variation over time in employment outcomes is also similar between the two groups.  The greater similarity of the target and comparison group means when marital status and education is used to define these groups, as compared to when education is used to define the groups, suggests that married women with similar education are a better comparison group than unmarried women with more education.

Results - Effects of Welfare Reform on Employment in U.S.
    Table 2 presents the DD estimates of the effect of AFDC waivers and TANF on the employment of unmarried women.  Analyses were done separately for three groups: native-born citizens, foreign-born citizens and foreign-born non-citizens.  The top panel of Table 2 presents estimates obtained using married women with similar education levels as the comparison group.  The bottom panel presents estimates obtained using more educated unmarried women as the comparison group.  Each row (within column) of Table 1 presents estimates from a separate regression.  So row 1 presents DD estimates of the effect of AFDC waivers and TANF on unmarried women with 12 or fewer years of education, and row 2 presents similar estimates for unmarried women with 12 years of education (i.e., high school degree).  We do not present separate estimates by education level for foreign-born citizens because of small sample sizes.  All estimates were obtained by ordinary least squares (OLS) regression using White’s (1980) correction for heteroscedasticity.
    Estimates in the top panel of Table 2 indicate that TANF increased the employment of low-educated unmarried native-born women in the U.S. by approximately four percentage points, which represents a relative effect of about six percent.  The effect of TANF was slightly larger for women without a high school degree than it was for women with a high school degree.  In contrast to TANF, AFDC waivers had little effect on the employment of native-born low-educated unmarried women.  Similar estimates using other data sources and methodologies have been reported by by Kaushal and Kaestner (2000) and the Council of Economic Advisers (1999).
   While the TANF effect may seem too small to be consistent with the large declines in AFDC/TANF caseloads, it should be recognized that this is a downward biased estimate of the effect of TANF on those who are actually at risk of welfare receipt.  Only a portion of the target group is truly at risk of welfare receipt and some portion of the comparison group is also at risk.  Thus, the four percentage point increase in the employment rate associated with TANF may mask a much larger increase in the employment rate of women truly at risk.  However, four percent of all low-educated unmarried native-born women represents approximately 500,000 women and suggests that employment increases related to TANF may have accounted for a large part of the decline in the welfare caseload.  According to CPS data, between 1994 and 1999, the number of unmarried women with 12 or fewer years of education who were receiving public assistance declined by 1,280,818.  Thus, we estimate that up to 38 percent of the decline in the caseload during this period among this group was due to increased employment related to TANF.   Obviously this is a rough estimate of the TANF-induced employment related decline in welfare caseloads, but it makes the point that the relatively small estimates in Table 2 can account for a significant portion of the change in the welfare caseload.
Estimates in the top panel of Table 2 also indicate that the employment of foreign-born women was also affected by welfare reform.  For these women, AFDC waivers and TANF had approximately the same size effect on employment and the magnitudes of the estimates did not differ greatly between foreign-born citizens and foreign-born non-citizens.  AFDC waivers and TANF are associated with between a five and eight percentage point increase in employment.  In relative terms, the largest effects are those pertaining to foreign-born non-citizens because they have the lowest levels of mean employment.  However, it is difficult, and perhaps inappropriate, to compare estimates across the three samples because the quality of the target and comparison groups may be sample specific.  If that is the case, the size of the downward bias associated with the DD estimates will be sample specific.  Larger estimates associated with the foreign-born women may simply reflect the fact that the target and comparison groups in these samples are less contaminated.  For example, a significant portion of the foreign-born non-citizen sample is undocumented and ineligible for AFDC or TANF benefits.  Therefore, this group is unaffected by welfare reform and thus the target group in this sample may consist of fewer women at risk of welfare receipt.  This would lead to a DD estimate that is more biased toward zero.  If this was the case, the larger estimates associated with the foreign-born non-citizens would unambiguously imply larger behavioral response.  The size of the bias for each sample, however, is unknown and not easily determined.  If we ignore this possibility, the estimates in Table 2 suggest that foreign-born low-educated unmarried women are more responsive to welfare reform than similar native-born women.
    Before moving on to discuss other results, it is necessary to comment on the estimates in the bottom panel of Table 2.  Few of these estimates are significant although many are positive.  In general they differ greatly from those in the top panel.  We believe this result stems from the inappropriateness of the comparison group used to obtain the estimates.  As we argued above, the differences in mean employment rates between this comparison group and the target group suggests that the underlying assumption of the DD analysis may not be valid.  In addition, the downward bias is likely to be greater for this group because of the composition of the target and comparison groups.  Therefore, we believe the estimates in the top panel are more credible and prefer to emphasize these estimates in our discussion.  Because there is no definitive way to resolve this issue, we present both sets of estimates throughout the paper, but we discuss only those obtained using married women with similar education as the comparison group.
    An important issue in the immigration literature is assimilation, or how fast immigrants start acting like native-born persons.  To address this issue we allowed the effect of welfare reform among foreign-born non-citizens to differ by the recency of immigration.   A similar analysis could not be performed for foreign-born citizens because of small sample sizes.  Table 3 presents the estimates from this analysis.  Estimates in Table 3 indicate that the effect of welfare reform, both AFDC waivers and TANF, was larger for more recent immigrants as compared to earlier immigrants.  Importantly, these models hold constant the effect of recency of immigration on the employment level so our estimates are not confounded by the effect of unmeasured characteristics associated with recency of immigration.  However, the immigrant provisions of TANF may have altered the composition of new immigrants and the estimates in Table 3 related to the newest immigrants may be a combination of a behavioral response and a selection effect.  Both effects are due to TANF and are therefore legitimately a TANF effect, but it is a factor to consider when comparing behavioral responses across groups in Table 3.  But this point reinforces the earlier one that it is difficult to know whether the larger estimated effects are due to a larger behavioral response by some groups or because of differences in the quality of the target and comparison groups.  As discussed above, differences in the behavioral response to welfare reform stem from different underlying causes of program participation.  Estimates in Table 3, if we assume that they represent different behavioral responses, suggest that recency of immigration signals something about the reasons why low-educated immigrant women require cash assistance.  Identifying those reasons is beyond the scope of this analysis because of data limitations, but future research may want to address this issue.
    Another issue related to immigrants and welfare reform that has received much attention is the “chilling” hypothesis.  In brief, the “chilling” hypothesis suggests that the anti-immigrant provisions of PRWORA have caused even eligible immigrants to forgo benefits.  To test this hypothesis, we exploit the variation in state policies that offset the PRWORA immigrant regulations.  Specifically, 19 states continued to offer TANF-like benefits to legal immigrants arriving after 1996.  Thus, the effect of TANF on new immigrants in these states should be smaller than the effect of TANF on new immigrants in states that cut off benefits to this group.  For this analysis, we allow the effect of TANF to differ by whether the state did or did not make TANF benefits available to new immigrants.  Again, it is important to note that recency of immigration is being held constant.  Thus, we are comparing the employment of recently arrived immigrants after TANF to the employment of recently arrived immigrants before TANF.  We allow that effect to differ by whether new immigrants were eligible for TANF.
    Table 4 presents the estimates of this analysis and they are revealing.  In states that made TANF benefits available to new immigrants, TANF had no effect on these new immigrants employment.  In contrast, TANF had a large effect on the employment of new immigrants in states that cutoff TANF benefits.  These results are inconsistent with the “chilling” hypothesis, which would suggest more equally sized effects.  The smaller estimates associated with the effect of TANF on immigrants who immigrated before 1996 are also inconsistent with the “chilling” hypothesis.  Again, the “chilling” hypothesis suggests more equal estimates since the important aspect of the law was the debate and controversy over the immigrant provisions that may have frightened eligible immigrants from obtaining benefits.  Differences in actual eligibility should not matter, but the estimates in Table 4 strongly suggest that it does matter.

Results - Effects of Welfare Reform on Marital Status in U.S.
    One of the primary objectives of welfare reform was to reduce dependency on government assistance by encouraging women to change behaviors that are the underlying cause of dependency.  Marital choices are one of the most important of these behaviors.  In this section we present estimates of the effect of welfare reform on marital choices.  One problem associated with this analysis is the quality of the target and comparison groups.  The use of education alone to define target and comparison groups is less than ideal and in the best case results in downward biased estimates, but may even lead to wrongly signed estimates if the marital experiences of more educated women are dissimilar to those of less educated women.  Table 5 presents some evidence related to this point.  The mean marriage rates of the target and comparison groups are similar only for the native-born women.  For foreign-born women, there are significant differences, both statistically and practically, between the marriage rates of low- and high-educated women.  These considerations imply a cautious interpretation of the estimates presented below.
    Table 6 presents estimates of the effect of welfare reform on marriage.  There are few statistically significant estimates in Table 6.  TANF appears to have no statistically significant effect on marriage rates of any of the demographic groups listed in Table 6.  Estimates of the effect of TANF on marriage rates of foreign-born women tend to be positive.  In contrast, AFDC waivers had a negative effect on marriage rates of native-born low-educated women and a positive effect on foreign-born low-educated women.  In both cases, some of the estimates are statistically significant.
    These results are difficult to reconcile theoretically.  AFDC waivers mostly reduced the generosity of cash benefits (e.g., time limits) and the utility (e.g., greater work requirements) of public cash assistance and should have encouraged behaviors that reduce the need for such assistance by both native-born and foreign-born women.  The heterogeneity of the estimates raises questions as to the reliability of these estimates and combined with legitimate questions related to the adequacy of the target and comparison groups suggests that the underlying identification strategy may be problematic.  Therefore, we believe the most appropriate interpretation of these findings is that they provide no evidence either way about the effect of welfare reform on marriage of low-educated women, either native- or foreign-born.

Results - Effects of Welfare Reform on Employment in New York Metropolitan Region
    As discussed above, the high concentration of immigrants in the New York metropolitan area makes it of special interest to our analysis of the effect of welfare reform on immigrants.  IN addition, New York and New Jersey, the states relevant to the New York metropolitan region did not extend TANF-like benefits to new legal immigrants.  Thus, it is of interest to compare the immigrant response in this area to that of other areas that did extend such benefits, for example California.
    The analysis of the New York metropolitan area parallels our national analysis with a few exceptions.  First, we focus only on TANF since New York did not have an AFDC waiver and New Jersey’s waiver was not that restrictive, for example it did not have time-limited benefits, so TANF was a substantial policy change in both New York and New Jersey.  Second, since sample sizes tend to be small, we are limited in our ability to obtain separate estimates of the effect of welfare reform for certain groups (e.g., foreign-born citizens) and by certain characteristics such as recency of immigration.  Finally, we do not present the estimates of the effect of welfare reform on employment or marriage obtained using more educated women as the comparison group.  The relatively small sample sizes associated with the New York analysis increases our doubts about the efficacy of the identification strategy underlying this analysis.  Therefore, we do not report the results.  Table 7 present sample sizes and descriptive information of the New York sample.
    Table 8 presents estimates of the effect of TANF on employment and marriage.   Surprisingly, estimates of the effect of TANF on the employment of low-educated native-born women are not statistically significant and somewhat smaller in magnitude than the national estimates.  The estimate in Table 8 indicates that TANF is associated with two percentage point increase in the employment of low-educated native-born women in New York.  The smaller estimated effect of TANF is surprising because New York City is known as a locality with a relatively strict implementation of TANF, which would suggest that effects would be larger for this sample since New York City residents dominate in the data.  However, the welfare caseload in New York City is also more disadvantaged than in the rest of the country and the caseload has fell significantly less in New York than in other states.  Between January 1994 and December 1999, the welfare caseload, as measured by recipients, fell 56 percent nationally and only by 39 percent in New York.
    The estimate of the effect of TANF on the employment of foreign-born non-citizens is positive and large, although not statistically significant.  The large estimated effect for this group is consistent with the national estimates and with the large policy change that TANF represents.  Interestingly, estimates in Table 8 strongly suggest that immigrants in the New York region have responded more positively (i.e., increased employment) to welfare reform than natives.  Whether the estimate reflects a “chilling” effect cannot be determined without additional information.  The lack of statistical significance of the estimate may reflect the relatively small sample sizes.

Results - Effects of Welfare Reform on Employment in California
    The second area we focus on is the state of California.  California is home to the largest number of immigrants in the U.S. and it provides relatively generous benefits to immigrants, particularly newly arrived legal immigrants who are eligible for TANF-like benefits.  The analysis of welfare reform in California is similar to that in New York.  We focus only on TANF because California had a waiver prior to 1994, we omit detailed analyses by recency of immigration and citizen status because of small sample sizes, and we drop the marriage analysis because of concerns over the identification strategy.  Table 9 presents sample sizes and descriptive information.
    Estimates of the effect of welfare reform on employment are presented in Table 10.  Similar to New York, TANF had no statistically significant effect on the employment of low-educated native-born unmarried women.  Unlike New York, California’s caseload decline was similar to the national average.  Therefore, the absence of a significant employment effect is surprising.   More interesting, however, is the estimate associated with the foreign-born non-citizen sample.  It is not significantly different from zero and is negative.  This contrasts sharply with the similar estimate in New York and together the New York and California estimates are inconsistent with the “chilling” hypothesis.  The effect of TANF on immigrants in California, which has a generous eligibility policy toward immigrants, is much smaller than the effect in New York, which has a much less generous eligibility policy.   Thus, eligibility appears to matter, which is inconsistent with the “chilling” hypothesis.

Conclusions
    In this paper, we have investigated the effect of welfare reform on the employment and marriage of three groups of low-educated unmarried women: native born citizens, foreign-born citizens and foreign-born non-citizens.  The results of this analysis suggest that welfare reform, particularly TANF, induced all three groups of women to increase their employment.  Moreover, the employment increase resulting from reform can explain a large part of the declining welfare caseload.  All three groups of women responded to welfare reform and estimated effects of foreign-born women were large than for native-born women.  One interpretation of this is that the behavioral response of foreign-born women was greater than that of native-born women.  Understanding the reasons for such different behavioral responses is beyond the scope of this paper, but provides a stylized fact that future research should investigate.  An alternative interpretation of the different size effects is that differences in the quality of the target and comparison groups used in the difference-in-differences analysis.  If this is the explanation, than different sized estimates do not necessarily imply different behavioral responses.
    We also investigated whether immigrant responses to welfare reform differed by recency of immigration.  Estimates related to this question indicated that more recent immigrants had larger behavioral responses to welfare reform than did earlier arriving immigrants.  Again, if these estimates are truly comparable, they indicate something about the underlying causes of welfare receipt of immigrants and suggest that recency of immigration is correlated with the these causes.
    The “chilling” hypothesis that has received so much attention in the popular press and professional literature is not supported by our results.  We found that actual eligibility for benefits is an important determinant of the behavioral response to welfare reform.  This is inconsistent with the “chilling” hypothesis, which suggests that the debate and controversey surrounding passage of TANF was the primary cause of immigrant behavior.  Our results strongly suggest otherwise.
    Finally, our estimates of the effect of welfare reform on marriage rates of native- and foreign-born low-educated women are uninformative.  We do not believe that the underlying identification strategy of this analysis associated with this outcome was credible.  Therefore, we remain agnostic as to the effect of welfare reform on marriage rates.

 References
Besley, Timothy and Stephen Coate. 1992. “Workfare versus Welfare Incentive Arguments for Work
Requirements in Poverty-alleviation Programs.” American Economic Review 82(1): 249-161.
 
Besley, Timothy and Stephen Coate. 1995. “The Design of Income Maintenance Programs.” Review of
Economic Studies 62(2): 187-221.

Borjas, George. 1995. “Immigration and Welfare, 1970-1990.” Research in Labor Economics,
14: 251-280.

Borjas, George. 1999. “Immigration and Welfare Magnets.” Journal of Labor Economics 17(4): 604-637.

Borjas, George and Lynette Hilton. 1996. “Immigration and the Welfare State: Immigrant Participation in
Means-tested Entitlement Programs.” Quarterly Journal of Economics 111: 575-604.

Council of Economic Advisers, 1999. "Technical Report: The Effects of Welfare Policy and
 Economic Expansion on Welfare Caseloads: An Update." Washington, D.C: Executive
 Office of the President of the United States.

Fix, Michael and Jeffrey Passel. 1999. “Trends in Noncitizens’ and Citizens’ Use of Public Benefits
Following Welfare Reform: 1994-97.” Unpublished manuscript Urban Institute, Washington, DC.

Immigration and Naturalization Service. 1996.  Immigration to the United States Fiscal Year 1996,
www.ins.gov/graphics/aboutins/statistics.

Kaushal, Neeraj and Robert Kaestner. 2000. “Welfare to Work: Has Welfare Reform Worked?”
Unpublished manuscript Department of Economics, The Graduate Center of the City University
of New York, New York.

Passel, Jeffrey and Rebecca Clark. 1998. “Immigrants in New York: Their Legal Status, Incomes, and
Taxes.” Unpublished manuscript Urban Institute, Washington, DC.

Round Two Summary of Selected Elements of State Programs for Temporary Assistance for
 Needy Families, National Governors' Association Center for Best Practices, March 14,
 1999.

Summary of Selected Elements of State Plans for Temporary Assistance for Needy Families as of
November 20, 1997, National Governors' Association Center for Best Practices.  http://www.nga.org/CBP/Activities/WelfareReform.asp.
 
Schoeni, Robert F. and Blank, Rebecca M, 2000. "What has Welfare Reform Accomplished?
 Impacts on Welfare Participation, Employment, Income, Poverty and Family Structure."
 National Bureau of Economic Research, Working Paper No. 7627.

Smith, James and Barry Edmonston, eds. 1997. The New Americans: Economic, Demographic and Fiscal
Effects of Immigration, Washington, DC: National Academy Press.

Zimmerman, Wendy and Karen Tumlin. 1999. “Patchwork Policies: State Assistance for Immigrants
under Welfare Reform.” Occasional Paper Number 24, Washington, DC: Urban Institute.