Background

The U.S. welfare system is currently undergoing major change as the result of 1996 legislation, the Personal Responsibility and Work Opportunity Reconciliation Act. However, because the research whose review is the main focus of this chapter entirely concerns the welfare system prior to this legislation, only the old system is described here. The relevance of this research to the future welfare system is discussed in the last section.

Chapter 3 contains a discussion of the welfare system that provides a general background. In this chapter, only the features of the system specific to marriage and fertility are outlined.

The most well-known aspect of the welfare system bearing on marriage and fertility is the set of of eligibility rules in the Aid to Families with Depandent Children (AFDC) program that result in a high concentration of single mothers among recipients, a relatively tiny fraction of married couples on the rolls, and no families or individuals without children (single mothers are defined as women with children under 18 in the household but no spouse or cohabiting partner present). This feature is a result of the basic eligibility requirement, laid out in the 1935 Social Security Act, which created the AFDC program, that the program is intended to provide cash support only to children living without at least one of their biological parents. Thus children for whom one parent has died are eligible, but so are children whose parents never married but are living apart or whose parents are divorced or separated. The mother, or other caretaker relative, is also supported by the grant. Children who are living with both parents are eligible, along with their parents, only for the AFDC-UP (unemployed parent) program, but eligibility for those benefits has additional conditions requiring that at least one parent be unemployed, that this parent have a significant history of employment, and that the family meet the same stringent income and asset requirements as a single-parent family. As a result, AFDC-UP families constitute only a snall fraction of the AFDC caseload.¹

The Food Stamp program provides food coupons to low-income families regardless of family structure and hence does not have the same ''bias" toward single-parent families as does AFDC. Eligibility and benefits for the program are based on the income and resources of a group of people who eat together, regardless of their relationship to each other. Thus two-parent as well as single-parent families are eligible, although the fixed upper income and asset limits knock more two-parent families than single-parent families out of eligibility.² Single individuals and childless families are also eligible.

The Medicaid program provides subsidized medical care assistance to poor families. Historically it has been made available primarily to AFDC recipients and therefore has the same bias toward single-parent families. However, in the last decade, eligibility for Medicaid benefits has been greatly broadened to include children in poor families even if both parents are present and the family is off AFDC. However, despite the growth of Medicaid recipients under these new eligibility rules, the program is still disproportionately composed of single-parent families.

Housing programs come in several different forms—public housing as well as subsidized private housing, for example—and provide housing at below-market rents to families with low income and assets. However, these programs are distinguished from the other programs so far discussed by their nonentitlement status. Expenditure allocations to local public housing authorities limit the amount of funds available and therefore limit the number of recipients that can be served. Eligible families who apply and are accepted but cannot be supported are put on waiting lists that can be quite long (e.g., several years). To choose from among the pool of eligibles, local housing authorities are required to give certain groups priority over others (called "preferences"). One of the preferred groups is AFDC recipients. This, along with the fact that family income (per family member) is lower among the single-parent population than the two-parent population, results in a high fraction of single-parent families receiving housing benefits. However, the preference is not absolute, and there have been been times in the history of the program when middle-income families were preferred, so there are sizable representations of two-parent families in the housing program.

In summary, therefore, the conventional perception of the U.S. welfare system as largely favoring single-parent families over two-parent families and childless couples and individuals is essentially correct.³ This favored treatment affects incentives to marry as well as incentives to have children. Fertility incentives are present in one additional way, however, which arises simply because benefits are based on the number of children present in the family unit. Hence the monetary cost of having an additional child is smaller in the presence of these welfare programs than it would be in their absence.

Than these marriage and fertility incentives may have an effect on behavior can be understood both with common sense and from a variety of theoretical perspectives. The most natural modern conceptual framework is the economic theory of marriage and fertility as developed by Becker (1981) because that model emphasizes the economic gains to marriage and the economic benefits and costs of having children. However, one could easily understand incentives induced by the welfare system without the formalization of the Beckerian theory, for almost any framework in which economic factors play a role will predict that, if all else is held fixed, a welfare system biased against marriage and toward childbearing will change behavior in that direction (although the magnitude of the effect can, of course, be large or small).

Although more complex theories can give different predictions, the only simple economic theory that does so is that which conceptualizes single parenthood as an unlucky outcome of an attempt at marriage (or union formation in general) and in which benefits play the role of insurance against that outcome. Standard economic theories imply that government provision of such insurance—welfare benefits—would induce more individuals to attempt marriage in the same way that providing insurance to protect checking accounts against bank failure encourages individuals to put their money in banks. The difficulty with this way of viewing the problem is that it ignores what is called the "moral hazard" problem in insurance terminology—the simple fact that individuals who are given insurance have an incentive to put themselves more at risk or even to cause the insured-against event to happen; this means, in the case of welfare and family structure, simply that individuals have an incentive to take actions that lead, directly or indirectly, to single motherhood as an outcome.

Welfare effects on marital and fertility behavior occur necessarily through one of a fixed set of routes. An unmarried childless woman entering adulthood may have an child out of wedlock, for example, and welfare may affect the probability of this outcome. She may later marry and possibly have additional children within marriage, but then separate or divorce, returning to a state of single motherhood; welfare may also affect the likelihood of this outcome. Alternatively, she may have married and begun childbearing within marriage but then divorce or separate, which is a different path to the same eventual outcome. Once divorced or separated, she may have additional children out of wedlock; and she may or may not remarry. Both of these behaviors may be affected by the presence of welfare and the level of benefits.

Whether welfare is more likely to influence some of these behaviors than others is an empirical matter, but it is often argued on intuitive grounds that some "routes" to single motherhood are more likely to be affected than others. For example, it is often argued that an unmarried woman's second and subsequent out-of-wedlock births may be more influenced by welfare benefits, especially if the woman is already on welfare, than the first birth because the latter is more likely to be "unintended" and because awareness of welfare is less acute before a woman has been on welfare. It is also often argued that divorce and separation are likely to be less affected by welfare than remarriage probabilities, because divorce and separation are heavily influenced by other factors—most notably, whether the marital "match" is a good one—while remarriage is (so it is argued) more subject to rational calculation. These notions are useful as a starting point in thinking about differential motivations for women in different positions, but they should be regarded initially only as hypotheses to be tested.

When other determinants of marriage and fertility are considered, a rich set of conceptual models developed over decades of research is available. Some of the more important factors posited to affect marriage propensities and fertility rates are economic opportunities for women; economic opportunities for men (often hypothesized to have the opposite effects of those of women); sex and sex-employment ratios in the population; neighborhood effects; and the influence of education, family background, and other factors on social norms and values. Although enumerating these factors in detail would take us too far afield from the review exercise, it is important to emphasize that there are many influences on marriage and fertility other than welfare benefits, a point that is often deemphasized in studies whose sole focus is a single-minded search for welfare effects. Moreover, even if these other factors are not examined in detail when testing for the effects of the welfare system, it is always necessary either implicitly or explicitly to parcel out their influence relative to that of welfare, which means in most cases controlling for these other factors statistically, a point to be discussed further in the next section. Since a single mother does, after all, have alternatives to welfare, it is only the influence of the welfare benefit relative to the alternatives that should affect her choices.

Unfortunately, the large number and diversity of these alternative factors make it difficult empirically to control for them all and often leave the door open to doubts as to whether it is welfare that is affecting behavior or some other omitted factor, as discussed below in the review of the empirical research literature.

Different Questions of Interest

In turning from theories of welfare effects to the more specific issue of what empirical questions are of interest, an important distinction necessary to make at the outset is between what may properly be called a "time-series" question and a "cross-sectional" question. An important time-series question is why marriage rates have declined and nonmarital childbearing rates have increased in the United States. The corresponding welfare-related question is whether the welfare system has contributed to these trends. An important cross-sectional question, on the other hand, is whether welfare, if eliminated or reduced in generosity (for example), would raise marriage rates and lower nonmarital fertility rates, if all else is held fixed.

The answers to these questions need not be the same. One may simultaneously conclude, for example, that welfare is not a major contributor to the time-series trends in marriage and fertility but also that welfare, if reduced in generosity, would have the effects mentioned above, if all else is held fixed. Differing answers to these two questions are not necessarily inconsistent because all else is not held fixed in time series; many other factors are changing at the same time, most notably, changes in the economic and social environment and in social norms. These other factors could have been primarily responsible for the marriage and fertility trends, and could have outweighed any welfare effect. However, if it is concluded that welfare would have had an effect if nothing else had changed, one must also conclude that the time-series trend would have been different if welfare had not trended the way it did.

Both questions are of importance. Some analysts argue that the only important question is the time-series question. That question does receive much of the attention of the public. However, the cross-sectional question is also important because it bears on what would happen in the future if the welfare system were altered, regardless of what might have caused marriage and fertility trends in the past. If welfare has undesirable effects, for example, it could be used as a tool to increase marriage rates and reduce nonmarital fertility rates in the future. In any case, as the review below shows, virtually the entire research literature on the effect of welfare on demographic outcomes has focused on the cross-sectional question, not the time-series question. The majority of analyses have attempted to hold everything else fixed in a cross-sectional sense. Indeed, those studies that have utilized data over multiple time periods, which could conceivably examine time-series questions, have, by and large, deliberately eliminated the influence of

time trends in the data and have based their welfare results on the cross-sectional variation in the data instead.⁴

Methodologies Used in Estimating Welfare Effects

Basic Time-Series Patterns in Welfare and Demographic Outcomes

Three of the methodologies—cross-state comparison of levels, cross-state comparison of changes, and time-series analysis—can be studied by examining trends over time in unadjusted state-level or national-level aggregates of demographic outcomes, on the one hand, and measures of welfare generosity, on the other. It is useful to present the basic patterns of these correlations with unadjusted aggregates before reviewing the multivariate analyses in the econometric literature. As it turns out, the patterns that appear in this analysis capture, in large degree, the patterns revealed by multivariate analyses. Consequently, much of the basic story is understandable in relatively simple terms and does not need recourse to controlling for additional variables or use of specialized statistical methods.

The pure time-series method involves a simple comparison of trends in welfare benefits and in demographic outcomes. Figure 4-1 shows the time trend in welfare benefits of different types in the United States over the period 1970–1993. It has been noted repeatedly that the time-series evidence for a welfare effect on marriage and fertility is weak because welfare benefits declined in real terms over the 1970s and 1980s while marriage rates declined and nonmarital childbearing increased; both trends have been noted in the overviews in Chapters 2 and 3. Figure 4-1 provides further confirmation, because it indicates that real AFDC benefits have fallen continuously since the early 1970s. Real Food Stamp benefits have remained roughly constant, primarily because they are indexed to inflation, while real Medicaid benefits were roughly fixed until the mid-1980s, when they began to rise. The sum of benefits therefore declined up to the late 1980s. It did begin to rise at that time, but this increase is too late to explain the secular trends in marriage and fertility. In addition, Medicaid benefits began to be available to many poor families off AFDC in the late 1980s, thereby weakening the link between welfare and the availability of medical care.

Figure 4-1.

Trends in real monthly welfare benefits per person. SOURCE: U.S. House of Representatives (1994:378, 782, 806).

The inconsistency between benefit and demographic trends could mask the presence of long lags (Murray, 1993). The generosity of the transfer system increased significantly in the late 1960s and early 1970s, as Food Stamps were mandated nationally and the Medicaid system was expanded. It is possible that this expansion of benefits resulted in increases in (say) nonmarital childbearing 10 years later, if the effect of the expansion took time to occur because social norms were slow to adjust. This is a difficult hypothesis to prove or disprove because the trends have been so universal. It is not possible to isolate specific communities where benefits increased much more than other communities, for example, and where the population was fixed for 10 years so that lagged effects could be measured. Consequently, the importance of this argument at the present time must rest to a great extent on whether one believes that low-income families react quickly or slowly to the monetary opportunities facing them.

As noted previously, the inconsistency between time trends in benefits and demographic outcomes may mean only that there have been other factors changing over time that masked the effect of welfare benefits; this is the major weakness of the method. There may have been changes in the other factors affecting marriage and divorce—economic opportunities for women and men, the availability of partners in the marriage market, and changes in social norms. More persuasive evidence on the effect of welfare per se might therefore be gained from cross-state comparisons of levels because these comparisons are made at a single point in time, across states, and hence are not complicated by such major time trends. Figure 4-2, drawn from Murray (1993), shows illegitimacy rates and welfare benefit levels among white women in different states in 1988.⁹ A positive relationship between benefits and illegitimacy is clear from the figure. Much of the relationship comes from the concentration of southern states with low benefits and low rates of illegitimacy, although the relationship would still be positive (but weaker) if the South were omitted. Thus some evidence for a positive effect of welfare on out-of-wedlock childbearing is yielded by these data.

Figure 4-2

Illegitimacy rates and benefit levels for white women, 1988. SOURCE: Murray (1993).

To ensure that this pattern is not special to the particular dataset, time period, and variables used by Murray, data from the Current Population Survey (CPS) for 1993 were obtained for this study, and tabulations of welfare benefits and rates of single motherhood by state were computed. Single motherhood rates rather than illegitimacy are examined because single motherhood is a broader and more inclusive measure of the demographic outcome of interest.¹⁰ Figure 4-3 shows the cross-state result for white women.¹¹ Interestingly, very little relationship between headship and benefits appears in this figure, contrary to the results of Murray. A least-squares regression line, also shown in the figure, confirms this visual impression of only a slight positive relationship between the two variables. However, when women 20—44 and without a high school diploma are examined instead (Figure 4-4)—a subpopulation with relatively high welfare participation rates—the positive correlation reappears with a greater magnitude. Illegitimacy rates are no doubt more concentrated among the less educated, low-income population than are single mothers, who are fairly common in higher-income groups as well; this may explain why the positive correlation appears for illegitimacy rates even without restricting the sample to young, less educated women. This positive covariation extends to an examination of rates of never-married single mothers—that is, the fraction of women who have children but have never been married (thus omitting divorced, separated, and widowed single mothers)—where the relationship is, if anything stronger (figure not shown).

Figure 4-3

Single motherhood rates and real AFDC benefits by state: CPS, 1993, white women.

Figure 4-4

Single motherhood rates and real AFDC benefits by state: CPS, 1993, white women 20–44 without high school diploma.

This simple analysis shows that the level of state welfare benefits is substantially correlated with single-motherhood rates. Many of the largest states such as New York, California, and Illinois have relatively generous welfare systems as well as high rates of single motherhood; another large state, Texas, has low benefits and low single-motherhood rates. Clearly, a major question is whether this simple correlation is the result of some other characteristic of the populations of these states or of their socioeconomic environments; however, as seen in the next section, this positive covariation persists even when other measurable influences are controlled for and therefore appears to be reasonably robust.

The positive relationship holds for other periods as well—all the way back to the 1960s, when CPS micro data are first available for these computations. It also holds when other measures of the welfare system—Medicaid, for example—are included. The relationship also appears in simple regional comparisons because the Northeast has high welfare benefits and high rates of single motherhood, while the South has the lowest benefits and lowest single motherhood rates. The Midwest and West have much higher benefits than the South and slightly higher rates of single motherhood.

To determine whether these comparisons of levels have the same implications as those from cross-state comparisons of changes, CPS data from a different year can be compared to the 1993 data. The tables discussed below use CPS information from 1977, a full 15 years prior to 1993, when benefit levels were quite a bit higher. Rates of single motherhood were lower in 1977 overall, but the issue here is whether those states that lowered their AFDC benefits the least—benefits fell in virtually all states—also had the largest increases in single motherhood; if so, this could be taken as evidence of an effect of welfare consistent with the pure cross-state comparison of levels.

As Figure 4-5 indicates, however, the relationship between benefit levels and single motherhood is extremely weak for young less educated white women when this type of comparison is made. Although different states lowered benefits over this period by different amounts, the increases in single motherhood across the states were fairly uniformly distributed. New Jersey, for example, which reduced its benefits by a very large amount over the period ($257 reduction per month) saw its single motherhood rate increase by about 5 percentage points, whereas Texas, which reduced its benefit by much less ($78 per month), saw about the same increase, 4 percentage points.

Figure 4-5

Never-married female headship rates and real AFDC benefits by state: CPS, 1993, white women 20–44 without high school diploma.

Mechanically, the difference in results between the different comparisons arises from two facts. First, over the 1970s and 1980s, states with high average welfare benefits had higher-than-average rates of single motherhood (as well as nonmarital fertility rates); this relationship held not just for 1993 but also for 1977 and other years. Second, over the course of the 1970s and 1980s, states that raised benefits more than others—or, more accurately, lowered them less than others—did not experience faster-than-average increases in single motherhood and fertility. Welfare benefits across U.S. regions converged slightly over the 1970s and 1980s, with the southern states lowering benefits the least and northeastern states lowering their benefits the most, for example, but this pattern does not correspond at all to rates of change of single motherhood among less educated white women (e.g., the Northeast experienced the greatest increase in single motherhood even though it, along with the industrial midwestern states, lowered benefits the most).

The difference between the results using these two sources of welfare variation may stem from the omission of factors in one or both of the two comparisons. One possibility is that the comparison of levels omits key state differences that affect both marriage and fertility behavior as well as benefits. For example, as mentioned earlier, southern states have strong promarriage social norms and also have low welfare benefits; the correlation between marriage and benefits may therefore arise coincidentally. The fact that the South did not lower its benefits very much over time, for example, and yet did not experience a high growth of single motherhood relative to regions like the Northeast, which lowered their benefits a great deal, suggests indeed that the cross-state levels relationship may have been spurious and due to other factors.¹² However, it may also be the case that the comparison of changes omits some factor that is causing benefits to change at different rates across states. Differences in rates of change in the economic performance of different states, in unemployment rates, and in related factors may have been responsible for both the change in benefits and the change in single motherhood. For example, the South experienced significant economic growth in the late 1970s and 1980s, and closed its economic gap with the rest of the country to some degree; this could have caused both its relative increase in welfare benefits and its relative decline in single motherhood. The place to begin in testing these hypotheses and attempting to reconcile the different forms of evidence (levels vs. changes) is to attempt to control for some of these omitted variables in a multivariate analysis. This is one of the roles of the econometric research to be described momentarily.

The patterns for black women are roughly similar. As shown in Figure 4-6, the levels of single motherhood for young, less educated black women are positively related to welfare benefits. Also, a comparison of changes in single-motherhood rates and benefit levels also shows no relationship between the two, if not a negative relationship, as shown in Figure 4-7. Overall, single-motherhood rates grew quite a bit faster for black women than for white women over this period but, as for white women, the rate of growth across states was not closely related to the magnitude of changes in welfare benefits in the state. Single-motherhood rates for less educated black women grew at about the same rate in the South, the Northeast, and the Midwest, for example, despite very different changes in welfare benefits in those regions.

Figure 4-6

Female headship rates and real AFDC benefits by region: CPS, 1993, white women 20–44 without high school diploma.

Figure 4-7

Change in female headship rates and real AFDC benefits by state from 1977 to 1993: CPS, white women 20–44 without high school diploma.

Results from Multivariate Econometric Models

The econometric studies in the literature are fairly large in number. A table listing many of the studies appears in the appendix to this chapter. A more detailed summary of each is available in Hudson and Moffitt (1997). These studies all use one of the four methods of obtaining welfare variation described in the discussion of methodologies, three of which have been discussed in graphical terms in the preceeding section (all except the use of within-state variation).

Relative to the graphical analysis, a simple question that can be answered here is whether the patterns of effects across states, over time and for different racial groups, is any different in a multivariate analysis where additional covariates are entered into the model and where more sophisticated methods of estimation are employed.

Table 4-1 summarizes the results of 68 different estimates from these studies, classified according to the method by which welfare variation was obtained and the nature of the result.¹³ Over all types of studies, a slight majority find a significantly negative effect on marriage or positive effect on fertility rather than an insignificant effect (the mixed estimates, which could be classified in either way, are ignored). This may seem surprising in light of what was taken to be the conventional wisdom approximately 10 years ago, when it was generally believed that the evidence did not support much of an effect of welfare on marriage and fertility at all. However, that consensus was based on studies from the 1970s, which indeed showed weaker results than the studies that have been conducted since then. Among analysts who work on the topic, there is now a rough consensus that the evidence does support some effect of welfare on marriage and fertility, although the magnitude of the effect remains in question. Whether this change in estimates has been a result of superior analysis methods in the later studies or an increase in the underlying effect of welfare on behavior is difficult to determine with certainty, but some evidence points to the latter (Moffitt, 1990).

TABLE 4-1

Counts of Studies of Effect of Welfare on Marriage and Fertility, by Nature of Findings and Source of Welfare Variation.

However, the overall counts of estimates are misleading because they are disproportionately concentrated among studies using cross-state comparisons of levels—a much smaller number have used cross-state comparisons of changes and only a handful have used either time-series or within-state methods—and because the results differ notably by race. As Table 4-1 shows, a majority of the estimates from cross-state comparisons of levels show that welfare benefits have an effect on marriage or fertility—negative on the former, positive on the latter—but when the results are disaggregated by race, the studies show more of an effect for white women than for nonwhite or black women. For white women, nine studies show effects of welfare while only two show no effect. For black and nonwhite women, however, the split is almost exactly 50–50 between those that find an effect and those that find none. Thus these multivariate analyses are quite similar to those revealed by the simpler graphical analyses reported in the previous section, at least for white women. The implication, perhaps surprising, is that the additional covariates added in these studies—typically variables like age, educational level, and family background, as well as (sometimes) variables for the state unemployment rate, labor market wages, etc.—do not explain away the cross-state differences for white women in the simple unadjusted cross-state comparisons. For black women, however, these variables do appear to explain much of the raw difference; black women of similar characteristics in different states do not have significantly different demographic outcomes, at least in many of the studies in the literature, despite the differences in benefit levels across those states. It is not possible to determine the precise set of measured influences that account for the unadjusted difference across states noted earlier, but differences in urbanization may be one factor.¹⁴

The weaker effect for black women is unexpected in light of their greater rates of participation in the welfare system compared to those of white women. In general, it is possible that there is some omitted factor that differs between the races (including possibly cultural differences), but no such factor has yet been identified in the literature. Murray (1993) hypothesizes that the low-income black population is more geographically concentrated than the low-income white population and that neighborhood effects lead to changes in social norms that increase illegitimacy rates (for example) even in the face of low benefits. Thus the South, with its concentration of the black population, has high illegitimacy rates. However, if this argument is correct, it implies that the variation in illegitimacy between black women in different states is indeed a result of something other than the welfare system. The racial difference therefore must still be regarded as an unresolved puzzle.

As shown in Table 4-1, many fewer studies have conducted cross-state comparisons of changes instead of levels. Of those that have used this method, however, the estimated strength of the welfare system is quite different than the results of the levels method. For white women, the estimated welfare effect is quite weakened: the studies that compare changes are evenly spread over those that find an effect and those that find none. For black and nonwhite women, the estimated welfare effect is actually somewhat stronger in the comparison of changes than it was in the comparison of levels, in terms of the relative numbers of studies finding an effect. These results are, therefore, once again quite consistent with the simpler analysis reported in the last section for white women but only roughly so for black women, although even for black women there is about a 50–50 split between studies finding a significant effect and studies finding no effect. In addition, an implication of this pattern of results is that the differences between the comparisons of levels and changes, and between the race differences within each, are not explained away by the typical covariates used in these analyses.

There has been considerable discussion in the research literature concerning the different results across methods but without any definitive resolution. Although it has been argued that the comparison of levels is subject to the biases noted earlier in the discussion of methodologies, the comparison of changes also has the defects noted there. In the absence of definitive evidence that either methodology is incorrect, an equal weighting of the two still leads to a conclusion that the welfare system has effects on marriage and fertility, even if not as strong as might be thought based on comparison-level methods alone.

There have been even fewer within-state and time-series studies, mainly for the reasons noted earlier: within-state comparisons must find some characteristic of women that affects their eligibility for benefits but does not independently affect their marriage and fertility outcomes, while time-series analyses inevitably have difficulty controlling for all alternative influences that are changing over time. For example, one study utilizing within-state variation did not examine benefits at all but found no effect of AFDC participation rates on demographic outcomes across races, a method that implicitly assumes there would be no racial difference in demographic outcomes in the absence of AFDC. Another study compared the divorce rates of women with and without children in states with high and low welfare benefits, thus implicitly assuming that divorce rates would be identical among women with and without children in the absence of AFDC. The implausibility of these assumptions shows the extreme weakness of the method. As for the time-series studies, most simply estimate a variety of bivariate relationships and find either no effect or mixed effects. The one study that found a negative effect regressed the illegitimacy ratio in a year on the lagged AFDC participation rate rather than on the AFDC benefit; yet the AFDC participation rate is an endogenous variable and is as much the product of time-series trends in illegitimacy as its cause. The within-state and time-series methods are sufficiently problematical that they should probably be dropped from any weighing of the evidence on the question.

While the discussion thus far has concentrated on what now appear to be unresolved differences between results using levels and changes comparisons, and between races, there is also a considerable variance of results within these types of studies and there are quite a few studies in each category that differ from the central tendency of the results for each type. Once again, without further analysis, it is difficult to resolve most of these differences. To be sure, there are a few studies that appear to suffer from a significant defect that could explain why their results differ from the central tendency. Many of these defects concern the use of endogenous variables either for the welfare benefit or in controlling for nonwelfare factors, where an endogenous variable is, roughly, a variable that is a result of women's marital and fertility choices themselves (rathern than a cause of them). Among the studies of levels that find a significant cross-sectional welfare effect for black and nonwhite women, for example, one study replaced the welfare benefit ("instrumented" it, in econometric parlance) with such endogenous variables, while others included in the regression variables of questionable exogeneity such as the labor force participation rates and earnings levels of men and women. Other defects in the studies arise as well: one constrained the welfare benefit coefficient to be the same as the coefficient on other income, while another defined the dependent variable as AFDC receipt, which could by itself and separately be expected to respond to benefit levels. However, the number of studies that can be dropped from consideration for these reasons is relatively small, and even for these, it cannot be determined conclusively that a correction of the problem would have had a major quantitative effect on the results. Thus most studies must be given some positive weight in a balancing of the evidence.¹⁵

One notable difference among the different studies behind Table 4-1 is their great diversity in the types of variables held fixed when estimating welfare effects. Duncan and Hoffman (1990), for example, control for differences in women's labor market opportunities and even for differences in the labor market opportunities of potential male marital partners. Schultz (1994) and Lundberg and Plotnick (1995) similarly attempt to control for labor market differences. Ellwood and Bane (1985) and Matthews et al. (1995) go the farthest in this direction, controlling for a large number of state characteristics, even including characteristics of state political systems. On the other hand, Murray (1993), in an intentional effort to keep his analysis simple and easy to understand, does not adjust for any other differences between women or across states besides welfare. Roughly speaking, the more variables that are controlled for in an analysis, the weaker is the estimated effect of welfare—although there is no logical reason why this need be so, which may also be responsible for some of the differences across studies. Determining whether this is the case would require reanalyzing some of the datasets under consideration and estimating similar specifications across datasets.¹⁶

In addition to these differences, however, the studies with different results vary in the dataset used, in the age range of the individuals examined, and in the calendar years covered by the data. A simple way to summarize these differences is by ordinary least-squares regression, using as a dependent variable the strength of the estimated effect and as independent variables the characteristics of the study. By taking only the studies in the first two rows of Table 4-1 (levels and changes studies) and defining a dependent variable (Y) equal to 1 if an effect was found, 0 if not, and .5 if a mixed finding was obtained, the following regression-based summary of the importance of study characteristics results:

where CHANGES is a dummy variable equal to 1 if the study used the changes rather than levels method; BLACK is a dummy equal to 1 if the estimate in question is for the black or nonwhite population; YEAR is the median year of the data; AGE is the median age of the individuals in the data; VITAL, NLSY, PSID, and CPS are dummies equal to 1 if the study used vital statistics, National Longitudinal Survey of Youth (NLSY), Panel Study on Income Dynamics (PSID), or CPS data, respectively (omitted category is all other datasets); and SM and REMDIV are dummies equal to 1 if the study-dependent variable was single-motherhood or divorce-remarriage transitions (omitted category is a dependent variable pertaining to fertility, almost always nonmarital). Standard errors, in parentheses, are large because of the small sample size. Interestingly, the result simply that the changes studies yield stronger, rather than lesser, effects when the other variables are controlled; that estimated effects are larger in samples of older women (contrary to some of the hypotheses in the literature) and grow overtime; and that the effects are stronger when vital statistics and NLSY data are use drather than CPS or PSID data.¹⁷ The summary also indicates that welfare effects are weaker in studies that examine single motherhood as a single state, or remarriage or divorce, than studies that examine the effects on nonmarital fertility. When these results are taken at face value, they imply that the strongest effect of welfare occurs in nonmarital fertility but that these effects eventually disappear, perhaps because a woman eventually marries and her subsequent demographic behavior is unaffected by her having experienced an out-of-wedlock birth previously.¹⁸ This finding bears further investigation because it means that the implications of early nonmarital childbearing for later family structure may not be as strong as imagined. Of course, there are many other differences in these studies that have not been controlled for. Once again, however, only a reanalysis of the various datasets and models can confirm these differences.

Studies that compare changes are thought by many analysts to be more reliable than studies that compare levels for the reasons noted previously—namely, that the level studies confound cross-sectional benefit variations and unobserved variations in economic, social, and political factors. If this view is taken, there are sufficiently small numbers of these studies to make more headway by making more detailed comparisons between specific individual studies. When the studies are examined at this more detailed level, many possible explanations for differences appear. For example, the stronger effects found by Jackson and Klerman (1995) hold only when effects on nonmarital fertility in isolation are estimaged; when effects on marital fertility are examined as well, no effect of welfare on their relative magnitudes is found. This should properly move the study from one reporting a significant effect to one reporting an insignificant effect in Table 4-1. Clarke and Strauss (1995), who also find a significant effect of welfare, obtain strong effects with a two-stage least-squares procedure using state per capita income (among other variables) as an "instrument" for the benefit, but per capita income probably belongs in the main equation. Rosenzweig (1995) argues that his significant estimated effects of welfare result from separating out the low-income population for analysis, but a similar separation was conducted by Hoynes (1996), Moffitt (1994), and Robins and Fronstin (1996), who all found either no change in effects because of this separation or insignificant changes even for the disadvantaged population; this suggests that some other difference between the Rosenzweig study and the other three studies explains their different findings.¹⁹ Finally, these studies differ dramatically in the extent to which other state-level influences are controlled in the regressions and the types of influences controlled for. Table 4-2 shows the different area-level controls used in the studies of changes. While some of the variables are potentially endogenous and therefore perhaps should be excluded, some of the studies control for no area-level variables at all, which could easily explain some of the differences in findings.

TABLE 4-2

State-Level Control Variables Used in Studies of Cross-State Changes.

A final important issue concerns the magnitudes of the estimated effects of welfare for those studies finding significant estimates. Not surprisingly, the estimated magnitudes have a wide dispersion as well. At the upper end are three studies (Fossett and Kiecolt, 1993; Hill and O'Neill, 1993; Rosenzweig, 1995) implying that a 25 percent reduction in welfare benefits would reduce the probability of a nonmarital birth by approximately .04 or .05.²⁰ If the mean probability is .16, this implies a reduction to a level of .11 or .12, or about a 30 percent reduction in the rate. In time series, the welfare benefit has indeed fallen by about 25 percent over the last 20 years (see Figure 4-1), while the nonmarital childbearing rate for this age group has doubled (U.S. Department of Health and Human Services, 1995: Figure II-1). One interpretation of these estimates is therefore that the historical increase in the nonmarital childbearing rate could have been cut by a significant amount if benefits had been reduced by twice the amount that they were. At the other end are studies obtaining estimates that are statistically significant but quite small in magnitude (Danziger et al., 1982; Duncan and Hoffman, 1990; Lichter et al., 1991, 1996). A typical and recent estimate is that of Lichter et al. (1996), who found that a 25 percent reduction in the welfare benefit would increase the percentage of women who are female heads by only .007, a small amount. Clearly, therefore, a resolution of the differences in these magnitudes is also a priority item for future research.

What Do We Need to Know?

This review of what we know about the effect of welfare on marriage and fertility has demonstrated that much has been learned from research regarding the basic patterns of relationship between welfare and the demographic outcomes, where a significant relationship appears and where it does not, and about the general robustness of the strength of the estimated relationship across different datasets and different methods. Based on this review, it is clear that a simple majority of the studies that have been conducted to date show a significant correlation between welfare benefits and marriage and fertility, suggesting the presence of such behavioral incentive effects. However, in addition to this finding not being able to explain the time-series increase in nonmarital fertility and decline in marriage, the majority finding itself is weakened by the sensitivity of the result to the methodology used and to numerous other differences in specifications across the studies. A neutral weighing of the evidence still leads to the conclusion that welfare has incentive effects on marriage and fertility, but the uncertainty introduced by the disparties in the research findings weakens the strength of that conclusion.

The resolution of the discrepancies between these studies is important for welfare policy at minimum because the issue of how demographic outcomes are affected by the overall level of welfare benefits is so basic to all discussions of welfare effects. It is also relevant to many of the reforms tested in the states in the past several years and to many of the changes enacted in the 1996 welfare legislation. Women who lose eligibility because of time limits or failure to comply with new rules, as well as women who do not choose to go onto welfare but would have otherwise, can be viewed as having suffered benefit reductions similar to those whose effects are studied by the research literature. More generally, the legislation is intended to reduce the welfare caseload and to lower the overall level of welfare benefits provided to low-income populations; it is explicitly

intended to have effects on nonmarital fertility of the type with which the research literature is concerned.²¹

Although much of the analysis of the 1996 legislation will be conducted with program evaluation methodologies using experimental-control or treatment-comparison-group frameworks rather than the econometric approach underlying the studies in the research literature, the latter has a role to play in understanding the former. Ideally, the econometric research should be consistent with demonstration and evaluation evidence and any differences should be reconciled. If, for example, the New Jersey family cap experiment shows little effect of a family cap on fertility, it would increase the confidence in that finding considerably if it could be concluded from the research literature that incremental benefits in the range tested in New Jersey also appear to have no effect on fertility. Even more important—to continue to follow the New Jersey case—the research literature should be capable of providing estimates of the effects of benefit changes of greater magnitudes than that in New Jersey and for a greater number of states with differing economic and social environments. Regardless of the number of demonstration evaluations conducted, there will never be a sufficient number of them to provide the same range of alternative programs that occur naturally over time and across states. Econometric research using secondary data can make use of that larger range.

Unfortunately, the diversity in findings in the research necessarily reduces the power of that research to play this role. Moreover, studies that attempt to resolve the discrepancies have not been conducted. Three different types of studies are necessary. The first type involves replication studies, or studies that reanalyze the same dataset used by each study (or the major studies) to determine whether the results were correct as reported. The second type, robustness studies, conduct sensitivity testing to the model reported in each study to determine if the results are robuts to variations in the specification. The third type—reconciliation studies—attempt to estimate common specifications across studies on common samples, in an attempt to reconcile why the studies achieved different findings. These types of studies—the three Rs of replication, robustness, and reconciliation—have not been applied to this literature. As in much other literature in which the main contributors are academic scholars, the lack of attention to the three Rs is primarily a result of the bias in academic publishing and research toward new findings, new techniques, and original analysis, and against mere replication of other researchers' results. This situation is unlikely to change without government or other funding to give researchers an incentive to conduct such studies.

The most likely cause of the discrepancy across studies is the omission of different alternative influences on marriage and childbearing. Very few studies control for the same factors, and the studies using the different methodologies outlined here almost never attempt to control for the confounding influences appropriate to the method in question (e.g., alternative influences across states, across states over time, or in time series). Relatively little attention has been paid to nonwelfare influences, particularly those that might be correlated with welfare, in the analyses. This defect could also be addressed with additional research.

Acknowledgments

The author would like to thank E. Michael Foster, John Haaga, Robert Haveman, Anne Hill, Hilary Hoynes, Daniel Lichter, Howard Rolston, and Barbara Wolfe for comments; Julie Hudson and Chris Ruebeck for research assistance; and grant R01-HD27248 from the National Institute of Child Health and Human Development (NICHD) for support. Assistance was also provided by NICHD grant P30-HD06268 to the Johns Hopkins University Population Center.

Footnotes

1

The eligibility rules have many other important facets which space does not permit discussing, especially rules governing eligibility of children living with cohabiting adults and whose caretaker parent has remarried. For details on these rules, see Moffitt et al. (1998).

2

AFDC recipient families are automatically eligible for Food Stamp benefits, so this also results in a disproportionate number of single-parent families actually on the Food Stamp rolls.

3

It is worth noting, however, that any program that provides benefits on the basis of the income of a family unit rather than the income of individuals will necessarily, and inherently, have at least a minimal amount of bias toward single-parent families. If bias is defined as occurring when the income gain to marrying, for example, is less in the presence of a government program than in its complete absence, then a welfare program will be nonbiasing only if benefits are completely unaffected if a single parent marries. But this violates the definition of a targeted transfer program, namely, one that concentrates its benefits on those with lower income. This is an example of the equity-efficiency economic principle.

4

In a regression framework, "eliminating the influence of time trends in the data" is meant to imply, for example, entering dummies for year or other time intervals into the equation.

5

There are exceptions, and more experimental evaluations examining demographic outcomes are under way at this writing. See Chapter 6 for a discussion of state-level experiments on demographic outcomes. Also, the negative income tax (NIT) experiments of the 1980s were used to examine the effect on an NIT on marital stability (Hannan and Tuma, 1990; Cain and Wissoker, 1990) but, aside from being troubled by small sample sizes and design problems in the experiments, their results cannot be generalized to the AFDC program.

6

Even the 1996 welfare legislation does not eliminate welfare entirely for anyone, because some minimum number of years of receipt is guaranteed.

7

This notion appears to have first been explicitly discussed and emphasized by Ellwood and Bane (1985).

8

A related possibility is that the comparison-of-changes method measures a short-term response, while the comparison-of-levels method measures a long-term response if it shows where marriage and fertility levels have ended up after several years of adjustment. Thus it may be that the two methods are simply not measuring the same thing.

9

The illegitimacy ratios are taken from vital statistics reports.

10

This is because single motherhood is an overall category that can be reached by any of the routes discussed earlier—nonmarital childbearing, divorce or separation, and failure to remarry. Thus it is a summary measure of all these routes taken together.

11

The March CPS is used. Single mothers are defined as women without a spouse in the household who have children under 18. Family and subfamily heads are included as separate observations. The rates are computed as a fraction of all women 18–64. The AFDC benefit variables are those for a family of four with no other income.

12

For example, Hoynes (1996) used data that followed individuals over time (i.e., panel data) to determine whether the correlation between changes in single motherhood at the individual level and changes in benefits is the same as that at the state level. She found this to be the case for white women. This supports the interpretation of the cross-state differences as traceable to differences in the types of women in those states.

13

The studies were located by searching the economics and sociology literature since 1970 and following references to other articles therein, as well as by a general search for articles on the subject since 1970 in a variety of other sources. All studies located were included that (1) had the estimation of the effect of AFDC on marriage, fertility, or a related demographic outcome as a significant, major focus of the study and (2) were either published or had been circulated in draft form by May 1996. No study was intentionally excluded that met these criteria. It should also be noted that there are 68 estimates but fewer individual studies than this because most studies provided estimates for both races. Estimates for outcomes other than marriage and fertility, (e.g., living arrangements) are excluded from the table but are indicated in the Appendix.

14

For example, see Moffitt (1994: Table 3). Adding age, education, urban residence, and a few other variables to an equation explaining black female headship rates changed a welfare-benefit coefficient from significance to insignificance. Urban residence, which is less common in the South than in other regions, for example, had a positive effect on headship rates.

15

These remarks are relevant to a common criticism of the "vote-counting" method used in Table 4-1—namely, that simply counting studies that have differing results without any adjustment for the quality of the study is misleading. The argument here is that only in rare instances can defects in the methodology in a study be determined to account for any nontrivial amount of the difference in estimates from another study, because too much else differs as well; hence the magnitude of the bias cannot be isolated.

16

Although in general it is to be desired that as many alternative influences be controlled for as possible, this does not extend to endogenous variables, which were discussed previously and should not be included. However, as important as this distinction is, it is not necessary to discriminate between exogenous and endogenous variables when one is attempting merely to answer the simpler factual question of whether differences in regressor sets across studies account for their differences in estimated welfare impacts. Only after it has been determined which variable sets lead to what results can the question of which is "best" be addressed.

17

Klerman (1996) argues that the sample sizes in all datasets save vital statistics are insufficient to detect effects of reasonable magnitudes. This is supported by the estimated coefficient on VITAL but not by the coefficient on CPS, which is the next largest dataset.

18

This conclusion necessarily follows because a young woman who has a premarital birth necessarily becomes a single mother, thereby driving up the fraction of the population who are single mothers; but if the overall rate of single mothers is not significantly affected by welfare, it must be the case that these young mothers later marry so that, on average and over all ages, the single-motherhood rate ends up not much different than it would have been if the early premarital childbearing had not occurred. It should be noted that the vast majority of studies (about three-quarters) are of nonmarital fertility and that there is only one study of divorce, which is why it is lumped in with remarriage (for which there were only two studies).

19

A notable difference, however, is that Rosenzweig stratified on the income of the family of origin, while the other three studies stratified on the education of the woman in question. Whether this could explain the differing results cannot be determined.

20

The different studies define their dependent variables slightly differently; for two of them it is approximately the probability of ever having had a nonmarital birth up to a particular age (which is higher than the annual probability of the event). The 4 percentage point number is scaled from the numbers actually given in the articles.

21

In addition, many of the states have adopted or will newly adopt family caps on payments to additional children, changes in the AFDC-UP program to encourage married-couple welfare participation, and other rules that directly affect fertility and marriage apart from simply reducing the caseload (see Chapter 6).