The Impact of Welfare Reforms on Welfare Caseload

The Impact of Welfare Reforms on Welfare Caseload
Eric C. H. Yu
Department of Political Science
Columbia University
420 W118th Street
New York, NY 10027
Phone: 646-358-2282
Fax: 212-222-0598
Email: [email protected]
Prepared for presentation at the State Politics and Policy Annual Meeting, East Lansing, MI,
May 13-14, 2005.
Abstract
This research adopts Bayesian multilevel models using annual state panel data to
investigate the change in welfare caseloads spanning the period from 1977 to 2003, with
particular attention to the impact of state welfare reforms on caseloads in the post-TANF era.
This paper starts with generating TANF policy indices by state and by year on a
two-dimensional policy space—namely, that the first dimension captures the eligibility policy
and the second dimension summarizes the flexibility policy. These two indices are used as a
proxy for state TANF programs in my following data analysis. Three major findings in this
paper are as follows: first, while the eligibility policy is negatively associated with caseloads,
the flexibility policy is positively associated with caseloads. Second, although both economic
and policy factors can explain substantive shares of the caseload decline in the mid 1990s,
these two factors cannot explain any share of the caseload decline in the most recent years.
Third, the raises of minimum wages and EITC rates by the federal government also help
reduce welfare caseloads
1
Introduction
Since the Clinton administration, the number of families enrolling in welfare cash
programs (i.e., Aid to Families with Dependent Children and Temporary Assistance for
Needy Families) has declined dramatically. Nationwide, the annual AFDC/TANF caseload
has decreased by nearly 60% between 1993 and 2003. During this period, the United States
has experienced the third longest economic expansion of the twentieth century. This has
resulted in a dramatic growth of new jobs and the lowest unemployment rate in a generation.
Simultaneously, the welfare system has undergone a major transformation, in which states
made changes to their welfare programs through federal waivers from AFDC requirement.
The rapid growth of caseloads in the early 1990s was frequently cited as a primary
motivation for these changes in state welfare programs (Blank 2001, 2002). The welfare
reform legislation passed in 1996 (i.e., PRWORA) furthered these changes by replacing
AFDC with TANF and considerably increased states' autonomy to experiment with their own
welfare programs.
The key question is whether the drop of cash assistance caseloads in the mid-1990s was
caused by the numbers of welfare reform polices or by the strong U.S. economy during in the
same time period. Some welfare reformers used caseload decline as an indicator of success
for new policies with a strong emphasis on work and independence. Others argued that the
strongest economic growth period of the 1990’s provided welfare recipients better
employment opportunities to leave cash assistance.
Spurred by a report by the Council of Economic Advisor, a substantial body of literature
uses a combination of national and administrative data to study the impacts of economy and
welfare reform. While there is general agreement that the economy and welfare reform are
both factors in explaining welfare caseload reduction in the mid-1990s, no agreement exists
as to the proportion of the decline attributable to each (Blank 2001; Council of Economic
Advisors 1997, 1999; Figlio and Ziliak 1999; Levine and Whitmore 1998; Moffitt 1999;
2
Wallace and Blank 1999; Ziliak et al. 2000). For instance, the CEA report (1997) estimated
that 30.9% of the caseload reduction between 1993 and 1996 was due to welfare reform,
44.1% to the strong economy. On the other hand, Ziliak et al. (2000) analyzed the caseload
decline during the same period and reached different conclusions by attributing only 6% of
the decline to welfare reform and 78% to the economy. Among other studies, the relative
importance of welfare reform and the economy on welfare caseload reduction generally
ranged between the estimates obtained from the above two analyses.
The ambiguity continues because numerous relevant studies have asked different
questions and assessed different populations. There are no generally agreed upon
specifications for the major variables across studies and overall methodologies vary widely
from one study to the next. For instance, Moffitt (1999) replicated the 1997 CEA study using
individual household interview data rather than aggregate administrative data. Ziliak et. al.
(2000) used monthly rather than annual data to distinguish between the short-term and
long-term effects of welfare reform. Bartik and Eberts (1999) examined a range of economic
measures to better represent employment opportunities for welfare recipients than the overall
unemployment rate in a state. Wallace and Blank (1999) and Blank (2001) added political and
demographic variables to the model and examined single- and two-parent households
separately. While these studies could distinguish the sources of their disagreement, neither
method has emerged as being distinctly superior.
Although the relevant literature utilized diverse methodologies and specifications, much
of this work was based on state panel data by state and year. Blank (2002) examined 16 major
studies since the 1997 CEA report and indicated that those studies using annual state panel
data actually produced reasonably consistent results. Blank concluded that in the AFDC era, a
1 percentage point increase in the unemployment rate appears to increase caseloads by about
5 to 7 percent while the implementation of a waiver program produced a 5 to10 percent
decline in caseloads. However, because states under the TANF block grant are free to design
3
more extensive programs to reduce caseloads, such as time limits, sanctioning policies, and
diversion policies, any estimate that uses historical evidence on the AFDC program to predict
future changes in TANF-funded programs are probably unreliable (Wallace and Blank 1999).
In the most recent years, the economy has cooled and welfare reform efforts have begun
to mature. Early results from analyses of the years between 1996 and 1998 suggested that
welfare reform policies may have a greater impact on caseload reduction than they did during
the period from 1993 to 1996 (CEA 1999; Wallace and Blank 1999; Rector and Youssef 1999;
Shoeni and Blank 2000). Meanwhile, the economy might have mattered less to caseload
changes during this time. This preliminary result suggested that the responsiveness of
caseloads to unemployment might decrease in the TANF era. However, the effectiveness of
the TANF program observed in those post-TANF analyses is also based on the data in an
economic expansion. It is still difficult to forecast what will happen to caseloads in a future
recession.
Figure 1 shows that the period of decreasing unemployment rate, or economic expansion,
ended in 2000. The average annual unemployment rate increased about 1.8 percentage points
in the nation between 2000 and 2003. However, it appears that the number of national TANF
caseloads did not respond to this increasing unemployment trend, which contradicts to the
prediction based on the historical data in the AFDC era.
Figure 1 about here
Beyond the methodological and specification differences in the existing research, a key
substantial question regarding the debate of “welfare reform vs. economy” remains
unanswered and should be our major concern—that is, whether the decline in caseloads in the
mid-1990s is permanent and how much of it would be reversed in a recession. In fact, no
study to date has suggested that reforms work better in a weak economy. Since TANF
changed the funding for public assistance from a matching grant system to a fixed block grant,
states now bear the residual financial risk of any change in economic need. Thus, how
4
effective the TANF program on reducing public assistance caseloads during a recession is
important for future adjustments on state policy and funding options.
Additionally, although some policymakers speculated that other policies such as the
1990 and 1993 expansion of the federal Earned Income Tax Credit (EITC), and the raise of
federal minimum wages may also affect welfare caseloads, few studies actually put these
policy variables in their equations simply because they are state-invariant in any given time.1
All existing regression analyses on state panel data used the year-fixed effects to remove any
common changes occurring in all states in the same year. This approach indeed removes the
effects of policies (e.g. expansion of EITC and changes of minimum wages) that were
implemented everywhere at once. However, it cannot explore or control the real effects of
these nationwide policies. Similarly, those analyses also utilized the state-fixed effects to
remove long-term state-specific difference. And again, some time-invariant effects such as
regional effects would not be specified in the model or effectively controlled.
The purpose of this paper is twofold. First, this paper aims to provide an updated
analysis of caseload changes by incorporating the post-TANF data from 1997 to 2003. This
time span includes an economic recession between 2000 and 2003. Thus, we can directly
investigate how the TANF program dealt with the recession.
Second, unlike simply using year- and state- fixed effects, this analysis utilize a
multilevel modeling approach by incorporating group-level predictors to explicitly control for
important effects of state- and time-invariant variables. Specifically, these group-level
predictors directly control for key policy changes in the federal level and geographical
characteristics specific to a state.
The multilevel models in this analysis will be estimated via Bayesian methods. The
Bayesian framework combines prior (non-sample) information with sample data to produce a
1
Grogger’s (2004) study on how EITC affected welfare entries and exits and Page, Spetz, and Millar’s (2005)
study on how minimum wages affected welfare caseloads are two exceptions. However, both of their analyses
utilized policy variations among states and did not directly focus on the policy changes in the federal level.
5
posterior distribution of parameter estimates. The mean and standard deviation of this
distribution is comparable to OLS and maximum likelihood parameter estimates.2
TANF Programs
Early regression analyses focusing on the effects of state welfare waivers on AFDC
caseload changes generally characterized policy shifts by dummy variables that “turn on”
when a policy is implemented in the state. The effects of welfare waivers are reasonably well
identified since different states adopted these waivers at different points in time. In contrast,
how to identify the policy effects in the TANF era has been a problem in the more recent
econometric studies. Some papers tried to identify effects by combining waivers and TANF,
coding a dummy variable that equals one if a state has a major waiver in effect or if it has
adopted a TANF program. This has the odd effect of forcing waivers and TANF programs to
have identical effects. Blank (2002) argued that this approach is almost surely not justified
given how much more extensive were the changes involved with state TANF plans.
This paper generates TANF policy indices by state and by year on a two-dimensional
policy space—namely, that the first dimension captures the eligibility policy and the second
dimension summarizes the flexibility policy. In other words, the first index measures the
degree to which states restrict access to welfare service, such as time limit, work requirement,
family cap, and others, while the second index measures the degree to which states enhance
access to welfare service, such as income disregards, asset disregards, income tax credits, and
others. These two indices refer to the extent of state TANF programs over time and should
have opposite effects on changes in caseloads. That is, a high level of the eligibility policy is
expected to reduce welfare caseloads while a high level of the flexibility policy is expected to
increase welfare caseloads.
2
Please refer to Gelman et al. (1995), Jackman (2000) and Western and Jackman (1994) for introductions to
Bayesian inferences.
6
Following Fellows and Rowe’s (2003) approach and coding scheme, I use 28 unique
state-eligibility rules to construct an eligibility policy index.3 Each state is assigned 1 point if
the rule is in place and 0 point if it is not. Total scores for each state and every year were
calculated, where higher numbers indicate a more strict approach to eligibility policy; or that
more rules are in place to limit welfare participation. The final eligibility policy index is a
value of the score divided into the maximum score value and multiplied by the percent of the
year in which the state has actually implemented its TANF programs.4
The second dimension, flexibility policy index, includes 12 state decisions (rules) on the
flexibility of work requirements for TANF programs. Each state is assigned 1 point if the rule
is in place and 0 point if it is not. Total scores for each state and every year were calculated,
where higher numbers indicate a more flexible policy approach to work requirements. Again,
the final flexibility policy index is a value of the score divided into the maximum score value
and multiplied by the percent of the year in which the state has actually implemented its
TANF programs.5
Figure 2 illustrates the average state eligibility and flexibility indices from 1997 to 2003.
Each index is multiplied by 100 for case of presentation. Note that the low values of both
indices in 1997 are simply due to the fact that most of the states had not implemented their
TANF programs until the middle of 1997.
Figure 2 about here
In Figure 2, we can observe that state TANF programs gradually changed over the past
several years. Specifically, while the flexibility policy index almost remained constant for the
period from 1998 to 2003, the eligibility policy index, measuring welfare policy strictness,
3
Please refer to Appendix A in Fellows and Rowe (2003) for details regarding the components of eligibility and
flexibility indexes.
4
Some of the states began their TANF programs in the late 1996 and most of the states began Their TANF
programs at some point of time in 1997. One exception is California, which had not implemented its TANF
program until January 1998.
5
The Cronbach Alpha coefficients for the eligibility and flexibility indices are .71 and .74 respectively, which
are slightly lower than those in Fellowes and Rowe’s original indices using the 1997-1999 data (i.e. .78 and .76).
7
dropped about 10 percent (e.g. 5.6 points in a 100-point scale) for the same period. It seems
that the decrease of welfare policy strictness happened simultaneously with the increase of
the unemployment rate between 2000 and 2003. This scenario is different from that in the
mid-1990s when the increase of welfare policy strictness (through AFDC waivers) happened
along with the decrease of the unemployment rate.
In the following regression analyses, I use both eligibility and flexibility indices as
policy indicators to proximate state-level TANF programs across states over years in the
TANF era.
Multilevel Modeling
Multilevel modeling accounts for the variation in an outcome variable measured at the
lowest level of analysis, while using explanatory information from that level as well as others
(Steenbergen and Jones 2002). Simply including group-level (e.g. state and year) indicators
representing “fixed effects” not only eats up valuable degree of freedom, but also leaves the
resulting estimates as unexplained. The multilevel modeling allows the researcher to combine
data sources from multiple sources and explain both group-level and individual-level
variations directly (Gelman 2002; Kreft and de Leeuw 1998; Shor et al. 2003).
Additionally, multilevel models are typically more conservative than their classical
counterparts because uncertainty in multilevel models is estimated at both the individual and
group levels. Besides better estimates of uncertainty, multilevel models also generate more
accurate parameter estimates. This increased accuracy mainly results from the inclusion of
context (i.e. group-level) effects (Western 1998).
Data
I use a panel (time-series cross-sectional: TSCS) data set across 51 “states” (i.e., the
District of Columbia is included) over 27 years, from 1977 to 2003. The data contain a total
8
of 1,377 individual-level observations. The multilevel data set has information at three levels
of analysis: an individual level (state-year) and two group levels (states and years). In other
words, data which changes across time and space is individual-level; whereas information
about the time context common to all states, or the state context common to all the
measurements of a particular state are both considered group level. The outcome variable is
measured at the individual level.
The outcome variable in this study is per capita AFDC/TANF caseload in a given
state-year. I first divided the number of AFDC/TANF caseloads by the total population in
each state-year to derive the per capita welfare caseload figure.6 Since there are still wide
variations in this caseload figure (from a low of 0.08 in Wyoming in 2003 to a high of 4.82 in
DC in 1994), I then took the natural log of the per capita caseloads to smooth out these
nonlinearities.
A number of state-year (i.e. individual level) explanatory variables were collected from
publicly available sources. These include several key economic and demographic variables
commonly used in the relevant studies—namely, unemployment rates, total population7, the
share of black population, the share of out-of wedlock births, and the share of newly admitted
immigrants. Previous studies commonly used both one-year and two-year lags of
unemployment rates to fully specify the association between dynamic economic conditions
and caseload changes (CEA 1997, 1999; Lavine and Whitmore 1998; Blank 2001; Wallace
and Blank 1999). In order to be comparable to the unemployment effects estimated in similar
annual panel data analysis, both lags of unemployment rates are included in this analysis as
economic variables. Following Blank’s (2001) specification, information on the party
affiliation of the governor and the majority party of the state House and state Senate are also
6
The results change little when I re-estimate my models using female population, age 15-44, as the
denominator.
7
The existing literature seldom put population in the model. I use this variable to control for small state bias.
The natural log of the total population for a state in a given year is the actual variable included in the model.
9
included as political variables. Additionally, instead of the current share of newly admitted
immigrants, one-year and two-year lags of share immigrants are included in the model.
Regarding AFDC/TANF policy variables, I use information on which states were
granted Federal waivers after 1991 to implement state wide changes in their AFDC programs.
The waiver variables are coded as the percent of the year in which the state has a major state
waiver approval for implementation.8 In the TANF era, the eligibility and flexibility indices
for each state in a given year replaced the AFDC waiver indicators. Additionally, information
on state maximum AFDC/TANF benefits for the family of three is also included as a key
policy parameter.
In the group level, three state-invariant and six time-invariant variables were collected to
specify year and state effects respectively. The former includes measures of yearly federal
EITC rates, federal minimum wages, and an indicator for the Republican presidency over the
27-year time span. The latter includes measures of land area and 5 regional indicators for all
51 “states.”9 The main purpose to include these random intercepts for region is to model a
potential source of spatial correlation of errors for this analysis. Because states seem to be
concerned about the implementation of welfare policy from their neighboring states (e.g.
“race to the bottom”), neighboring states may have correlated errors. We can pick up some of
this affect with the random intercepts for regions. Because this approach lets the state-level
variance be estimated from the data rather than be fixed, it is not the same as simply
including fixed effects for regions in an OLS framework (Shor et al. 2003).
Source information and summary statistics for all variables can be found in Appendix I
and II respectively.
8
The coding scheme for the wavier variables is the same as that in the 1997 CEA report, which has been
adopted in most relevant research.
9
I provided Washington DC with its own regional code because most measures of Washington DC have extreme
values that could unduly influence the results for the other states.
10
Models
I begin my empirical analysis by specifying a basic model. The general from of the basic
model is as follows:10
y it = ( Xβ ) it + α i + δ t + ε it
(1)
where y it stands for estimated caseload in a given state i (1 to 51) and year t (1 to 27). X is a
vector of state-year predictors.βis a vector of the coefficients to be estimated. The
hyperprior σ yit is vague with mean 0 and a variance that is estimated from the data. The
multilevel nature of the model is revealed in theαandδterms.
α ~ N ( µ α , σ α ), µ α = ( Sγ ) i , σ α ~ Γ −1 (0.001, 0.001)
(2)
δ ~ N ( µ δ , σ δ ), µ δ = (Tλ ) t , σ δ ~ Γ −1 (0.001, 0.001)
(3)
αis the indicator for state i andδis the indicator for year t. They are normally distributed
with means derived from the product of the matrices of state-and year-level predictors (S and
T) multiplied by the state- and year-level coefficients (γandλ) respectively. The group-level
variance σ α and σ δ are random and are given very vague inverse gamma priors.
Note that these group level indicators can be considered “random intercepts with fixed
slop” since they modify the individual-level constant but do not vary with the level of any
individual-level parameters. The assumption here is that the state- and year- predictors
affecting welfare caseloads do not differ in different years and states. However, allowing for
the possibility of fully varying slopes can be an extension of this methodology. For now, this
“fixed-slop” assumption for the group-level parameters seems reasonable enough, given the
high similarity of states to each other under the U.S. federal system.
In addition to the state and year effects, I also extend the basic model with a
unit-specific time trend for every state to account for the possible serial correlation. This is
done in a linear fashion for simplicity as illustrated in the following from:
10
I follow Shor et al.’s (2003) notation for Bayesian multilevel models with TSCS data.
11
y it = ( Xβ ) it + α i + δ t + ρ i t + ε it
(4)
The time trend model is very similar to the base model except for the addition of ρ i ,
which is the estimated time trend for caseloads in a given state i.
The basic and time trend models simply ignore the persistence effect of past caseload
levels on current caseload sizes. Ziliak et al. (2000), Figlio and Ziliak (1999), and Bartik and
Eberts (1999) considered sheer persistence in caseload levels as an alternative explanation on
caseload changes.11 They observed that once the persistence effect is taken into account, the
falling caseloads of the mid-1990s are no longer systematically associated with the
implementation of the waiver reforms as in other specifications. In short, the persistence
model suggested that persistence of past levels plays a role, not just changes in current
circumstance.
To adjust this persistence effect, I expanded the analysis to allow current caseloads to
dynamically adjust to past caseload levels. Specifically, the persistence effect is added to the
model through the inclusion of a lagged dependent variable as an independent variable. This
is also an alternative approach to deal with serial correlation. The estimated equation of the
persistence model is as the following:
y it = ( Xβ ) it + α i + δ t + y it −1 + ε it
(5)
where y it −1 stands for the caseload in a given state i and year t-1.
The multilevel models introduced here were estimated via Bayesian software WinBUGS
(Spiegelhalter et al. 1999) as called from R (R Development Core Team 2003) using
Gelman’s (2003) Bugs.R.. The models were run with 6 chains of 25,000 iterations each.
One-half of the iterations were discarded as a burn-in. The chains converged successfully for
all the parameters of the model, as measured by the Gelman-Rubin-Carlin Rhat statistics.
11
The persistence effect is modeled either by including the previous year’s (or month’s) caseload as an
explanatory variable in a regression model or by analyzing year-to-year changes in caseloads in a “first
difference” framework that assumes the current caseload mirrors last year’s caseload except due to changes in
other measured factors.
12
Estimates
Table 1 shows all the estimates from four different specifications. I start with an
extremely sparse specification in Model 1, which only includes economic variables, policy
variables, as well as state- and year- fixed effects. This specification is comparable to much
of the other relevant literature, including the 1997 and 1999 CEA reports. The first column of
Table 1 shows the results of this estimation.
Table 1 about here
The results in Model 1 are similar to those seen elsewhere in the literature. A one point
rise in the unemployment rate raises caseloads by 6.8 percent over a three year period. This is
comparable to the unemployment effects estimated in similar annual panel data analysis.12
The other results in Model 1 are also similar to earlier work. States with higher
AFDC/TANF benefits have higher caseloads. Additionally, states that implemented any major
AFDC waiver experienced a 9 percent decline in caseloads. The magnitude of this effect is
consistent with the estimates elsewhere in the literature.
The results in Model 1 also indicates that TANF-related policy can affect
caseloads—namely, that a one standard deviation (i.e. approximately 14 points in a 100-point
scale) rise in the eligibility index decreases caseloads by 15.6 percent while a one standard
deviation (i.e. approximately 16 points in a 100-point scale) rise in the flexibility index
increases caseloads by 11.3 percent. These results show that the eligibility and flexibility
policies have opposite effects on welfare caseloads.
Model 2 utilizes a richer set of explanatory variables in both individual and group levels
to estimate caseloads. The inclusion of these other variables has little effect on the estimated
effect of unemployment. A one-point rise in unemployment raises the caseload share by 6.9
percent over three years, similar to the estimates in Model 1.
While the results for AFDC/TANF benefits and TANF-related policy in Model 2 are also
12
Please refer to Table 6 in Blank’s (2002) comprehensive literature review.
13
similar to those in Model 1, the estimated effect of the implementation of any major AFDC
waiver on the reduction of caseload share declines from 9 percent in Model 1 to 5.3 percent
in Model 2. Additionally, unlike that in Model 1, the 95% posterior distribution of the
coefficient for waiver implementation in Model 2 includes zero. These results suggest that
when we control for more variables in the model, the waiver effect on caseloads decreases
and the degree of uncertainty over this effect increases substantively.
Among the demographic variables in Model 2, only the share of non-marital births is
important and has a 95% posterior distribution that does not cross zero. Because a larger
share of non-marital births indicates an increase of single-parent households, the group which
is potentially eligible for welfare benefit if they find themselves in economic difficulty, it is
not surprising that the larger the share of non-marital births, the larger the caseload share will
be.13
The share of black population is weakly negatively correlated with caseloads although it
has a 95% posterior distribution that includes zero. Two lags of share new immigrants in a
state has mixed effects on caseloads; that is, while the one-year lag of share new immigrants
has a negative effect on caseloads, the two-year lag has a positive effect. However, neither
parameter estimates has a 95% interval that does not cross zero. In short, these findings
suggest that both the share of black population and the share of newly entered immigrants in
a state do not have significant effects on caseloads.
Blank (2001) argues that Republican state administrations may implement more
restrictive welfare policies. Her empirical analysis showed that the party of the governor is
strongly significant, with lower caseloads under Republican governors. Furthermore, Blank
also argued that because welfare policy is often a crucial political issue in state legislatures,
13
One may speculate that there is a two-way flow of influence between the formation of single-mother
households and the presence of welfare benefits. Thus, the variable of share non-marital birth is endogenous in
the equation and its coefficient is biased upward. However, previous evidence suggests this may not be a major
problem (Moffitt 1992; Blank 2001).
14
states with split party control may have more difficulty passing welfare reform legislation.
Her empirical findings suggested that states where both the House and Senate are controlled
by the same party also have lower caseloads, regardless of whether that party is Democratic
or Republican.
The results for the political variables in Model 2 do not fully support Blank’s findings.
While states with Republican governors are indeed more likely to have lower welfare
caseloads, states with unified-controlled legislatures (i.e., both the House and Senate are
controlled by the same party.) do not necessarily have lower caseloads. Instead, states with
Republican-controlled legislatures are more likely to have lower welfare caseloads while
states with Democratic-controlled legislatures are more likely to have higher welfare
caseloads. This evidence suggests that Republicans may be consistently in favor of more
restrictive welfare policies to reduce caseloads, regardless of which institutions they control.
On the other hand, Democrats may be consistently in favor of less restrictive welfare policies
that may cause an increase in caseloads, regardless of which institutions they control. These
findings conform to intuitive policy differences between Republicans and Democrats (or
conservatives and liberals).14
The year level predictors in Model 2 provide some empirical evidence supporting the
speculation that the raises of minimum wages and EITC rates by the federal government may
help reduce welfare caseloads. That is, both predictors are negatively correlated with
year-fixed effects on caseloads though the minimum wage predictor has a 95% interval that
includes zero. A one dollar increase in the federal minimum wage decreases caseloads by 6.5
percent, and a one percentage point increase in the EITC rate decreases caseloads by 2.4
percent. In addition to these two policy variables in the federal level, the party of the
president is also strongly significant, with lower caseloads under Republican presidents.
14
Fellows and Rowe (2003) analyzed state TANF programs and also found that state welfare generosity
increases as the proportion of Democrats in government increases. The similar pattern regarding the party
difference in welfare policy can also be seen in Soss et al. (2001)
15
The regional indicators in the state level in Model 2 do not have substantial impacts on
welfare caseloads through the state-fixed effects. The variance for each regional parameter
estimate is considerably large and it is difficult to conclude whether there exists any regional
effect on welfare caseloads.
Model 3 uses the same specification as Model 2, but also includes state-specific time
trends in the estimations. It is not surprising that the inclusion of state time trends reduces the
magnitude of most of the estimated coefficients.15 But the signs of most variables remain
unchanged (except for the share of black population and the Northeast region indicator). In
short, because the results for the key variables including the economic and policy variables in
Model 3 are similar to those in Model 2, the inclusion of state time trends changes few of the
major conclusions drawn from Model 2.
Model 4 includes a lagged dependent variable as an independent variable to adjust the
persistence effect of past caseload levels on current caseload levels. The inclusion of the
lagged dependent variable substantively reduces the magnitude of almost all the estimated
coefficients (except for the current and one-year lag of unemployment rates). Furthermore,
the lag dependent variable even drives the sign of the two-year lag of unemployment rates
flipped. As Achen (2000) argued, lagged dependent variables can absorb so much explanatory
power that the effects of typically influential covariates are biased and driven to zero.
In short, the inclusion of the lagged dependent variable may be a preferred method to
deal with serial correlations (Beck 2001). And indeed the persistence model may match up
well with this heavily trending caseload data. However, we cannot substantively explain the
change in caseloads based on real economic and policy variables when we utilize the
persistence model.16 Shor et al.(2003) suggested that the addition of a lagged dependent
variable may require an adjustment to the theory that drives the model. Thus, in addition to
15
The inclusion of state time trends may subsume the effects of the variables that are simply trending up (or
down) in a linear way.
16
See Bell (2001) for Blank’s discussion on the downsides of the persistence model.
16
show how persistence effect works on caseload changes, we need a theory to explain why
persistence effect exists at the first place.
Relative Contributions of Economy and Welfare Reforms
In this section, I followed the method used in the 1997 and 1999 CEA report to estimate
the relative contributions of economic and policy factors to the change in welfare caseloads
during three periods: 1993-1996 (the AFDC waiver period under the Clinton Administration),
1996-2000 (the TANF period under the Clinton Administration), and 2000-2003 (the TANF
under the Bush Administration). Specifically, the change in the national average of each
variable (obtained by weighting by the state population) is multiplied by its respective
coefficient estimate to determine the predicted change induced by the factor. Then the ratio of
the predicted change to the actual change in welfare caseloads is calculated during the three
periods respectively. Table 2 presents the results of this exercise for each model specification
in Table 1.
Table 2 about here
The results in Panel A of Table 2 are similar to the findings in previous literature.
Between 1993 and 1996, while the AFDC waivers and welfare benefit levels accounted for
about 22 to 42 percent of the caseload decline (based on the estimates in models that ignore
the persistence effect), the lower unemployment rate was responsible for about 38-41 percent
of the decline (depending on the model).17 When the persistence effect is taken into
consideration (i.e. Model 4), the AFDC waivers and welfare benefit levels accounted for only
about 6 percent of the caseload decline from 1993 to 1996; the lower unemployment rate
contributed about 22 percent of the decline.
17
The estimated percent of caseload decline due to the policy factor tends to be higher than those in earlier work.
It is because I included the measure of maximum welfare benefits when I calculated the contribution of policy
factor. Without considering the persistence effect, the sole contribution of AFDC waivers to the caseload decline
during this period ranged from 14% to 24%, which is similar to most previous studies using annual panel data.
17
The results in Panel B of Table 2 indicate that in the early TANF era from 1996 to 2000,
both economic and policy factors contributed relatively small toward the decline in welfare
caseloads. The estimates based on Model 1-3 suggest that roughly one-fifth of the decline
during this period was due to the policy factor and one-eighth of the decline was due to the
economic factor. During the later TANF period between 2000 and 2003, neither economic nor
policy factor had any major contribution to the change in welfare caseloads. Each of them in
fact explained a negative share of the caseload decline during this period, as shown in Panel
C of Table 2. Because the unemployment rate increased and state TANF programs were less
restrictive between 2000 and 2003, the caseload would have increased during this period if
only the economic and policy factors had contributed toward the change in welfare caseloads.
In short, the decline in welfare caseloads between 2000 and 2003 was not due to the change
in either economic or policy factor but may be due to other changes in policy practice or
habitual behavior.
Conclusion and Discussion
This paper investigates changes in welfare caseloads over a 27-year period, with
particular attention to the impact of state TANF programs on changes in caseloads between
1997 and 2003. Its primary findings and conclusions include:
First, state TANF programs are composed of two distinct policies—namely, the
eligibility (strictness) and flexibility policies. While the eligibility policy is negatively
associated with caseloads, the flexibility policy is positively associated with caseloads.
Additionally, state TANF programs changed over time. Specifically, state welfare policies are
less restrictive in the most recent years than they were in the late 1990s. This decreasing trend
of policy strictness happened along with the increasing trend of unemployment between 2000
and 2003.
Second, the results of the regression analyses in this paper are similar to those in the
18
previous studies using annual state panel data. Specifically, state economies have a significant
effect on caseloads; political and policy variables also matter.
Third, the use of multilevel models in this research allows us to investigate some
important effects of state- and time-invariant variables on welfare caseloads. The results
suggest that the raises of minimum wages and EITC rates by the federal government have
certain impacts on reducing welfare caseloads.
Fourth, both economic and policy factors contributed substantively toward the decline in
welfare caseloads between 1993 and 1996 if we exclude the persistence effect on caseloads.
My estimation of relative contribution of each factor on the caseload reduction during this
period is consistent with previous relevant research. However, some preliminary results here
suggest that neither economic nor policy factor plays a major role in explaining the caseload
decline in the most recent years.
Several issues are prime candidates for future research in understanding the impact of
welfare reforms on welfare caseloads:
First, it is important to develop more nuanced measures of TANF policy and policy
implementation. This paper utilized the most intuitive method and a limited number of policy
items to construct eligibility and flexibility policy indices that summarize state TANF
programs. Because the Welfare Rules Database maintained by The Urban Institute contains
numerous rules that states have in place at every stage of welfare reform, we can measure
state TANF programs in greater details if other advanced scaling techniques can be applied to
that extensive dataset.
Second, although both economic and policy factors can explain substantive shares of the
caseload decline in the mid 1990s, they cannot explain any share of the caseload decline in
the most recent years. The current caseload decline might be due to some other long-term
factors that cannot be measured by traditional economic and policy variables. For instance,
we may speculate that the long-term practice of the AFDC waivers and TANF programs have
19
led to some habitual behavior or institutional barriers for welfare participation, which in turn
cause an increase in the exit rate and a decrease in the entry rate. It may result in a long-term
decline in caseloads, regardless of any short-term economic or policy change (Grogger 2004).
Third, it has been a long term debate in the relevant literature regarding whether a model
should include any lagged dependent variable as an independent variable to adjust the
persistence effect of past caseload levels on current caseload levels. One of the model
specifications in this research (Model 4) shows that the persistence effect indeed exists and
the lagged dependent variable substantively explains the changes in caseloads. Future studies
may focus on decomposing this persistence effect and exploring the factors that cause
persistence to be present.
20
References
Achen, Christopher H. 2000. “Why Lagged Dependent Variables Can Suppress the
Explanatory Power of Other Independent Variables.” Paper presented at the 2000
Annual Meeting of the Political Methodology Section of the American Political
Sceince Association, UCLA.
Beck, Nathaniel. 2001. “Time-Series-Cross-Section Data: What have We Learned in the Past
Few Years.” Annual Review of Political Science 4: 271-93.
Bertik, Timothy J., and Randall W. Eberts.. 1999. "Examing the Effect of Industry Trends and
Structure on Welfare Caseloads." In Sheldon H. Danziger, eds., Economic Conditions
and Welfare Reform, pp. 119-58. Kalamazoo, MI: W.E. Upjohn Institute for
Employment Research.
Blank, Rebecca M. 2001. "What Causes Public Assistance Caseloads to Grow? Journal of
Human Resources 36: 85-118.
Blank, Rebecca M. 2001. "Evaluating Welfare Reform in the United States” Journal of
Economic Literatures 40: 1105-66.
Shor, Boris, David Park, Joseph Bafumi, and Andrew Gelman. 2003. “Examining Time
Series Cross Sectional Data with Bayesian Multilevel Models.” Paper presented in the
2003 Annual Meeting of American Political Science Association, Philadelphia, PA.
Council of Economic Advisors. 1997. "Explaining the Decline in Welfare Receipt,
1993-1996." Washington DC.
Council of Economic Advisors. 1999. "The Effects of Welfare Policy and the Economic
Expansion on Welfare Caseloads: An Update." Washington DC.
Fellowes, Matthew C. and Gretchen Rowe. 2004. "Politics and the New American Welfare
States." American Journal of Political Science 48: 362-73.
Figlio, David N., and James P. Ziliak. 1999. "Welfare Reform, The Business Cycle, and the
Decline in AFDC Caseloads." In Sheldon H. Danziger, eds., Economic Conditions
21
and Welfare Reform, pp. 17-48. Kalamazoo, MI: W.E. Upjohn Institute for
Employment Research.
Gelman, Andrew, John B. Carlin, Hal S. Stern, and Donald B. Rubin. 1995. Bayesian Data
Analysis. New York: Chapman and Hall/CRC.
Gelman, Andrew. 2002. Draft Notes for Multilevel Models. (Available from
http://www.stat.columbia.edu/~gelman.)
Gelman, Andrew. 2003 Bugs.R: Functions for Calling Bugs from R. (Available from
http://www.state.columbia.edu/~gelman/bugsR.)
Grogger, Jeffrey. 2004. “Welfare Transitions in the 1990s: The Economy, Welfare Policy, and
the EITC.” Journal of Policy Analysis and Management 23: 671-95.
Jackman, Simon. 2000. “Estimation and Inference via Bayesian Simulation: An Introduction
to Markov Chain Monte Carlo.” American Journal of Political Science 44: 375-404.
Kreft, Ita G. G., and Jan de Leeuw. 1998. Introducing Multilevel Modeling. London; Thousand
Oaks, Calif: Sage.
Levine, Philip B. and Diana M. Whitmore. 1998. “The Impact of Welfare Reform on the
AFDC Caseload.” National Tax Association Proceedings-1997: 24-33.
Moffitt, Robert A. 1999. "The Effect of Pre-PRWORA Waivers on AFDC Caseloads and
Female Earnings, Income, and Labor Force Behavior." In Sheldon H. Danziger, eds.,
Economic Conditions and Welfare Reform, pp. 91-118. Kalamazoo, MI: W.E. Upjohn
Institute for Employment Research.
Page, Marianne E., Joanne Spetz, and Jane Millar. 2004. “Does the Minimum Wage Affect
Welfare Caseloads?” Journal of Policy Analysis and Management 24: 273-95.
R Development Core Team. 2003. R: A Language and Environment for Statistical Computing.
Vienna, Austria: R Foundation for Statistical Computing. (Available from
http://www.R-project.org.)
Rector, Robert, and Sarah Youssef. 1999. "The Determinants of Welfare Caseload Decline"
22
Washington DC: Heritage Foundation.
Schoeni, Robert F., and Rebecca M. Blank. 2000. “What Has Welfare Reform Accomplished?
Impacts on Welfare Participation, Employment, Income, Poverty, and Family
Structure.” NBER work paper 7627, Cambridge, MA.
Soss, Joe, Stanford Schram, Tom Vartanian, and Erin O’ Brien. 2001. “Setting the Terms of
Relief: Explaining State Policy Choices in the Devolution Revolution.” American
Journal of Political Science 45: 378-95.
Spiegelhalter, David, Andrew Thomas, and Nicky Best. 1999. WinBugs Version 1.4.
Cambridge, UK: MRC Biostatistics Unit.
Steenbergen, Marco R., and Bradford S. Jones. 2002. “Modeling Multilevel Data Structures.”
American Journal of Political Science 46: 218-37.
Wallace, Geoffrey, and Rebecca M. Blank. 1999. "What Goes Up Must Come Down?
Explaining Recent Changes in Public Assistance Caseloads." In Sheldon H. Danziger,
eds., Economic Conditions and Welfare Reform, pp. 49-89. Kalamazoo, MI: W.E.
Upjohn Institute for Employment Research.
Western, Bruce, and Simon Jackman. 1994. “Bayesian Inference for Comparative Research.”
American Political Science Review 88: 412-33.
Western, Bruce. 1998. “Causal Heterogeneity in Comparative Research: A Bayesian
Hierarchical Modeling Approach.” American Journal of Political Science 42:
1233-59.
Ziliak, James P., David N. Figlio, Elizabeth E. Davis, and Laura S. Connolly. 2000.
"Accounting for the Decline in AFDC Caseloads: Welfare Reform or the Economy?"
The Journal of Human Resources 35: 570-86.
23
Appendix I
The following list documents the sources for all variables included in the regression
analyses.
AFDC/TANF caseloads
Monthly data by state from 1977-1996 were acquired
electronically from the U.S. Department of Health and Human
Services.
State population
U.S. Census Bureau: State Population Estimates.
Unemployment rates
U.S. Department of Labor, Bureau of Labor Statistics: State
and Area Current Statistics
Percent black
Number blacks divided by total state population. Data for
1977-1996 were acquired from Blank (2001). Number black
for 1997-2003 is available through U.S. Census Bureau: State
Population Estimates.
Percent non-marital births
Number of live births to unmarried women divided by the total
number of live births. Data for 1977-1996 were acquired from
Blank (2001). Data for 1997-2003 are available through the
Center for Disease Control’s National Center for Health
Statistics: National Vital Statistics Reports.
Percent new immigrants
Number of newly-arrived immigrants divided by total state
population. Data for 1977-1996 were acquired from Blank
(2001). Data for 1997-2003 are available through U.S.
Immigration and Naturalization Service: The Statistical
Yearbook.
Indicator for Republican
governor
Coded from information in various editions of The Book of the
States. D.C. is considered Democratic.
Indicator for
Republican-controlled
state Senate
Coded from information in various editions of The Book of the
States. In years where there is an exact tie between the number
of Republican and Democratic senators, the variable is coded
0.5. D.C. is considered Democratic. Nebraska, which has a
unicameral and nonpartisan legislature is coded Republican.
Indicator for
Republican-controlled
state House
Same source as the previous variable.
AFDC/ TANF maximum
benefit levels for a family
of 3 (in 2003 dollars)
Data for 1977-1996 were acquired from Blank (2001). Data
for 1997-2003 are available through U.S. House of
Representatives: The Green Book
AFDC waivers
Equals one for all years (until 1996) after the Federal
24
government approved a major state request for a waiver to the
national AFDC program rules. For waivers granted in the
middle of the year, the variable equals the share of the year
after the waiver was approved. Data were acquired from Blank
(2001)
Eligibility Index
Relevant rules are available from Urban Institute: Welfare
Rules Database
Flexibility Index
Same source as the previous variable.
Minimum Wage (in 2003
dollars)
Data are available from U.S. Department of labor, Division of
State Standards Programs, Wage and Hour Division,
Employment Standards Administration.
Earned Income Tax Credit
rate
Data are available through U.S. House of Representatives: The
Green Book
State area
Data are available through U.S. Census Bureau.
Indicator for Republican
president
Equals one for 1981-1992 and 2001-2003.
Indicators for census
region
The coding scheme is followed by the census region (i.e.
Northeast, Midwest, West, and South) plus D.C.
25
Appendix II
Summary Statistics for the Data
Variable Description
Obs
Mean
Std Dev
Min
Max
Log (per capita AFDC/TANF caseloads)
1377
-4.473
.537
-7.136
-3.032
Log (state population)
1377
14.915
1.026
12.904
17.384
Unemployment rates %
1377
5.987
2.033
2.275
17.450
Unemployment rates % (t-1)
1377
6.045
2.066
2.275
17.450
Unemployment rates % (t-2)
1377
6.144
2.125
2.275
17.450
Share black %
1377
10.734
12.036
.221
71.881
Share non-marital births %
1377
25.879
9.737
3.640
68.800
Share new immigrants (t-1) %
1377
.197
.210
.021
2.409
Share new immigrants (t-2) %
1377
.192
.209
.020
2.409
Dummy variable for Republican governor
1377
.445
.497
0
1
Dummy variable for Republican-controlled
state legislature (House and Senate)
1377
.254
.437
0
1
Dummy variable for Democratic-controlled
state legislature (House and Senate)
1377
.522
.500
0
1
Log (AFDC/ TANF maximum benefit
levels for a family of 3)
1377
6.198
.432
4.889
7.142
Any major AFDC waiver
1377
.056
.215
0
1
Eligibility Index
1377
14.174
24.207
0
75
Flexibility Index
1377
16.169
29.099
0
100
Minimum Wage
27
5.641
.541
4.80
6.81
Earned Income Tax Credit rate %
27
20.337
10.443
10
34
Dummy variable for Republican president
27
.556
.506
0
1
Log (state area)
51
10.624
1.429
4.224
13.405
Dummy variable for Northeast
51
.176
.385
0
1
Dummy variable for Midwest
51
.235
.428
0
1
Dummy variable for South
51
.314
.469
0
1
Dummy variable for West
51
.394
.460
0
1
Dummy variable for DC
51
.019
.140
0
1
26
Table 1
Multilevel Models on AFDC/TANF Caseloads: Short Model vs. Basic Model
(Dependent Variable: Log (Per Capita AFDC/TANF Caseloads)
Model 1 (Short)
Parameter
Coef.
(Std Err)
Model 2 (Basic)
95% Posterior
Distribution
Coef.
(Std Err)
95% Posterior
Distribution
Individual Level (State-Year) Predictor
Unemployment rates %
.011
(.009)
-.007
.029
.015
(.009)
-.003
.033
Unemployment rates %
(t-1)
.012
(.013)
-.013
.036
.013
(.013)
-.013
.039
Unemployment rates %
(t-2)
.045
(.009)
.027
.062
.041
(.009)
.024
.058
-.007
-.015
.0001
Share black
(.0003)
Share non-marital births
.025
(.002)
.021
.030
Share new immigrants
(t-1)
-.031
(.065)
-.156
.102
Share new immigrants
(t-2)
.035
(.067)
-.097
.166
Log (state population)
.155
(.032)
.095
.222
Republican governor
-.065
(.011)
-.087
-.045
Republican-controlled
state legislature
-.025
(.018)
-.058
.011
Democratic-controlled
state legislature
.050
(.017)
.018
.083
Log (AFDC/ TANF
maximum benefit)
.379
(.052)
.279
.478
.309
(.050)
.218
.408
Major AFDC waiver
-.090
(.033)
-.157
-.027
-.053
(.032)
-.112
.010
Eligibility Index
-.011
(.001)
-.013
-.009
-.012
(.001)
-.014
-.009
Flexibility Index
.007
(.0005)
.006
.008
.007
(.0005)
.006
.008
27
Lag DV
Constant
-7.198
(.342)
-7.818
-6.539
-7.714
(.626)
-9.084
-6.479
Group Level (State-invariant) Predictors
Minimum Wage
-.065
(.061)
-.176
.071
EITC rate
-.024
(.005)
-.033
-.015
Republican president
-.274
(.077)
-.431
-.126
Group Level (Time-invariant) Predictors
Log (state area)
-.089
(.028
-.158
-.039
Northeast
-.025
(.092)
-.237
.126
Midwest
.024
(.087)
-.141
.194
South
.029
(.089)
-.117
.226
West
-.034
(.092)
-.026
.106
Washington D.C.
.029
(.106)
-.158
.306
State effects
Yes
Yes
Year effects
Yes
Yes
State time trends
No
No
DIC*
-642.6
-781.4
# of observations
1377
1377
*DIC is an estimate of expected predictive error (lower deviance is better).
28
Table 1 (continue)
Multilevel Models on AFDC/TANF Caseloads: Time-Trend Model vs. Lag DV Model
(Dependent Variable: Log (Per Capita AFDC/TANF Caseloads)
Model 3 (Time Trends)
Parameter
Coef.
(Std Err)
95% Posterior
Distribution
Model 4 (Lag DV)
Coef.
(Std Err)
95% Posterior
Distribution
Individual Level (State-Year) Predictor
Unemployment rates %
.010
(.007)
-.004
.023
.014
(.004)
.006
.022
Unemployment rates %
(t-1)
.011
(.010)
-.009
.034
.014
(.006)
.003
.025
Unemployment rates %
(t-2)
.044
(.006)
.031
.056
-.013
(.004)
-.020
-.005
.006
-.003
.009
-.001
-.002
.0005
Share black
(.005)
(.0007)
Share non-marital births
.016
(.003)
.010
.022
.002
(.0009)
.0005
.004
Share new immigrants
(t-1)
-.048
(.052)
-.150
.058
.0016
(.027)
-.053
.055
Share new immigrants
(t-2)
.022
(.054)
-.079
.134
.011
(.028)
-.042
.067
Log (state population)
.048
(.042)
-.038
.132
.002
(.005)
-.008
.012
Republican governor
-.055
(.010)
-.076
-.036
-.019
(.005)
-.028
-.010
Republican-controlled
state legislature
-.030
(.016)
-.060
.001
-.008
(.008)
-.023
.006
Democratic-controlled
state legislature
.014
(.016)
-.015
.044
.014
(.007)
.001
.028
Log (AFDC/ TANF
maximum benefit)
.156
(.054)
.044
.258
.065
(.017)
.033
.102
Major AFDC waiver
-.055
(.026)
-.103
-.004
-.011
(.013)
-.039
.013
Eligibility Index
-.009
(.0001)
-.011
-.007
-.001
(.0004)
-.002
-.0006
Flexibility Index
.006
(.0005)
.005
.007
.001
(.0002)
.0007
.0016
29
Lag DV
Constant
-5.393
(.763)
-6.801
-3.928
.927
(.011)
.903
.948
-.360
(.256)
-.883
.114
Group Level (State-invariant) Predictors
Minimum Wage
-.115
(.069)
-.229
.0314
-.070
(.028)
-.124
-.008
EITC rate
-.023
(.005)
-.033
-.013
-.006
(.002)
-.010
-.003
Republican president
-.285
(.088)
-.445
-.112
-.022
(.034)
-.091
.049
Group Level (Time-invariant) Predictors
Log (state area)
-.0312
(.025)
-.078
.014
-.003
(.004)
-.012
.006
Northeast
.108
(.156)
-.117
.506
-.011
(.020)
-.057
.023
Midwest
.116
(.163)
-.091
.536
.008
(.020)
-.033
.047
South
.002
(.144)
-.261
.345
.014
(.019)
-.020
.055
West
-.053
(.151)
-.322
.292
-.019
(.022)
-.070
.016
Washington D.C.
-.049
(.193)
-.489
.314
.003
(.028)
-.048
.059
State effects
Yes
Yes
Year effects
Yes
Yes
State time trends
Yes
No
-1426.4
-3021.2
1337
1337
DIC
# of observations
*DIC is an estimate of expected predictive error (lower deviance is better).
30
Table 2
Percentage of Change in Welfare Caseloads Attributable to Different Factors
A. 1993-1996: Actual Percent Change in Per Capita Caseload: -17.21%
Factor
Model 1
Model 2
Model 3
Model 4
Economy Factor (%)*
41.18
44.17
38.81
22.26
Policy Factor (%)**
42.33
29.05
22.23
6.07
B. 1996-2000: Actual Percent Change in Per Capita Caseload: -80.56%
Factor
Model 1
Model 2
Model 3
Model 4
Economy Factor (%)*
12.99
13.09
12.45
2.36
Policy Factor (%)***
19.37
22.15
17.05
1.49
C. 2000-2003: Actual Percent Change in Per Capita Caseload: -11.01%
Factor
Model 1
Model 2
Model 3
Model 4
Economy Factor (%)*
-45.65
-53.49
-42.23
-42.55
Policy Factor (%)***
-37.25
-42.97
-37.65
-4.27
Notes: *Unemployment rates over a 3- year period are included.
**Maximum welfare benefit levels and any AFDC waiver are included.
***Maximum welfare benefit levels and both TANF indices are included.
31
4
40000
60000
80000
Number of Cases
Unemployment Rate (%)
5
6
7
8
9
100000
Figure 1
AFDC/TANF Total Caseloads vs. Unemployment Rate
1970
1980
1990
Year
Unemployment Rate
2000
2010
Welfare Caseload
32
45
Policy Index (0-100)
50
55
60
65
Figure 2
Changes in State TANF Programs
1996
1998
2000
Year
Eligibility
2002
2004
Flexibility
33