Assessing the Impact of Welfare
Reform on Single Mothers
Hanming Fang Michael P. Keane
Department of Economics
• U.S. welfare system has been substantially reformed since 1993.
• Many States obtained AFDC waivers in 1993-1996.
• AFDC was replaced by TANF in 1996-1997. Under TANF:
Block grants replaced federal matching funds;
5-year time limit on welfare from federal funds;
Maintenance of Effort Requirement.
• From 1993 to 2002,
Welfare participation rate among single mothers dropped
from 32% to 9%;
Work participation increased from 68% to 79%.
Goal of the Paper
• To ascertain what features of welfare reform,
Child support enforcement, etc.
have been most responsible for the decline in welfare
participation and increase in work among single
States have great leeway in the design of their TANF
programs, resulting in substantial program heterogeneity,
making it very difficult to develop of set of variables that
comprehensively characterize the nature of different State
Contemporaneous economic and policy changes:
strong macroeconomy of 1996-2000;
Significant expansion of the Earned Income Tax Credit
(EITC) after 1993;
Dramatic increases in child care development fund (CCDF)
expenditures since 1996;
Additional Purpose of this Study
One could write down a structural dynamic model of single
mother labor supply and welfare participation, estimate it using
pre-reform data, and predict the effects of time limits, work
requirement, EITC etc.
In fact, Keane (1995), Keane and Wolpin (2003), Fang and
Silverman (2003) etc. have attempted to make predictions of
various policies using such structural approach.
We still need to rely on such structural model to predict the
effects of other policies that are yet to be implemented.
But which of the various structural models should we use in
We need to compare the predictions of currently proposed
models with evidence from the data.
Prior Literature …
• … often looked at one policy at a time, and took a
Difference-in-difference (DD) approach.
• Find a control group that is not affected by the
policy, but otherwise similar to the treatment group.
• E.g. Grogger (JHR 2000) uses single mothers whose
children are over 13 as the control group, since a 5-
year time limit would not affect them.
• Assumption that other aspects of the reform, like daycare
expansion, affect the two groups similarly is implausible;
• The two groups have very different baseline participation
Table 1: Welfare Participation Rates of Single Mothers, By Age of Youngest Child
Age of Before After Changes in Changes
Youngest Time Time Percentage in Percent
Child Limits Limits Points
0-6 41.3 23.8 -17.5 -42%
13-17 16.0 11.0 -5.0 -31%
DD Effect -12.5 -4.5
DD Effect -30% -11%
Note: Reproduced from Grogger (2000, JHR), Table 2. Data is the March CPS from 1979-1999.
We collect detailed data about all the time varying
factors that we think are relevant in explaining
welfare and work participation of single mothers.
We combine these economic and policy data with
individual data from March CPS 1981-2003.
To allow for policies to have different effects on
single mothers with different characteristics, we
include interactions between demographic
characteristics and economic and policy variables.
Prior Literature (cont’d)
Grogger and Michalopoulos (2003, JPE) used data from a
randomized experiment in Florida (Family Transition Program,
FTP is a fairly small experiment in which welfare recipients in
one county were randomized into a treatment group that was
subject to a 2-year time limit and a control group that was not.
[Also a childcare subsidy was provided to both groups alone.
Thus the DD approach still needs to assume no interaction b/w
childcare subsidy and time limit.]
They show that the 2-year time limit lowered welfare
participation rates among single mothers with youngest children
in 3-5 year old range by 7.4 percentage points (from a base rate
of 40.3 percent).
The main problem is that the way FTP was implemented was at
odds with the national norm (examples).
Prior Literature (cont’d) …
Prior work attempting to look at many policies simultaneously
includes Blank (1997), Council of Economic Advisors (1997,
Dependent variable - State level caseloads or welfare
Puzzling results. E.g., Blank estimates that TL is insignificant
(wrong sign); WR insignificant; biggest factor is the family cap
(which seems implausible). And the models don’t fit caseloads
limited duration of reforms;
collinearity between the strong macroeconomy and the timing of
failure to exploit how demographic groups are differentially
affected by policies.
Some Institutional Background Explaining the
Rise of Child Care Subsidy
Section 409 of PRWORA stipulates that DHHS can
reduce States’ Federal TANF block grant if the State
failed to maintain its level of assistance for needy
families at 75% of their historical peak levels (usually
the level at 1994).
This Maintenance of Effort (MOE) requirement was
designed to prevent a feared “race to the bottom”
where many States might start to cut assistance once
Federal AFDC matching funds vanished. Clinton
signed PRWORA only after MOE was added.
MOE and Multiple Equilibria
MOE has had some dramatic (and unexpected) consequences:
Welfare caseloads for almost all States dropped dramatically
after 1996, causing expenditure on TANF cash assistance to fall;
Thus to satisfy MOE, States are forced to redirect money into
other qualified programs – mainly to subsided daycare;
Subsidized daycare faciliates more women to work and possibly
leave welfare, freeing more money for daycare subsidy (under
Possibility of a switch in multiple equilibria.
Economic and Policy Variables
AFDC/TANF benefit level
Work Requirement and Exemptions;
20 States had an immediate work requirement; while most
others allows for 24-month work requirement time limit;
States may exempt women with young children from work
requirement (up to 2 years old);
States also allow for exemptions if they are disabled, if
theire household member is disabled, if they had difficulty
finding care for children under 6;
Sanctions if WR is not satisfied also vary by States;
Benefit Reduction Rate and Earnings Disregard;
Diversion programs (receiving a few months benefits up front
for agreeing not be in TANF for some period of time)
Child Support Enforcement and Treatment of Child Support
Child care subsidies;
EITC: both Federal and State EITC
The Federal EITC phase-in rate for families with one child
increased from 10% in 1980-1984 to 18.5% in 1993, 26.3%
in 1994 and 34% from 1995 onward
Since 1991, the EITC phase-in rate and max credit differ for
families with one vs. two or more children.
Medicaid and S-CHIP: Medicaid expansion since 1988
State Unemployment Rate
Federal income tax
Minimum wage (real)
20 percentile wage by state (real)
Identifying the Effects of Macroeconomy
We have an advantage over earlier studies because
we have data from the 2001-2002 recession
This allows us to better identify the separate effects
of welfare reform and macroeconomy.
We found that the macroeconomy, as measured by
the local unemployment rate, has relatively small
effects on welfare participation, but big effects on
This is consistent with the simple observation that:
Welfare participation continues to drop in 2001-
Work participation rate actually declined in the
Identifying the Effects of Time Limits
NY, MI and VT do not enforce time limits. Yet, these States
have had caseload drops comparable to other States.
Moreover, many States, such as CA, only reduce the benefit
level (child portion kept) once TL is exhausted.
Case studies suggest that many States provide liberal
extensions and exemptions.
Thus it seems implausible that TL were a key factor in reducing
More on Identification
We are careful about exploiting individual level variation
WR: women with young kids could be exempted;
EITC: EITC phase-in rates differ according to the number of
Child care subsidy: should have no effect on women whose
children are older than 12;
Child Support Enforcement: should have no effect on the
Medicaid: large variation across State and time in the oldest
age of children covered by the Medicaid expansion.
Includes demographics, and policy variables
and rich interactions between the two;
A total of 245 terms.
A lot of terms, but …
Some Defense for Our Specification
We are effectively trying to explain the welfare and work
participation rates for women with different demographics in
different States in different years;
How many such participation rates are there?
For the sake of argument, consider a simple case:
Suppose we observe women in
Race (3 categories);
educational attainment (4 categories);
marital status (4 categories);
4 age categories; (4 categories)
youngest child in 3 age intervals; (3 categories)
Oldest child in 3 age intervals; (3 categories)
Number of children is either 1, 2 or 3+; (3 categories)
State of residence (51 categories)
Urban/rural (2 categories);
Year (23 years)
Thus our specification is remarkably parsimonious.
What will be the specification of the prior
Much in the prior literature has relied on specifications that
State specific quadratic time trends.
This gives 51 + 23 + 100 = 174 parameters;
Then it would include a measure of a single policy, such as a
time-varying dummy variable for whether a State has yet
imposed time limits. This adds 51 parameters.
Note such model has no hope of explaining the evolution of
welfare and work participation rates by demographics.
Moreover, including State dummy makes OLS estimates more
likely to be biased.
Interpreting the Estimates
The main problem of having so many interaction terms is that
the regression coefficients are impossible to interpret by
We instead try to give an intuition for what the estimates mean
by presenting the predicted probability of welfare participation
for a set of single mothers with different demographic
characteristics when time limits and work requirements are
[TL]: No Time Limit: Suppose that no State
implemented time limit, while all other economic and
policy variables evolved as they actually did;
[WR]: No work requirement;
[EITC]: No EITC expansion (stayed the same as that
[UNEMP]: No unemployment rate change (stayed the
same as that in 1993);
[CCDF]: No CCDF expenditure;
[MEDICAID]: No further Medicaid expansion since
According to our model, the key economic and policy variables
that drove the overall 23 percentage-points decrease in the
welfare participation rate from 1993-2002 were:
Work requirements (57%);
Time limits (11%);
The key economic and policy variables that drove the overall
11.3 percentage-points increase in the work participation rate
from 1993-2002 were:
Work requirement (17%);
Time Limit (10%).
Work requirements are very effective at getting
people off welfare, but not so effective at inducing
EITC is rather effective at both.
Our model implies some important differences across
demographic groups in the impact of various policies.
The macro economy and EITC largely explain the
increase in work among relatively well educated
Work requirements were a much more important
factor for high school drop outs.
Confession: Problems with our Approach
Assumed education, marriage and fertility decisions to be
unaffected by the welfare reform itself;
Though evidence of welfare policy on marriage and fertility
decisions are rather mixed, these are definitely endogenous in
More damagingly, in a dynamic framework, women’s
expectations about the future welfare policies, not just the
current welfare system, should affect their decisions to work
Such expectations are not available; and any available
measurement would be noisy.
Keane and Wolpin (2003) assumed perfect foresight.