530
[Journal of Political Economy, 2003, vol. 111, no. 3]
᭧ 2003 by The University of Chicago. All rights reserved. 0022-3808/2003/11103-0003$10.00
Welfare Dynamics under Time Limits
Jeffrey Grogger
University of California, Los Angeles and National Bureau of Economic Research
Charles Michalopoulos
Manpower Demonstration Research Corporation
Among the most important changes brought about by the Personal
Responsibility and Work Opportunity Reconciliation Act of 1996 are
time limits, which provide consumers with an incentive to conserve
their welfare benefits for future use. Among forward-looking,
expected-utility-maximizing consumers who face liquidity constraints
and earnings uncertainty, economic theory predicts that the incentive
to conserve should be strongest among families with the youngest
children. We test this prediction using data from Florida’s Family
Transition Program, a randomized welfare reform experiment. Our
estimates generally exhibit the predicted age dependence, which suggests
that time limits affect welfare use before they become binding.
Our estimates indicate that, in the absence of other reforms that
increased welfare use, FTP’s time limit would have reduced welfare
receipt by 16 percent.
We thank two anonymous referees, the editor, Dan Bloom, Janet Currie, Peter Gottschalk,
V. Joseph Hotz, David Levine, Phillip Levine, Robert Moffitt, David Ribar, Seth
Sanders, David Zimmerman, and seminar participants at Manpower Demonstration Research
Corporation, University of California at Davis, the Institute for Research on Poverty,
the Joint Center for Poverty Research Conference on Tax and Transfer Programs for LowIncome
Participants, and the NBER Labor Studies Program meeting for helpful comments.
The views expressed herein are those of the authors, and do not represent the positions
of MDRC or the State of Florida. Any errors are the responsibility of the authors.
welfare dynamics 531
I. Introduction
The U.S. welfare system has changed dramatically as a result of recent
welfare reform. Among the most important of these reforms is the imposition
of time limits. Under the old Aid to Families with Dependent
Children (AFDC) program, welfare benefits were an entitlement: all
poor, single-parent families with at least one child under 18 years of age
were eligible to receive aid.1
The duration of welfare receipt was limited
only by the age of the youngest child in the family. Under the new
Temporary Aid to Needy Families (TANF) program, which replaced
AFDC after the passage of the Personal Responsibility and Work Opportunity
Reconciliation Act of 1996 (PRWORA), benefits remain payable
largely to poor, single-parent families with children under 18. However,
the duration of welfare receipt is sharply limited: PRWORA allows
families to receive federally funded benefits for no more than five years.
Many states set even stricter limits.2
One of the implications of time limits is that consumers’ current
welfare choices affect their future opportunity sets. This represents a
substantial departure from the old entitlement regime. It gives consumers
an incentive to conserve their benefits for future use.
Economic theory predicts that the strength of this incentive should
vary among families. For forward-looking, expected-utility-maximizing
consumers who face credit constraints and unstable job prospects, this
incentive should be stronger, the younger the youngest child in the
family. The reason is that welfare acts as insurance in the consumer’s
lifetime utility-maximization problem, and the value of retaining one’s
eligibility for such insurance rises, the longer the horizon over which
it could be used to smooth consumption. Since a family can no longer
receive welfare once its youngest child turns 18, families with the youngest
children have the longest eligibility horizons and thus the greatest
incentive to preserve their future eligibility by reducing their current
welfare use.3
1
Some poor, married-couple families were also eligible under the Unemployed Parent
component of AFDC, but this program never accounted for more than a small fraction
of the AFDC caseload (Moffitt 1992).
2
States can exempt up to 20 percent of their caseload from the five-year time limit. In
addition, states can allow more individuals to continue receiving cash assistance for more
than five years as long as they do not use federal funds for payments past the five-year
limit. Several states have adopted plans that use this option. Several other states, including
California, eliminate only the adults’ portion of welfare grants when the five-year limit is
imposed, allowing the families to continue receiving the children’s portion of the grants.
3
This prediction is derived from a formal model presented in Grogger and Michalopoulos
(1999). There we show that families should be less likely to use welfare, the greater
the difference between their remaining eligibility horizon and their available stock of
benefits. Since the available stock of benefits is equal to the time limit less welfare use
since the time limit was imposed, it is endogenous. In contrast, the remaining eligibility
horizon is largely determined by the age of the youngest child when time limits are
532 journal of political economy
We test this prediction by using data from a randomized experiment,
the Florida Family Transition Program (FTP). However, because the
FTP treatment group was subject to other policy reforms in addition to
the time limit, randomization by itself does not identify the effects of
the time limit. To distinguish the effects of the time limit from the
effects of the other reforms, we make use of a further implication of
age dependence: a family whose youngest child exceeds a threshold age
(which would be 13 in the case of a five-year limit) should be unaffected
by the time limit. The reason is that such families will become ageineligible
for welfare before they could reach the time limit. Put differently,
for a family whose youngest child is old enough, the time limit
is a nonbinding constraint. We use such families to construct an estimate
of the effect of FTP’s other reforms. If the effects of those other reforms
are age-invariant, then we can identify the effects of time limits using
difference-in-difference methods. We present evidence from several
other welfare reform experiments that supports the age-invariance
assumption.
The estimated effects of time limits exhibit age dependence that is
largely consistent with the prediction. Moreover, age dependence appears
well before the families could have exhausted their benefits. The
estimates indicate that, in the absence of FTP’s other reforms, which
tended to increase welfare use, the time limit would have caused welfare
receipt to fall by as much as 16 percent.
Other researchers who have considered the effects of time limits include
Council of Economic Advisers (1997, 1999), Swann (1998), Moffitt
(1999), and Ziliak et al. (2000).4
Swann formulates a stochastic dynamic
programming model and estimates its parameters with data on AFDC
use from the Panel Study of Income Dynamics. The other authors have
estimated the effects of time limits using aggregate state-level caseload
data. None of them consider the age dependence issue, which proves
to be essential to estimating the effects of time limits.5
Section II of the paper provides background on the FTP program
and our data. Section III elaborates on the identifying assumptions under
which we can distinguish the effects of FTP’s time limit from its
other policy reforms. It also provides evidence on the key age-invariance
assumption. Section IV presents our main estimates. Section V provides
some checks on the robustness of those estimates and some additional
imposed, which is exogenous. The age-dependence results that we focus on here can be
thought of as “reduced-form” predictions of the model.
4
Bloom et al. (1997, 1998, 2000) have studied the FTP program using a subset of the
data we analyze here, but they did not attempt to isolate the effects of the time limit from
the effects of the other policy reforms.
5
See Grogger (2000) for further discussion and a reconciliation of the results presented
here with those of the studies listed above.
welfare dynamics 533
evidence on the age-invariance assumptions. Section VI presents
conclusions.
II. Data from Florida’s Family Transition Program
A. Background
Florida’s experimental FTP program was implemented in Escambia
County (Pensacola) starting in May 1994 under a waiver from AFDC
program rules.6
All welfare recipients in the county were randomized
into one of two groups: the experimental FTP group, which was subject
to time limits, or the control group, which was enrolled in the AFDC
program. New entrants were randomized at the time they applied for
benefits. Ongoing recipients, that is, persons already receiving welfare,
were randomized at the time of their semiannual recertification interviews.
Recruitment into the experiment continued until October 1996.
Our sample includes all single-parent families ever drawn into the experiment.
We follow them for 24 months, beginning with the month
after random assignment. As we explain below, the time limit was such
that none of the sample members could have exhausted their benefits
within our 24-month follow-up period.
The data analyzed here come primarily from two sources: Florida
administrative records and a short survey instrument known as the Background
Information Form. Florida administrative records provide data
on monthly benefit receipt. The survey was administered to welfare
applicants (and those recertifying) at the time they applied (or were
recertified) for benefits, prior to random assignment. It contains baseline
demographic characteristics, which we summarize in table 1.
Comparing mean characteristics of the treatment and control groups
indicates that the initial randomization was properly executed, since the
characteristics of the two groups are so similar. Like welfare recipients
elsewhere, those in our sample are disproportionately black and have
relatively low levels of education, high levels of welfare use in the period
prior to random assignment, and low levels of recent labor market
experience.
B. FTP Policy Reforms and Conditions of AFDC Receipt
Although time limits were arguably the central component of the FTP
program, members of the treatment group were also subject to a number
of other policy reforms. This complicates our efforts to isolate the effects
of time limits. We summarize the full set of FTP reforms, as well as the
6
Bloom et al. (1997, 1998, 2000) provide detailed information about the program and
its impacts on numerous aspects of behavior.
534 journal of political economy
TABLE 1
Summary Characteristics of the Treatment and Control Groups at the Time
of Random Assignment
Characteristic FTP Group AFDC Group Total
Mother black 49.3
(50.0)
50.1
(50.0)
49.7
(50.0)
Mother’s age 29.4
(7.4)
29.5
(7.4)
29.5
(7.4)
Years of schooling 11.1
(1.6)
11.2
(1.5)
11.1
(1.5)
Age of youngest child 5.1
(4.2)
5.1
(4.3)
5.1
(4.3)
Months of welfare utilization
in 24 months prior to random
assignment
10.0
(9.7)
10.0
(9.7)
10.0
(9.7)
Quarters of employment in
year preceding random
assignment
1.3
(1.6)
1.3
(1.6)
1.3
(1.6)
Number of persons 2,238 2,235 4,473
Note.—Table entries are sample means; figures in parentheses are standard deviations.
corresponding conditions imposed on the AFDC control group, in table
2.
The FTP families were subject to three different categories of reforms:
the time limit, financial work incentives, and enhanced welfare-to-work
services. Most FTP families faced a 24-month time limit, after which
their benefits could be, and nearly always were, terminated. Particularly
disadvantaged families, who made up 48 percent of the sample, received
a 36-month time limit; AFDC recipients faced no time limit.
The FTP families also enjoyed relatively generous financial incentives:
the first $200 of monthly earnings was disregarded from income in
determining their monthly benefits, and earnings in excess of $200 were
subject to a benefit reduction rate (i.e., a tax rate) of 50 percent. The
AFDC group faced conventional AFDC financial incentives, which were
substantially less generous, as table 2 shows. By themselves, FTP’s financial
incentives should have increased the family’s welfare use, since
they provided an incentive to combine welfare and work by raising the
income eligibility threshold and by increasing payments to persons earning
less than the threshold.
The FTP and AFDC groups both were required to spend 30 hours
per week in mandatory work-related activities (MWRAs). These requirements
could be satisfied by either working or participating in a welfareto-work
program. The FTP welfare-to-work program had a lower ratio
of recipients to case managers than the AFDC welfare-to-work program
and provided enhanced employment and training services. To the extent
that FTP’s enhanced services were valuable and were tied to welfare,
welfare dynamics 535
TABLE 2
Summary of FTP Treatments and Corresponding Conditions of AFDC Receipt
Type of Treatment FTP Group AFDC Group
Time limits 24-month default time
limit; 36-month time
limit for particularly
disadvantaged
recipients
No time limits
Financial work incentives $200 earned income disregard
and 50 percent
benefit reduction rate
$120 disregard and 67
percent benefit reduction
rate for first four
months of work; 100
percent benefit reduction
rate after four
months; $90 disregard
after 12 months
Mandatory work-related
activities (MWRAs)
30 hours per week of either
work or workrelated
activities
30 hours per week of either
work or workrelated
activities.
Mandate could be satisfied
by participating in
a welfare-to-work program
that provided enhanced
employment
and training services
Mandate could be satisfied
by participating in
a welfare-to-work program
that provided
conventional AFDC
services
Exemptions only for
mothers with infants
under 6 months of age
Exemptions for mothers
with children under 3
years old
they also could have increased welfare use, all else equal.7
The control
group received conventional AFDC services.
For families with children under 3 years old, there was a further
difference in the conditions faced by the treatment and control groups.
Whereas AFDC mothers were exempted from the MWRAs if they had
children younger than 3, FTP mothers were exempt only if they had
children younger than 6 months old. Because of this, we are unable to
isolate the effects of time limits on families in the 6-month to 2-year
age group.8
The conditions under which we can isolate the effects of
7
The low staffing ratio in the FTP welfare-to-work program raises the possibility that
FTP recipients may have been “coached” on how to deal with the time limit. However,
Bloom et al. (1998, chap. 3) report that caseworkers were much more likely to counsel
recipients to use their time on aid to get training than to counsel them to bank their
months of eligibility for future use.
8
Both the FTP and AFDC groups were eligible for child care subsidies. Although these
subsidies were intended to be equally available to both groups, early evidence gave us
concern that in fact they had been more readily available to the treatment group than
to the control group (Grogger and Michalopoulos 1999). However, more recent evidence
suggests that any such differences were minor (Bloom et al. 2000). As a result, we do not
include child care subsidies among the differences between the treatment and control
groups.
536 journal of political economy
TABLE 3
Monthly Probability of Receiving Aid during First 24 Months after Random
Assignment, by Age of Youngest Child
Age Group
FTP Group AFDC Group
Age-
Specific
Impact
of FTP
(Col.
2ϪCol.
4)
(5)
Effect
of
Time
Limit
(6)
Person-
Months
(1)
Mean
(2)
Person-
Months
(3)
Mean
(4)
6 months to 2 years
(group 0)
14,299 .473
(.014)
14,052 .473
(.014)
.000
(.020)
3–5 (group 1) 18,506 .404
(.012)
18,913 .403
(.012)
.001
(.017)
Ϫ.085
(.054)
6–10 (group 2) 12,158 .344
(.016)
11,964 .375
(.016)
Ϫ.031
(.021)
Ϫ.117
(.056)
11–14/15 (group 3) 6,457 .368
(.021)
5,855 .313
(.021)
.055
(.030)
Ϫ.031
(.057)
15/16–17 (group 4) 1,703 .399
(.036)
2,242 .313
(.033)
.086
(.052)
Total 53,123 .404
(.008)
53,026 .401
(.008)
.003
(.012)
Note.—Standard errors (in parentheses) are adjusted for the presence of multiple observations per person.
the time limits from the effects of the other reforms are discussed further
in the next section.
III. Identification and Evidence on the Age-Invariance Assumption
A. Identification and Age-Specific Estimates of the Effect of FTP
Table 3 presents age-specific estimates of the effect of FTP and helps
illustrate the conditions under which we can identify the effects of the
time limit. Distinguishing the effects of the time limit from the effects
of the other reforms requires two assumptions: (i) the effects of the
individual reforms are additive, and (ii) the effects of the financial work
incentives and enhanced services are age-invariant. It also requires time
limits to have no effect on families whose youngest children exceed the
threshold age, as discussed above.
In table 3, the sample has been divided into five age groups. Group
0 consists of families whose youngest children are less than 3 years old.
As mentioned above, because the exemptions from the MWRAs differed
between the treatment and control groups for families with children
under 3, we are unable to identify the effects of time limits on this age
group.
Group 1 includes families whose youngest children are between 3 and
5. Group 2 consists of families whose youngest children are 6–10 years
welfare dynamics 537
old. Group 3 includes families whose youngest children range from 11
to the threshold age, which is 16 for families assigned a two-year limit
and 15 for families assigned a three-year limit. Group 4 consists of families
whose youngest children exceed the threshold age.9
The unit of observation in table 3, as in most of the other tables
below, is the person-month. The entries in columns 2 and 4 give the
mean rate of monthly welfare receipt during the first two years after
random assignment. We focus on the first 24 months after entry into
the program to ensure that none of the families could have actually
reached their time limit and become ineligible for further aid during
the sample period.10
Column 2 gives treatment sample means, and column
4 gives control sample means for each age group j, where j p 0,
1, … , 4.
The last entries in columns 2 and 4 show that, in a typical month,
the treatment and control groups were equally likely to use welfare. The
final entry in column 5 presents the estimated “impact” of the program,
defined as the difference in mean welfare utilization rates between the
treatment and control samples. It indicates that FTP left welfare use
virtually unchanged in the full sample, increasing the monthly probability
of welfare use by only 0.003 (with a standard error of 0.012). This
replicates and extends the finding of Bloom et al. (1998). On the basis
of a subset of the data that we analyze here, they concluded that FTP
had essentially no effect on welfare use during the first two years after
random assignment.
The other entries in column 5 present age group–specific estimates
of the impact of FTP. Since these impact estimates reflect the combined
effects of the time limit, the financial incentives, and the enhanced
services, their signs generally cannot be predicted. However, the estimates
suggest that the null effect of the program in the full sample
masks some variation in its effects by age group. The program decreased
welfare use slightly among age group 2 but increased it among age
groups 3 and 4.
The next to last entry in column 5 shows that FTP increased the
monthly welfare utilization rate among members of age group 4 by 0.086
(0.052). Since members of this age group should not have been affected
by the time limit, we take this as an estimate of the combined effects
9
The age thresholds for the two oldest groups depend on the length of the family’s
time limit. Our classification scheme is based on an imputed time limit, which can be
constructed for all families. The actual time limit, which was determined by county welfare
officials, is available only for the FTP group. When we replace the imputation with the
actual time limit for the FTP group, we obtain similar results.
10
Most families contribute 24 monthly observations to the sample; families may progress
from one age group to the next as their youngest child ages. Families whose youngest
children turn 18 during the sample period contribute observations only until the youngest
child turns 18.
538 journal of political economy
of the financial incentives and the enhanced services. The fact that it
is positive is consistent with our expectation that the financial incentives
should have increased welfare use, all else equal.
Imposing the assumption that the financial incentives and enhanced
services had age-invariant effects, we can estimate the age-specific effects
of time limits by subtracting the impact estimate for age group 4 from
the other age group–specific impact estimates. These difference-indifference
estimates of the effects of the time limit are presented in
column 6.
These preliminary estimates appear to provide mixed evidence as to
whether time limits have greater effects among families with younger
children. On the one hand, all the estimates are negative, as expected.
Moreover, the estimate for age group 2 is significant and is significantly
more negative than the estimate for age group 3 (the t-statistic for the
test of equality is Ϫ2.39). On the other hand, the estimate for age group
2 is more negative than the estimate for age group 1. Potentially, this
may represent a contradiction of the prediction from the model discussed
above. Alternatively, it may indicate that child care issues pose
constraints on the extent to which treatment group members with preschool
age children can respond to the time limit. Finally, some data
analysis revealed that FTP families in the 6–10 age group spent fewer
months on aid during the 24 months prior to random assignment than
AFDC families in the same age group. Since presample aid use is a
strong predictor of current aid use, the regression-adjusted estimates
presented in Section IV, which control for presample welfare use, provide
evidence that is more consistent with the expected effects of time
limits.11
Of course, both the preliminary difference-in-difference estimates in
table 3 and the main regression estimates below rely on the assumption
that FTP’s other reforms had age-invariant effects. If this assumption is
violated, then none of our estimates isolate the effects of the program’s
time limit. Because this assumption is crucial to our estimation strategy,
we use data from other welfare reform experiments to test it before
proceeding to our main estimation results.
11
Properly executed randomization ensures that the treatment and control groups will
have identical ex ante characteristics on average. The figures in table 1 indicate that
randomization was properly executed. However, because randomization is not carried out
within strata defined by the age of the youngest child, there is no guarantee that treatment
and control group characteristics will be identical within strata. Age group–specific treatment
control differences in months on aid during the 24 months prior to random assignment
for age groups 0–4, respectively, are 0.2, 0.3, Ϫ1.0, 0.6, and 0.7.
welfare dynamics 539
B. Evidence on the Age-Invariance Assumption
The evidence on age invariance comes from a set of welfare reform
experiments that, like FTP, involved financial incentives, MWRAs, or
both. However, unlike FTP, none of these other experiments involved
time limits. Our motivation in presenting these results is straightforward.
If these other experiments had largely age-invariant effects, then that
would reflect favorably on our assumption that the financial work incentives
and enhanced services in FTP likewise had age-invariant effects.
The analogy is not perfect because the financial incentives and MWRAs
(and external conditions, such as the unemployment rate) varied among
programs. However, if several experiments yield age-invariant results,
despite the differences between them, it may help to provide support
for the critical assumption underlying our analysis.
Table 4 presents age group–specific impact estimates from three different
welfare reform experiments on the monthly probability of welfare
receipt during the first 24 months after random assignment.12
The experiments
we analyze are the National Evaluation of Welfare-to-Work
Services (NEWWS), the Vermont Welfare Restructuring Project (WRP),
and the Minnesota Family Investment Program (MFIP).
The NEWWS program involved 11 subexperiments in seven different
locations around the country; we present pooled estimates based on
data from all 11 programs. In each subexperiment, the treatment group
was subject to MWRAs. Both WRP and MFIP were dual-treatment experiments;
that is, each program randomly assigned participants either
to the control group or to one of two treatment groups. In both programs,
one treatment group (referred to as the “incentives only” group)
was subject to financial incentives. The other treatment group was subject
both to the financial incentives and to MWRAs.
The generosity of the financial incentives varied between WRP, MFIP,
and FTP. As in FTP, the MFIP financial incentive was unambiguously
more generous than that of AFDC. In contrast, whether WRP was more
generous than AFDC depended on the consumer’s work history and
current hours of work.13
The estimates in column 6 of table 4 show that, overall, both NEWWS
and MFIP had significant effects on welfare use. The MWRAs in NEWWS
reduced welfare use by five percentage points, on average, whereas the
MFIP financial incentives increased welfare use by eight to nine per-
12
Like the FTP sample, the samples for the other programs are limited to single parents.
13
All three experiments imposed additional minor reforms on the treatment groups in
addition to those described above, such as higher asset limits and, in the case of MFIP, a
food stamp cash-out. Grogger, Karoly, and Klerman (2002, chap. 3) provide a more complete
comparison of these and several other welfare reform experiments. Detailed descriptions
and results from NEWWS, WRP, and MFIP are available in Freedman et al.
(2000), Bloom, Hendra, and Michalopoulos (2000), and Miller et al. (2000), respectively.
540 journal of political economy
TABLE 4
Age-Specific Impacts from Other Welfare Reform Demonstrations on Monthly
Welfare Receipt during Years 1 and 2 after Random Assignment
Age of Youngest Child
0–2
(1)
3–5
(2)
6–11
(3)
12–14
(4)
14–18
(5)
0–18
(6)
A. NEWWS Program: Mandatory Work-Related Activities
(x2
p10.28)*
Number of persons 9,914 17,656 12,283 5,996 2,607 48,456
Control group mean .71 .73 .71 .58 .47 .69
Impact Ϫ.05
(.01)
Ϫ.05
(.01)
Ϫ.06
(.01)
Ϫ.03
(.01)
Ϫ.04
(.01)
Ϫ.05
(.01)
B. WRP: Financial Work Incentives and Mandatory WorkRelated
Activities (x2
p3.56)
Number of persons 1,422 999 1,019 508 210 4,158
Control group mean .64 .66 .64 .62 .47 .64
Impact Ϫ.03
(.02)
.02
(.02)
.03
(.03)
Ϫ.02
(.04)
.01
(.06)
.00
(.01)
C. WRP—Incentives Only: Financial Work Incentives
(x2
p1.46)
Number of persons 714 517 498 266 101 2,096
Control group mean .67 .64 .62 .65 .45 .64
Impact .01
(.03)
.04
(.03)
.03
(.03)
Ϫ.01
(.04)
Ϫ.03
(.07)
.02
(.02)
D. MFIP: Financial Work Incentives and Mandatory WorkRelated
Activities (x2
p2.30)
Number of persons 2,659 1,529 1,325 668 333 6,514
Control group mean .63 .63 .60 .58 .48 .61
Impact .10
(.01)
.07
(.02)
.07
(.02)
.06
(.03)
.07
(.04)
.08
(.01)
E. MFIP—Incentives Only: Financial Work Incentives
(x2
p3.87)
Number of persons 2,107 1,197 1,039 482 262 5,087
Control group mean .64 .66 .62 .59 .49 .63
Impact .11
(.02)
.08
(.02)
.09
(.02)
.08
(.03)
.02
(.05)
.09
(.01)
Note.—Standard errors are in parentheses.
* statistic for homogeneity (five degrees of freedom).2
x
centage points, irrespective of the program’s MWRAs.14
The WRP’s less
generous financial incentives had no significant effect on welfare use,
either alone or in combination with the program’s delayed work
requirements.15
14
These effects are similar to the effects of FTP’s financial work incentives and enhanced
services, as measured by the impact of FTP on age group 4.
15
In contrast to the other programs, WRP’s work requirement became effective only
after the recipient had received welfare for 24 months.
welfare dynamics 541
Regardless of the overall effect of the program, however, there is little
age variation in the impacts of the programs. In neither of the MFIP
or WRP experiments do the treatment effects vary significantly by the
age of the youngest child. The NEWWS estimates, which are based on
data from over 48,000 individuals, differ by age at the 10 percent level,
but not at the 5 percent level. Moreover, despite the marginal statistical
significance of the differences, the substantive differences among the
age-specific impact estimates are slight. Taken as a whole, these results
suggest that age invariance is more the norm than the exception in
reform programs involving various combinations of financial incentives
and MWRAs. This provides some justification for the age-invariance
assumption that is central to our estimation strategy.
IV. Regression Estimates
A. Two Regression Models
In this section we return to the FTP experiment, using regression methods
to estimate the effect of time limits on welfare receipt. The regression
estimates have three advantages over the simple difference-indifference
estimates. First, they control for personal characteristics of
the experiment participants, such as different levels of presample welfare
use. Second, they should be more precise, because the covariates
in the model should reduce the variance of the error term. Third, they
allow us to consider alternative functional forms for the key age-FTP
interaction.
As in table 3, we analyze monthly welfare use, so the unit of observation
is the person-month. We first allow the effect of time limits to
depend on age according to a step function specification, which yields
estimates that are directly comparable to the difference-in-difference
estimates. This model is given by
TL
y p a ϩ a A ϩ a A ϩ a A ϩ a A ϩ t A E ϩ t A Eit 0 0it 1 1it 2 2it 3 3it 0 0it i 1 1it i
TL TL
ϩ t A E ϩ t A E ϩ tE ϩ X b ϩ m ϩ e (1)2 2it i 3 3it i i it i it
for and where n is the number of personsi p 1, … , n t p 1, … , 24,
in the sample. The dependent variable yit equals one if family i utilized
welfare in period t, and equals zero otherwise. The age group dummies
are defined as if the youngest child in the ith family at the timeA p 1jit
of random assignment falls into age group j at time t and oth-A p 0jit
erwise, for The variable Ei is the FTP, or experimental,j p 0, 1, … , 4.
dummy; if family i is in the FTP group and if family i isE p 1 E p 0i i
in the AFDC group. The vector Xit includes a number of exogenous
regressors including the mother’s age at time t; a dummy equal to one
542 journal of political economy
if she is black and zero otherwise; a dummy equal to one if she had a
three-year time limit and equal to zero if she had a two-year limit; the
number of children in the family; the mother’s years of schooling; the
number of months in which the family had received welfare during the
24 months prior to random assignment; the number of quarters of
employment during the year preceding random assignment; and a vector
of year dummies. Time is measured in months from the date of
random assignment, so is the month of entry into the program.t p 0
The a, b, and t terms are parameters to be estimated from the data.
Under the assumption that the financial work incentives and the enhanced
social services have age-invariant effects, their joint effect is given
by t, the coefficient on the FTP dummy, since we exclude the interaction
between the FTP dummy and the dummy for age group 4. The effects
of time limits for the younger age groups, and are given byTL TL TL
t , t , t ,1 2 3
the coefficients on the interactions between the corresponding age
group dummies and the FTP dummy.
In our second specification, we allow the effects of the time limit to
depend continuously on age, at least for ages from 3 to the threshold
age. This “linear interaction” model is given by
TL
y p a ϩ a A ϩ a A ϩ t A E ϩ t A E ϩ tEit 0 0it 1 it 0 0it i it i i
ϩ X b ϩ m ϩ e , (2)it i it
where Ait is the age at time t of the youngest child in family i; A pit
if and otherwise; and denotes the thresh-¯ ¯ ¯A Ϫ A 3 ≤ A ! A A p 0 Ait i it i it i
old age for family i.16
Under this specification, the coefficient on the
FTP dummy again gives the combined effect of the financial work incentives
and the enhanced social supports.17
The effect of the time limit
depends on the age of the youngest child via the interaction between
and the FTP dummy. Under this specification, we would expect toAit
find that sinceTL
t 1 0 A ≤ 0.it
In both specifications, the error term consists of two components, mi
and eit. The former is a family effect, which gives rise to dependence
among the monthly observations from a particular family. Since this
groupwise dependence may cause conventional ordinary least squares
standard errors to be biased downward, we employ a Huber-White covariance
matrix estimator that corrects for such dependence. It also
16
Because eq. (2) includes rather than the age group dummies, it implicitly imposesAit
constraints on the intercepts for age groups 1, 2, and 3. When we relaxed these constraints
by adding the age group dummies to the model, their coefficients were small and insignificant.
Since including them had no effect on the other coefficients in the regression,
we excluded them from the specifications reported below.
17
This is achieved by subtracting the threshold age from the age of the youngest child
in constructing A .it
welfare dynamics 543
accounts for heteroskedasticity, which is an inherent property of the
linear regression estimator when applied to a binary dependent variable.
B. Main Estimates
Results of these specifications are reported in table 5. (Summary statistics
for all variables included in the regression model appear in App. table
A1.) The coefficients on the FTP dummies, in the sixth row of the table,
are positive. Under our identifying assumptions, they suggest that the
combined effects of the financial work incentives and enhanced services
act to raise welfare use by about six percentage points. This is slightly
smaller than the corresponding difference-in-difference estimate from
table 3.
The coefficients on the interaction between the FTP dummy and the
dummy for age group 0 appear in the first row of the table. This age
group includes families with children under age 3, for which the effects
of time limits cannot be distinguished from the effects of MWRAs. The
estimates are negative and at least marginally significant. They show that
the combined effects of time limits and MWRAs reduce aid use by about
seven percentage points. This is smaller than the corresponding estimate
for age group 1, whereas we would expect time limits by themselves to
have larger effects on the younger age group. This suggests that members
of age group 0 satisfied their MWRAs by means that were more
likely to keep them on welfare. Indeed, analysis of data from the TwoYear
Client Survey (TYCS), which was administered to about 600 program
participants during the first half of 1997, showed that members
of age group 0 were more likely to take part in vocational training than
members of age group 1.
In the step function specification, shown in column 1, the coefficients
on the interactions between the FTP dummy and the age group dummies
provide estimates of and the age group–specific effectsTL TL TL
t , t , t ,1 2 3
of time limits. Given the discussion in Section I, our expectation is that
The regression estimates are more consistent with thisTL TL TL
t ! t ! t ! 0.1 2 3
expectation than the difference-in-difference estimates in table 3. Adding
the regressors to the model has reduced the estimated group 2
effect, as expected, with the result that Moreover,TL TL TL
ˆ ˆ ˆt ! t ! t ! 0.1 2 3
although and are only marginally significant, they are both sig-TL TL
ˆ ˆt t1 2
nificantly different from (t-statistics are Ϫ2.42 and Ϫ2.06, respec-TL
ˆt3
tively). However, even though their difference is small. OneTL TL
ˆ ˆt ! t ,1 2
potential explanation for this is that, despite the child care subsidies
available to both the treatment and control groups, nonprice barriers
to obtaining child care prevented families with preschool age children
from fully responding to the time limit.
The linear interaction specification, reported in column 2, provides
TABLE 5
Linear Regression Estimates of the Effects of Time Limits on the Monthly
Probability of Receiving Aid (Sample Size 106,149)
Variable
Step Function
Specification
(1)
Linear Interaction
Specification
(2)
FTP dummy#youngest child between
6 months and 2 years
Ϫ.069
(.042)
Ϫ.066
(.028)
FTP dummy#youngest child between
3 and 5 ( )TL
ˆt1
Ϫ.074
(.041)
FTP dummy#youngest child between
6 and 10 ( )TL
ˆt2
Ϫ.067
(.042)
FTP dummy#youngest child between
11 and 14/15 ( )TL
ˆt3
Ϫ.012
(.042)
FTP dummy# ( )TL
ˆA tit .007
(.002)
FTP dummy .063
(.039)
.060
(.022)
Youngest child between 6 months
and 2 years
.094
(.032)
.029
(.015)
Youngest child between 3 and 5 .054
(.030)
Youngest child between 6 and 10 .047
(.030)
Youngest child between 11 and 14/
15
.034
(.029)
Age of youngest child Ϫ.004
(.002)
36-month time limit .055
(.010)
.052
(.010)
Mother black .072
(.010)
.073
(.010)
Mother’s age Ϫ.002
(.001)
Ϫ.001
(.001)
Number of children .004
(.004)
.004
(.004)
Years of schooling Ϫ.013
(.003)
Ϫ.013
(.003)
Months of welfare use in 24 months
prior to random assignment
.019
(.001)
.019
(.001)
Quarters of employment in year
preceding random assignment
Ϫ.011
(.003)
Ϫ.011
(.003)
1995 dummy Ϫ.144
(.009)
Ϫ.144
(.009)
1996 dummy Ϫ.288
(.011)
Ϫ.287
(.011)
1997 dummy Ϫ.375
(.012)
Ϫ.374
(.012)
1998 dummy Ϫ.454
(.014)
Ϫ.454
(.014)
Constant .543
(.051)
.614
(.041)
Adjusted 2
R .280 .280
Note.—The term is defined as for and otherwise, where is the age of the youngest¯ ¯A A p A Ϫ A 3 ≤ A ! A A p 0 Ait it it i it i it it
child in family i at time t and is the threshold age that pertains to family i. Huber-White standard errors (in parentheses)¯Ai
account both for groupwise dependence arising from multiple observations per person and for heteroskedasticity.
welfare dynamics 545
the same fit to the data, as judged by the adjusted but the key2
R ,
parameter estimate provides clearer support for the prediction that time
limits should decrease welfare use the most among families with the
youngest children. The coefficient on the interaction between the FTP
dummy and is positive, as expected, and significant. Each one-yearAit
increase in the age of the youngest child increases the reduction in the
likelihood of welfare use attributable to time limits by 0.7 percentage
point, on average. This implies that, for a family with a 36-month time
limit (and consequently a threshold age of 15) whose youngest child is
5 years old, the time limit reduces welfare use by seven percentage
points. For a family with a 36-month time limit whose youngest child is
13 years old, the time limit reduces welfare use by 1.4 percentage points.
These numbers are in line with the estimated effects of the time limit
from the step function specification.
In both specifications, the other coefficients in the model are generally
in accord with expectations based on previous research. The presence
of young children increases the likelihood of welfare use, as do
low levels of maternal education. Black mothers are more likely to use
welfare than white mothers, all else equal. The mother’s past welfare
use, past employment, and assignment to a 36-month time limit are also
associated with higher welfare use. The effect of welfare use prior to
random assignment is positive and particularly significant. In conjunction
with the fact that presample welfare use was lower among families
in the 6–10 age group than among those in the 3–5 age group, this
explains why adding the covariates to the model results in estimates
more in accord with our expectations than the simple difference-indifference
estimates.
These regression-adjusted estimates suggest that time limits substantially
reduce welfare use among families with young children. If we
weight the estimated age-specific reductions in welfare use by the age
distribution of the sample, we calculate that time limits by themselves
would have led to a 16 percent reduction in welfare receipt among the
FTP group.18
These results are particularly striking because they represent
anticipatory responses to the time limit. None of the FTP families
could have reached the time limit within our 24-month sample period.
The estimates suggest that time limits substantially reduced welfare use,
at least among families with younger children, well before any of the
families in our sample could have exhausted their benefits.
18
This calculation is limited to families with youngest children age 3 and over, for which
we can estimate the effects of time limits.
546 journal of political economy
V. Additional Estimates
In this section, we provide additional evidence on the effects of time
limits. First, we stratify the sample according to recipients’ level of exposure
to welfare prior to random assignment. Next, we consider estimates
from shorter follow-up periods. We also report on other results
designed to check the overall robustness of our findings. Finally, we
discuss some results that shed further light on the age-invariance
assumption.
A. Estimates by Prior Welfare Exposure
We stratify the results by prior welfare exposure for two reasons. First,
one might be concerned that the age-FTP interaction was somehow
picking up the effects of prior welfare exposure in a complicated way.
Second, stratifying the sample allows us to see whether welfare recipients’
response to time limits varies according to their level of welfare
dependency. This is a particularly interesting question in light of concerns
that recipients facing the greatest job market barriers might be
the least likely to respond to time limits, leaving behind a more disadvantaged
welfare population that would ultimately exhaust their benefits
(Sawhill and Zedlewski 1995; Duncan, Harris, and Boisjoly 2000;
Moffitt and Pavetti 2000). Unfortunately, as interesting as these questions
are, our ability to answer them is seriously limited by sample size
issues.
In table 6, we stratify the sample according to presample welfare use
in two different ways. First, we divide the sample into those with 12
months or less of prior welfare use and those with more than 12 months.
Second, to focus on the most welfare-dependent group, we divide the
sample into those with less than 24 months of prior aid receipt and
those with exactly 24 months.
For the first stratification scheme, estimates from the step function
specification, reported in columns 1 and 3, are inconclusive. All the
estimates are insignificant, the result of splitting the sample. For both
groups of families—those with 12 or fewer months of presample welfare
use and those with more—the regression coefficients are partially consistent
with expectations, but partially inconsistent as well.
Estimates from the linear interaction specifications, in columns 2 and
4, are more precise. For families with both greater and lesser levels of
presample welfare exposure, they are positive and at least marginally
significant. They suggest that the effects of time limits are similar for
both groups.
The results from the second stratification scheme are presented in
columns 5–8. For families with fewer than 24 months of presample
welfare dynamics 547
TABLE 6
Linear Regression Estimates of the Effects of Time Limits on the Probability
of Receiving Aid, by Extent of Welfare Receipt during the 24 Months prior to
Random Assignment
12 Months or Less vs. More
than 12 Months
Less than 24 Months vs. Exactly
24 Months
12 Months or
Less
More than 12
Months
Less than 24
Months 24 Months
Variable (1) (2) (3) (4) (5) (6) (7) (8)
FTP dummy#youngest
child between 3
and 5 ( )TL
ˆt1
Ϫ.045
(.049)
Ϫ.129
(.076)
Ϫ.069
(.044)
Ϫ.105
(.116)
FTP dummy#youngest
child between 6
and 10 ( )TL
ˆt2
Ϫ.013
(.048)
Ϫ.158
(.078)
Ϫ.043
(.045)
Ϫ.194
(.118)
FTP dummy#youngest
child between 11
and 14/15 ( )TL
ˆt3
.028
(.048)
Ϫ.089
(.080)
Ϫ.006
(.045)
Ϫ.031
(.122)
FTP dummy#Ait
( )TL
ˆt
.006
(.002)
.007
(.004)
.006
(.002)
.007
(.006)
FTP dummy .022
(.044)
.046
(.023)
.136
(.073)
.063
(.034)
.051
(.041)
.050
(.019)
.130
(.111)
.070
(.052)
Observations 63,447 63,447 42,702 42,702 88,251 88,251 17,898 17,898
2
R .173 .173 .138 .137 .225 .225 .133 .131
Note.—The term is defined as for and otherwise, where Ait is the age of the youngest¯ ¯A A p A Ϫ A 3 ≤ A ! A A p 0it it it i it i it
child in family i at time t and is the threshold age that pertains to family i. Huber-White standard errors (in parentheses)¯Ai
account both for groupwise dependence arising from multiple observations per person and for heteroskedasticity. In
addition to the variables shown, all regressions include all other variables shown in table 5.
welfare exposure, the step function estimates, in column 5, are similar
to those from the full sample, as reported in table 5. The linear interaction
coefficient, in column 6, is likewise positive and significant. Both
sets of estimates are consistent with the notion that time limits should
have greater effects on families with younger children.
Results for the most welfare-dependent families are presented in columns
7 and 8. In the step function specification, all the age-FTP interaction
terms are negative, but the estimate for age group 2 is substantially
more negative than the estimate for age group 1. Moreover, the linear
age-FTP interaction coefficient is completely insignificant for this group,
in contrast to the other groups. This suggests that the most welfaredependent
families may be unresponsive to the time limit, although the
imprecision of the subsample estimates makes it difficult to draw any
conclusions too strongly.
B. Estimates from Shorter Follow-up Periods
Above we interpreted our main results as indicating that time limits lead
welfare families to leave welfare even before the time limit could begin
to bind. Here we present some further evidence to ask more directly
whether our results truly represent anticipatory responses to the time
548 journal of political economy
TABLE 7
Linear Regression Estimates of the Effects of Time Limits on the Monthly
Probability of Receiving Aid during Various Sample Periods Following Random
Assignment
Variable
Sample Period (Following Random Assignment)
6
Months
(1)
12
Months
(2)
18
Months
(3)
6
Months
(4)
12
Months
(5)
18
Months
(6)
FTP dummy#youngest child
between 3 and 5 ( )TL
ˆt1
Ϫ.068
(.057)
Ϫ.084
(.049)
Ϫ.071
(.046)
FTP dummy#youngest child
between 6 and 10 ( )TL
ˆt2
Ϫ.063
(.058)
Ϫ.078
(.051)
Ϫ.068
(.047)
FTP dummy#youngest child
between 11 and 14/15 ( )TL
ˆt3
Ϫ.040
(.062)
Ϫ.022
(.052)
Ϫ.008
(.048)
FTP dummy# ( )TL
ˆA tit .003
(.003)
.007
(.003)
.007
(.003)
FTP dummy .091
(.053)
.090
(.047)
.067
(.043)
.059
(.030)
.081
(.026)
.065
(.024)
2
R .302 .292 .285 .302 .292 .285
Sample size 26,764 53,384 79,849 26,764 53,384 79,849
Note.—The term is defined as for and otherwise, where is the age of the youngest¯ ¯A A p A Ϫ A 3 ≤ A ! A A p 0 Ait it it i it i it it
child in family i at time t and is the threshold age that pertains to family i. Huber-White standard errors (in parentheses)¯Ai
account both for groupwise dependence arising from multiple observations per person and for heteroskedasticity. In
addition to the variables shown, all regressions include all other variables shown in table 5.
limit. We do this by presenting estimates from shorter follow-up periods.
The results are reported in table 7.
The general qualitative patterns that appear in our main results are
present even in the models based on sample periods that extend for
only 12 months after random assignment. By that time, the estimates
are nearly the same as those from the full 24-month follow-up period.
The estimates from the linear interaction specification are significant.
This suggests that families subject to time limits did not wait until the
last minute to leave the welfare rolls. Rather, they began leaving substantially
earlier, consistent with our earlier interpretation of the fullsample
results.
C. Further Estimation Results
In addition to the results reported so far, we estimated a number of
other specifications to test the robustness of our findings. We included
an urban/rural indicator in the regressions both in levels and interacted
with the FTP dummy to accommodate the possibility that job markets
may differ between Pensacola and outlying parts of Escambia County.
We added a post-TANF dummy and FTP-TANF interactions to account
for the possibility that the introduction of Florida’s TANF program in
October 1996 might have changed the behavior of program partici-
welfare dynamics 549
pants.19
In another specification, we replaced the year dummies that we
use to control for period effects with month dummies. We also allowed
for linear age effects within the oldest age group. None of these changes
to the model substantially affected the main results.
We also estimated the effect of FTP on the likelihood of leaving Escambia
County. Such mobility is important because it could change the
family’s experimental status. If an FTP group member left Escambia
County before late 1996, when Florida implemented its TANF program
statewide, then she would no longer have been subject to a time limit.
After late 1996, however, her time limit “clock” would have followed her
to any other jurisdiction in the state. Conversely, after late 1996, a control
group member who left Escambia County would have become subject
to the TANF program time limit (Bloom et al. 2000, p. 13).
The key question for us, of course, is whether FTP resulted in differential
out-of-county mobility between treatment and control groups
that varied significantly by age. To answer this question we make use of
data from the Four-Year Client Survey (FYCS), which was administered
in late 1999 to about 1,700 program participants. Compared to the
TYCS, the larger sample size of the FYCS makes it more likely that we
would reject the null of no age variation if age variation were indeed
present. The disadvantage of the FYCS is that it captures behavior occurring
over four years after random assignment, whereas all the results
above focus on the two-year period preceding the time in which families
could have exhausted their benefits.20
Table 8 reports results from a regression in which the dependent
variable is equal to one if the family ever relocated outside of Escambia
County. In addition to the variables shown, the regression includes all
the variables included in the models from table 5. Neither specification
yields any evidence either of differential mobility between the treatment
and control groups or of significant age differentials in mobility.
D. Further Evidence on the Age-Invariance Assumption
Finally, we computed some estimates based on the FTP data designed
to provide additional evidence on the validity of the crucial age-invariance
assumption.21
To test for age invariance in the effects of FTP’s
financial incentives, we estimated models like those in table 5 in which
the dependent variable was replaced with a dummy that was equal to
one if the consumer both received welfare and worked. The motivation
19
The FTP control group members remained subject to the old AFDC rules, even after
Florida’s TANF plan went into effect.
20
The TYCS included no questions on out-of-county mobility.
21
The full set of estimates discussed in this subsection is available from the authors on
request.
550 journal of political economy
TABLE 8
Linear Regression Estimates of the Effects of Time Limits on the Probability
of Living Outside Escambia County at Some Point within Four Years after
Random Assignment (Sample Size 1,726)
Variable
Step Function
Specification
(1)
Linear Interaction
Specification
(2)
FTP dummy#youngest child between
3 and 5 ( )TL
ˆt1
.007
(.013)
FTP dummy#youngest child between
6 and 10 ( )TL
ˆt2
.010
(.013)
FTP dummy#youngest child between
11 and 14/15 ( )TL
ˆt3
.006
(.014)
FTP dummy# ( )TL
ˆA tit Ϫ.000
(.001)
FTP dummy Ϫ.010
(.012)
Ϫ.004
(.005)
2
R .005 .003
Note.—Unit of observation for these regressions is the person. The term is defined as for¯A A p A Ϫ A 3 ≤ A !it it it i it
and otherwise, where is the age of the youngest child in family i at time t and is the threshold age that¯ ¯A A p 0 A Ai it it i
pertains to family i. Huber-White standard errors (in parentheses) account both for groupwise dependence arising
from multiple observations per person and for heteroskedasticity. In addition to the variables shown, all regressions
include all other variables shown in table 5.
for this regression is simple: in order for the financial incentive to affect
the consumer’s income, she must combine work and welfare. If she
receives welfare but does not work, or works but does not receive welfare,
she cannot benefit from FTP’s financial incentive. In these regressions,
the coefficients on the FTP dummies were positive and significant, as
expected. However, the age-FTP interaction terms were completely insignificant,
which is consistent with age invariance in the effects of the
financial incentive.
Our tests for age invariance in the effects of the enhanced services
were less direct. They were based on the idea that, if these services
included valuable training opportunities, then human capital theory
predicts that they should be more valuable to younger mothers than to
older mothers. Thus we added an interaction between the FTP dummy
and maternal age to the model. The coefficients on the maternal age
interactions were insignificant, and including them in the model had
no effect on the estimated effects of time limits.
VI. Conclusions
Time limits are among the most fundamental of the recent reforms to
the U.S. welfare system. Because they limit families’ eligibility for assistance,
they provide families with an incentive to conserve their benefits.
Economic theory predicts that, among forward-looking, expected-utilitymaximizing
consumers who face liquidity constraints and earnings un-
welfare dynamics 551
certainty, these incentives should vary by the age of the youngest child
in the family. Whereas families with older youngest children have short
eligibility horizons and relatively little chance of prematurely exhausting
their benefits, families with younger youngest children have longer eligibility
horizons and greater risk of exhausting their benefits. Thus
families with younger children should be more reluctant to utilize their
benefits than families with older children.
This prediction is largely borne out by our empirical results. Indeed,
our estimates suggest that the time limit in the FTP experiment had
sizable effects, which, in the absence of other features of the program
that worked to increase welfare use, would have reduced welfare receipt
by as much as 16 percent. This accords in a general way with the finding
that the states’ TANF programs, most of which involve time limits, have
had substantial effects on the welfare caseload in recent years (Council
of Economic Advisers 1999; Wallace and Blank 1999; Schoeni and Blank
2000).
The age-varying reductions in welfare use that arise from time limits
may have substantive implications for the well-being of children in poor
families. They imply that any reductions in welfare payments that result
from the imposition of time limits are more likely to occur earlier, rather
than later, during childhood. Recent studies suggest that the adverse
effects of family poverty on children’s educational attainment are
greater, the younger the child is when the family experiences poverty
(Duncan et al. 1998; Guo 1998). If age-related reductions in welfare
payments result in corresponding reductions in income, then time limits
could have greater adverse consequences for children than policies with
more age-neutral effects. The key question is whether the reductions in
welfare use that arise from time limits translate into reductions in income.
The one existing study on the topic suggests that they may not,
but more research on the question is needed (Grogger 2003).
Because the time limit was not the only reform included in FTP, our
results are based on assumptions regarding the effects of those other
reforms. In particular, we assume that the effects of the program’s financial
work incentives and enhanced welfare-to-work services are ageinvariant.
Tests of those assumptions do not reject them, but such tests
could never be as compelling as evidence from an experiment explicitly
designed to isolate the effects of time limits. In the absence of any such
experiment, it is encouraging that estimates based on two nationwide
surveys, and a very different set of identifying assumptions, are similar
to those presented here (Grogger 2002, 2004).
552 journal of political economy
Appendix
TABLE A1
Means of Variables from FTP Sample
Variable Mean
Welfare utilization indicator (dependent
variable) .403
36-month time limit .480
Mother black .498
Mother’s age 30.4
(7.3)
Number of children 2.0
(1.1)
Years of schooling 11.1
(1.5)
Months of welfare utilization in 24
months preceding random assignment
10.0
(9.7)
Quarters of employment in year preceding
random assignment
1.3
(1.6)
1995 dummy .301
1996 dummy .367
1997 dummy .199
1998 dummy .071
Youngest child between 6 months and 2
years .267
Youngest child between 3 and 5 .353
Youngest child between 6 and 10 .227
Youngest child between 11 and 14/15 .116
Age of youngest child 6.1
(4.1)
Ait Ϫ5.9
(4.8)
FTP dummy .500
FTP dummy#youngest child between 6
months and 2 years .135
FTP dummy#youngest child between 3
and 5 .174
FTP dummy#youngest child between 6
and 10 .115
FTP dummy#youngest child between 11
and 14/15 .061
FTP dummy#Ait Ϫ3.0
(4.5)
Number of monthly observations 106,149
Note.—Standard deviations are in parentheses.
References
Bloom, Dan, et al. The Family Transition Program: Implementation and Early Impacts
of Florida’s Initial Time-Limited Welfare Program. New York: Manpower Demonstration
Res. Corp., 1997.
———. The Family Transition Program: Implementation and Interim Impacts of Florida’s
welfare dynamics 553
Initial Time-Limited Welfare Program. New York: Manpower Demonstration Res.
Corp., 1998.
———. The Family Transition Program: Final Report on Florida’s Initial Time-Limited
Welfare Program. New York: Manpower Demonstration Res. Corp., 2000.
Bloom, Dan; Hendra, Richard; and Michalopoulos, Charles. Vermont’s Welfare
Restructuring Project: Key Findings from the Forty-Two Month Client Survey. New
York: Manpower Demonstration Res. Corp., 2000.
Council of Economic Advisers. Technical Report: Explaining the Decline in Welfare
Receipt, 1993–1996. Washington: Council Econ. Advisers, May 1997.
———. The Effects of Welfare Policy and the Economic Expansion on Welfare Caseloads:
An Update. Washington: Council Econ. Advisers, August 1999.
Duncan, Greg J.; Harris, Kathleen Mullan; and Boisjoly, Johanne. “Time Limits
and Welfare Reform: New Estimates of the Number and Characteristics of
Affected Families.” Soc. Service Rev. 74 (March 2000): 55–75.
Duncan, Greg J.; Yeung, W. Jean; Brooks-Gunn, Jeanne; and Smith, Judith R.
“How Much Does Childhood Poverty Affect the Life Chance of Children?”
American Sociological Rev. 63 (June 1998): 406–23.
Freedman, Stephen, et al. National Evaluation of Welfare-to-Work Strategies: Evaluating
Alternative Welfare-to-Work Approaches: Two-Year Impacts for Eleven Programs.
New York: Manpower Demonstration Res. Corp., 2000.
Grogger, Jeffrey. “The Behavioral Effects of Welfare Time Limits.” A.E.R. Papers
and Proc. 92 (May 2002): 385–89.
———. “The Effects of Time Limits, the EITC, and Other Policy Changes on
Welfare Use, Work, and Income among Female-Headed Families.” Rev. Econ.
and Statis. (2003), in press.
———. “Time Limits and Welfare Use.” J. Human Resources (2004), in press.
Grogger, Jeffrey; Karoly, Lynn A.; and Klerman, Jacob A. Consequences of Welfare
Reform: A Research Synthesis. RAND DRU-2676-DHHS. Santa Monica, Calif.:
Rand Corp., May 2002. http://www.acf.hhs.gov/programs/opre/welfare_
reform/reform_cover.html.
Grogger, Jeffrey, and Michalopoulos, Charles. “Welfare Dynamics under Time
Limits.” Working Paper no. 7353. Cambridge, Mass.: NBER, September 1999.
Guo, Guang. “The Timing of the Influences of Cumulative Poverty on Children’s
Cognitive Ability and Achievement.” Soc. Forces 77 (September 1998): 257–87.
Miller, Cynthia, et al. Reforming Welfare and Rewarding Work: Final Report on the
Minnesota Family Investment Program. Vol. 1. Effects on Adults. New York: Manpower
Demonstration Res. Corp., 2000.
Moffitt, Robert A. “Incentive Effects of the U.S. Welfare System: A Review.” J.
Econ. Literature 30 (March 1992): 1–61.
———. “The Effect of Pre-PRWORA Waivers on AFDC Caseloads and Female
Earnings, Income, and Labor Force Behavior.” In Economic Conditions and
Welfare Reform, edited by Sheldon H. Danziger. Kalamazoo, Mich.: Upjohn Inst.
Employment Res., 1999.
Moffitt, Robert A., and Pavetti, LaDonna A. “Time Limits.” In Finding Jobs: Work
and Welfare Reform, edited by David E. Card and Rebecca M. Blank. New York:
Russell Sage Found., 2000.
Sawhill, Isabel V., and Zedlewski, Sheila. “Assessing the Personal Responsibility
Act.” In Welfare Reform: An Analysis of the Issues, edited by Isabel V. Sawhill.
Washington: Urban Inst., 1995.
Schoeni, Robert F., and Blank, Rebecca M. “What Has Welfare Reform Accomplished?
Impacts on Welfare Participation, Employment, Income, Poverty, and
554 journal of political economy
Family Structure.” Working Paper no. 7627. Cambridge, Mass.: NBER, March
2000.
Swann, Christopher A. “Welfare Reform When Agents Are Forward-Looking.”
Manuscript. Charlottesville: Univ. Virginia, December 1998.
Wallace, Geoffrey L., and Blank, Rebecca M. “What Goes Up Must Come Down?
Explaining Recent Changes in Public Assistance Caseloads.” In Economic Conditions
and Welfare Reform, edited by Sheldon H. Danziger. Kalamazoo, Mich.:
Upjohn Inst. Employment Res., 1999.
Ziliak, James P.; Figlio, David N.; Davis, Elizabeth E.; and Connolly, Laura S.
“Accounting for the Decline in AFDC Caseloads: Welfare Reform or the Economy?”
J. Human Resources 35 (Summer 2000): 570–86.