اثرات بازار کار یک برنامه ضد فقر: نتایج حاصل از مدل سلسله مراتبی خطی

Using Hierarchical Linear Modeling (HLM), this paper estimates the impact of random assignment into an anti-poverty program, New Hope, on earnings. The use of HLM allows for an examination of program impacts on levels, as well as trends, in quarterly data spanning more than two years of earnings. Results show that New Hope is associated with higher final levels of earnings. Additionally, results indicate that experimental group members experienced rapid earnings growth immediately after the quarter of random assignment, but that their earnings growth did not increase at a significant rate over the subsequent two-year period. Like many welfare recipients today, New Hope participants were required to work in order to receive the program’s benefits. The results from this paper indicate that such a policy can lead to a rapid increase in participants’ earnings, but that earnings may not continue to increase over time and may not increase enough to lift families out of poverty.

Welfare-to-work programs often have dual goals—encouraging work among welfare recipients, and promoting economic self-sufficiency. Researchers have struggled to identify components of anti-poverty, job training, and other programs that can achieve these goals for various subgroups of the poor population. Answers to these questions are best found in randomized experiments, in which eligible participants are randomly assigned to experimental and control groups and in which treatment effects can be assessed free of selection bias. One such program is New Hope, a random assignment anti-poverty program taking place in Milwaukee in the mid-1990s. This paper uses hierarchical liner modeling (Raudenbush and Bryk, 2002) to supplement the New Hope evaluation of two-year impacts of the program (Bos et al., 1999). Traditionally, program evaluations assess program impacts by comparing outcome means (often regression-adjusted) for experimental and control groups. Hierarchical linear modeling allows one to compare not only means (levels) between groups, but also rates of change in outcomes of interest, in this case earnings, measured quarterly over the two-year period following random assignment, as well as three quarters prior to random assignment. Additionally, the within-person aspect of HLM allows for an examination of within person patterns of earnings trends at different time periods.

Table 4 presents the results from a series of unconditional models: that is, no covariates are included in the Level 2 model. Such a model produces average final levels of and changes in quarterly earnings, not conditional on the control variables. Results from this model show the average level of, and rates of change in, earnings across the two-year period. Table 4. Results from unconditional model Quarterly earnings Average intercept (π0i) 2265.75***(67.47) Reliability .94 Average pre-random assignment slope (π1i) 35.92* (20.05) Reliability .18 Average random assignment slope increment (π2i) 246.27*** (25.86) Reliability .67 Average post-random assignment slope increment (π3i) 97.43*** (9.98) Reliability .81 Coefficients and standard errors come from model including auto-correlated errors. DF=1356. Data source. Manpower Demonstration Research Corporation. *** Indicates p<.01. Table options The intercept results show that earnings averaged $2266 in the final quarter (π0i) of this two-year period. The pre-random assignment slopes (π1i) indicate earnings had been increasing by about $36 per quarter prior to random assignment. Immediately after random assignment (π2i) earnings increased at a significant rate—a $246 earnings increase during this quarter. Post-random assignment slopes, (π3i) indicate that earnings increased by about $97 per quarter during the two-year post random assignment period. Finally, this table presents the reliabilities for the Level 1 parameters. The majority of these reliabilities are greater than .50, suggesting that there is a substantial “signal-to-noise” ratio in measuring individual differences in these parameters. Higher reliabilities mean that relationships between person-level variables and these parameters may be detected, because most of the variability is due to true variance, rather than error (Raudenbush and Bryk, 2002). The PRETIME (π1i) measure is the only one with a potential reliability problem, meaning that it may be difficult to reliably predict this parameter.