This effort has produced some important findings, but has also fallen short of its objectives in some respects. Future work that would build on the findings of this project may benefit from the following recommendations.
1. Refinement of the "expected participation" variables should be considered. The single most important explanatory variable in the participation model is the change in the logarithm of our "expected participation" measure. This measure is a weighted sum of the number of women in the state in age groups that are at-risk for participation, with weights equal to national participation rates by age-group in 1990 (see Chapter 3). This variable is intended to capture the size and age distributions of the population of women who are at-risk for participation. Michael Wiseman has suggested that the widening over the sample period between actual caseloads and predicted caseloads is in part due to the fact that this is an index with weights fixed by observed participation rates in one year. An alternative would be to compute the change for each state in each year using weights based on national participation by age group in the previous year--i.e., to update the weights when computing each year's change.(13) Obtaining different weights for each year from SIPP may be problematic, but it should be possible to obtain weights for many other years. An even more desirable solution would be to use each state's own age-specific participation rates in the previous year to predict the change in participation due to population change, but the information needed to construct state-specific indices is not available.
2. Development of models with alternative dependent variables should be considered. Possibilities include:
- Openings and closings--we initially considered using case openings and closings as dependent variables, but elected not to in keeping with the mainstream of the existing literature and because of concerns about the openings and closings data. The extra effort needed to estimate such models given the existence of the database for the explanatory variables is small, however, and might well pay off.
- Disaggregated dependent variables--the ACF participation data are reported by program (Basic vs. UP), and cannot be directly disaggregated further. It might be possible, however, to use the ACF quality control (QC) data to estimate disaggregated caseloads. We think that disaggregation by citizenship, reason for deprivation, child-only vs. other, ethnicity, age, or sex of household head would be especially helpful if feasible. Disaggregation by state regions might also be very useful. Reviewers have provided three examples of how caseload behavior in some regions within states have varied substantially from those for the rest of the state.(14)
- Aggregated dependent variables--we developed separate participation models for the Basic and UP programs, but it may be that the separation of the two programs created some problems for the analysis. We have learned, for instance, that a large share of UP cases in California eventually convert to Basic cases. Such conversions might help explain caseload growth in the Basic program during the 1984 -1992 period following the large increases in the UP program during the earlier recession. While this could be modeled directly, it may be problematic to build models in which lagged participation measures for one program appear as explanatory variables in participation models for the other. Combining the caseloads and estimating a single model may be a more useful alternative.
- Joint AFDC/SSI models--we attempted to estimate the effect of changes in SSI child recipients on AFDC caseloads, but were not successful. The likely reason is that participation in both programs is, in part, determined by a common set of factors. It might be fruitful, instead, to develop models of "welfare recipients," including those on either AFDC or SSI. We think it would be feasible to develop a model for AFDC or SSI child recipients, and another for adults age 15 - 64. The adult model could be substantially improved by separating female and male recipients; this is possible for the SSI recipients, but may not be possible for the AFDC recipients--depending on the QC data. Garrett and Glied (1997) have pursued this approach.
- Alternative functional forms for the participation variables--we used the logarithm of the participation variable or the log of participation rates. The logarithmic specification implies that a change in an explanatory variable is proportional to the level of participation. Use of the level of the participation rate (caseload per adult in the relevant age group, say) is a defensible alternative.(15) We have noticed that several states with relatively small UP caseload rates at the beginning of the sample period experienced extraordinarily large percent age increases in their UP caseload rates in the first three years of the sample period relative to the percent increases observed in other states; the percentage point changes in the caseload rates themselves, however, are much more comparable to those in other states.(16)
3. Development of additional explanatory variables should be considered. Possibilities include:
- Wages for low-skill workers--one promising explanation of growth that is not accounted for over the sample period is the relative decline in wages for low-skill workers. Substantial research has been done by labor economists on understanding the causes of this decline, and some of that work has been done using state or regional data. Several economists familiar with this research have suggested that, with some effort, reasonably accurate state-level estimates of wages for low-skill workers could be constructed, following work that has already been done in this area.(17) We did not pursue this under the current project because of the level of effort required;
- Unemployment insurance benefits--Dr. Jonathan Gruber (MIT) has developed a computer model for simulating UI benefits to workers with specified characteristics in each state for every year since 1968. This model could be used to construct a measure of each state's benefits that would be analogous in origin to the AFDC program parameters we used in our research. This would be a better measure of the generosity of the state programs than the UI measure we used in the estimates reported here;
- Immigration variables--we have not included a variable for refugees in our models. Even though the number of refugees is small, high participation rates among refugees may mean that they are an important explanation of caseload growth in some states. More generally, the strength of our findings for IRCA legalizations suggests that more effort in modeling the effects of immigration on caseloads may be fruitful;
- Female headed households--our efforts to produce state-level estimates of this explanatory variable were not very successful. We found that even subregional estimates based on the March CPS exhibited substantial random variation from one year to the next. We have not had time to examine the reasons for this variation--whether it is just sample size or whether it might be some other aspect of the CPS methodology. It may be that better estimates can be produced, but significant effort will be required to determine how this might be done. Ongoing efforts by ASPE and others to develop state-level estimates of many health and welfare indicators may prove useful here;
- Medicaid measures--we were surprised not to find more significant findings for the Medicaid variables we tried. It is puzzling that previous analyses of micro data have produced much stronger results. A substantial effort to understand the reason behind these inconsistent findings might produce better measures for use in the state-level analysis;
- Neighbor variables--Shroder (1995) used measures of labor market and AFDC program variables in neighboring states in his models. We did not have the time to construct such variables for this project, but they can be constructed from the data we have already assembled. Given concerns about states "racing to the bottom" under devolution, it may be especially valuable to examine how reductions in benefits in one state affect caseloads in neighboring states;
- Minimum wage--we did not include the minimum wage in our models. Most states use the federal minimum, but some states have had higher values in at least some years. A recent analysis by Burkhauser, Couch and Wittenburg (1996) of the effect of minimum wage increases on employment of "at-risk" youth in 1990 and 1991 suggests that this would be fruitful.
4. Simultaneous models of benefit levels and program participation should be considered. Shroder (1995) found much stronger results for the impact of increases in benefits on participation when he modeled them as simultaneously determined with participation levels. As he and others have argued, an increase in the burden of a state's welfare expenditures on taxpayers is likely to make cuts in benefits politically popular; i.e., higher participation results in legislated reductions in benefit levels. This negative, reverse relationship between benefits and participation may result in an underestimate of the presumably positive effect of benefits on participation. We are, however, concerned about the quality of possible instrumental variables needed to identify the effects of benefits on participation, especially because there are three benefit variables that would require instrumentation.(18) Three instruments would be required at a minimum, and all three would need to be important determinants of the benefit parameters that do not have a direct impact on program participation. Additional resources would be needed to pursue this further.
5. Extension of the sample period should be considered. Data for the participation variables in our model are available from the fourth quarter of 1974 on. Data through the end of the federal program (third quarter of 1996) should soon be available. The sample period constraints we faced primarily came from the explanatory variables. Earlier and, especially, later data for many key variables can, however, be obtained or constructed, but may require substantial effort.
6. Further examination of the stability of the parameter estimates across subperiods should be considered. We were limited in our ability to produce subperiod estimates by both resources and software. More effort in this area would be helpful in determining whether changes in the program or the characteristics of participants have affected the sensitivity of participation to the business cycle, program parameters, or other factors. For instance, the introduction of the UP program could have reduced the sensitivity of participation in the Basic program to business cycles.(19) Other major programmatic changes that occurred during the period under investigation could have had similar effects--especially the provisions of OBRA-81 and the introduction of JOBS.
7. Further examination of asymmetries in business cycle effects and, more generally, model dynamics should be considered. Many have suggested that the effects of business cycles on the caseload are asymmetric, with rapid growth during recessions and only gradual declines during recoveries. We explored this idea by estimating models with different distributed lags for unemployment in growth and recession periods, but found only very small, insignificant differences. Such models are known as "switching regression" models, with the applicable specification for a period depending on some decision rule. Our models were crude in this respect, with the applicable regression depending on whether or not the national unemployment rate was increasing in the current period. Steven Thompson recently developed a more sophisticated version of a switching regression model for his monthly time-series model of Maryland's caseload, with some success in identifying asymmetric effects (Regional Economic Studies Institute, 1994). Maryland's caseload (see Chapter 6) displays a pattern of rapid increases after recessions and gradual declines during the following recoveries. Thompson's initial experience was similar to our own; using the change in the current national or regional unemployment rate yielded results that did not seem to capture the asymmetric relationship very well. After some experimentation, however, he discovered that lagging the "switch" from the recession model to the recovery model by 8 to 12 months after the peak in the caseload yielded a much better fit for Maryland. We do not know whether a similar lag would fit other state experiences well.
Another reviewer, Michael Wiseman, has suggested developing a dynamic model in which the speed of adjustment of the caseload depends on the difference between the actual caseload and a long-run equilibrium level. Such a model would put a different structure on the lagged variables and include the lagged caseload variable as an explanatory variable. This seems to be echoed, but with a twist, in the comments of a third reviewer, Don Winstead, who suggests that lack of program capacity to deal with the large numbers filing for benefits during the Florida recession resulted in larger caseloads than would have been realized had capacity been greater.
Don Winstead also suggested that changes in the unemployment rate lag, rather than lead, caseload changes in a recession, but lead them in a recovery. The theory is that low-skill workers are the first fired in a downturn and last hired in a recovery.
8. More extensive analyses of actual and predicted series in individual states might be useful. The independent reviews of our findings for four selected states were useful in both understanding growth in those states and understanding the strengths and weaknesses of the model. We did not have an opportunity to revise the model to incorporate information we obtained from these reviews, but this would be feasible. For instance, Steven Thompson reported that he found a substantial interaction in his time-series model between an 1115 waiver for the "Up-Front Job Search" program and labor market variables. The estimated coefficients of the labor market variables increased substantially after the program's implementation. This essential elements of this specification could be implemented in the pooled model. Detailed assessments for additional states might also be very useful.