By Dallin Overstreet

Introduction: ​

Welfare continues to be a subject of much deliberation in politics. Understanding the effect that the level of welfare benefits has on the unemployment rate would be necessary to know how to design policy in the best manner possible. There have been multiple studies conducted on this topic, but many researchers have come to differing conclusions about the effects that welfare has on unemployment. A study of unemployment conducted by Maria Malkerson and Jan Saarela suggests that: “unemployment has a clear impact on both the probability of welfare participation and on welfare dependence among recipients.” Knowing that unemployment has a noticeable effect on welfare participation, we would like to see if somewhat
of the opposite holds true. In our study we have tried to model this question to the best of our ability. We have collected data on the cash benefit of 3 types of welfare benefit programs: SNAP (Food Stamps), TANF (Temporary Assistance to Needy Families), and Unemployment Insurance. These programs have been designed to help those in poverty, those that can’t provide all goods necessary for life, and to help those that have fallen on hard times and need a little extra help. That being said, unintended consequences could be happening due to the level of benefits given out. The SNAP Program is designed to provide the necessary amount of food for individuals and families to survive. It aims to provide necessary nutrition to those who otherwise could not obtain it themselves. TANF program is aimed at helping families achieve self-sufficiency by providing block grants to the states, and then the states further distribute the money. The stated goals of TANF is to provide for needy families so that children can be cared for in their own homes, reduce the dependency of needy parents by encouraging them to work and preparing them to do so, and to increase the formation of two-parent families rather than out-of-wedlock births. Unemployment insurance is designed to help those that lose their jobs
temporarily. People who receive unemployment insurance receive a certain percentage of what they were being paid before, contingent upon them searching actively for new work. Understanding the effects of these welfare programs on the unemployment rate is the purpose of our research.

Research Question:​

How does the increase or decrease of welfare benefits received affect the observed unemployment rate?

Data with summary stats table:

We are using data from The University of Kentucky Center for Poverty Research and the Bureau of Labor Statistics. Both of these sources are trustworthy, thorough research centers that draw from multiple official sources like data collection from bureaus running these assistance programs. This is collected data from the year 1980 til 2015 on statistics of government aid programs. AFDC/TANF are welfare aid programs, where we have data measured per year, per state. The states are coded with numbers 1-51 (Washington DC is included because the original data sets measured it along with the other 50 states), and years are let 1980-2015. We have information on the Employment Rate (er), Unemployment Rate (uer), Total TANF Recipients per state, Total SNAP Recipients per state, Total TANF Benefits (tanfben), Total SNAP Benefits (snapben), and Total Unemployment Insurance Benefits (uiben). There are 1,836 observations, given the 35 years of data for 51 places. The data follows these groups through the years, and thus we have panel data to run regressions on. Below is a summary of the data set.

Screen Shot 2018-01-31 at 12.12.47 PM

Data Scrubbing:

We used data from two credible sources so we combined them over years and states. The states needed to be coded by to make sure the data aligned correctly along states. To do so, one of the data sets already had integers assigned to the states to give them a numeral code. The other data set for the Unemployment benefits had only state abbreviations so we went through and renamed all of the states with the number that corresponded with the same state in the other data set. After coding, we sorted the data to match up with the correct year and state and deleted the excess data (i.e. years before 1980, data from Puerto Rico that was only in one data set, ect). The data set that included SNAP and TANF had information for households in 1, 2, 3, and 4 member increments so we combined them and took the average to get the average benefits per household for each program. These categories are the variables we ran regressions on so we could get an outlook for the effect of the programs as a whole.

Theory & Model Specification:

Theoretical Model: Part of change in Unemployment = f(Food Stamps, Unemployment Insurance, AFDC/TANF), where we expect higher levels of welfare benefits to be positively correlated with higher levels of unemployment.

Econometric Model:

Unemployment Rate = 𝞫0 + 𝞫1 (TANF Benefits)t + 𝞫2 (TANF Benefits)t-1                                  + 𝞫3 (TANF Benefits)t-2 + 𝞫4 (SNAP Benefits)+ 𝞫5 (SNAP Benefits)t-1
+ 𝞫6 (SNAP Benefits)t-2 + 𝞫7 (UI Benefits)+ 𝞫8 (UI Benefits)t-1 + 𝞫9 (UI Benefits)t-2 + u + ai

We used a distributed lag model with fixed effects to evaluate the data. Due to the fact that we had panel data on states over a 35 year span, a fixed effects model was the best model type to use to control for unobserved time constant factors that could bias the results. We also used lags in the model because we believe that people may not react to the changes in welfare benefit levels as soon as they are put in place. It would take some time for people to decide what to do before they actually started using more of the welfare programs.

The parameters of interest are 𝞫1, 𝞫2, 𝞫3, 𝞫4,𝞫5, 𝞫6, 𝞫7, 𝞫8, and 𝞫9. Each coefficient will tell us the effect that a change in the benefit levels of the welfare programs has on the unemployment rate the year it is put in place, a year after, and then two years after. If the coefficients are positive, it will mean that raising benefit levels increases the unemployment rate. If the coefficients are negative, it would mean that raising benefit levels actually decreases unemployment. Lastly, if the coefficients are equal to zero, equivalent to zero, or not statistically significant, it would mean that raising benefit levels has no effect on the unemployment rate.

Results:

We regressed using the fixed effects method of regression to evaluate our panel data directly.

Screen Shot 2018-01-31 at 12.21.30 PM

Hypothesis Tests:

H0 : 𝞫1 , 𝞫2 , 𝞫3 , 𝞫4 , 𝞫5 , 𝞫6 , 𝞫7 , 𝞫8 , and 𝞫9 = 0

HA : 𝞫1 , 𝞫2 , 𝞫3 , 𝞫4 , 𝞫5 , 𝞫6 , 𝞫7 , 𝞫8 , and 𝞫9 ≠ 0

Interpreting Coefficients:

  • 𝞫​1​ has a coefficient of -.0007, meaning a $1 increase in TANF benefits in the same year decreases the unemployment rate by .0007. It is statistically significant, so the null hypothesis is rejected.​
  • 𝞫2 has a coefficient of -.000099, meaning a $1 increase in TANF benefits decreases unemployment by .000099 one year later. It is not statistically significant, the null hypothesis cannot be rejected.
  • 𝞫3 has a coefficient of -.000045, meaning a $1 increase in TANF benefits decreases the unemployment rate by .000045 two years later. It is not statistically significant, the null hypothesis is not rejected.
  • 𝞫​4 ​has a coefficient of .00142, meaning a $1 increase in SNAP benefits increases unemployment by .00142 in the same year. It is statistically significant, the null hypothesis is rejected.​ ​
  • 𝞫​5​ has a coefficient of .00211, meaning a $1 increase in SNAP benefits increases the unemployment rate by .00211 a year later. It is statistically significant, the null hypothesis is rejected.​ ​
  • 𝞫​6​ has a coefficient of -.00332, meaning a $1 increase in SNAP benefits decreases unemployment by .00332 two years later. It is statistically significant, the null hypothesis is rejected. ​
  • 𝞫7 has a coefficient of -.000018, meaning a $1 increase in Unemployment Insurance benefits decreases the unemployment rate by .000018 in the same year. It is not statistically significant, the null hypothesis is not rejected.
  • 𝞫​8​ has a coefficient of -.0000389, meaning a $1 increase in Unemployment Insurance benefits decreases the unemployment rate by .0000389 a year later. It is statistically significant, the null hypothesis is rejected.
  • 𝞫​9​ has a coefficient of -.0000413, meaning a $1 increase in Unemployment Insurance benefits decreases the unemployment rate .0000413 two years later. It is statistically significant, the null hypothesis is rejected.

Result on parameters of interest:

The only coefficient on TANF benefits that was statistically significant in the same year the change is made. In this case, an increase in benefits caused a decrease in unemployment. The coefficients on all three betas for SNAP benefits were statistically significant. In the year the change is made and the year after it could increase unemployment, but the second year after it could decrease it. The coefficients for lag one and two of Unemployment Insurance benefits are statistically significant. They show that one and two years after the change is made, an increase in benefits can reduce unemployment.

Diagnostics:

1) Random sampling and missing data: Our analysis should have no problem with a bias from random sampling or missing data. We got our data from reliable sources that always use random sampling and we were not missing any data in our regression.

2) Omitted variable bias: There may exist some sort of bias in our analysis because there are so many variables that contribute to the unemployment rate, not just the variables we have included in the regression. Other welfare type programs like housing programs could create a positive bias on the coefficients, due to the fact that a person using any of these three programs would be more likely to use the housing program, which could have a positive correlation to the unemployment rate. Other welfare programs may have the same correlation to both the unemployment rate and these three welfare programs.

3) Time Trend and Stationarity: Due to the fact that we used panel data and a fixed effects regression, any possible time trend that may have existed on our independent variables has been accounted for and removed.

Screen Shot 2018-01-31 at 12.34.23 PM

According to the table above, by running the Levin-Lin-Chu unit-root test on our panel data we conclude that our data is stationary. The statistics are very significant, so there is no problem with non-stationarity in our data.

4) Outliers: To test for outliers, we standardized the residuals of the regression and calculated the leverage for the regression. After obtaining the standardized residuals and leverage, we re-ran the regression without any outliers.

Screen Shot 2018-01-31 at 12.35.42 PM

After running the regression without outliers, it seems that outliers are not having that great of an effect on our regression. TANF does however have statistically significant coefficients on the base year and the second lag, while only the second lag of Unemployment Insurance is now significant. The coefficients on SNAP Benefits are all still significant. These changes in coefficients may be due to outliers. However, the R squared is much bigger on the regression that includes the outliers, so we continue to use them.

5) Functional form specification: In order to check if our model was specified correctly, we ran both a Ramsey RESET test and a link test in stata to see if we had any sort of misspecification. The results of these tests are as follows:

Screen Shot 2018-01-31 at 12.37.09 PM

Unfortunately, according to both tests, it appears that our model suffers from a significant degree of misspecification. This is likely related to the omitted variable bias inherent in our specified model, as there exist a multitude of variables that directly or indirectly impact the unemployment rate. We would not be able to account for this bias without a considerable alteration of our model, so we will keep this in mind in continuing with our analysis.

6) Small samples and normality: The extensive amount of observations we had in our data, totaling 1836 unique observations, gave us a very large sample size, meaning we would not have to worry about any possible setbacks due to an insufficient sample size. This includes any normality relating to our error term.

7) Serial Correlation: Running a test for serial correlation in panel data, we found that there was serial correlation in our analysis. The table below shows it.

Screen Shot 2018-01-31 at 12.38.27 PM

To correct for serial correlation, we calculate the robust standard errors.

Screen Shot 2018-01-31 at 12.49.58 PM

After calculating the robust standard errors, we see that all the coefficients that were significant before are still significant, and the base year for Unemployment Insurance Benefits has now become statistically significant as well.

8) Heteroskedasticity: Due to the fact that we calculated the robust standard errors to correct for serial correlation, any heteroskedasticity that may have existed in our regression has been corrected for already. The Newey West robust standard errors automatically correct for it.

Conclusion:

From our research we can conclude that some of these welfare programs can increase unemployment, some may reduce it, and some may have no effect at all. The SNAP program was the most statistically significant, increasing unemployment up until two years after the increase in benefits is made. Unemployment Insurance and TANF seem to reduce unemployment. That being said, these increases and decreases are very small and may not be economically significant unless there are larger increases and decreases in the benefit amounts. An increase of $100 in SNAP benefits would somewhat increase unemployment for the first two years, but would still have close to zero effect for Unemployment Insurance and TANF. If only a few dollars are added into these programs every year, the effect would be close to zero on the unemployment rate. After going through all the diagnostics tests, we found that there were a few errors with our first model. There may exist an upward omitted variable bias which could call into question our conclusions about the effects of these welfare programs on the unemployment rate. This could possibly take away any of the effects that we’ve really concluded here. We also found evidence of a model misspecification that could potentially bias the coefficients on the welfare programs. Non-stationarity and outliers were not a problem in our original analysis. There was evidence for serial correlation, but we corrected for that using Newey-West robust standard errors, which in turn corrected for any heteroskedasticity that may have existed in our analysis. After correcting for all these, we have basically the same conclusion as before, that the SNAP Program has the most statistically significant coefficients and it increases unemployment, while the other two programs may actually decrease unemployment, but are not as statistically significant. Thus, if we put into perspective these programs, this evidence of only slight changes in the unemployment rate shows that the goals of these programs cannot be to affect the overall economy in drastic ways, but to further personal utility and wellbeing of the citizens within the economy. Analysis on programs like these need to take more than just numbers into account when evaluating the worth of a program. Statistically, these programs make only small incremental changes to the economy, but larger ones to individuals.

Bibliography:

Hoynes, Hilary W., and Diane Whitmore Schanzenbach. “Consumption Responses to In-Kind Transfers: Evidence from the Introduction of the Food Stamp Program.” American Economic Journal: Applied Economics, vol. 1, no. 4, 2009, pp. 109–139., http://www.jstor.org/stable/25760184.

Melkersson, Maria, and Jan Saarela. “Welfare Participation and Welfare Dependence among the Unemployed.” Journal of Population Economics, vol. 17, no. 3, 2004, pp. 409–431., http://www.jstor.org/stable/20007919.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s