Cause & Effect: Ma Health Insurance Mandate

...

The estimate of INTERACTION should give us the effect of the policy, but there are some assumptions we have to take into consideration when dealing with differences in differences. The estimate will give us an unbiased estimate if and only if the parallel trend assumption holds. The parallel trend says that if the policy had not taken place, the average change in Y would have been the same for treatment and control.

There are a few ways to measure the parallel trend. The first is called the eyeball test. Although it is not very accurate, we can plot data across time and check whether or not the control and treated groups are trending together prior to the policy effect date. After the policy is implemented, we expect to see a change in the treatment group and no longer be trending of that with the control group.

The second and more reliable way to measure parallel trend is through a placebo experiment. I will use time periods before the real policy date and create a fake treatment date. This placebo test will show us if there was any difference between the treatment group and control group. We expect the differences to be zero if parallel trend holds to be true. If the difference is non-zero, chances are that the difference in difference estimate is not a good method to find causal effect of the policy change.

Another assumption is that there may be a likelihood of a random effect that makes the control and treatment groups move together over time. If not careful, this could lead to a downward bias, or a notion that the policy has no effect at all. To control for this, I will use several pre-treatment time periods to calculate the variance of the difference in difference estimator and adjust the standard errors for it. In order to correct this, I will run my regression in Stata with the cluster command.

The last test will be for heterogeneity. Was everyone affected the same way? Did it matter if the individual was a male or female, young or old, and married or not? This analysis will concentrate on race, specifically if the individual was white or not.

4. Data

For this research I used data from the Integrated Public Use Microdata Series[3] (IPUMS) Current Population Survey (CPS). The two online resources collaborate and harmonize their data to be user friendly. It’s a popular resource for researchers and is deemed reliable. Table 1 below shows the means for the variables that will be discussed respectively:

The individual-level data I gathered was for all adults (aged 18-64) from 2002 to 2010. The treatment years are identified as 2008 and later. The control group are years 2006 and before. Although information was gathered for 2007, I dropped it from this analysis. This decision was made because the implementation of the policy was mid year, and the data used from IPUMS is gathered in March. Thus, I have no way of knowing if the data was before or after the policy for that particular year. Table 1 below shows the means for the variables that will be discussed respectively.

The second identifying statistic is the person's state of residence. In this set, I used 10 different states that had the highest population density per square mile[4] in 2012. These states (ranked from highest to lowest) are New Jersey, Rhode Island, Massachusetts, Connecticut, Maryland, Delaware, New York, Florida, Ohio, and Pennsylvania. I also used the 11th highest state, California, to account for the dropped fixed effects state.

The first dependent variable is if the individual has health insurance coverage. I acknowledge those that are insured as the treatment group. Those without insurance are marked as the control group. Being insured means having any health insurance at all, whether it’s public or private. It doesn’t take into consideration why a person doesn’t have coverage. For example, maybe the data classified the person as a resident of a state, but they moved within the 63-day grace period that the law mandated them to get covered. Maybe they were unaware of the new policy. The reform did allow individuals to petition unique situations where coverage may not have been warranted. Again, this variable doesn’t distinguish the reasons behind not being insured. It would make good data though.

The second dependent variable is if the person is healthy. Individuals were asked to conduct a self-assessment on their current health and rate it on a five-point scale, as excellent, very good, good, fair, or poor. For those that identified themselves with excellent or very good health, I made them the treatment group. For those that felt like they had only good, fair, or poor health I classified them as the control group.

The third dependent variable represents part-time or full-time work status conditional on working at all. There were a few examples in the data set that were either unemployed or had an unknown work status. Those observations were dropped accordingly. Part-time workers are my treatment group. Part-time is defined as working 34 hours a week or less on average during the previous calendar year. Full-time employment is my control group.

Finally, I will use race as an independent variable to measure for heterogeneity. In 2003, respondents to CPS could report more than one race. In this study, the treatment group is classified as only white. If the individual reported more than one race, then they would be in the control group.

5. Results

It is worth mentioning that my estimates are reported as percentage points and not just a percentage. A percentage point is the difference between two percentages. A quick example is to think about data at 5% and it increases to 10%. We can say, “The data increased by 5 percentage points.” We wouldn’t report the results in this section as saying, “The data increased by 100%.”

5.1 Effect on Health Insurance Coverage

After running the regression for equation 1, the DD estimates say that the policy increased the number of individuals insured by 6.8 percentage points. This is shown on Table 2. This estimate is statistically significant at the 1% level. While being statistically significant does not suggest the results are conclusive, it does imply that the statistic is reliable.

So what does this mean? How do we know that the reform really impacted the number of insured by 6.8 percentage points? One can argue that this increase would have happened with or without the implementation of the policy. Maybe people in Massachusetts are smarter (it does have Harvard and M.I.T after all) than other people in different states. They know the value of health insurance.

That’s