Academic Articles Health Insurance Medicaid/Medicare

Crowd Out

In recent years, the federal government has attempted to increase access to government provided health insurance. Between 1984 and 2004, the percentage of non-elderly individual with government provided health insurance rose from 13.5% to 17.5%. Over the same time period, however, the percentage of American without health insurance also rose from 13.7% to 17.8%.

In their 1996 paper, Cutler and Gruber claim that increasing access to public health insurance plans crowds out private health insurance. It is an important policy question to understand whether expanding public health insurance is reducing the amount of uninsured individuals or simply shifting Americans from private to public insurance rolls. In Cutler and Gruber (QJE 1996), the authors estimate that a 10% increase in Medicaid coverage reduced private health insurance rates by 5%; this represents a 50% crowd out level.

Subsequent studies have argued that this 50% crowd out figure is an overestimate. Card and Shore-Sheppard (2004) use SIPP data (instead of the CPS) and found no crowd out with the 1990 OBRA Medicaid expansion. A paper a year later by Ham and Shore-Sheppard (2005) in Industrial and Labor Relations Review claims that by adding state*year interaction terms to the Cutler Gruber (1996) econometric specification changes the crowd out estimate to zero. Other studies, such as LoSasso and Buchmueller (2004) and Dubay and Keeney, have found crowd out estimates on the magnitude of the Culter/Gruber paper.

To combat these critics, Gruber and Simon have released a 2007 NBER working paper to re-estimate crowd out figures using updated data. The data used are the 1996-2002 SIPP data. Despite the panel nature of the data, Gruber and Simon have decided to treat the data as if it were simply a pooled cross-section, thus losing the ability to fully control for individual or household characteristics. Their econometric estimation technique is as follows:

  • INSijt = α + ELIGijt + νj + Ï?t + εijt

They authors also use an instrumental variables approach similar to the one employed in Currie and Gruber (1996). A random sample of 300 children of each age (and their families) is taken from each year of the SIPP. Eligibility rules for each state are applied to this sample for each of the 12 months of each of the years to calculate the fraction of the national sample eligible (in state j, time t) that is eligible for Medicaid. This effectively weights the rules in each state by their effects if applied nationally. Eligibility is instrumented by this ‘simulated percent eligible‘ variable. The authors also later include state*year interaction terms as well.

Using an individual level of observation, the authors find 20% to 40% crowd out, although the authors can not rule out that these estimates are statistically different from zero. Using family level estimates, crowd out is larger 60% to 80%, and these results are more statistically significant. One problem of using the family level estimates are that families where all household members are eligible for Medicaid of SCHIP and families where none of the household members are eligible for Medicaid are composed of vastly different income levels. SCHIP health insurance is available for all children part of a household with income below 133% of the federal poverty line and some children between 133% and 350% of the federal poverty line depending on the state, whereas adult generally need to be below the poverty line to qualify for Medicaid. Also, I find it unintuitive that the paper finds more crowd-out for individuals with employer-provided health insurance compared to non-group policies. Could these individuals be switching jobs to higher paying jobs without insurance and then taking up Medicaid coverage?

The authors also examine anti-crowd out measures such as mandatory waiting times between when private insurance is dropped and when Medicaid insurance is taken up. Gruber and Simon find that the waiting times are ineffective against preventing crowd out but these estimates are not precise.