The contribution of insurance coverage and community resources to reducing racial/ethnic disparities in access to care - Impact of Health Care Context

Health Services Research, June, 2003 by J. Lee Hargraves, Jack Hadley

Need. We included a measure of general health status (poor, fair, good, very good, or excellent) to account for individual needs for medical care. The final models include two community-level demand measures: the percentage of persons below the FPL and the percentage who were African American or Hispanic.

Regression-Based Decomposition

We used regression-based decomposition (Oaxaca 1973) to separate observed differences in access to medical care into two parts: that due to measured personal and community characteristics (i.e., factors that we can explain), and that which cannot be explained by differences in observed characteristics. The second component can be thought of as measuring differences between whites and ethnic minorities attributable to differences in the "returns" on their characteristics, because they are based on differences in regression coefficients.

This approach requires estimating linear models using ordinary least squares regression, even though the dependent variables are binary measures. Linear models have the desirable property that the mean of the dependent variable equals the sum of the mean values of the independent variables multiplied by their respective coefficients. Even though linear probability models can yield predicted probabilities outside the 0/1 range, the parameter estimates are consistent, which is the critical property for the decomposition analysis (Acs 1995; Acs and Danziger 1993).

Using comparisons between whites and Hispanics as an example, the mean values for each access indicator (Y) for whites (w) and Hispanics (h) evaluated at the means of the independent variables can be represented by:

[Y.sub.h] = [X'sub.h][[beta].sub.h]

and

[Y.sub.w] = [X'sub.w][[beta].sub.w](1)

Thus, differences between whites and Hispanics in the access indicator can be expressed as:

[Y.sub.w]-[Y.sub.h] = [X'.sub.w][[beta].sub.w]-[X'.sub.h][[beta].sub.h] (2)

We add and subtract [X.sub.w][[beta].sub.h] to obtain:

[Y.sub.w]-[Y.sub.h] = (X'[[beta].sub.w] [X'.sub.w][[beta].sub.h]) - ([X'.sub.w][[beta].sub.h]-[X'.sub.w][[beta].sub.w]) (3)

and rearrange terms to decompose the overall differences into an explained and an unexplained component:

[Y.sub.w]-[Y.sub.h] = [X.sub.w]([[beta].sub.w]-[[beta].sub.h]) ([X'.sub.w]-[X'.sub.h])[[beta].sub.h] (4)

The first term on the right hand side of equation 4 is the difference in the "returns" to personal characteristics evaluated at the white's mean characteristics. Here, for example, personal characteristics could include insurance coverage, income, education level, and health status. in effect, we simulate a model in which everyone has the characteristics of the average white person, and then analyze whether or not there would be a difference in the returns to those same characteristics for Hispanics.

This is the portion of the difference that is not explained by differences in measured characteristics. Presumably, they result from differences in unobservable characteristics such as care-seeking behavior, attitudes, or discrimination. They suggest that the behavioral effect of any observed characteristic, such as insurance coverage or income, is unequal among different ethnic or racial groups. This portion is unexplained because we have controlled for observable differences between whites and Hispanics.