contains descriptive results on our quality measures as well as demographic, socioeconomic and health factors potentially affecting these measures. In general, blacks fare worse on the quality measures, experiencing more physical restraints, catheterizations, more anti-psychotic medication, and, in the largest quantitative difference, much more frequent use of feeding tubes. All of these differences are statistically significant within our very large sample size. The descriptive results from will be used to construct a first measure of unfair differences by race.

The demographic and socioeconomic status characteristics also differ between the races. Blacks are younger, more likely to be male, and less likely to be currently or previously married. Blacks are more likely to be currently on Medicaid and to be in the lower educational groups. Importantly, there are a large number of observations with missing information on education—for both blacks and whites. However, blacks are less likely to have missing education information. The length of time in the nursing home differs little between the races. Blacks have slightly higher ADL and CPS scores, indicating they are both more physically and cognitively disabled (despite being younger on average).

displays results from a series of regressions based on specifications (

1) and (

3) above. Extensive demographic and health status measures are included in the models, but we report only those coefficients pertinent to our disparity calculations: the estimate of the race effect and the SES variables through which racial disparities might be mediated. There are two models for each quality indicator. Odd number models do not include nursing home fixed effects, while the even ones do; otherwise, the models are identical.

| **Table 2**Regression Results from Basic Specification |

Recall that we will use the odd-numbered models to compute two measures of disparities, one based on the race coefficient and a second one adding payer and educational status mediated effects of race. Because “more is worse” for all of our quality indicators, a positive coefficient on the race variable is consistent with disparities. Thus, the small, negative coefficients on race in models 1, 3 and 5 do not indicate disparities. For example, the estimated race effect of −.00232 for Restraint use in Model 1 means that blacks are 0.232 percentage points less likely to have restraints applied adjusting for all SES and health status variables in the model. The only quality indicator with meaningful disparities as measured by the estimated race coefficient is feeding tube use where blacks are 6.2 percentage points more likely to have feeding tubes after adjusting for all other variables. This is somewhat less than the 10.2 percentage point unadjusted mean difference in feeding tube use from .

The results in are used in the calculation of disparities, allowing for SES mediation as recommended by the IOM. The specific variables used for mediation are the payer status and education variables listed in the table. All are indicator variables. For example, private-pay status is a binary measure (0,1) of whether a resident paid out-of-pocket or used insurance during that quarter. Medicaid is the omitted category and we expect both private-pay and the less frequently observed other payer to be associated with higher quality because of the generally higher prices nursing homes are paid from these sources. We see exactly this relationship in the case of physical restraint use where both variables have a negative coefficient.

The mediation analysis works as follows: the estimated effect of private pay is −.00104, suggesting there is a small (0.1 percentage point) decrease in the likelihood of restraint use for a resident with private sources of payment compared to Medicaid. How much does this contribute to disparities? From , we know that 34.0% of whites pay privately, but only 11.1% of blacks do. Thus, the effect of private payment on disparities is the black-white disadvantage in this favorable variable (.111–.340) times the effect of the variable (−.001). This product is positive, meaning it does contribute to disparities, but the magnitude, in this case, is small at +.02 percent. All of the indicator variables in are expected to be favorable in terms of an impact on quality in relation to the omitted categories, Medicaid in the case of the payer variable, and “not completing high school” in relation to the education variables. The negative sign on all of these variables in Model 1 is consistent with this expectation. The negative sign pattern holds for physical restraints, antipsychotics and feeding tubes, but not so clearly for catheter use.

Blacks have a less favorable payer and education profile than whites; therefore, because the black-white difference for these variables is always negative, a negative sign on the variable indicates that mediation contributes to disparities. Thus, in general, with the mainly negative coefficients on these variables in , the recognition of mediation will increase measured disparities. The calculation of the mediation impact of disparities is straightforward following from formula (

2) above. The magnitude of these increments to the disparity measures is depicted in , along with the mean differences from .

The mean differences for all four measures were positive, consistent with black-white disparities, but were small in all cases except for feeding tube use. The basic finding with respect to physical restraints, catheters and anti-psychotics is that after adjustment for health status measures, there are no black-white disparities when measured either by the coefficient on black race or when allowing for mediation through SES variables. For these three measures, the race coefficient is negative and therefore not consistent with disparities. Mediation makes the disparity measure slightly less negative in the case of physical restraints and anti-psychotics, but not positive. Mediation actually makes the disparities estimate slightly more negative for catheter use.

The largest mean difference in is for feeding tube use: 15.9% for blacks and 5.7% for whites. Adjustment for health status and SES reduces the black coefficient to a 6.2 percentage point difference, but this is still quite large in relation to the mean for whites, suggesting a rough doubling of the risk of feeding tube use. All of the estimated coefficients in Model 7 are negative, indicating that there are also some disparities mediated through payer and educational status. As indicates, however, the magnitude of these measured mediators is small, increasing estimated disparities by only about 1 percentage point.

Results reported in can also be used to attribute disparities to within facility versus across facility racial differences. Because we find no disparities in the first three measures, the issue of decomposition is not interesting for these. In the case of feeding tubes, inclusion of nursing home fixed effects reduces the race coefficient from .062 to .037. The .037 is an estimate of the average “within nursing home” race effect adjusting for all other variables. Thus, by our accounting, 60% (.37/.62) of the coefficient-based disparities in feeding tube use are within nursing home disparities, and the balance is across nursing home disparities. Mediation has essentially no effect on the within nursing home estimate of disparities.

Finally, we turn to our fully interactive model for estimating disparities, the results are displayed graphically in for each of the four quality measures. Regression estimates for the fully interactive model are reported in . The race-payer status interaction terms are statistically meaningful across all four outcomes, while the race-education terms are generally significant in all cases except for the catheters model. We compare the fully interacted results from (

4) to those from the basic model (

1). Panel (a) of shows results for physical restraint use. The coefficient estimate of −0.2 percentage points was reported in and is reproduced here to be compared to the RDE estimate, which gathers the main effect and all the interaction effects with black race. The disparity becomes slightly more negative in Panel a than in Panel c. With respect to catheter use (Panel b) and feeding tubes (Panel d), the measured disparity increases quite a bit when comparing the RDE to the simple coefficient estimate. The RDE measure includes effects such as the following: if illness severity as measured by ADL makes it more likely that a resident uses a feeding tube, but that effect is greater for a black resident, the interaction of black and severity will reflect this. The RDE includes all interaction effects, with health status, SES and demographic variables as part of the effect of black race.

| **Table 3**Regression Results, Fully Interacted Model |

The second two bars in each panel compare the mediated disparity from the basic model with the mediated estimate in the fully interacted model describe by

equation (2"). In effect, the fully mediated estimate excludes differences due to health status differences between blacks and whites using the estimated coefficient for whites. Differences due to SES are included to be consistent with the IOM approach. The difference between the basic and fully interactive mediated estimates are that the basic model weights the differences by the effect of health status for both groups on average, and the fully interacted uses the white coefficients as the standard. This change in coefficient to weight racial differences in health status does not have a major effect on any estimate. It does, however, increase the estimated disparities in the feeding tube case, the one process indicator where we consistently find a black-white disparity.