A confounder variable may affect the estimated relationship between two other variables. In the present analysis we are interested in the association between caregiving arrangements and the outcomes of mortality or entering a nursing home. We may observe two kinds of confounding. one is a situation in which there is a significant bi-variate association between caregiving arrangements and the outcomes. For example, receiving help is related to mortality and entering a nursing home. If a confounding variable were associated with both caregiving arrangements and with the outcomes, then in a multivariate analysis where all the variables are included so that caregiving and outcomes are adjusted for the presence of the confounder, the original association might disappear or even be reversed. In the case of receiving help and mortality, controlling for severity of illness should cause the positive association between help and mortality to disappear. That is to say, there would be no true association between receiving help and dying; rather, both help and mortality are a reflection of an underlying progression of disease.
A second kind of confounder is a suppressor variable. This describes a situation in which there is no observed bivariate association between caregiving arrangements and the outcomes. This we have observed for most of the caregiving variables. In this instance, a variable that is associated with caregiving and with the outcomes in opposite ways could result in an observed association between caregiving and the outcomes when the confounder is included in a multivariate analysis. For example, if income decreased the likelihood of mortality, yet increased the likelihood of unpaid help, then the association between unpaid help and mortality may be seen to be positive after adjustment for income.
As described in our methods section, we have two major categories of confounding variables: social variables -- including demographic, socioeconomic, and social network variables -- and health status variables. Our analyses of confounders identified the sets of these variables that are related to mortality and to entering a nursing home, respectively.
As we have noted, for a confounder to affect the association of caregiving and either of the outcomes, it must be related both to the outcomes and to the caregiving. Thus, for each of caregiving comparisons we might determine which of the confounders that are related to mortality or entering a nursing home are also related to that particular caregiving arrangement. By doing this, we would have a unique set of confounders for each pair of caregiving arrangements and outcomes. To avoid the inefficiency of such an approach, however, we have decided to use all of the predictors related to each of the outcomes. As it turns out, each of the confounders is related to at least one of the caregiving arrangements, making it reasonable to include all of them. An additional rationale for keeping in a predictor related to the outcomes even though it is not related to the caregiving arrangements is for precision of estimation (Kleinbaum et al., 1980). We have included in Appendix B a table of associations between the confounder variables and the caregiving arrangements. The odds ratios in the table indicate the strengths of the associations. Although the associations differ slightly between the mortality and the institutionalization analyses because of the differing number of cases, the two sets of associations are nearly identical. Thus, we have reported only those for the mortality data set, the larger data set of the two.