Some Methodological Remarks on Self-Rated Health

In analyses concerned with self-rated health it is common to dichotomise an ordinal-scale health measure and compare different subgroups of a population on basis of odds ratios from logistic regression models. Selfrated health is often explored also in wide age intervals. Since people’s health correlates strongly with their age, that approach is empirically problematic, particularly when it comes to determining the cut-off point for dichotomisation and the role of age-dependent covariates. We set out to investigate these issues, and prove them to be practically relevant.

Using a highly representative data set, the Health 2000 survey in Finland, we focus on ages 35-64 years. Separate analyses are undertaken for this whole age interval, and for three shorter age intervals.

Self-rated health was in the survey measured on a five-point ordinal scale. We dichotomised the responses in two alternative ways (bad health categorised as “poor” or “fairly poor”, and as “poor”, “fairly poor” or “average”), and explored the estimated effects of some standard covariates.

When the whole age interval was analysed, the choice of cut-off point for health dichotomisation had only a modest impact on the estimated effects of the covariates. However, with a narrower categorisation of poor health, the effect of educational level, as well as of marital status, was found to be highly age-dependent.

Researchers and health policy practitioners should be aware of the risks for drawing misleading or even incorrect conclusions from studies of self-rated health based on wide age intervals that do not explicitly account for agedependent covariates.