Introduction Self-rated health is a widely used indicator of overall health status. It is most often reported on a Likert scale of three to five values in surveys. To facilitate presentation and interpretation, it is common practice to simplify the variable by dichotomizing it; however, little documented reflection has been done on how this should be done. Methods We use regression analysis to explore the validity and reliability of all four possible dichotomizations of self-reported health in the Canadian Community Health Survey in 2013-2015 by mapping them to a validated health measure: the Health Utility Index Mark 3 (HUI). We posit that more valid cutpoints in self-rated health are associated with larger changes in HUI. We posit further that more reliable cutpoints are associated with similar changes across sociodemographic variables, including age, sex, education, marital status, geography and income. We also provide descriptive statistics to contextualize our analysis. Results The greatest proportion of respondents reported having "very good" health, although the proportion of the population reporting "excellent" or "very good" health decreased with age. Similarly, Canadians tend to score highly in HUI. Our regression results suggest that HUI tends to be higher for younger, richer, married, educated and urban populations. However, these associations are muted as the cutpoint used to dichotomize self-reported health is raised. The model with the lowest cutpoint, distinguishing between poor health and all other health statuses, was associated with the greatest and most consistent negative changes in HUI among different sociodemographic groups. Conclusions Dichotomizing self-rated health using lower cutpoints captures more pronounced differences in health status measured by HUI and tends to capture more consistent differences across sociodemographic variables. That is, lower cutpoints produce more valid and reliable results. However, lower cutpoints isolate less commonly reported health levels and may lead to less accurate results in smaller populations.