Effect of dichotomising a continuous variable on the assessment of familial aggregation: an empirical study using body mass index data from the Busselton Health Study.

BACKGROUND: Continuous measures are often dichotomised according to some meaningful threshold. When measures of association are the study focus, such as in familial aggregation studies, we may use an odds-ratio (OR), instead of a Pearson correlation coefficient, to measure correlation between outcomes. In this paper, we examine the effect of using different threshold values on the resulting OR estimates and their statistical efficiency.

METHODS: We use an example based on obesity data, in the form of BMI measurements on family members, to guide a study of the OR based on four-fold probabilities which result when a pair of normally distributed outcomes is dichotomised at a certain value. This leads to a study of outcomes from simulated nuclear family data. As a possible alternative, we also assess the performance use of a simple tetrachoric correlation coefficient, the dichotomous analogue of Pearson's correlation coefficient.

RESULTS: The studies indicate that dichotomisation at values close to the mean leads to results more comparable to the continuous equivalent. The results show that, in our setting, there is a positive relationship between threshold value and resulting OR, and between threshold value and resulting standard errors. The value of the tetrachoric correlation coefficient decreases, relative to Pearson's correlation coefficient, as threshold values deviate from the mean.

DISCUSSION: Familial OR that result from dichotomisation are influenced by the choice of threshold value. We discuss implications of this and related issues on the interpretation of results.