Polytomous logistic regression as a tool for exploring heterogeneity across birth defect subtypes: an example using...

In birth defect epidemiology, phenotypic subgroups are often combined into a composite phenotype in an effort to increase statistical power. Although the validity of using composite phenotypes has been questioned, formal evaluations of the underlying assumption of effect homogeneity across component phenotypes have not been conducted.

Polytomous logistic regression was used to assess effect heterogeneity of several generally accepted neural tube defect (NTD) risk factors across the component phenotypes of anencephaly and spina bifida. Data for these analyses were obtained from the National Birth Defects Prevention Study.

The use of a composite phenotype has the potential to mask associations specific to a component phenotype and in some cases the effect of a variable may be misattributed to the composite phenotype. For example, an association between infant sex and anencephaly (adjusted odds ratio [AOR], 1.5; 95% CI, 1.1-1.9) was masked when data from all NTDs were analyzed (AOR, 1.1; 95% CI, 0.9-1.3), whereas an association with maternal body mass index that was specific to spina bifida (AOR, 1.9; 95% CI, 1.6-2.4) was attributed to all NTDs (AOR, 1.6; 95% CI, 1.4-2.0). Furthermore, conclusions regarding effect heterogeneity based on ad hoc comparisons, rather than some formal assessment, may be vulnerable to considerable subjectivity, as was the case for the association of maternal Hispanic ethnicity with spina bifida (AOR, 1.4; 95% CI, 1.2-1.8) and anencephaly (AOR, 2.0; 95% CI, 1.5-2.8).

Polytomous logistic regression provides a useful tool for evaluating putative risk factors for which there is no a priori basis for assuming effect homogeneity across component phenotypes.