We read with interest Garriock and colleagues’ article “A Genomewide Association Study of Citalopram Response in Major Depressive Disorder” (1). This work is an important step in analyzing the Sequenced Treatment Alternatives to Relieve Depression (STAR*D) genomewide SNP data. Given STAR*D’s complexity, many additional analyses will undoubtedly be required to realize the full potential of this valuable resource. In this spirit, we quantitatively model response as measured by the Quick Inventory of Depressive Symptoms (QIDS). The gains to quantitatively modeling response are evident in our results—unlike the initial analysis, we identify multiple genomewide significant SNPs.
We apply three well-established psychometric insights to improve treatment response measurement. First, dichotomizing a quantitative variable sharply reduces statistical power (2, 3). For instance, in regression involving two normally distributed variables, dichotomizing the predictor at its median reduces variance explained by 38%, with further reductions as the dichotomization point moves away from the median (3). Thus, dichotomizing citalopram response likely substantially decreases power to detect SNP effects. Second, when measuring change (i.e., treatment response), increasing number of assessments per subject monotonically increases estimate precision, with dramatic gains when assessments per subject are numerous, as in STAR*D (4, 5). Thus, modeling all QIDS assessments (~7.2 per subject) increases response estimate precision more than four-fold compared to using only first and last assessments (5). Third, as questionnaire items generally vary in how well they describe underlying constructs (6), factor analyses can improve measurement through estimating the association of each QIDS item to the depressed affect construct. Cumulatively, these techniques help maximize power to find genetic effects, which is crucial in pharmacogenomics given tremendously expensive clinical trials and relatively small sample sizes.