Stephen J.LurieMD, PhD, Senior EditorIndividualAuthorJody W.ZylkeMD, Contributing EditorIndividualAuthor
In Reply: Dr Kyriakides and Mr Guarino raise several important issues regarding the analysis of data derived from observational cohorts. Analysis of linear data, such as CD4 cell counts or viral loads, is more statistically powerful when the data are kept continuous, rather than categorized. In general, categorization of data (such as CD4 cell count <200/mm3) reduces statistical power and can underestimate effects that might be found with linear data. This is only a problem if a negative result is obtained when analysis of continuous data would yield a positive result. Our analysis found important differences by the CD4 cell strata we used, so no underestimation of effect resulted. In addition, categorization of laboratory data is important in clinical practice, for physicians need thresholds on which to base therapeutic interventions. Rather than being arbitrary, our cut points were based on widely used and accepted thresholds of CD4 cell counts and viral loads that have been validated both prognostically and therapeutically. While an analysis that reports that the relative hazard of treatment failure is 1.001 for each 1-cell decline in the CD4 cell count below 500/mm3 might be statistically powerful, it would have little utility for clinicians. We believe our analysis was sufficiently powered and clinically framed to provide both valid and useful results.
Chaisson RE, Moore RD. Timing of Antiretroviral Treatment Initiation—Reply. JAMA. 2001;285(13):1702-1703. doi:10.1001/jama.285.13.1702