Editors' ChoiceBiostatistics

Improved Analysis

See allHide authors and affiliations

Science Translational Medicine  23 Jun 2010:
Vol. 2, Issue 37, pp. 37ec100
DOI: 10.1126/scitranslmed.3001360

Many chronic conditions—such as migraines and asthma—manifest as events that occur repeatedly over time. Clinical trials to investigate treatments for such conditions often involve repeated treatments over the study period, resulting in outcome data that are recorded as a sequence of treatment successes and failures for each participant. These data are similar to classic longitudinal data in that there are repeated measurements taken on a participant over time, but different because the number of data points varies for each participant for reasons other than the participant dropping out of the study. It is well known that the analysis strategy for repeatedly measured data should account for the correlation that exists between the successes/failures for a single individual. What has been less clear is whether applying standard analysis strategies is correct when the number of data points varies randomly across the individuals.

Pullenayegum and Cook now show that in studies of episodic events, the mechanisms that lead to varying amounts of data on each participant differ enough from conventional dropout mechanisms that traditional longitudinal analysis methods, such as generalized estimating equations (GEEs) or analysis of the average response, lead to biased results. Fortunately, they show that other existing but less standard approaches offer solutions to the two major statistical issues: (i) participants that experience the most events are different than those with fewer events, and (ii) the effect of treatment changes over time, and thus those with more events have more (or less) favorable outcomes. Pullenayegum and Cook show that if these characteristics differ between the treatment and control groups, classic analyses can fail miserably. Two of the alternative analysis strategies they show to have better performance include cluster-based resampling, in which the resampling is such that a single measurement is randomly drawn from each subject and analyzed with standard analysis methods, and weighted GEE, in which the probability of the number of measurements is incorporated into the analysis.

E. M. Pullenayegum, R. J. Cook, The analysis of treatment effects for recurring episodic conditions. Stat. Med. 29, 1539–1558 (2010). [Abstract]

Stay Connected to Science Translational Medicine

Navigate This Article