Data monitoring: a hypothesis-testing approach for treatment-outcome research

Traditional inferential statistics require that hypotheses be evaluated at only 1 sample size. That is, researchers must choose how many participants will be included in a study before conducting analyses; they are not allowed to add data if initial results are not significant. This requirement forces researchers to choose among including more participants than necessary, risking inconclusive results, or violating the requirement by adding participants. This study presents a more flexible approach, called data monitoring, that allows repeating an analysis as the sample increases. First, the cost of the uncorrected data monitoring that researchers sometimes do is estimated. Second, the correction that is needed to allow data monitoring while holding an overall alpha at a desired level is estimated. Third, the power of data monitoring is compared with traditional approaches. This study also provides an example of the use of data monitoring. At least in some circumstances, data monitoring can reduce Type II error or the number of participants needed without sacrificing Type I error.