Design Suggestions for Comparative Effectiveness Studies
Choosing the right statistical model reduces the risk of obtaining biased, misleading results from randomized trials and observational comparative effectiveness studies. Here are three examples.
Take Advantage of Within-Subject Correlation in Longitudinal Studies
In a longitudinal study, the outcome variable is measured repeatedly over time. Missing outcome measurements are a potential source of bias, but this potential can be mitigated by leveraging the positive correlation among each subject’s series of measurements.
For comparing longitudinal outcomes between treatment groups, it is best to choose a method that accounts for this within-subject correlation, such as a linear mixed model or generalized linear mixed model, rather than a method that disregards within-subject correlation, such as a t-test, chi-square test, or simple ANOVA.
Match on Estimated Propensity Score in Observational Studies
In an observational comparative effectiveness study, assignments to treatment and control groups are determined by the choices of patients and providers. These choices can be affected by numerous covariates, leading to imbalances in the distributions of covariates between groups, a potential source of confounding bias.
Constructing treatment and control groups by matching patients on estimated propensity scores is a very effective means of balancing the distributions of observed covariates between groups and is generally more effective at reducing confounding bias than simple regression adjustment.
Address Competing Risks in Studies with Time-to-Event Outcomes
A competing risk is an event that alters the probability of the event of interest, when the study outcome is the time-to-event of interest (survival time). If a competing event is clinically associated with the event of interest, it is a potential source of bias.
For comparing survival times between treatment groups, it is best to use a method that separately models the cumulative incidences of the event of interest and the competing event, rather than a method that simply assumes the events are entirely independent, such as a log-rank test or Cox proportional hazards regression.