Revista de biometría y bioestadística

Efficient Monte Carlo Resampling of Continuous Data with a Dichotomous Treatment Variable

Abstract

John J Rogus, Shu-Fang Lin and Eduardo K Lacson Jr

Introduction: Methods such as generalized least squares regression and linear mixed models have traditionally been used for analyzing repeated measurement data. However, the computational burden for these procedures can be prohibitively high for large data sets. We propose an efficient, non-parametric method for the analysis of a continuous outcome variable with intrapatient correlation and a dichotomous predictor variable.
Methods: The patient-level values of the dichotomous variable of interest are randomized to generate sets of equally likely permutations of the data under the null hypothesis. For each replication, the test statistic for the dichotomous variable is calculated and the collection of all such test statistics forms an empirical reference distribution used to assign a p-value to the actual test statistic from the original data. Efficient calculation of the reference distribution is possible by operating on the level of sufficient statistics for the outcome variable, as the dichotomous nature of the predictor variable then allows for rapid recalculation of the tests statistic at each replicate. An example based on 629,452 measurements of systolic blood pressure in 39,313 dialysis patients is used for illustration.
Results: The Monte Carlo p-value for a decrease in systolic blood pressure following a decrease in dialysate sodium was 0.04. Other computationally feasible, but inefficient, approaches such as data aggregation and year-overyear comparisons were unable to find a significant association.
Discussion: Monte Carlo simulation offers a valid approach to analyze a continuous outcome variable with intrapatient correlation and a dichotomous predictor of interest. This method can accommodate other predictors through a two-step procedure involving an initial regression analysis. Future work is needed to characterize the power of this approach relative to other methods and to study whether weighting strategies may be helpful in the situation where not all patients contribute the same number of data points to the analysis.