Different approaches to dealing with missing values in quantitative variables and survey of their effects on the results of a clinical trial

Background and Objectives: A major challenge that affects the longitudinal studies is the problem of missing data. Missing in the data may result in the loss of part of the information which reduces the accuracy of the estimator and obtain the results will be biased and inaccurate. Therefore, it is necessary to evaluate the missing data mechanism from a longitudinal research and to consider this fact, an analysis should be performed on the data with or without imputation. The purpose of this study was to identify the Missing data mechanism from a clinical trial Related to the efficacy of non-surgical method of joint distraction in improving performance in patients with severe knee osteoarthritis, and the effect of different approaches to dealing with the missing on the results the analysis of this data set.

Methods: In order to obtain reliable results in the analysis of the data associated with the longitudinal clinical studies involving missing values, there are different approaches for dealing with such data, including complete case data approach, available data, LOCF and multiple imputation methods such as predictive model based method, propensity score method (propensity score) and predictive mean matching method (Predictive mean matching method). This approaches were used on the data obtained from clinical trial of the effectiveness of non-surgical joint distraction by two methods physiotherapy and joint distraction in addition to physiotherapy in improving patients performance with severe knee osteoarthritis were measured in three times, beginning of treatment, end of treatment and one month later and the results were compared. SAS 9.1 software was applied for fitting the marginal model and SOLAS software version 4 for missing data imputation.

Results: By using the complete data, available data and LOCF imputation approaches time effect was significant. All the three multiple imputation methods had the same results. In none of the approaches used, the effect of treatment group was not significant, although there were significant differences in the length of confidence intervals and coefficients of the model.

Conclusion: According to the missing data mechanism and results obtained from coefficients of the model, standard errors and confidence interval length, to use of multiple imputation approach in the data is more appropriate and LOCF imputation method is not appropriate for interventional longitudinal data.