Impact of Correlation between Multiple Time Point Measurements on Pooled Effect Measures in Meta-analysis

Abstract

Introduction: Effect estimates obtained from multiple time points based on the same set of subjects are observed to be correlated. There is a need to integrate these correlations in the derivation of pooled summary measures to improve the precision of estimates. The conventional meta-analysis does not consider this dependency into account. Aim: To compare the results obtained from meta-analysis which incorporate various levels of correlation in repeated measures data to the traditional meta-analysis. Materials and Methods: The present statistical analytical study was conducted at Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India, from January 2021 to February 2022 on data from a systematic review that studied the effect of analgesics in reducing orthodontic pain using Visual Analogue Scale (VAS) pain score measured at three time points was used for demonstration. This study attempted to illustrate two distinct approaches to deal with dependency between measurements obtained from different follow-ups by adopting constant and degenerating correlation structures. Results: The pooled effect estimates and confidence intervals obtained from models which incorporated correlation were different from the results of traditional approach. Naproxen fared to be better when compared to other two treatments. Pooled effect estimates and confidence intervals from Model 2 and Model 3 hovered around the same values. Non significant difference was observed in the Akaike Information Criterion (AIC) values of Model 2 and Model 3 for all three treatments. The between study variance ranged from 0.07 to 1.46, 1.25 to 3.17 and 0.01 to 0.98 for Acetaminophen, Naproxen and Ibuprofen, respectively. Conclusion: The models which took care of dependency had a better fit to the data over conventional meta-analysis.