Methods for combining cohort data set having different variables for pooled-analysis

摘要

BACKGROUND : When we conduct pooled-analysis from different cohort data, we often face data set having different variables. In addition, the system of collecting cohort data may different among studies. Therefore, we need to find how to reduce bias taking advantage of statistical and mathematical methods. Further, we will discuss the way to set the thresholds of effects in cohort study. METHODS : We used cohort data of infants from BICCA(Birth cohort consortium of Asia) which develops a core set of exposure and outcome measures. Since the cohort data have various variables, we used multivariate analysis. We also used randomized sampling methods and Monte-Carlo Variance Reduction Principle for proof. RESULTS : When extracting from original data set using SRSWR(Simple Random Sampling With Reputation) with stratified method, we found Monte-Carlo Principal was applicable. The 1st stratum was center and 2nd stratum was inspector. We found that variance was reduced, which means bias was reduced. Also, when we compared 3 countries with final regression model, converting each variable for normalization fit better than just setting original variable. For instance, to compare Infant's neurodevelopment, there is different estimator among 3 countries. When we normalized similar category of Taipei's and Japan's to Korea's Bayley, we got better fitting.(Variance reduced almost to 1/4) CONCLUSION : We obtained more reliable result using new variables when combining data of different cohort studies. Also, by normalization, we can combine, similar but not exactly same estimator. In this manner, countries which couldn't do this cohort study because of the expansive cost of neurodevelopment, can now do it, giving possibility that international cohort study will be opened. For this reason, I also submit it to the Symposium.