Grant: Travel Awards - Institute Community Support

Background: Missing data on important variables in large clinical databases is a challenge for primary care researchers. Including only complete cases (with no missing data) in an analysis may decrease the accuracy of results. Different statistical approaches were suggested to deal with this problem. There is, however, no general rule on how to evaluate which of them provides more accurate results. A so-called "plasmode simulation" technique is a promising modern method to deal with this problem. Objective: Using a "plasmode simulation" approach, we compare the accuracy and precision of results between the two statistical techniques. The first technique includes only complete cases in the analysis (CC). The second replaces missing data with data obtained by another statistical method called "multiple imputation by chained equations (MICE)". We will apply these techniques to a large real-life primary care dataset. Methods: This study uses population-based data, the national Primary Care Practice-based Electronic Medical Records Database, for 2011-2016. We will measure weight, antidepressants (AD) prescribing, age, sex, comorbidities, smoking. A statistical method called logistic regression will be used to analyze the "plasmode" (part real, part generated) data in order to estimate whether obese patients are more likely to be prescribed AD than normal weight patients. Results: Of 120381 patients, 46.8% lacked data on weight, 0.02% on sex, and 64.2% on smoking. Patients with obesity were 1.16 to 1.26 times more likely to have AD prescription than patients with normal weight when using the CC analysis. With the MICE technique, obese patients were 1.10 to 1.21 times more likely to receive AD. Use of a "plasmode simulation" will show which of these methods is more accurate and precise. Conclusion: In the original data, results differed between CC and MICE analyses. Plasmode simulations will help to choose the best way to handle missing data in primary care databases.