2.3. Data Processing and Statistical Analysis

To perform the medication risk stratification, a webservice interface and customized scripts were used. The MRS^TM were generated by processing prescribed drug claims using National Drug Codes (NDCs) as drug identifiers. Medication data were extracted and cleaned from errors and inconsistencies through quality and integrity analyses. Since NDCs can also be assigned to non-medications (e.g., medical devices and consumables), active medication data were further filtered to exclude such NDCs. Afterward, active medication data for each participant was filtered based on prescription dates and days’ supply, which include any possible refills.

Descriptive population characteristics were measured, including means, standard deviations, and proportions as appropriate. For comparing the MRS™ and composite individual risk factors before and after addition of repurposed drugs into participants’ drug regimens, the Wilcoxon signed-rank test analysis was performed. To determine the statistical significance of participants moving to a higher risk stratification category (Low-to-Moderate, Low-to-High, or Moderate-to-High), the McNemar test and the Stuart–Maxwell test of marginal homogeneity were utilized (no participants moved to a lower risk score category). To determine if the MRS™ or a risk category were more influenced by one drug than by others, we used a Friedman test, followed by paired comparisons with the Wilcoxon signed-rank test. For all the Wilcoxon signed-rank tests, the ranks of zeros were included in calculating the statistic (implemented as zero method = ‘zsplit’ in SciPy 1.4.1). For the statistical analysis assessing the difference between female and male groups, the effect size was calculated using the method denoted f.

The diseases were identified by finding the ICD-10 codes with the highest number of participants, limiting the ICD-10 codes to the first three digits. When ICD-9 codes were still used, these codes were translated to the appropriate ICD-10 code, if possible.

For statistical significance, we considered p-values below 0.05 to be significant. A population’s mean score change of at least 1 point was considered significant, as reported by Bankes et al. [21]. To adjust for multiple comparisons, the Benjamini/Hochberg adjustment was applied. When the p-values were too low, a value of p < 0.0001 is indicated. Statistical analysis was performed in Python 3.7.6 using the pandas (v. 1.0.1) (open-source software fiscally sponsored by NumFOCUS, Austin, TX, USA), SciPy (v. 1.4.1)) (open-source software fiscally sponsored by NumFOCUS (Austin, TX, USA)), statsmodels (v. 0.11.0) (open-source software sponsored by Google (Menlo Park, CA, USA) and AQR Capital Management (Greenwich, CT, USA)), Matplotlib (v. 3.1.3) (open-source comprehensive library sponsored by NumFOCUS (Austin, TX, USA)), and seaborn packages (v. 0.10.0) (a Python data visualization library created by Michael Waskom (New York, NY, USA) and in R (v. 1.2.5019) (R foundation for statistical computing. RC Team. (Vienna, Austria) with the dplyr, data.table, sqldf, scales, ggplot2, and igraph packages. Microsoft SQL Server (v. 15) was used to manipulate and analyze large datasets.