OBJECTIVES: Prospective clinical trial (PCT) brings the rigor of clinical trial research to the real world setting and makes findings more relevant for physicians, payers, and patients. This study was conducted to compare the accuracy of casewise deletion (CD) of missing data and Predictive Mean Matching (PMM) imputation in R, Multivariate Imputation by Chained Equation-MICE).

METHODS: Simulated data was constructed containing four independent (IV) and one dependent variables (DV). Two levels of correlation (Low: 0.1 - 0.3 and Medium: 0.5) among IVs were created using Cholesky’s decomposition algorithm. The data missing completely at random (MCAR) assumption formed the creation of missing values. Percent of missingness varied from 2-60% for 1 IV and 2 IVs. Simulations use a sample size = 1,000. Analysis was linear regression for 4 IVs and 1 DV. Comparisons were made over beta coefficients, Residual Standard Error (RSE), and Adjusted (R2). PMM was compared to CD on performance in reproducing results from the gold standard analysis using complete data.

RESULTS: While PMM reproduced Intercepts, RSEs and IVs with complete data more accurately when 2 IVs had missing values, for IVs with missing values, CD better estimates of intercepts and RSE than PMM in 23 of 40 simulations (57.5%). Regardless of the percent of missing values, CD was more accurate when only one IV had missing values (62.5%). R values were more accurate with PMM (81.3%) for low missingness (≤8%) but there was no advantage to PMM as compared to CD with missingness ≥10% (50/50 percent).

CONCLUSIONS: Results suggest there is no absolute advantage to PMM over CD when the percent of missingness exceeds 10%, or when the IVs of interest are the variables with missing data. When missingness are ≤8% PMM work better than CD but only for variables with no missing values and for RSE.