OBJECTIVES: Missing data is a common challenge in healthcare research, and various imputation methods have been developed to address this issue. Markov-chain Monte Carlo (MCMC) imputation is a well-accepted method for handling continuous missing data in primary studies. Our primary objective is to propose an alternative parametric imputation using empirical cumulative distribution function (eCDF) curve (PIE) and compare its performance with the standard MCMC method.

METHODS: Data on diastolic blood pressure (DBP) and 2-hour serum insulin levels (SIL) for 768 patients were selected for analysis from a publicly available dataset (Smith 1988). We randomly removed 10% of the data from this primary dataset to create a missing value dataset (MVD). Different parametric distributions (Normal, Cauchy, Log-normal, Gamma, Weibull, Exponential, Pareto etc.) were fitted on MVD. Model fit was compared using Akaike’s Information Criteria (AIC) and eCDF. Each missing value was imputed with the mean of one hundred samples from the best-fitting distribution. Root-mean-square error (RMSE) was estimated to compare the accuracy of the two techniques. Sensitivity analysis for the estimated parameters of the best-fitting distribution was performed using bootstrap sampling to check for parametric uncertainty and results were summarized using RMSE.

RESULTS: Log-normal distribution was the best-fitting distribution for both the scenarios. RMSEs were estimated to be lower for PIE compared to MCMC for both DBP [PIE: 7.9; MCMC (five-seeds): 9.8 to 11.9] and SIL [PIE: 108.1; MCMC (five-seeds): 196.6 to 209.5]. Bootstrapped sampling yielded RMSE ranges of 7.5 to 8.2 for DBP and 103.1 to 120.8 for SIL. The estimated RMSE to check the parametric uncertainty in PIE imputation was lower compared to that of MCMC imputation.

CONCLUSIONS: The results of this comparative analysis suggest that PIE imputation outperforms MCMC imputation in terms of RMSE, indicating its potential as a viable alternative.