A Note on the Required Sample Size of Model-Based Dose-Finding Methods for Molecularly Targeted Agents

Abstract

Random forest has proven to be a successful machine learning method, but it also can be time-consuming for handling large datasets, especially for doing iterative tasks. Machine learning iterative imputation methods have been well accepted by researchers for imputing missing data, but such methods can be more time-consuming than standard imputation methods. To overcome this drawback, different parallel computing strategies have been proposed but their impact on imputation results and subsequent statistical analyses are relatively unknown. Newly proposed random forest implementations, such as ranger and randomForestSRC, have provided alternatives for easier parallelization, but their validity for doing iterative imputation are still unclear. Using random-forest imputation algorithm missForest as an example, this study examines two parallelized methods using newly proposed random forest implementations in comparison with the two parallel strategies (variable-wise distributed computation and model-wise distributed computation) using language-level parallelization from the software package. Results from the simulation experiments showed that the parallel strategies could influence both the imputation process and the final imputation results differently. Different parallel strategies can improve computational speed to a variable extent, and based on simulations, ranger can provide performance boost for datasets of different sizes with reasonable accuracy. Specifically, even though different strategies can produce similar normalized root mean squared prediction errors, the variable-wise distributed strategy led to additional biases when estimating the mean and inter-correlation of the covariates and their regression coefficients. And parallelization by randomForestSRC can lead to changes in both prediction errors and estimates.