Likelihood methods for treatment noncompliance and subsequent nonresponse in randomized trials

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Causal Inference

Background:

• Randomized trials often face challenges with treatment noncompliance and missing data.
• Simultaneously addressing both noncompliance and nonresponse is crucial for accurate causal effect estimation.
• Existing methods rarely tackle these dual issues concurrently.

Purpose of the Study:

• To develop a statistical method that simultaneously accounts for noncompliance and nonresponse in randomized trials.
• To construct a maximum likelihood estimator (MLE) for the causal effect of treatment assignment.
• To evaluate the performance of the proposed MLE against existing estimators.

Main Methods:

• Utilized a maximum likelihood estimator (MLE) for causal effect estimation in two-armed randomized trials.
• Employed the Expectation-Maximization (EM) algorithm for parameter estimation.
• Incorporated a latent compliance state covariate to model subject behavior and missing data mechanisms.

Main Results:

• The proposed MLE demonstrated favorable performance compared to method-of-moments (MOM) and intention-to-treat (ITT) estimators.
• The MLE showed robustness under both normal and non-normal data distributions.
• The method performed well even with departures from latent ignorability and compound exclusion restriction assumptions.

Conclusions:

• The developed MLE provides a robust approach for analyzing randomized trials with both noncompliance and nonresponse.
• This method offers improved causal effect estimation in complex trial settings.
• The approach was illustrated using data from a clinical trial comparing antipsychotics for schizophrenia.