The Missing=Smoking Assumption: A Fallacy in Internet-Based Smoking Cessation Trials?

Introduction: In this study, penalized imputation (PI), a common approach to handle missing smoking status data and sometimes referred to as "missing=smoking," is compared with other missing data approaches using data from internet-based smoking cessation trials. Two hypotheses were tested: (1) PI leads to more conservative effect estimates than complete observations analysis; and (2) PI and multiple imputation (MI) lead to similar effect estimates under balanced (equal missingness proportions among the trial arms) and unbalanced missingness.

Methods: First, the outcomes of 22 trials included in a recent Cochrane review on internet-based smoking cessation interventions were reanalyzed using only the complete observations, and after applying PI. Second, in a simulation study outcomes under PI, complete observations analysis, and two types of MI were compared. For this purpose, individual patient data from one of the Cochrane review trials were used. Results of the missing data approaches were compared with reference data without missing observations, upon which balanced and unbalanced missingness scenarios were imposed.

Results: In the reanalysis of 22 trials, relative risks (RR = 1.15 [1.00; 1.33]) after PI were nearly identical to those under complete observations analysis (RR = 1.14 [0.98; 1.32]). In the simulation study, PI was the only approach that led to deviations from the reference data beyond its 95% confidence interval.

Conclusions: Analyses after PI led to pooled results equivalent to complete observations analyses. PI also led to significant deviations from the reference in the simulation studies. PI biases the reported effects of interventions, favoring the condition with the lowest proportion of missingness. Therefore, more sophisticated missing data approaches than PI should be applied.