Keywords: Causal inference, smoking cessation, Internet interventions, sensitivity analysis
Evaluating the effects on smoking abstinence of adherence to tobacco dependence interventions delivered via Internet is crucial to assessing their public health impact. Estimates of these effects are subject to selection bias in the use of intervention components, namely skills training, community use and medication use. Inverse Probability of Treatment Weighing can adjust for differences in measured participant characteristics under the assumption of no unmeasured confounders. Sensitivity analyses were performed to assess the influence of hidden bias on estimated intervention effects on 3297 website users. The abstinence benefit when community features were utilized was very robust to the omission of any confounders as strong as the ones observed. Comparisons with medication use were found to be sensitive to the omission of confounders as important as race, nicotine dependence level and past use of cessation medication at baseline. Thus, medication effect estimates can potentially be sensitive to hidden bias due to such unmeasured determinants of medication use. The statistical approach serves as a model for evaluating the sensitivity of effect estimates in Internet interventions.