Missing data are a ubiquitous aspect of behavioral intervention research. There are a number of reasons why data could be missing, some of which are controllable by the researcher and some that are not, and each prompting a different inference. For instance, researchers who study rural populations often employ community-based participatory research techniques to get a buy-in from the community they are studying. Stakeholders and community advocates are mobilized to collect data and to ensure that participants regularly attend all of the intervention sessions and also come back for follow-up visits. However, often even after much effort, researchers can collect follow-up data on only a subset of the initial study sample. Although list-wise deletion continues to be offered in popular software programs, in recent years there is growing recognition that failure to address issues of missing data can lead to biased parameter estimates and incorrect standard errors. Researchers have now been relying on a multitude of techniques for dealing with missing or incomplete data that are currently available to behavioral interventionists and run the full gamut of sophistication and effectiveness, with full-information maximum likelihood and multiple imputation methods deemed to be the most effective strategies for analysis with incomplete data (Little & Rubin, 2014). However, there is no consensus as to which methods should be used in managing missing data, and the best policy is to try to minimize this occurrence as much as possible.
Another issue specific to all randomized trials is the issue of compliance. Intention to treat (ITT) utilizes the data of every participant who was randomly assigned to condition, essentially ignoring what treatment the participant actually received. This approach provides a conservative estimate of treatment effect, but eliminates bias from protocol deviations when persons drop out, for example, because of their lack of response to treatment (Gupta, 2011). Complier-average causal effect (CACE) estimation is an alternative approach that provides robust estimates of a treatment effect among compliant patients (Little & Rubin, 2000) and is becoming increasingly acceptable as an accompaniment to ITT to more fully understand treatment effects.
CACE analysis builds upon Rubin’s causal modeling framework to yield causal estimates of the effects of intervention for individuals who comply with treatment (Little & Yau, 1998). The main challenge in CACE modeling is identifying the proportion of individuals who fall under the four compliance subgroups in the study population, namely, compliers, always-takers, never-takers, and defiers. These compliance subgroups are defined on the basis of how participants would comply with an assigned treatment under random assignment. Compilers are those who will use the treatment if they are assigned to the intervention arm of the study, but not if they are assigned to the control arm of the study. Always-takers will use the treatment irrespective of their intervention assignment. Never-takers will never use the treatment even if it is provided to them in the intervention arm, and defiers will do the opposite of their assigned treatment. Once researchers are able to account for the sample proportions for each of the four subgroups and verify necessary assumptions, they could determine an unbiased estimate of the difference in outcomes for compliers in the intervention group with those in the control group who would have complied with treatment given the opportunity to do so.