In both experimental and quasi-experimental studies, sometimes it becomes known to the researcher that one or more participants have been compromised with regard to the study goals. This "contamination" can occur when a member of the control/comparison group is exposed to factors (eg, receives treatment) that are similar to the experimental group. It can also occur when members of the experimental group unintentionally receive a different or additional type of treatment than the original study design intends. Does removal or censoring of a limited number of participants impact the reliability of the study results?
Removal of "protocol deviation" subjects from analysis reduces the statistical power of the research study for the control or comparison group but I don't know why this would compromise the internal validity of the study. If such "protocol deviation" events become more than simple isolated occurrences (e.g., a cluster at a single treatment center) then it does raise questions regarding the capability of the investigator(s) at that institution to participate and comply with the requirements of the research study and the potential of other (hidden) "protocol deviations" that might alter the results of the study findings.
Thanks for your response, James. The concern with removing participants from the study who have deviated from the protocol is that the groups, previously matched or randomized, may now be no longer equivalent. For example, if 10 participants in the experimental group were removed for non-compliance with protocol, would it be necessary to also remove 10 similar control group subjects in order to preserve internal validity? I would think this action may be necessary with a relatively small matched group (eg, 100 in each group), but not so important in a total study group of 1000 or larger?
Yes, agree, especially with "intent-to-treat" principle. When reporting the results, it seems best to report results of both analyses: one analysis with no exclusions ("intent-to-treat") and a second analysis (that excludes protocol deviationsviolations). The first analysis is based on a purist approach ("intent-to-treat") but to gain insight into the scientific question of whether the actual intervention improves the outcome or not, it still seems the second analysis (that excludes protocol deviations/violations) would be needed. Both results would be useful.
Levente, thanks for providing the link to the Statistical Principles for Clinical Trials. This is a very helpful high-level guide for designing and analyzing trials.
Is there a means of quantifying the impact, of removal of randomized participants, from either control or experimental groups? Certainly, the location of the participant removed in the sample distribution would be a factor (location toward the middle of distribution would have much less impact that one toward either end of distribution), but would having an a priori procedure for removing participants help lessen the impact on internal validity? For example, would a rule that requires removal of a "matched" participant from the corresponding group (control or experimental) be a sound statistical procedure?
Hi Mr Krohn
It's not at all valid to try and reachieve balance by further excluding 'matched' participants. The power issue is bad. Worse is that you just can't be certain that protocol violations are independent of randomization leading to biases in unknowable directions. Worser, your putative attempt to rebalance is likely to introduce a whole new set of biases.
Stick with Intention-to-treat approach: once randomized always analyzed.
As a secondary analysis you can do a per-protocol analysis, excluding just your violators.
Finally, if you have too many protocol violations, you could a tertiary analysis comparing characteristics of the violators (prognostic, diagnostic and therapeutic) with those who didn't violate protocol to try and convince readers that violations weren't systematically related to the treatment (i.e. they had to have an extra treatment B because intervention treatment A had side effect. Or there was unblinding or unmasking and folks crossed over deliberately).
I'm guessing you're doing an internal membership trial, say of wellness programs, with real small numbers as a pilot and you're dealing with some intractable protocol issues. I'd recommend starting over - you won't be able to convince a clinical or management audience otherwise.
Good luck!
- Badges
- Science topic
- Similar topics
- Medicine
- Disease
- Infectious Diseases