I am working on a SysRev about an infection prevention strategy for intubated patients. In most of the included RCT's, the randomization takes place after the intubation. Therefore, some of the patients actually receive the control intervention (standardcare) for up to 48hrs. I am adjusting this issue on the "other potential sources of bias"-section. But I don't know if I should rate those studies as "high risk of bias". Because it seems obvious to me that if there is any bias, it would be that the effect of the intervention is underestimated. So it would not seem fair to conduct e.g. a sensivity analyses omitting those studies because they score high on the RoB assessment if, in fact, the bias is towards the null.
Can anybody help me with that? Thanks!
It seems to me that bias is bias. If bias toward the null then the intervention will not get the credit (for good or bad) results. Do you have the raw data? I would exclude those subjects or calculate the statistics with three groups instead of two. IMO.
Thanks! I don't have the raw data but I can include it in the RoB assessment!
Whether there is bias depends on the exact question you are trying to answer.
If your question is something like: 'How large is the effect of the prevention strategy?' then including results of trials where some participants did not receive the prevention as you want to measure it, leads to a bias toward 'no effect'.
If your question is more like; "Does this prevention strategy work?" then including those trials may make your answer more accurate.
Consider an extreme example: You have 1 small trial without control treatment in the intervention arm which shows no significant effect and you have 10 large trials with some control treatment in the intervention arm, each of which show a significant effect. I would say that in this example there is obviously an effect and it would not be the best answer to your question to remove the 10 large trials from your analysis and conclude that there is no significant effect.
Note that it is possible that the amount of control treatment in the intervention arm is so large that adding those trials to the analysis predominantly increases noise.
If you are interested in a best estimate of effect size, you may want to consider correcting for the bias you described, see e.g.: https://www.ncbi.nlm.nih.gov/pubmed/20196067 However, that would go well beyond the Cochrane practice.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis