Hi Riccardo,

An underlying principle of systematic reviews (SRs) is "garbage in, garbage out", hence, assessing the quality of all included studies is very important. Risk of bias assessment (ROB) should be done for all eligible studies identified for inclusion in your SR.

Assessing ROB for the domain of "incomplete outcome data" is tackled across the different versions of the Cochrane Handbook of Systematic Reviews for Interventions. Chapter 8.13 of Version 5.1 ( https://handbook-5-1.cochrane.org/chapter_8/8_13_incomplete_outcome_data.htm ) gives a good discussion of how to assess this domain.

Side note: "adherence to intervention" is an intermediate outcome. In most cases, outcomes that are important to end users of SRs are hard outcomes. For example, in an anti-hypertensive trial, adherence to medication is an intermediate outcome, whereas prevention of cardiovascular disease is a hard outcome.

Best Regards,

Elenore

Should you exclude studies with ITT only?

- That depends on what type of study you are looking at. ITT is more suitable for superiority studies than noninferiority studies. I see no reason why superiority studies whose only ITT was performed should be excluded.

How should you manage the missing outcome domain?

- You can see if the assumption that the missing outcomes are MCAR or MAR is true or not. Reasons for the missing data must be provided.

Hope this helps.

I think you have posed an interesting question. I think it can only be sensibly answered by fully understanding why you want to determine the effect of adhering to an intervention. ITT is generally preferred for assessing 'real world' effectiveness in an unbiased fashion.

However, if you want to assess a fully adherent to treatment effect (or equivalent) you could still regard an intent to treat analysis as providing a potentially biased but still informative estimate. The likely extent and direction of bias can also be assessed - based on treatment dropouts and data provided on compliance. If such data is not available the risk of bias remains unclear. As ITT studies are likely to give a downward bias (assuming an effective treatment) unless non-compliant participants take more effective treatment, your estimate of effect from studies that are unbiased in other respects can provide you with a lower bound point estimate of the 'on-treatment effect'.

Under all circumstances that I can imagine this is better than selecting a small number of studies that report on treatment effects. This sample is likely to be biased in other ways but, in any case, the resulting estimates will be more error-prone because of the smaller samples (participants in studies and studies across the review).

I hope that helps.

Peter