if ~2% of my data is missing on the outcome (continuous scale), out of a total of 200, two in control and three in intervention group, do I need to impute? Or can I make a case that with such small missing data per-protocol (PP) approximates Intention-to-treat (ITT).  If I do impute, am I better off doing multiple imputation (MI) or can I do last observation carried forward (LOCF) 

The purpose of the analysis is to basically compare the improvement in a score from baseline to time point 5 and there are 3 other time points between baseline and time point 5, where scores were taken. So, I can carry the observation at 4th time point forward instead of MI for the 5 missing data points (LOCF). 

So which of the following is preferable: 1. Multiple Imputation 2. LOCF 3. Make a case PP was as good as ITT because the missing data was minimal and there was nothing extra ordinary about the 3 patients that had missing data at time point 5. 

  • 1
    $\begingroup$ To add to the case-deletion strategy, you could do a sensitivity analysis where you set their values to the worst case scenario and see if your results change. If they don't, you can report this in a footnote and it will bolster your claim that case-deletion is valid. $\endgroup$
    – Noah
    Aug 2, 2018 at 17:47
  • $\begingroup$ Hi, welcome. 1) to begin with: what are PP, ITT, MI and LOCF? (please keep the general user in mind). 2) "The study basically compares..." which study? $\endgroup$
    – Jim
    Aug 2, 2018 at 18:01
  • $\begingroup$ I recommend to perform a complete case analysis according to ITT as (quasi) main analysis, and then LOCF and MI analyses as sensitivity ones. If you do have cross-overs then perform according to PP complete case, LOCF and MI analyses. Hope you get similar results, and then you can simply report in detail the main complete case ITT analysis leaving the rest for the busy reader. $\endgroup$ Aug 3, 2018 at 10:22

1 Answer 1


Per protocol/Intent-to-treat has no bearing on how to handle missing-data.

A PP analysis may disagree with ITT when the patient discontinues or does not comply with the investigational treatment: the key difference is how treatment is coded and when the patient is considered at risk. PP is never "as good" as ITT, it answers a different question.

For instance, suppose a drug is administered on 4 cycles and pre/post measurements are taken at each cycle. However subject 123 responds poorly to treatment after cycle 1, so they discontinue. A good protocol PP analysis will recode their treatment to SOC or no-treatment and continue to administer measurements during the remaining cycles. However, many trialists neglect this point and fail to encourage the patient to attend remaining visits, so these data are "missing". We cannot predict how they would have responded to future cycles because they are now contraindicated to this treatment, we have no data to inform the basis of such predictions, gathering it would be unethical.

Using imputation or LOCF to "project" what the patient outcomes would have been had they kept on the study drug requires a rather strong and unverifiable set of assumptions. A better approach is to use a sound data analysis strategy for the imbalanaced classes that are likely to arise due to attrition--(discontinuation of/noncompliance with the study drug or loss to follow-up, study withdrawal). Mixed models are shown to perform much better in this regard. Dialysis and chemotherapy are two examples of treatments applied in cycles as in the above example.

For a PP analysis, I would suggest using all available cases, and performing a worst/best possible imputation to provide bounds for the estimate. Worst case carried forward and Best case carried forward are discussed for doing this (Google this to find relevant citations).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.