How does missing data affect analysis of clinical trials? I was helping a friend with some problems for a clinical trials administrations course they are taking, most of it was simple statistics the last question was a little more tricky. We got the answer from google but it made me think a bit the quesiton was to do with patients dropping out of a trial with time without achieveing the requisite follow-up period. I was wondering about a simper case.
Consider a simple trial where I compare two different interventions. Patients are assigned to one of two interventions randomly, 100 into each arm. My end point is a single continuous variable. If, for whatever reason, only 94 patients in one arm and 97 in the other are suitable for analysis does this affect my analysis beyond the obvious increase in random uncertainty?
 A: It depends on why they are missing. Missing data mechanisms have been divided into three sorts of missingness:
Missing completely at random: The reason for missingness has nothing to do with anything, it is random. 
Missing at random: The reason for missingness may have something to do with the dependent variable, but that relationship can be entirely accounted for with data that you  do have.
Not missing at random: Neither of the above is true.
If the data are MCAR then you  just lose power; in the other two cases, you will need to do more, and, in the NMAR case, nothing you do can be shown to definitely result in the right numbers.
A: Just to add to Peter Flom's comment, if you're interested in reading more, Alan Acock wrote an excellent primer on missing data which is available outside the paywall.
A: I like Peter Flom's answer but i could add something about how missing data are handled in the drug industry where there is a fondness for convention. If the patients have no post-baseline data at all (your example suggests only one timepoint for data collection on the primary outcome) then nothing can be done, aside from maybe some slight-of-hand using multiple imputation methods (see eg john graham's book). The ICHE9 guideline defines the 'full analysis set' which would exclude patients who have no post-baseline data. If they have some data, last-observatin-carried-forward (locf) was typical in the industry for many years, even though everyone knew it could produce biased results. Eg in alzhiemer's where carrying the last value forward would be anti-conservative (because patients would drop out before they would have deteriorated further, thus making a toxic drug look good). These days mixed modelling is favoured. In any case, the regulatory authority (eg FDA) would likely request 'sensitivity analyses' to explore the effect of missing data. There are more difficult scenarios eg how do you handle patients who switch treatment (not uncommon in oncology trials). I think they are treated as failures at the time of switching. Probably every book ever written on missing data ends with this advice: the best way to handle missing data is to minimise it, eg thorough followup of patients. As you can imagine, missing data are a greater issue in trials run by academics where they do not have the same resources big pharma have to monitor data collection. Etc
