There I am using a specific RCT design described by Zelen et al. where randomization occurs prior to consent and only those subjects who were allocated to the treatment group would be consented, all others would be passively followed. The problem arises when more than 40% of those randomized in the treatment group deny participation. If you do an intention to treat analysis, the intervention effect would be dramatically diluted. This problem has been characterized as a form of non-compliance (when some patients refuse treatment or miss follow up). Other types of analyses such as a per protocol analysis or an as treated analysts are regularly used. However, these types of corrections introduce selection bias into the randomization. Other methods like CACE or propensity score analyses are useful as they correct both the treatment and control groups thus maintaining the randomization unbiased. However, these methods are explained in the literature in a mathematically and statistically dense way, which would be too time consuming for me to tackle.
Does anyone know about a practical and easy way to account for non-compliance in RCT that would yield a statistical significance (i.e. p value)?
Non compliance results in loss of power which has been addressed in by Wittes J 2002
see below
https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=3&ved=0CDcQFjAC&url=http%3A%2F%2Fwww.researchgate.net%2Fprofile%2FJanet_Wittes%2Fpublication%2F11258732_Sample_size_calculations_for_randomized_controlled_trials%2Flinks%2F02bfe5130b66e4e3e1000000.pdf&ei=YCooVaG1H4-V7AbBqIC4BQ&usg=AFQjCNGVVADn5buXZ0MllzWt4uHvcXoI5g&bvm=bv.90491159,d.ZGU
Alejandro,
I hate to tell you there's no easy answer. Propensity methods and CACE methods are the appropriate response to your question, but it's hard to recommend using methods you don't understand--and I agree, they're not simple.
I recommend looking at some of Elizabeth Stuart's recent work on propensity scores for a fairly accessible treatment. Alternatively, try to invite a collaborator/co-author/consultant who can help with this aspect of the analysis in an informed way.
Pat
I agree with Patrick. There is no straightforward answer. As I understand it, the principle is that dropouts (or non-consenters) are an effect of the treatment, i.e., the treatment causes some people to drop out because of toxicity, etc. and thus the treatment is "ineffective" for those people (even though they may not have started the actual treatment). One solution I've heard of is to consent everyone, start the treatment, then have a "run-in" period where those who want to discontinue may do so, then complete the analysis only those who stay in the trial.
And your terminology may be a little confusing: In a regular RCT, everyone is "consented" (with an informed consent form) even if they are in the placebo arm, at least in the U.S. You may be using a clinical population where certain people are selected for an experimental treatment and the regular, non-experimental clinical population is used as a control population (and this may include historical patients who were patients before the experiment is taking place). Those types of studies are certainly done, and can be a grey area (about getting consent) but it is generally allowed to use historical controls if it involves just record review.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling