In a simple experiment we randomize to group, manipulate the independent variable (treatment) and look for the effect on the dependent variable (measuring only after treatment). However, with clinical trials we usually measure both before and after the treatment. Logically this isn't necessary, as if cases are randomized to group, with a sufficiently large sample, there should be no pre-treatment difference between groups anyway. And in fact doing before and after treatment risks effects caused by repeat testing, such as regression to the mean. So why do we do it?
There are two reasons. First, randomization is often not perfect and the two groups are not starting off the same. Second, and more importantly, if you have baseline measurements, you can run repeated measures (aka dependent samples) statistical procedures, which are more powerful. These procedures examine the change scores, which reduce the problems associated with between-subject variability in baseline scores.
I agreed with Gordon on his two points.
It is important to measure baselines. In our modeling practice we are interested in two areas: one is how to estimate to which extent change is related to the baseline; the other is how to estimate change in a variable of interest in relation to other variables, given the baseline of the variable of interest.
In response to Gordon's response to my initial question. I suspected it was due to power issues, but still, the analysis of interest will be the interaction that still has a between subjects factor, doesn't this negate the advantage of the within subjects factor (of pre and post treatment assessments). Also, I agree that randomization is not perfect, but in large studies, say N=1000, the differences between randomized groups would be minimal. I'm not saying you are wrong, and I appreciate your response, I'm still just exploring some of these issues.
Having baseline data in clinical trials is important since the control and experimental groups may not statistically be different but they can be different from an external group. This is useful information in a clinical context for medical providers since it allows them to better translate findings into clinical decision making. For example randomized control trials looking at clinical outcomes of patients with chronic kidney disease done in Australia (very low Aboriginal or Black population) would not have as much external validity when compared to North Carolina, USA (high African American or Black population) as they would compared to a different geographic region, say Minnesota, USA (demographics fairly similar to Australia).
In practicality even though randomization is used in clinical, randomized control trials, there are instances were the generalizability of the results may depend on the baseline characteristics, particularly when the research is handed to the clinician. This is true in large scale, randomized, clinical trials for new pharmaceuticals, since they clinical sites are selected and then patients from that site are then randomized. This is done because it is not feasible to randomly select sites for participation geographically.
Baseline is a significant predictor. I always put Baseline in the model, and when it's not significant, I wonder about the results. I typically analyze Change From Baseline, as it is easily interpreted by all without change in p-values, etc. Baseline also has variability that can be removed.
Is there some references regarding to the baseline should be put in model or not?
Yes, I would be interested in knowing whether there are references on this as well
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation