[StATS]: The paired availability design (May 31, 2005)
In the quest to finish my book on Statistical Evidence, I had to leave some material on the cutting room floor. One of the nicer descriptions was about the paired availability design. Here’s what I had written.
If you have a large group of hospitals, each of which has seen a change over time in the availability of a new therapy, then you can pool the effects in these hospitals in a way that avoids some of the biases in a simpler historical controls study. The trick is that your before group were all patients when availability was low, recognizing that some of these patients will still be lucky enough to get the new therapy. The after group were all patients when availability was high, again recognizing that some of these patients will be unlucky and will be stuck with the old therapy. This dilutes the estimates of effectiveness, but you can adjust directly for this dilution effect. By comparing all patients when availability was low to all patients when availability was high, you can avoid some of the covariate imbalance that occurs due the differing demographic characteristics of those patients who seek out the new therapy versus those that stay with the old therapy.
This pooled analysis is known as a paired availability study (Baker 2001). You have to assume that the population being studied, the concurrent treatments being given, and the evaluation of the outcome is stable over time. You also have to assume that patient preferences do not change over time. This means that no widely publicized media reports change the dynamics of patient demand for the new therapy. Finally, you have to assume that the intervention itself does not become more or less effective when it becomes more readily available.
Example: In a study of breast cancer mortality (Baker 2004), deaths due to breast cancer were compared in six counties in Sweden over a time range when mammography became more readily available. Adjusting for the limited screening done early and the missed opportunities for screening later, the researchers estimated that 9 fewer women per 100,000 died when mammography screening was used (95% CI, 4 to 14 per 100,000).
The other design that I didn’t have room to discuss was a patient preference trial where patients who consent to be randomized are compared to patients who prefer a particular treatment. I’ll try to describe the patient preference trial and give an example in a future weblog entry.
You can find an earlier version of this page on my website.