Evidence factors in observational studies

Some experiments involve more than one random assignment of treatments to units. An analogous situation arises in certain observational studies, although randomization is not used, so each assignment may be biased. If each assignment is suspect, it is natural to ask whether there are separate pieces of information, dependent upon different assumptions, and perhaps whether conclusions about treatment effects are not critically dependent upon one or another suspect assumption. The design of an observational study contains evidence factors if it permits several statistically independent tests of the same null hypothesis about treatment effects, where these tests rely on different assumptions about treatment assignments at several levels of assignment. Two designs and two empirical examples are considered, one example of each design. In the dose-control design, there are matched pairs of a treated subject and an untreated control, and doses of treatment vary between pairs for treated subjects; this yields two evidence factors. In the varied intensity design, there are matched sets with two treated subjects and one or more untreated controls, where the two treated subjects within the same matched set receive different doses of treatment, and in a technically different way, the design yields two evidence factors. Copyright 2010, Oxford University Press.