I have some questions concerning a study on cancer patients, which seems to require the use of linear mixed effects modeling in conjunction with propensity scores. The study description is provided first and the questions are listed at the end.
The study includes 2 sites (A and B), located in different health care jurisdictions. The level of health care and the access to health care resources may be different across sites.
When they enter the study, patients at site A are known to have chosen one and only one of 3 different curative-intent therapies (based on radiation), whereas patients at site B are known to have chosen a single curative-intent therapy (based on surgery). It is assumed that this choice will not change during the study.
In the first year of the study, n = 80 patients from site A and n=80 patients from site B are allocated to a Standard of Care group.
In the second year of the study, n = 80 patients from site A and n = 80 patients from site B are assigned to an Intervention group. (The Intervention is "developed" during the first year and utilized during the second year.)
The patients in the Intervention group will have access to a specialized set of online educational resources regarding their type of cancer, which will hopefully enable them to better understand and manage the side effects of their chosen curative-intent therapy.
The patients in the Standard of Care group will not have access to these specialized resources, but will have access to resources usually available to patients with this particular type of cancer.
At pre-defined time points (e.g., baseline, 26 weeks, 52 weeks), patients will be assessed on their self-efficacy via a total self-efficacy score.
A first analysis will concern the patients at site A. For these patients, we need to determine whether the treatment effect on the total self-efficacy score depends on time. If it does, we need to evaluate the treatment effect at each time point.
For this first analysis, should propensity scores be used as a means to make the Intervention and Treatment groups comparable? If yes, is it reasonable to compute propensity scores only at baseline using a variety of baseline covariates (given that treatment assignment doesn't change over time)? Furthermore, is it reasonable to include the type of curative-intent therapy in the propensity score model and also in the outcome model? (The outcome model would use the total self-efficacy score as the outcome variable. In addition to including time, treatment and a time by treatment interaction term, the outcome model would also include the baseline propensity score and the type of curative intent therapy as covariates. The outcome model would be formulated as a linear mixed effects model with a random intercept.)
A second analysis will try to combine patients at the two sites, A and B. Does it make sense to formulate a single statistical model for patients at the two sites? If yes, should the propensity score model include the type of curative intent therapy as one of the covariates that may possibly explain assignment to Intervention? Should the outcome model include time, treatment and type of curative-intent therapy and all of their interactions? In addition, should the outcome model also include the baseline propensity score and the type of curative intent therapy as covariates? (It is not clear to me whether the available data would enable meaningful estimation of such a complex model. For practical reasons, the outcome model could be a linear mixed effects model with just a random intercept.)
Thanks in advance for any feedback you are able to provide.
------------------------------
Isabella Ghement
Ghement Statistical Consulting Company Ltd.
------------------------------