I am performing a propensity score analysis with weights on different medications. I'm obtaining a very low misclassification rate based on the fact that these medicines were introduced at different times and, the data seem to say, were most often prescribed right after their introduction to the market. By including when the patient first got their prescription I'm able to predict with a fairly high degree on certainty what drug she got.
The problem is that other variables pivotal to the clinical outcomes, such as measures of cardiovascular or overall health, are not as predictive and so remain unbalanced after propensity score weighting.
Since it seems to me that the outcomes (stroke and adverse bleeding) are not dependent upon when the drug was prescribed am I right in thinking that I can exclude the date of the first script as a predictor, settle for a much higher misclassification rate and conduct a more rigorous propensity score analysis with only those confounders included that I think will have an effect on the clinical outcomes?
These links may help you
https://pdfs.semanticscholar.org/a821/c9f67ba35f89dbdc63610087ce92b8bf4c16.pdf
https://catalyst.harvard.edu/docs/biostatsseminar/RobertGlynn01-09-2013.pdf
Article Radical Cystectomy versus Bladder-Preserving Therapy for Mus...
Article Misclassification in Administrative Claims Data: Quantifying...
Article An application of propensity score weighting to quantify the...
Article Propensity Score Weighting for Causal Inference with Clustered Data
Thank you Mohamed! You shared a lot of good material.
I think I resolved my own question. Propensity Score analysis relies on the assumption that all variables which had an effect on the treatment group and the outcome are part of the propensity determination. Any variable which simply had an effect on treatment selection, but not on the outcome, can be excluded from the analysis so that balance is achieved on those variables which did have an effect on the outcome measures.
Achieving misclassification rates of 30% or less is attractive, but if it is done by adding in variables with no effect on the outcome it can be detrimental to the analysis. Essentially, these variables are the same as random variation so far as propensity score analysis is concerned.
Fellow researchers, please feel free to comment if anything I said above is wrong.
- Badges
- Science topic
- Similar topics
- Chemistry
- Pharmacology
- Pharmaceuticals
- Drugs