Using Regression for Causal Analysis on Non-Experimental Data I have a small hypothetical dataset(D1) with X unique subjects that contains data on the smoking habits of the subjects, their age, gender, etc. and whether they have cancer or not. Let's assume all subjects did not have cancer.
In another dataset(D2) I have data on 0.5*X subjects from the first study, with the same data recorded as the first time but 5 years apart. Let's assume some of the subjects here have cancer now.
My objective is to perform a causal analysis to uncover what caused the cancer after 5 years. 
My current approach has involved anonymizing the dataset(removing the subject IDs), merging both datasets to include only those who were present in both surveys, creating a dummy variable for the year in which they were surveyed(coded as 0 for the first contact and 1 if they were involved in the study 5 years later).
I subsequently run a multiple regression analysis based on my hypothesis about the main causes of cancer from available variables in the dataset. 
My questions are:


*

*Is this an approach than can be published? Let's also assume I had no input in the analysis and I only had access to just the results obtained from the method described exactly as above. 

*How can I use propensity score matching in this case? 

*From the regression results I observe that 3 main variables are significant in predicting cancer. Would these variables be my causal factors, and is that a sensible interpretation?

 A: 1) I imagine you could publish this if the findings were novel or surprising. Many studies use secondary data sets. If you got the data set from a specific individual or group, they might have requests for how you acknowledge their data collection efforts (e.g., by adding them as co-authors).
2) You cannot. Propensity score matching can tell you the effect of a given cause, but not the causes of observed variability. You are asking a different question than what propensity scores are designed to answer.
3) It's very hard to claim that these are the causal factors. There may be other confounding variables that are the true causes and are just correlated with your predictors. There may also be mediators which provide a better causal explanation. In general, it's challenging to make a valid claim about what the causes of observed variability. You can only say that your factors are associated with variability in cancer survival unless your variable list contains a reasonably exhaustive set of variables and you don't want to include mediating causes in your analysis.
