A peer-reviewer wants more complex statistical modeling - is it reasonable? I've submitted a manuscript to a journal and the associate editor wrote that his readers will want to know more about how the predictor of main interest (occupation) is related to death, and this will require, according to the editor, more complex statistical modeling.
What is it all about?
I examined how occupation relates to death in individuals with coronary heart disease. I've, as conventional, performed survival analysis with the following characteristics:


*

*Cox regression.

*I have multiple observations for each individual and virtually all covariates are updated at each observation! The extended Cox model, in counting process syntax, performed with rms package by Harrell, has been used.

*Continuous predictors are modeled with restricted cubic splines, using 4 knots.

*Large cohort >50,000 observations, >10,000 individuals.

*I have adjusted for all known, and additionally some presumed, risk factors for the outcome of interest. After doing this, occupation remains a strong predictor of death.

*So I arrived in the conclusion that occupation is an independent predictor of death, regardless of known risk factors and presumed risk factors. (I've adjusted for up to 15 covariates).


What more can I do? What "complex statistical modeling" is the editor seeking?
I'd be grateful for some help.
 A: It is a bit hard to say without more context, but it seems that the editor is looking for a mediating factor or the actual cause of death that would explain higher mortality with certain occupations. For example, unless librarians are dying because bookshelves are falling on them, his or her occupation is not the agent responsible for his or her death. I would suggest 3 possibilities, listed in order of difficulty/expertise:


*

*Identify which occupations are the most hazardous, and determine what type of deaths they are more likely to succumb to. Or see what risk factors they tend to have (higher blood pressure?).

*Perform a mediation analysis (there are many methods, and perhaps some that can be implemented with Cox models).

*Perform a causality analysis (i.e. certain Bayesian network analyses, Granger causality tests, or structural equation modelling).   
The first option may suffice, but if this is what the editor has in mind, I would be surprised if he or she called it "more complex statistical modelling." Option 3 is probably much more advanced than the editor was thinking, and would require assistance from an expert in these approaches or at least some serious homework. Option 2 is probably the best, as it would offer some interesting insights and would not be too difficult to incorporate into your analysis.
Or you could respond to the editor and ask for clarification.
