Xtile - method for choosing cutpoints

[Roberto Vidri]

I'm revising a paper for a peer-reviewed medical journal. Part of our analysis included creating "quartiles" out of two continuous variable (distance and volume). We used the function "xtile" and allowed it to create these automatically

The reviewer is requesting clarification on how these cut points were chosen, and if they are the ideal outpoints for the data.

I have not been able to find the exact method by which xtile selects the cut points. My understanding is it chooses the optimal cut points automatically, based on the distribution of the data - I may be completely wrong.

Is there a way in which I can convey this to the reviewer. And/or is there a more robust methods to determine these cut points?

Thanks!

[Clyde Schechter]

The exact method is set out in the PDF documentation that comes installed with your Stata. Run -help xtile-. In the Viewer window that opens, click on the blue link to the PDF manual entry. When that opens, click on "Methods and Formulas." All the details are there. You can copy and paste it into a response to the reviewer if you prefer not to explain it yourself. I will also add that this definition of the cutpoints is standard, and except for different ways of breaking ties, I think all statistical software does it this way. Are you sure your reviewer isn't questioning why you are using quartiles in the first place?

[Nick Cox]

It's hard to say anything about quantile binning that has not been said many times here. https://www.stata-journal.com/articl...article=dm0095 offers a fairly concise but also moderately detailed discussion of pitfalls, including references to advice against ever doing this any way.

[Roberto Vidri]

The exact method is set out in the PDF documentation that comes installed with your Stata. Run -help xtile-. In the Viewer window that opens, click on the blue link to the PDF manual entry. When that opens, click on "Methods and Formulas." All the details are there. You can copy and paste it into a response to the reviewer if you prefer not to explain it yourself. I will also add that this definition of the cutpoints is standard, and except for different ways of breaking ties, I think all statistical software does it this way. Are you sure your reviewer isn't questioning why you are using quartiles in the first place?

Dr. Schechter,

Thank you (as always) for your response.

I have gone back to reading the reviewers comments, quoted below:
"For travel distance, it appears that patients were grouped based on quartile (not stated in the methods). What statistical test was performed to demonstrate that these were indeed the best cut-offs for travel distance? Any sensitivity analysis to determine if other distance groups performed better? Just b/c half of that patients travelled less than 18 miles does not mean that is the optimal cut-off for outcomes."

These are my thoughts:
1. The grouping was stated in the methods, and results
2. I'm wondering if he wants to know if there is meaningful literature regarding adequate distance "cutoffs" for patients similar to our analysis?
3. We did perform a sensitivity analysis comparing the lowest to the highest quartile groups.
4. I don't think I know of any statistical methods to demonstrate these are the best cutoffs.

Would appreciate your thoughts.

Best,
Roberto

[Roberto Vidri]

It's hard to say anything about quantile binning that has not been said many times here. https://www.stata-journal.com/articl...article=dm0095 offers a fairly concise but also moderately detailed discussion of pitfalls, including references to advice against ever doing this any way.

Dr. Cox,

Thanks for sharing that article - it is very useful!

Roberto

[Clyde Schechter]

Well, at this point, I have to ask you to explain what kind of analyses you carried out with these quartiles. Without knowing that, it's not possible to say what kinds of sensitivity analyses might have been appropriate.

And I will also express my concurrence with Nick Cox' point that usually the best approach to continuous variables is not to group them into categories at all but to use them as they are (or perhaps with some transformation). But, again, without knowing the context, I can't say more.

[Roberto Vidri]

Thanks for you help and explanation, Dr. Schechter.