how do i know what the cut off was following -xtile-

[Martin Imelda Borg]

hello i splint my continusous variable into quintiles using:

xtile quintilescore=score [pw=weight], n(5)

///When I enter
tab quintilescore

I get the following table


However, is there a way I can see what the ranges at which for eg the values within 1 consist of. ?

[Clyde Schechter]

tabstat score, by(quintilescore) statistics(min max)

[Nick Cox]

Not the question, but the tabulation shows how uneven the quintile groups are!

[Leonardo Guizzetti]

You can use the related command to see the quantiles that were used, but matching these up with Clyde's suggested tabulation will help you see how they are used.

_pctile score, nq(5) // for quintiles -- use the same options as you did for -xtile-
return list

Or you can read the manual to see how -xtile- defines them under the "Methods and formulas" subsection.

[Nick Cox]

As in https://www.statalist.org/forums/for...rtile-question I always recommend a plot so that you can jjudge for yourself how well this kind of binning works.

[Martin Imelda Borg]

Nick Cox from previous stat posts on this forum it seems you aren’t keen on the quintiles.

with regards to the uneven distribution, I expect this as it’s a patient score , some people are happy after a treatment and others aren’t

it’s also weighted following iptw

[Nick Cox]

On your behalf I am always curious: does quantile binning really help your objectives?

[Martin Imelda Borg]

Fair question. Here’s my response and I’m keen to hear your thoughts following this

yes there’s so much literature against categorisation

however in the real world eg clinician world we don’t work without categorisation. The people with really bad low scores are certainly different to those with the good scores. So we need a cut off level of whats excellent - fair - good- poor. This helps determine if there is a difference is terms of outcomes for eg complications in those in the excellent category vs poor category

now, your question is prob what level should one pick ?
great q … the ans… we could pick either one which categorisation has already been done, published in a journal but perhaps not as reputable as the lancet … one is depending on that authors research quality OR

opt for a unanimous unbiased decision and just use quintiles…

[Clyde Schechter]

Well, yes, in clinical practice, one has to make yes-no decisions. So, even though the relationship, say, between blood pressure and stroke risk is understood to be continuous, after determining a patient's blood pressure one has to ultimately either put them on blood pressure lowering medications or not. To facilitate this process, a spurious "disease", hypertension, has been invented based on an arbitrary cutoff in blood pressure. In a world with rational decision making, however, this decision making would be based on an actual decision analysis that takes into account the incidence rates of all of the sequelae of all possible levels of blood pressure in patients of given age and sex, as well as the way those change in response to medication, and also any adverse effects associated with taking medication. As a practical matter, that is more complicated than anybody is really able to do. So we use cutoffs selected by expert opinion instead to guide treatment decisions.

That said and acknowledged, it is highly unlikely that using quintiles, especially in a distribution that does not actually lend itself to division into 5 nearly equal segments, will produce cutoffs that make clinical sense. I think that the use of quintiles stems from a common practice a few decades back in epidemiology, particularly contrasting the 1st and 5th quintiles of the distribution. This had the advantage of providing a clear separation between the "low" and "high" that was more likely to show statistically significant relationships to outcomes than the "grey area" in between. In my opinion, though, it's one of those time honored practices whose time has passed. It dates from an era where most statistical calculation was done on a hand calculator because computers were largely inaccessible and good software for more sophisticated analytic approaches had not been developed.

Except in situations where there really is some discontinuous jump in the outcome at some point in the distribution of a continuous predictor, present-day studies should stick with the continuous variables in most cases. Yes, at some point a decision-making threshold will have to be arrived at--but that should be a separate later step, one that takes the much better developed decision-analysis infrastructure available in the present. Those decision analyses will still be somewhat incomplete because the real world is too messy, but we can do a much better job now than in the past.

Secondary analyses using quintiles, to be comparable to previous publications, might be a reasonable thing to include in the online-only appendix of an article. But they shouldn't be in the spotlight anymore.