2
$\begingroup$

I have a set of physician data which I have split into 5 somewhat equal quintiles (n's range from 44-53) based on the number of "HIV_patients_diagnosed_annually". I've then created a table showing the mean number of "years_in_practice" for each of the quintiles. Quintiles 1-4 are similar (19-21 years), while quintile 5 somewhat lower (16 years). I want to test if this is statistically significant so I used a grouped t-test and found the only significance is between Q5 and Q2.

  • Is the the correct test to use?
  • If so, are the results interpenetrated as it is not significantly different than all of the other quintiles, just one?

UPDATE (BACKGROUND)

The purpose of the exercise is to create a financial model to project drug sales. Since going door-to-door to generate sales is not a feasible approach, I've decided to segment the market based on number of patients diagnosed so a company would focus on the top one or two tiers first. Therefore, in my analysis I want to be able to highlight anything differences which stand out between the five segments.

Improve this question
$\endgroup$
5
  • 3
    $\begingroup$ This seems at best an indirect and awkward way to approach the question of how two variables are related. You have two variables, so plot them and then think about their relationship. If you show us the data, or at least a graph, we might be able to suggest a model. In contrast, division into quintile bins is arbitrary and loses information. What is that you are comparing, the means of # patients diagnosed for different bins of years in practice? There is also a problem of multiple comparison if you are thinking of 10 possible t tests, or even if you are not. $\endgroup$
    – Nick Cox
    Commented May 18, 2016 at 15:14
  • $\begingroup$ I want to segment the physician universe based on the number of HIV patients diagnose annually and then look to see if there are any specific variables which are significantly different in one segment vs the others. $\endgroup$
    – pheeper
    Commented May 18, 2016 at 15:42
  • $\begingroup$ I don't have good news for you: that sounds artificial, indirect and highly problematic. You'd find it hard to distinguish side-effects of the way you approach the problem from genuine differences. More simply put, why "segment the universe" at all? If there are pre-defined categories, that's fine; otherwise not. $\endgroup$
    – Nick Cox
    Commented May 18, 2016 at 16:16
  • $\begingroup$ I appreciate and understand your feedback. I should have given more background information upfront (see update above). $\endgroup$
    – pheeper
    Commented May 18, 2016 at 16:43
  • $\begingroup$ Thanks for the extra detail. If I were a consultant, I would still advise looking at the relationship between your variables both treated as essentially continuous. I've edited the title to quintile from quartile which seems uncontentious. $\endgroup$
    – Nick Cox
    Commented May 18, 2016 at 16:49

1 Answer 1

Reset to default
0
$\begingroup$

You could also run an ANOVA using your 5 groups. If you use effects coding for your dummy variables, you can see which groups are different from the overall average rather than do so many pairwise comparisons.

That way, you might more reasonably show that the 5th group is different from the overall average whereas the others are not.

Improve this answer
$\endgroup$
3
  • 1
    $\begingroup$ This isn't self-evidently a step in the right direction, as you'd need to work hard to respect even the ordering of the different levels of the different experience groups. Suppose people's height is a control: would you recommend dividing height into bins and then letting the bins be distinct levels of a factor variable? $\endgroup$
    – Nick Cox
    Commented May 18, 2016 at 16:18
  • $\begingroup$ You're right this approach does not take into account the ordinal nature of the variable, and thus is not a good generalizable solution. Given the initial information I suspected that he would find that only group 5 is significantly different from the overall average, a result I'd consider intuitive, reportable and useable in the context of the problem. I'm not classically trained as a statistician, so I'm here to learn - what would you suggest? $\endgroup$
    – ShainaR
    Commented May 22, 2016 at 12:46
  • 1
    $\begingroup$ I've already made broad suggestions in my comments on the question, essentially not to use bins at all. $\endgroup$
    – Nick Cox
    Commented May 22, 2016 at 12:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.