grey-zone model with STATA

[cyril cosse]

Dear all,

Does anyone know the command to find two cutoffs for a marker by fitting a grey-zone model that will define high, intermediate, and low risk groups when the outcome is binary ?

It is equivalent to "greyzone.funcs" in R software.

Thank you for your help

Best regards

Cyril

[Bruce Weaver]

For the benefit of others, the same basic question was posted to ResearchGate a few days ago.

Cyril, here's what you wrote in one of those RG posts:

I want to propose an algorithm helping to distinguish patients with or without ischemia. I previously used a marker (Procalcitonin, PCT) to do so. My aim is to calculate PCT thresholds to obtain 3 populations :

- the first for which PCT<threshold 1 : so the diagnosis of ischemia is negative

- the second for which threshold 1 < PCT < threshold 2 : the diagnosis is uncertain and we should add another marker

- the third for which PCT > threshold 2 : so the diagnosis is positive

I think what you're saying is that you want one cut-point that yields very good specificity, and another that yields very good sensitivity. But you've not said that exactly. And if that is what you want to do, you've not said what the desired levels of sensitivity and specificity are.

HTH.

[cyril cosse]

Dear Bruce

Thank you for your answer.

You're right : I want one cut-point with very good specificity (>90%), and another with very good sensitivity (>90%)

Best regards

[Bruce Weaver]

Something like this, then?

clear *
sysuse auto
* Treat foreign as 'disease' and mpg as quantitative 'marker'
roctab foreign mpg, graph table detail

generate byte cut1 = mpg >= 17 // Ensure sens >= 90%
generate byte cut2 = mpg >= 28 // Ensure spec >= 90%
tabulate cut1 foreign, col
tabulate cut2 foreign, col

generate byte testval = 1 if !cut1
replace testval = 2 if cut1 & !cut2
replace testval = 3 if cut2
label define testlabs 1 "Domestic" 2 "Uncertain" 3 "Foreign"
label values testval testlabs
tabulate testval foreign, row col

Output from the final -tabulate- command:

. tabulate testval foreign, row col

+-------------------+
| Key               |
|-------------------|
|     frequency     |
|  row percentage   |
| column percentage |
+-------------------+

           |       Car type
   testval |  Domestic    Foreign |     Total
-----------+----------------------+----------
  Domestic |        13          1 |        14
           |     92.86       7.14 |    100.00
           |     25.00       4.55 |     18.92
-----------+----------------------+----------
 Uncertain |        34         15 |        49
           |     69.39      30.61 |    100.00
           |     65.38      68.18 |     66.22
-----------+----------------------+----------
   Foreign |         5          6 |        11
           |     45.45      54.55 |    100.00
           |      9.62      27.27 |     14.86
-----------+----------------------+----------
     Total |        52         22 |        74
           |     70.27      29.73 |    100.00
           |    100.00     100.00 |    100.00

Using these cut-points, 4.55% of foreign cars would be misidentified as domestic, and 9.62% of domestic cars would be misidentified as foreign.

As you can see above, I hard-coded the cut-off values (17 and 28) when generating cut1 and cut2. If this is a one-time task, hard-coding the values may suffice. If it is not a one-time task, there is likely a way to grab those values from post-estimation info. Or maybe someone knows of a program that does all of this more efficiently.

HTH.

[cyril cosse]

Thank you Bruce for your help

Best regards

Cyril