Sensitivity and specificity using roctab

[Marijn Verwijs]

Hi everybody,

I am using Stata to calculate the sensitivity and specificity of a diagnostic test (Amsel score) compared to the golden standard test Nugent score.

tab BVbyAmsel highnugent, chi2
roctab BVbyAmsel highnugent, detail

The resulting table is

           |      highnugent
 BVbyAmsel |         0          1 |     Total
-----------+----------------------+----------
         0 |       192         66 |       258 
         1 |         6        137 |       143 
-----------+----------------------+----------
     Total |       198        203 |       401

If I use roctab, the results given are:

                                            Correctly
Cutpoint      Sensitivity   Specificity   Classified          LR+          LR-
------------------------------------------------------------------------------
( >= 0 )          100.00%         0.00%       35.66%       1.0000     
( >= 1 )           95.80%        74.42%       82.04%       3.7451       0.0564
( >  1 )            0.00%       100.00%       64.34%                    1.0000

So, a sensitivity of 95.80% and a specificity of 74.42%. But if I calculate by hand, I get the following results:

True positive: 137
False positive: 6
True negative: 192
False negative: 66

Sensitivity: TP / (TP + FN) = 137 / (137 + 66) = 67.49%
Specificity: TN / (TN + FP) = 192 / (192 + 6) = 96.97%
Positive predictive value: TP / (TP + FP) = 137 / (137 + 6) = 95.80%
Negative predictive value: TN / (TN + FN) = 192 / (192 + 66) = 74.42%

I have the impression I must have made a mistake somewhere, but I cannot find out how and why. Thank you for your help! I am using Stata 13.0.

Kindly, Marijn

[Clyde Schechter]

The -roctab- command requires the reference variable ("golden standard") be listed first. You have it reversed.

By the way, you don't have to hand-calculate the sensitivity and specificity that way. When you run -tabulate- specify the -col- option. For positive and negative predictive values, specify the -row- option.

[Marijn Verwijs]

Thank you so much! That is very useful.

[Inaamul Haq]

The -roctab- command requires the reference variable ("golden standard") be listed first. You have it reversed.

By the way, you don't have to hand-calculate the sensitivity and specificity that way. When you run -tabulate- specify the -col- option. For positive and negative predictive values, specify the -row- option.

Can you please tell us how we can get the 95% CI for Positive and negative predictive values and the LR in STATA?

[Clyde Schechter]

So, first you have to compute the total number of true positives, false positives, true negatives, and false negatives. Then you need two denominators: ppv_denominator = true_positives + false_positives, and npv_denominator = true_negatives + false_negatives. Then for the predictive values and confidence intervals you run

cii proportions ppv_denominator true_positives
cii proportions npv_denominator true_negatives

Read -help cii-. There are a number of different approaches to calculating confidence intervals for proportions, and you can chose the one you prefer by specifying it as an option to the -cii- command.

The positive likelihood ratio is just sensitivity/(1-specificity). The likelihood ratio negative is just specificity/(1-sensitivity).

The details of how you would calculate these various totals, etc. depends on the data you are starting with, and as you don't show any example data I can't provide them.