Stata 13 - Errors simulating confidence intervals when indexing on cut points

Question

I'm finalizing a short research note and having difficulty with my code in Stata 13.

I am estimating the effect of several variables on an ordinal dependent variable. I've run the initial estimation using ordered probit and generated predicted probabilities. For the purpose of this research note, I am required to generate confidence intervals for the predicted probabilities by hand (using simulation). Unfortunately, margins is not in my immediate future.

I've been trying to run the simulation but whenever I include the index function [i']` for each cut point of the dependent variable, I receive the error:

equation cut1][1 not found

The code I am using is:

set seed 23
mat b = e(b)
mat V = e(V)
drawnorm b_term b_pres b_acc cut1 cut2 cut3 cut4 cut5 cut6, mean(b) cov(V) n(1000) clear

gen p_1a_mean = .
forvalues i = 1/1000 {
    gen p_1a_`i' = normal(_b[/cut1][`i'] - (term*b_term[`i'] + pres*b_pres[`i'] + acc*b_acc[`i']))
    summarize p_1a_`i', meanonly
    replace p_1a_mean = r(mean) in `i'
}

drop p_1a_1-p_1a_1000

I find that when I do not index the cutpoint, I am able to run the code perfectly. This is strange, because I was let to believe that failing to index the cutpoint will bias the results.

Does anyone have any insight as to where I'm going wrong?

This is an international forum, so goodness knows how many people what UMD is. It doesn't matter for the question, however. — Nick Cox
Hey Nick, absolutely. I just wanted to give credit where credit was due, so that it didn't look like I came up with this on my own! — Yasha
A side-comment is that generating 1000 new variables here is quite unnecessary. You just need to initialise a single p_la outside the loop and replace it with new simulation results inside the loop. — Nick Cox
Hi Nick - that's pretty interesting; I didn't know that you could do that. I was given this code and told to use it, so I didn't want to deviate from the assignment, but in future attempts, I will try your suggestion. Thank you! — Yasha

Nick Cox Nick Cox · Accepted Answer · 2014-12-17T22:02:12

You don't show all your code but you are evidently fitting an ordered probit upstream. Consider the simplest kind of example:

. sysuse auto
(1978 Automobile Data)

. oprobit rep78 weight

[stuff omitted]
------------------------------------------------------------------------------
       rep78 |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      weight |  -.0005881   .0001729    -3.40   0.001     -.000927   -.0002492
-------------+----------------------------------------------------------------
       /cut1 |  -3.797833   .6421343                     -5.056393   -2.539273
       /cut2 |  -2.954918   .5957484                     -4.122563   -1.787272
       /cut3 |  -1.582673   .5471056                      -2.65498   -.5103658
       /cut4 |  -.6635675   .5192685                     -1.681315      .35418
 ------------------------------------------------------------------------------

. di _b[/cut1]
-3.7978328

The coefficient you are using is a single estimate. It isn't to be subscripted and there are certainly aren't one thousand members of the same family, as your subscripting implies.

Stata 13 - Errors simulating confidence intervals when indexing on cut points

1 Answers