Svy function - using svylogitgof

[Jacquie Bowden]

Hi there, I am running a logistic regression, using svy: logit etc and then trying to do a post test using svylogitgof, however unfortunately as I have missing values in some of my variables, I am getting an error message after I type svylogitgof which says matrix has missing values r(504); - I would appreciate any assistance in how I might do a post-test using the svy function with missing values?
Thanks

[Carlo Lazzaro]

Jacquie:
one step behind: are your missing values ignorable or not?
Providing more details about your missing data will increase the likelihood of helpful replies.
Kind regards,
Carlo

[Jacquie Bowden]

Hi there Carlo
I am comfortable ignoring the missing values (its a pen and paper questionnaire and the missing seem to be fairly randomly distributed)
Thanks
Jacquie

[Richard Williams]

Personally, I can't reproduce the error. Perhaps you could show your commands and output or provide a replicable example. Show the estimation commands and output too; the problem may be there rather than with svylogitgof.

[Steve Samuels]

The only matrix that appears in svylogit.ado is e(V), the variance-covariance matrix of the decile mean residuals, and it appears in only one statement:

matrix chivalm = e(b) * inv(e(V)) * e(b)'

I found this out by typing:

viewsource svylogitgof.ado

in the command line.

The error follows the attempt to execute that statement. You might have discovered this for yourself if you had

set trace on
svylogitgof
set trace off

The error can arise if all of the values of a residual in a "decile" group are constant; this leads to missing standard errors and to e(V) being less than full rank. In the example that follows I've induced this problem by deliberately increasing the number of groups to 20 instead of the default 10 in svylogitgof. That led to three groups with missing standard errors and to e(V) having rank 16.

When e(V) is less than full rank, svylogit cannot compute the F statistic. The solution, if there is a solution, is to make your model more complex, so as to reduce the likelihood of tied groups.

sysuse auto, clear
svyset _n [pw = turn]
svy: logistic foreign length
predict double phat
gen double resid = foreign - phat
codebook resid
xtile byte decile = resid [pw = turn], nquantiles(20)

svy, over(decile): mean resid

/* Check rank of e(V) */
mata: rank(st_matrix("e(V)"))

matrix chivalm = e(b) * inv(e(V)) * e(b)'

[Steve Samuels]

In Stata 13, you can use estat gof after svy: logit. estat gof is somewhat tolerant of problems like the one you encountered. There is no longer a good reason for using svylogitgof.

[Jacquie Bowden]

Thank you very much Steve, estat gof worked perfectly - much appreciated!

[michele puech]

Hi there, I am running a logistic regression on survey data, using svy: logit etc; I have STATA 12; I am trying to do a post test: could you please tell me how to get svylogitgof ? I typed "findit svylogitgof" in search bar above but do not get the first entry to click on "st0099_1"? Could you please help me locating it? Also I read Steve Samuel 's entry above saying that estat gof is ok to use with svy:logit in stata 13; what about using it in stata 12?
Thank you for your help.
Michele

[Steve Samuels]

Welcome to Statalist, Michele! Type findit in the Command area (bottom of window), not in the search bar.

[Steve Samuels]

Just to add: I don't have Stata 12 to test estat gof after svy: logit or svy: logistic. Just try it yourself.

[Richard Williams]

I do have Stata 12 and it is there. But yes, the simplest thing is just to try it or type things like -help estat gof-.