Problem with sampling weights in multilevel logistic regress

[virgiliomariano]

Hi,

I am running a two-stage multilevel logistic regression. The data that I am using comes from a survey with a multi-stage cluster sampling design. So I have tried to include the sampling weights in the analysis. However, when I run the commands stata only shows the ORs without the standard errors, z values or p values. Also, the 95% CI are not computed only sowing the same ORs values. The programs run fine if sampling weights are not included.

This is the syntax that I use:

runmlwin UNINPREG cons CONTROLBEHAV SexualVioEXP EVERPhysSex AREAmulti EDADMULTI QUINTILmulti2 QUINTILmulti3 QUINTILmulti4 QUINTILmulti5 EDUMULTI4 EDUMULTI5 EDUMULTI6 EVERMODERNCONT , level2(DEPMUN: cons) level1(CUES:, weightvar(aw1)) discrete(distribution(binomial) link(logit) denominator(cons) ) nopause or

And this is the output that I get:

MLwiN 2.28 multilevel model Number of obs = 5347
Binomial logit response model
Estimation algorithm: IGLS, MQL1

-----------------------------------------------------------
| No. of Observations per Group
Level Variable | Groups Minimum Average Maximum
----------------+------------------------------------------
DEPMUN | 142 4 37.7 370
-----------------------------------------------------------

Run time (seconds) = 1.87
Number of iterations = 6
------------------------------------------------------------------------------
UNINPREG | Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cons | .4451047 0 . . .4451047 .4451047
CONTROLBEHAV | 1.28416 0 . . 1.28416 1.28416
SexualVioEXP | 1.259928 0 . . 1.259928 1.259928
EVERPhysSex | 1.43794 0 . . 1.43794 1.43794
AREAmulti | .6532072 0 . . .6532072 .6532072
EDADMULTI | 1.016698 0 . . 1.016698 1.016698
QUINTILmul~2 | 1.053013 0 . . 1.053013 1.053013
QUINTILmul~3 | 1.13982 0 . . 1.13982 1.13982
QUINTILmul~4 | .9388392 0 . . .9388392 .9388392
QUINTILmul~5 | .6092721 0 . . .6092721 .6092721
EDUMULTI4 | .9125686 0 . . .9125686 .9125686
EDUMULTI5 | .9004646 0 . . .9004646 .9004646
EDUMULTI6 | .794364 0 . . .794364 .794364
EVERMODERN~T | 1.419933 0 . . 1.419933 1.419933
------------------------------------------------------------------------------

------------------------------------------------------------------------------
Random-effects Parameters | Estimate Std. Err. [95% Conf. Interval]
-----------------------------+------------------------------------------------
Level 2: DEPMUN |
var(cons) | .0731012 .0231148 .027797 .1184054
------------------------------------------------------------------------------

.
end of do-file

.

Thanks

[GeorgeLeckie]

Dear virgiliomariano,

It looks like MLwiN has failed to estimate the parameter sampling covariance matrix. MLwiN has therefore posted zeros for all the parameter standard errors. The z-ratios, associated p-values and 95% confidence intervals are therefore all also incorrect (since they are all based on the incorrect standard errors).

We will look into what is causing the problem and get back to you

Best wishes

George

[virgiliomariano]

Dear George ,

Thank you for your reply. There seems to be something related to the last version of the software. I used the previous version (27) and it worked fine.

Best

[ChrisCharlton]

Thank you for this additional information. It appears that a bug was introduced in version 2.28 of MLwiN that affected the calculation of the fixed part robust standard errors (used by default when weights are specified). This will be corrected in the next release.