[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

From |
Steven Samuels <sjhsamuels@earthlink.net> |

To |
statalist@hsphsun2.harvard.edu |

Subject |
Re: st: svy, subpop ( ) with version 10 |

Date |
Sat, 7 Jun 2008 09:37:07 -0400 |

Conditional logistic regression will not produce predictions that can be compared to those of ordinary "fixed effects" or random effects logistic regression. This was my answer to your original query.

However it appears that prediction is not your primary aim. So, conditional logistic regression (use -clogit-) sounds like the correct approach.

Good luck!

-Steven

On Jun 6, 2008, at 10:46 PM, Leda Inga wrote:

Unfortunately I don't have cluster specific variables. I'm worried

that you say fixed-effects logit will not solve the problem. I've read

the reference manual and it says that clogit ( which is the same as fe

logit) can be used for matched-case controles studies. In my data

every individual in a certain cluster for which the dependet variable

is zero would be a control for the rest (for which the dependent is

one). It also says that it doesn't matter if each cluster has a

different number of observations.

Finally, in the book of Wooldridge of panel data (p 496) it says that

an unobserved effect framework can be used for cluster samples. The

author cites the work of Geronimus and Korenman (1992) where they "use

sister pairs to determine the effects of teenage motherhood on

subsequente economic outcomes". They use a fixed-effects approach when

the outcome is binary.

In my model the most important omitted variables correlated with the

regressors are geographical access and cultural barriers.

Please, I would like to be confirmed if my approach is really incorrect.

2008/6/6 Steven Samuels <sjhsamuels@earthlink.net>:

If you want a 3-level or higher model with estimated variances at each

level, then your only recourse is -gllamm-. You would have to compute your

own weighted predictions.

In any case, -xtlogit- will not solve your problem; its "fixed- effects"

option makes sense only for longitudinal data.

-svylogit- will give you the standard goodness of fit statistics; will

permit you to model multiple stages of sampling, but will not provide

estimators of variance for each level. You can include cluster- specific

covariates and test the interaction of cluster-covariates and individual

covariates.

-Steven

On Jun 6, 2008, at 9:29 PM, Leda Inga wrote:

I'm still trying to find if there's a good measure of fit for a

fixed-effect logit model. As I answered in the last email, I'm really

interesed in getting one.

2008/6/6 Steven Samuels <sjhsamuels@earthlink.net>:

The numbers of PSU and strata (before subsetting) are different.

Therefore

either the -svyset- commands or the data sets differ. By the way, how

did

you resolve your cluster problem?

Steven

On Jun 6, 2008, at 8:04 PM, Leda Inga wrote:

Hi,

I'm running a logit using svy and found a difference in the

coefficients and pvalues reported by version 9 and 10 of Stata. I

thought that it could be because of the observations that each one

takes into account. As you can see below the version 10 considers 4178

observations but there are actually 8471 and the results are the same

as using the "if" after svy. Can anybody explain this?

The command I used for both versions was:

svy, subpop(urbano): logit PS n_hijos educ seguro V157 NSE rural

_ISREGION_3 _ISREGION_5 CPci CPci2 labor_prol

VERSION 10:

Survey: Logistic regression

Number of strata = 24 Number of obs =

4178

Number of PSUs = 451 Population size =

2920.9792

Subpop. no. of obs = 4178

Subpop. size

= 2920.9792

Design df

= 427

F( 10,

418) = 30.43

Prob > F

= 0.0000

------------------------------------------------------------------ ------------

| Linearized

PS | Coef. Std. Err. t P>|t| [95% Conf.

Interval]

------------- +----------------------------------------------------------------

n_hijos | -.0627022 .0233953 -2.68 0.008 -.1086865

-.0167179

educ | .0995025 .0176722 5.63 0.000 .0647671

.1342379

seguro | .2691795 .1430446 1.88 0.061 -.0119797

.5503387

V157 | .2184025 .0961237 2.27 0.024 .0294679

.4073371

NSE | .5893834 .0859681 6.86 0.000 .4204101

.7583568

_ISREGION_3 | -.8503491 .2073588 -4.10 0.000 -1.25792

-.4427781

_ISREGION_5 | -.7629466 .2529318 -3.02 0.003 -1.260093

-.2658003

CPci | .298412 .0547975 5.45 0.000 .1907057

.4061184

CPci2 | -.0109094 .0043379 -2.51 0.012 -.0194356

-.0023832

labor_prol | .415821 .1035528 4.02 0.000 .2122843

.6193578

_cons | -2.961981 .3026012 -9.79 0.000 -3.556754

-2.367207

------------------------------------------------------------------ ------------

Note: 25 strata omitted because they contain no subpopulation members.

VERSION 9:

Survey: Logistic regression

Number of strata = 49 Number of obs =

8471

Number of PSUs = 1118 Population size =

6948.092

Subpop. no.

of obs = 4212

Subpop. size

= 2945.3116

Design df

= 1069

F( 10,

1060) = 31.19

Prob > F

= 0.0000

------------------------------------------------------------------ ------------

| Linearized

PS | Coef. Std. Err. t P>|t| [95% Conf.

Interval]

------------- +----------------------------------------------------------------

n_hijos | -.0624413 .023198 -2.69 0.007 -.10796

-.0169225

educ | .100353 .0178213 5.63 0.000 .0653842

.1353218

seguro | .2789113 .1415214 1.97 0.049 .0012202

.5566025

V157 | .2192173 .0962575 2.28 0.023 .0303421

.4080924

NSE | .5896208 .0851999 6.92 0.000 .4224427

.7567989

_ISREGION_3 | -.8230625 .2058183 -4.00 0.000 -1.226916

-.4192088

_ISREGION_5 | -.7412179 .2491071 -2.98 0.003 -1.230012

-.2524236

CPci | .2921813 .04976 5.87 0.000 .194543

.3898195

CPci2 | -.0103967 .0037064 -2.81 0.005 -.0176694

-.003124

labor_prol | .4267068 .1025795 4.16 0.000 .2254268

.6279867

_cons | -2.97932 .293572 -10.15 0.000 -3.555362

-2.403277

------------------------------------------------------------------ ------------

Thanks.

*

* For searches and help try:

* http://www.stata.com/support/faqs/res/findit.html

* http://www.stata.com/support/statalist/faq

* http://www.ats.ucla.edu/stat/stata/

* * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/* * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/* * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

**References**:**st: svy, subpop ( ) with version 10***From:*"Leda Inga" <ledainga@gmail.com>

**Re: st: svy, subpop ( ) with version 10***From:*Steven Samuels <sjhsamuels@earthlink.net>

**Re: st: svy, subpop ( ) with version 10***From:*"Leda Inga" <ledainga@gmail.com>

**Re: st: svy, subpop ( ) with version 10***From:*Steven Samuels <sjhsamuels@earthlink.net>

**Re: st: svy, subpop ( ) with version 10***From:*"Leda Inga" <ledainga@gmail.com>

- Prev by Date:
**Re: st: /* */ comments inside Stata program** - Next by Date:
**Re: st: how to estima with Swamy Method or Swamy and Mehta with Stata** - Previous by thread:
**Re: st: svy, subpop ( ) with version 10** - Next by thread:
**st: System of Seemingly Unrelated Poisson Regressions (with Constraints)** - Index(es):

© Copyright 1996–2015 StataCorp LP | Terms of use | Privacy | Contact us | What's new | Site index |