[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: Very low sensitivity of fixed-effects logit |

Date |
Thu, 5 Jun 2008 21:57:46 -0400 |

Leda's fixed-effects probabilities would also have been distorted if she had specified the default prediction after fixed-effects - xtlogit-. The default is the probability of a positive response, conditional on there being exactly *one* positive response in the cluster.

-Steve

On Jun 5, 2008, at 9:13 PM, Steven Samuels wrote:

I can think of several reasons why the predictions differ. One is primary: predictions from the fixed-effects model are nonsensical. The model omits a constant (see the output), and the predictions omit fixed panel effects. On the other hand, the random effects model includes a constant and the predictions (with the pu0 optioons) omit a zero-average random panel effect.

Some other thoughts:

1. iweights cannot vary within a panel (cluster), so they are ordinarily a poor substitute for pweights.

2. I don't know what you mean about wanting to exclude "the effect of certain unobserved cluster

variables, as geographical access, in X4". If you want to know if there are cluster differences, a more direct approach would be to fit the random effects model and test whether the between-cluster SD (sigma) is zero. You can also include cluster-level variables in the prediction equation.

Good luck!

-Steve

On Jun 5, 2008, at 7:22 PM, Leda Inga wrote:

Thank you very much. I've used the following commands:

Fixed-effects models:

# delimit;

xi: xtlogit Y X1 X2 X3 i.X4 if [rural==1] [iw=pwi], fe i(V001) nolog or

;

delimit cr

Random-effects models.

# delimit;

xi: xtlogit Y X1 X2 X3 i.X4 region3 region5 if [rural==1] [iw=pwi],

re i(V001) nolog or

;

delimit cr

I didn't use svy: because it didn't allow me to use xtlogit (I'm

interested in excluding the effect of certain unobserved cluster

variables, as geographical access, in X4 and get a consistent

estimator of its impact on Y). I know that when using survey data it's

not correct to use the "if", but svy, subpop( www). Nevertheless, I

couldn't think of another way.

To calcute the predictions I used:

predict Yhat if e(sample), pu0

I didn't use lsens because it can't be used after xtlogit.

I hope this information helps to clarify my problem.

2008/6/5, Steven Samuels <sjhsamuels@earthlink.net>:

Leda,

You need to tell us exactly which logistic commands you ran. - svy: logit-

would be appropriate for fixed-effects logistic regression with stratified,

weighted, and clustered, data. -logistic- itself will not take stratum

variables, but will take weights and clustering.

-xtmelogit- cannot handle survey weights, but would otherwise be the proper

program for random effects logistic regression. The user-written package

-gllamm- can do multi-level logistic regression with weights. With random

effects models, several types of predictions are possible, depending on

whether random-effects are permitted in the prediction.

The user-written command -rocss- was written for Stata Version 8. It does

not accept weights and shouldn't be used with weighted survey data. After

-logistic-, -logit-, or -svy: logistic-, you should try Stata's built-in

commands: -lsens- and -lroc- .

After -gllamm- , you could compute your own tables of specificity and

sensitivity by printing out the weighted tables of probabilities. If there

are too many distinct values, round the probabilities to the nearest 0.01

before tabling.

Steve

On Jun 5, 2008, at 5:40 PM, Leda Inga wrote:

Hi,

I'm runnig a fixed-effects and random-effects logit with DHS

(Demographic Health Survey) data. The groups are the clusters within

which each female belongs to. Given the recomendation in a previous

statalist mail

(http://www.stata.com/statalist/archive/2007-06/msg00818.html), I calculated the percentege correctly predicted for both models.

Nevertheless, I'm very surprised because the sensitivity (% correctly

predicted of positive outcomes) is very low for the fixed-effects

logit while the specificity (% correctly predicted of negative

outcomes: zeros) very high. On the other hand, both measures are more

acceptable for the random effects models. Besides, the pvalue of the

hausman test is zero.

Are this measures (sensititvity and specificity) the most appropiate

for measuring the quality of the results of this kind of models? And

are the results I've gotten frequent when comparing a fixed effects

versus a random effects model?

Here are the results given by rocss, a command the calculates the

sensitivity (sens) and specificity (spec) for different cutoffs:

Fixed-effects logit:

cutoff sens spec omspec cclass carea

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

1. 0.000 1.0000 0.0000 1.0000 47.5149 0.0000

2. 0.100 0.7250 0.6656 0.3344 69.3837 0.5741

3. 0.200 0.4250 0.9026 0.0974 67.5660 0.7104

4. 0.300 0.2313 0.9637 0.0363 61.5734 0.7304

5. 0.400 0.1219 0.9908 0.0092 57.7961 0.7352

6. 0.500 0.0580 0.9951 0.0049 54.9844 0.7356

7. 0.600 0.0209 0.9995 0.0005 53.4507 0.7358

8. 0.700 0.0078 1.0000 0.0000 52.8543 0.7358

9. 0.800 0.0036 1.0000 0.0000 52.6555 0.7358

10. 0.900 0.0006 1.0000 0.0000 52.5135 0.7358

11. 1.000 0.0000 1.0000 0.0000 52.4851 0.7358

+------------------------------------------------------+Random-effects logit: cutoff sens spec omspec cclass carea ------------------------------------------------------ 1. 0.000 1.0000 0.0000 1.0000 48.9190 0.0000 2. 0.100 0.9791 0.1930 0.8070 57.7572 0.1910 3. 0.200 0.9291 0.3791 0.6209 64.8135 0.3685 4. 0.300 0.8490 0.5321 0.4679 68.7099 0.5046 5. 0.400 0.7528 0.6726 0.3274 71.1808 0.6171 6. 0.500 0.6353 0.7809 0.2191 70.9670 0.6923 7. 0.600 0.5036 0.8721 0.1279 69.1851 0.7442 8. 0.700 0.3871 0.9293 0.0707 66.4053 0.7697 9. 0.800 0.2448 0.9674 0.0326 61.3923 0.7817 10. 0.900 0.0971 0.9930 0.0070 55.4764 0.7861 11. 1.000 0.0000 1.0000 0.0000 51.0810 0.7864 Any help would be very appreciated. * * 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: Very low sensitivity of fixed-effects logit***From:*"Leda Inga" <ledainga@gmail.com>

**Re: st: Very low sensitivity of fixed-effects logit***From:*Steven Samuels <sjhsamuels@earthlink.net>

**Re: st: Very low sensitivity of fixed-effects logit***From:*"Leda Inga" <ledainga@gmail.com>

**Re: st: Very low sensitivity of fixed-effects logit***From:*Steven Samuels <sjhsamuels@earthlink.net>

- Prev by Date:
**Re: st: Too many variables** - Next by Date:
**RE: st: Fracplot after cnreg, intreg and nbreg** - Previous by thread:
**Re: st: Very low sensitivity of fixed-effects logit** - Next by thread:
**st: Is there any test to examine disitribution overlap with Stata?** - Index(es):

© Copyright 1996–2014 StataCorp LP | Terms of use | Privacy | Contact us | What's new | Site index |