Turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

- Home
- /
- Analytics
- /
- Stat Procs
- /
- Multinomial Logit and reversed signs

Options

- RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Mute
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content

Posted 07-04-2009 09:46 AM
(1248 views)

I am estimating a model of CEO turnover using panel data. For each year, a CEO can : 1. be forced from office, 2. have a normal retirement, or 3. remain in office. Initially, I modeled the event as the forced turnover (treating retirement as a non-event) using Proc Genmod and GEE to control for repeated observations from the same firms over time.

I am now using Proc CATMOD to estimate the likelihood of each type of event occuring, with no control for repeated observations from the same firm. The outcomes of the two estimations for the forced turnovers are qualitively the same, with the exception of a few of the coefficients of binomial variables, which are now nonsensical. For example, the results of the multinomial logit show that the likelihood of a forced turnovers is positively related to the CEO being age 62 or older and the likelihood of a voluntary turnover is negatively related to the CEO being age 62 or older. If I replace that variable with a continuous variable of the CEO's age, the signs reverse (and make sense). I have this problem with two other of the binomial variables as well.

Any suggestions as to what is wrong and solutions would be greatly (GREATLY!) appreciated. Message was edited by: svandals

I am now using Proc CATMOD to estimate the likelihood of each type of event occuring, with no control for repeated observations from the same firm. The outcomes of the two estimations for the forced turnovers are qualitively the same, with the exception of a few of the coefficients of binomial variables, which are now nonsensical. For example, the results of the multinomial logit show that the likelihood of a forced turnovers is positively related to the CEO being age 62 or older and the likelihood of a voluntary turnover is negatively related to the CEO being age 62 or older. If I replace that variable with a continuous variable of the CEO's age, the signs reverse (and make sense). I have this problem with two other of the binomial variables as well.

Any suggestions as to what is wrong and solutions would be greatly (GREATLY!) appreciated. Message was edited by: svandals

1 REPLY 1

- Mark as New
- Bookmark
- Subscribe
- Mute
- RSS Feed
- Permalink
- Report Inappropriate Content

It sounds like you have a nominal (unordered) response variable that is measured repeatedly on subjects. You can fit a logistic model to such data using PROC SURVEYLOGISTIC with the CLUSTER statement to indicate which observations come from the same subject, and the LINK=GLOGIT option in the MODEL statement. See "Cluster model with variance adjustment" in this usage note: http://support.sas.com/kb/22871 . GENMOD cannot model a nominal response.

To avoid problems with seemingly incorrect parameter signs, use the REF= response variable option in the MODEL statement to select the reference (denominator) level for the generalized logits that are being modeled. For instance, if you want the two generalized logits in your data to use the probability of "remain in office" as the reference level, and assuming that it is coded as 3 in your data, then specify: model turnover(ref="3") = ... / link=glogit; Also, if age is a predictor in the model and you want to treat it as categorical rather than continuous, be aware that the default coding imposed by the CLASS statement is "effects" coding which may not be what you are expecting. The typical 0,1 or "dummy" coding is used when you specify the PARAM=REF option. You can also use the REF= option to select a reference level for each predictor. For example, this statement selects dummy coding with age=1 as the reference level: class age(ref="1") / param=ref; For details on the different types of coding and the parameter interpretation that they induce, see "CLASS Variable Parameterization" in the SURVEYLOGISTIC documentation: http://support.sas.com/onlinedoc/913/getDoc/en/statug.hlp/surveylogistic_index.htm

The answers to many questions can be found in the Samples and SAS Notes in our searchable knowledgebase, http://support.sas.com/kb. You can use the search engine there to find the answers you need.

To avoid problems with seemingly incorrect parameter signs, use the REF= response variable option in the MODEL statement to select the reference (denominator) level for the generalized logits that are being modeled. For instance, if you want the two generalized logits in your data to use the probability of "remain in office" as the reference level, and assuming that it is coded as 3 in your data, then specify: model turnover(ref="3") = ... / link=glogit; Also, if age is a predictor in the model and you want to treat it as categorical rather than continuous, be aware that the default coding imposed by the CLASS statement is "effects" coding which may not be what you are expecting. The typical 0,1 or "dummy" coding is used when you specify the PARAM=REF option. You can also use the REF= option to select a reference level for each predictor. For example, this statement selects dummy coding with age=1 as the reference level: class age(ref="1") / param=ref; For details on the different types of coding and the parameter interpretation that they induce, see "CLASS Variable Parameterization" in the SURVEYLOGISTIC documentation: http://support.sas.com/onlinedoc/913/getDoc/en/statug.hlp/surveylogistic_index.htm

The answers to many questions can be found in the Samples and SAS Notes in our searchable knowledgebase, http://support.sas.com/kb. You can use the search engine there to find the answers you need.

Build your skills. Make connections. Enjoy creative freedom. Maybe change the world. **Registration is now open through August 30th**. Visit the SAS Hackathon homepage.

What is ANOVA?

ANOVA, or Analysis Of Variance, is used to compare the averages or means of two or more populations to better understand how they differ. Watch this tutorial for more.

Find more tutorials on the SAS Users YouTube channel.