News from the SPSS Community

[Jon K Peck]

Two new extension commands have been posted to the Extensions Collection on the SPSS Community website (www.ibm.com/developerworks/spssdevcentral).

STATS PROPOR REGR estimates regression models where the dependent variable is in (0,1).  It fits a beta distribution, and the regressors estimate the parameters of that distribution.

STATS PROPOR REGRPRED calculates predicted values for new data using a model produced by STATS PROPOR REGRPRED.

Users of Statistics 22 can download and install these commands using Utilities > Extension Bundles > Download and Install Extension Bundles.  Users on older versions of Statistics can get these from the website via Downloads for SPSS Statistics and then Extension Commands.

These commands require the R Essentials, which is also available via the website.

We hope you find these commands useful.


Jon Peck (no "h") aka Kim
Senior Software Engineer, IBM
[hidden email]
phone: [phone]

[Andy W]

Also for those interested, A better lemon squeezer? Maximum-likelihood regression with beta-distributed dependent variables. By Smithson, Michael; Verkuilen, Jay (2006) has code in the supplementary material on how to fit beta regression models using CNLR is SPSS.

See http://supp.apa.org/psycarticles/supplemental/met_11_1_54/met_11_1_54_supp.html for the supplementary materials. A few pdfs of the article are floating around the web, http://psychology3.anu.edu.au/people/smithson/details/Pubs/Smithson_Verkuilen06.pdf

[Bruce Weaver]

Thanks Andy.  That looks useful.

Andy W wrote

Also for those interested, A better lemon squeezer? Maximum-likelihood regression with beta-distributed dependent variables. By Smithson, Michael; Verkuilen, Jay (2006) has code in the supplementary material on how to fit beta regression models using CNLR is SPSS.

See http://supp.apa.org/psycarticles/supplemental/met_11_1_54/met_11_1_54_supp.html for the supplementary materials. A few pdfs of the article are floating around the web, http://psychology3.anu.edu.au/people/smithson/details/Pubs/Smithson_Verkuilen06.pdf

[Andy W]

FYI for those interested - Jon responded off list that the R package has more options for various link functions and estimating the variance as a separate parameter (which I did not even know was possible). So there are some obvious advantages to the R extension command over simply replicating the Smithson & Verkuilen CNLR routine.

Also as an FYI - Smithson and Verkuilen argue to use beta regression for not just proportion data, but Likert scale data with floor or ceiling effects. It seems if you have alot of ceiling effects though beta regression is probably not appropriate.

I've also seen suggestions for quantile regression and Tobit regression in such situations - see this related crossvalidated discussion http://stats.stackexchange.com/a/49450/1036. Jon also mentioned for these their is a Heckman regression approach as well (which I am not familiar with). I'm pretty sure Jon has already written R extension commands for all of these mentioned (so another thanks to Jon for putting this work in).