|
A Report on the Joint Meetings of the American Statistical AssociationBy Craig Henderson, Research and Statistical Support Services ConsultantI recently had the opportunity to attend the national joint meeting of the American Statistical Association, mainly because it was held in Dallas this year. Not to belabor a point, but I am sure you are aware of the financial condition of graduate students. Fortunately, funding was available and I was able to attend. Unfortunately, I was only able to attend one full day of the conference. During my time there, I attended a short course presented by Nathan Curtis and Russell Wolfinger of the SAS Institute entitled "Using Mixed Models and SAS PROC MIXED for Longitudinal Data," attended some of the papers presented by faculty members and graduate students, and spent a good deal of time in the exhibitors hall making contacts with book and software publishers. I would like to take some time to report some of the things that I learned and observed at the conference. Statistics in the "real world"Although many of the papers and short courses had a theoretical focus, I tended to choose sessions that were more applied. From the sessions I attended, I came away with the conclusion that the field in general is moving toward the development and implementation of more general statistical procedures, thereby dealing with the problem of restrictive assumptions that more traditional procedures such as the general linear model require. For example, many of the presentations dealt with nonlinear statistics and random coefficient modelling or mixed model procedures. This is not to say that people had a dismissive attitude toward traditional procedures such as the general linear model. Instead, the emphasis of many was that assumptions such as linearity and homogeneity of regression slopes are often not satisfied in "real-world" data sets. Furthermore, many presentations addressed alternatives to null hypothesis statistical testing such as the Bayesian approach to testing hypotheses. "Using Mixed Models and SAS PROC MIXED for Longitudinal Data"This short course had a strongly applied focus and discussed use of the SAS procedure PROC MIXED for analyzing repeated measures data. Compared to PROC GLM, PROC MIXED uses a maximum likelihood or restricted maximum likelihood estimation technique as opposed to ordinary least squares; therefore, subjects with missing data on some of the measurement points are not automatically deleted from the analysis. In addition to this, PROC MIXED allows the researcher to test both fixed and random effects, does not assume homogeneity of regression in models with both continuous and categorical predictors, and allows the researcher to specify a covariance structure for the data, thereby providing the flexibility of modelling variances and covariances as opposed to assuming their homogeneity. Wolfinger (1993) advocates the comparison of several models employing different error covariance structures and selecting one that is reasonable. Some of the covariance structures accommodated by PROC MIXED include the Type H and unstructured matrices used also in PROC GLM, compound symmetry, autoregressive and other time series structures, random coefficients models, and spatial correlations (Wolfinger & Chang, 1998). PROC MIXED includes all of the functionality now available in PROC VARCOMP, and we were informed that consequently, PROC VARCOMP will be discontinued in upcoming releases of SAS. Please see Appendix A at the end of this paper for an example of PROC MIXED CODE. If PROC MIXED interests you, I have a copy of the course notes along with papers written by Wolfinger (Wolfinger, 1993; Wolfinger, 1997; Wolfinger & Chang, 1998) and the SAS PROC MIXED manual (Littell, Milleken, Stroup, & Wolfinger, 1996) in the Research and Statistical Support Office that would provide you with greater details on this topic. I would also recommend the book Linear Mixed Models in Practice: A SAS-Oriented Approach (Verbeke & Molenberghs, 1997). What's new in Statistical Software?I also was able to make contacts with several software vendors. As you may know, the Research and Statistical Support office is now supporting the software S-Plus. I was able to learn that Release 5.0 for UNIX is due to be released any time and that the Windows version will most likely be released in the spring. As yet, we have only purchased a site license for the Windows version, although we are researching prices on the UNIX version as well.In addition, I was also able to obtain demo copies of SUDAAN, (a software package designed to analyze correlated data by using generalized estimating equations), SPSS, and Minitab. One of my goals in going to the conference was to explore software packages that could be implemented in introductory statistics classes for teaching purposes. Among the two packages that interested me was a package named Fathom, a software package that includes functionality for descriptive statistics, hypothesis testing, estimations, and simulation. The graphics are completely interactive, and are very user friendly. I was also intrigued by the teaching potential of Statlets, a compilation of Java scripts that operate via a web browser. Both of these products are currently free for academic institutions. Fathom and Statlets can be accessed via their Websites, http://www.keypress.com/fathom and http://www.statlets.com. Please feel free to contact me at the Research and Statistical Support Office of the Computing Center if you are interested in learning more about these packages. I have a copy of Fathom along with the manual and release notes. You can reach me, Craig Henderson, at x2140 or craigh@unt.edu. ConclusionIn summary, attending the conference was a day well spent. I would encourage everyone to attend next year's conference in Baltimore if you can find the time and funding to get away for a week. I know it's statistics, but learning can be fun, and you can skip out of dry afternoon sessions to enjoy the city. At the very least, you can get out of a week of the Texas summer. ReferencesCurtis, N., & Wolfinger, R. (1998, August). Using mixed models and SAS PROC MIXED for longitudinal data. Short course presented at the Annual Joint Meeting of the American Statistical Association. Littell, R. C., Milliken, G. A., Stroup, W. W., & Wolfinger, R. D. (1996). SAS system for mixed models. Cary, NC: SAS Institute Inc. Verbeke, G., & Molenberghs, G. (1997). Linear mixed models in practice: A SAS-oriented approach. New York: Springer-Verlag. Wolfinger, R. D. (1993). "Covariance structure selection in general mixed models." Communications in Statistics, Simulation, and Computation, 22, 1079-1106. Wolfinger, R. D. (1997). "An example of using mixed models and PROC MIXED for longitudinal data." Journal of Biopharmaceutical Statistics, 7, 381-500. Wolfinger, R. D., & Chang, M. (1998). Comparing the SAS GLM and MIXED procedures for repeated measures. Cary, NC: SAS Institute Inc. Appendix A - SAS PROC MIXED Code(Taken from Curtis & Wolfinger, 1998)
1. Heterogeneous General Linear Model
2. Heterogeneous Compound Symmetry
3. Heterogeneous Random Coefficients
4. Heterogeneous Unstructured
|