Hi everyone, Can someone explain me how to calculate SAS type III sums of squares in R? Not that I would like to use them, I know they are problematic. I would like to know how to calculate them in order to demonstrate that strange things happen when you use them (for a course for example). I know you can use drop1(lm(), test="F") but for an lm(y~A+B+A:B), type III SSQs are only calculated for the A:B interaction. If I specify lm(y~A+B), it calculates type II SSQ for the main effects (as type III only differs from type II if interactions are included). Thus, this approach doesn't help. Another approach is the Anova(, type="III") function within the library(car). But somehow it produces strange results. Somebody told me I have to change the contrast settings using options(contrasts=c("contr.helmert", "contr.poly")) But I had the impression that my results are still impossible. Are the calculations dependent from the version of R I use? I am currently using R2.1.1 The only thing that seems to work is a trick: Specify a separate column AB that codes a new variable for the interaction of A:B. Now you can fit A,B, and AB (as 3 main effects) in 3 different sequential models with each one of them in the end once. For the term in the end you then get type III SSQ which seem to be correct. Cheers Steffi -- --------------------------------- Stefanie von Felten Doktorandin ETH Z??rich Institut f??r Pflanzenwissenschaften Universit??tstrasse 2, LFW A2 8092 Z??rich Telefon: 044 632 85 97 Telefax: 044 632 11 53 e-mail: stefanie.vonfelten at ipw.agrl.ethz.ch Universit??t Z??rich Institut f??r Umweltwissenschaften Winterthurerstrasse 190 8057 Z??rich Telefon: 044 635 61 23 Telefax: 044 635 57 11 e-mail: sfelten at uwinst.unizh.ch
"Stefanie von Felten, IPW&IfU" <sfelten at uwinst.unizh.ch> writes:> Hi everyone, > > Can someone explain me how to calculate SAS type III sums of squares in > R? Not that I would like to use them, I know they are problematic. I > would like to know how to calculate them in order to demonstrate that > strange things happen when you use them (for a course for example). I > know you can use drop1(lm(), test="F") but for an lm(y~A+B+A:B), type > III SSQs are only calculated for the A:B interaction. If I specify > lm(y~A+B), it calculates type II SSQ for the main effects (as type III > only differs from type II if interactions are included). Thus, this > approach doesn't help. > > Another approach is the Anova(, type="III") function within the > library(car). But somehow it produces strange results. Somebody told me > I have to change the contrast settings using > options(contrasts=c("contr.helmert", "contr.poly")) > But I had the impression that my results are still impossible. > Are the calculations dependent from the version of R I use? I am > currently using R2.1.1 > > The only thing that seems to work is a trick: Specify a separate column > AB that codes a new variable for the interaction of A:B. Now you can fit > A,B, and AB (as 3 main effects) in 3 different sequential models with > each one of them in the end once. For the term in the end you then get > type III SSQ which seem to be correct.I think you need to qualify "correct", "strange", and "impossible"... In particular, since you're hinting at a bug in "car", I'm sure its author would like to know what it is you think is wrong. Also, in the "trick" you don't specify *how* you code the three terms. I don't think it is immaterial, and I'm not sure which of the possible versions (if any) matches up with e.g. SAS's version of type III for main effects in the presence of interaction. -- O__ ---- Peter Dalgaard ??ster Farimagsgade 5, Entr.B c/ /'_ --- Dept. of Biostatistics PO Box 2099, 1014 Cph. K (*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918 ~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk) FAX: (+45) 35327907
If you use the Anova function in the car package and specify contr.sum or contr.SAS for the contrasts for your categorical factors, you will get the same results as outputted by SAS. I've tried this with a variety of data sets and it works. On Thu November 24 2005 08:27, Stefanie von Felten, IPW&IfU wrote:> Hi everyone, > > Can someone explain me how to calculate SAS type III sums of squares in > R? Not that I would like to use them, I know they are problematic. I > would like to know how to calculate them in order to demonstrate that > strange things happen when you use them (for a course for example). I > know you can use drop1(lm(), test="F") but for an lm(y~A+B+A:B), type > III SSQs are only calculated for the A:B interaction. If I specify > lm(y~A+B), it calculates type II SSQ for the main effects (as type III > only differs from type II if interactions are included). Thus, this > approach doesn't help. > > Another approach is the Anova(, type="III") function within the > library(car). But somehow it produces strange results. Somebody told me > I have to change the contrast settings using > options(contrasts=c("contr.helmert", "contr.poly")) > But I had the impression that my results are still impossible. > Are the calculations dependent from the version of R I use? I am > currently using R2.1.1 > > The only thing that seems to work is a trick: Specify a separate column > AB that codes a new variable for the interaction of A:B. Now you can fit > A,B, and AB (as 3 main effects) in 3 different sequential models with > each one of them in the end once. For the term in the end you then get > type III SSQ which seem to be correct. > > Cheers > Steffi-- Michael W. Sears, Ph.D. Postdoctoral Fellow Department of Biology/MS 314 University of Nevada, Reno Reno, NV 89557 Assistant Professor Department of Zoology Southern Illinois University Carbondale, IL 62901 phone: 775.784.8008 cell: ??775.232.3520 web:????????http://www.unr.edu/homepage/msears ????????????????http://www.science.siu.edu/zoology/sears "Natural selection is a mechanism for generating ??an exceedingly high degree of improbability." ????????????????????????????????????????????????-Sir Ronald Fisher (1890-1962)
Reasonably Related Threads
- tapply with unequal length of arguments
- Antw: Re: Confidence interval for the difference between proportions - method used in prop.test()
- Manova and specifying the model
- discriminant analysis
- Confidence interval for the difference between proportions - method used in prop.test()