Re: [R] Crossed random effects in lme
Søren Højsgaard [EMAIL PROTECTED] writes: Dear all, In the SASmixed package there is an example of an analysis of a split-plot experiment. The model is fm1Semi - lme( resistance ~ ET * position, data = Semiconductor, random = ~ 1 | Grp) where Grp in the Semiconductor dataset is defined as ET*Wafer. Is it possible to specify the grouping directly some way, e.g. like fm1Semi - lme( resistance ~ ET * position, data = Semiconductor, random = ~ 1 | ET*Wafer) Actually that does work with lme from the as-yet-unreleased version 0.6-1 of the lme4 package. See the slides from my presentation at useR!2004 http://www.ci.tuwien.ac.at/Conferences/useR-2004/Keynotes/Bates.pdf I'm back at my home now and will endeavor to get lme4_0.6-1 released as soon as possible. -- Douglas Bates[EMAIL PROTECTED] Statistics Department608/262-2598 University of Wisconsin - Madisonhttp://www.stat.wisc.edu/~bates/ __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
[R] Crossed Random Effects ?
I have been asked how to handle the following situation in R: Given an unbalanced design of 3 crossed random effects, such as subject, rater and item, how to estimate the variance components? I know how to do it using lme, but this seems to be limited to the nested case; or to use aov with error strata, when the design is balanced. -- Dipl.-Math. Wilhelm Bernhard Kloke Institut fuer Arbeitsphysiologie an der Universitaet Dortmund Ardeystrasse 67, D-44139 Dortmund, Tel. 0231-1084-257 __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
Re: [R] Crossed Random Effects ?
Wilhelm B. Kloke [EMAIL PROTECTED] writes: I have been asked how to handle the following situation in R: Given an unbalanced design of 3 crossed random effects, such as subject, rater and item, how to estimate the variance components? I know how to do it using lme, but this seems to be limited to the nested case; or to use aov with error strata, when the design is balanced. With the current version of lme it is difficult and inefficient to work with crossed random effects. A new version of lme being developed makes this much easier. I expect to release new versions of the Matrix and lme4 packages with R-1.9.0 (early April). I will describe the new lme capabilities in more detail at the useR!2004 conference in May. -- Douglas Bates[EMAIL PROTECTED] Statistics Department608/262-2598 University of Wisconsin - Madisonhttp://www.stat.wisc.edu/~bates/ __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html
[R] crossed random effects- clarified version
Hi, here is a clarified version of my problem. I have a total of 4*74 observations on 74 different mums in 5 different populations of mums, subject to 6 treatments (2 moisture levels*3 substrate types). I want to know if mum interacts with moisture level or substrate type. Population, moisture and substrate are fixed effects and mum is a random effect within population. Plot is a random whole-plot error for a split-plot design. The data set is called fm. This is the formula I used and the error message I got: fm$pmu - getGroups(fm, ~1|pop/mum, level=2) fm$grp = as.factor(rep(1,nrow(fm))) fm$pl - getGroups(fm, ~1|plot) fm$mo - getGroups(fm, ~1|moist) fm$su - getGroups(fm, ~1|sub) fm1 - lme(sqrt(mass) ~ iheight + moist*sub*pop, data=fm, random=list(grp=pdBlocked(list(pdIdent(~pl - 1), pdIdent(~pmu - 1), pdIdent(~pmu:su - 1), pdIdent(~pmu:mo - 1) Error in chol((value + t(value))/2) : non-positive definite matrix in chol The model works if the interaction terms: pdIdent(~pmu:su - 1), pdIdent(~pmu:mo - 1), are removed, so they are causing the problem. The model also works if I test mums from one population at a time so that mum no longer needs to be nested, i.e. if I replace pmu with mu: fm$mu - getGroups(fm, ~1|mum) fm1 - lme(sqrt(mass) ~ iheight + moist*sub, data=fm, random=list(grp=pdBlocked(list(pdIdent(~pl - 1), pdIdent(~pmu - 1), pdIdent(~mu:su - 1), pdIdent(~mu:mo - 1) It would be a lot faster if I can test all of the populations at once instead of individually. Any help would be much appreciated. Thanks, Sarah http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help
Re: [R] crossed random effects
Have you studied Pinhiero and Bates (2000) Mixed Effects Models in S and S-Plus (Springer)? Also, have you tried simplifying your lme call until you get something that works, then start adding back terms in various configurations until it breaks? Have you tried to compute how many coefficients are estimated in both fixed and random terms and evaluate whether all are estimable? For example, with 2 factors at 2 levels each, if you don't have all 4 possible combinations, you can't estimate the interaction -- even if you have thousands of replications of each. Finally, you can always try to read the code. I've learned a lot about S-Plus / R by doing that -- and solved a lot of my own problems that way. hope this helps. spencer graves Sarah Mclean wrote: Hi, if I have posted this twice, please ignore this. I'm not sure if I sent it to the correct e-mail address the first time. I have a data set on germination and plant growth with the following variables: dataset=fm mass (response) sub (fixed effect) moist (fixed effect) pop (fixed effect) mum (random effect nested within population) iheight (covariate) plot (random effect- whole plot factor for split-plot design). I want to see if moist or sub interacts with mum for any of the pops, but I am getting an error message. This is the formula I used: fm$pmu - getGroups(fm, ~1|pop/mum, level=2) fm$grp = as.factor(rep(1,nrow(fm))) fm$pl - getGroups(fm, ~1|plot) fm$mo - getGroups(fm, ~1|moist) fm$su - getGroups(fm, ~1|sub) fm1 - lme(sqrt(mass) ~ iheight + moist*sub*pop, data=fm, random=list(grp=pdBlocked(list(pdIdent(~pl - 1), pdIdent(~pmu - 1), pdIdent(~pmu:su - 1), pdIdent(~pmu:mo - 1) Error in chol((value + t(value))/2) : non-positive definite matrix in chol I know the problem is with the random interaction terms, but I don't know how to overcome this. Any advice would be greatly appreciated. I'm new to R and analysis such as this. Thank you, Sarah Mclean [EMAIL PROTECTED] http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help
Re: [R] crossed random effects
Hi, thanks for the advice. I have looked at the Pinheiro and Bates book and I've tried simplifying my model. I've narrowed the problem down to having mum nested within pop. If I run the analysis on each population separately, the interaction between mo and su with mum works fine. If I could analyse all of the pops at once this would be preferable because I have multiple responses and pops to test so it would take a bit of time. I'm hoping there is any easier way. Thanks Sarah --- Spencer Graves [EMAIL PROTECTED] wrote: Have you studied Pinhiero and Bates (2000) Mixed Effects Models in S and S-Plus (Springer)? Also, have you tried simplifying your lme call until you get something that works, then start adding back terms in various configurations until it breaks? Have you tried to compute how many coefficients are estimated in both fixed and random terms and evaluate whether all are estimable? For example, with 2 factors at 2 levels each, if you don't have all 4 possible combinations, you can't estimate the interaction -- even if you have thousands of replications of each. Finally, you can always try to read the code. I've learned a lot about S-Plus / R by doing that -- and solved a lot of my own problems that way. hope this helps. spencer graves Sarah Mclean wrote: Hi, if I have posted this twice, please ignore this. I'm not sure if I sent it to the correct e-mail address the first time. I have a data set on germination and plant growth with the following variables: dataset=fm mass (response) sub (fixed effect) moist (fixed effect) pop (fixed effect) mum (random effect nested within population) iheight (covariate) plot (random effect- whole plot factor for split-plot design). I want to see if moist or sub interacts with mum for any of the pops, but I am getting an error message. This is the formula I used: fm$pmu - getGroups(fm, ~1|pop/mum, level=2) fm$grp = as.factor(rep(1,nrow(fm))) fm$pl - getGroups(fm, ~1|plot) fm$mo - getGroups(fm, ~1|moist) fm$su - getGroups(fm, ~1|sub) fm1 - lme(sqrt(mass) ~ iheight + moist*sub*pop, data=fm, random=list(grp=pdBlocked(list(pdIdent(~pl - 1), pdIdent(~pmu - 1), pdIdent(~pmu:su - 1), pdIdent(~pmu:mo - 1) Error in chol((value + t(value))/2) : non-positive definite matrix in chol I know the problem is with the random interaction terms, but I don't know how to overcome this. Any advice would be greatly appreciated. I'm new to R and analysis such as this. Thank you, Sarah Mclean [EMAIL PROTECTED] http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help
Re: [R] crossed random effects
How many mum's and pop's do you have, and how many observations do you have of each mum-pop combination? If you want mum nested within pop, do I infer correctly that each mum has mated with only one pop, but that each pop may have offspring by multiple mums? The table of mum-pop combinations might help explain why you got, Error in chol((value + t(value))/2) : non-positive definite matrix in chol. If you can get an answer ignoring pop, then you might be able to get an answer with pop as a separate random term without specifying mum nested within pop. Also, I'd check very carefully the specification of nesting: I've messed that up more than once, and I'm bald now, because I tore all my hair out before I figured out what I was doing wrong. (Well, there is a slight exageration there.) Have you tried a very simple toy problem (or a published example) with nesting to make sure you can get the correct answer? hope this helps. spencer graves Sarah Mclean wrote: Hi, thanks for the advice. I have looked at the Pinheiro and Bates book and I've tried simplifying my model. I've narrowed the problem down to having mum nested within pop. If I run the analysis on each population separately, the interaction between mo and su with mum works fine. If I could analyse all of the pops at once this would be preferable because I have multiple responses and pops to test so it would take a bit of time. I'm hoping there is any easier way. Thanks Sarah --- Spencer Graves [EMAIL PROTECTED] wrote: Have you studied Pinhiero and Bates (2000) Mixed Effects Models in S and S-Plus (Springer)? Also, have you tried simplifying your lme call until you get something that works, then start adding back terms in various configurations until it breaks? Have you tried to compute how many coefficients are estimated in both fixed and random terms and evaluate whether all are estimable? For example, with 2 factors at 2 levels each, if you don't have all 4 possible combinations, you can't estimate the interaction -- even if you have thousands of replications of each. Finally, you can always try to read the code. I've learned a lot about S-Plus / R by doing that -- and solved a lot of my own problems that way. hope this helps. spencer graves Sarah Mclean wrote: Hi, if I have posted this twice, please ignore this. I'm not sure if I sent it to the correct e-mail address the first time. I have a data set on germination and plant growth with the following variables: dataset=fm mass (response) sub (fixed effect) moist (fixed effect) pop (fixed effect) mum (random effect nested within population) iheight (covariate) plot (random effect- whole plot factor for split-plot design). I want to see if moist or sub interacts with mum for any of the pops, but I am getting an error message. This is the formula I used: fm$pmu - getGroups(fm, ~1|pop/mum, level=2) fm$grp = as.factor(rep(1,nrow(fm))) fm$pl - getGroups(fm, ~1|plot) fm$mo - getGroups(fm, ~1|moist) fm$su - getGroups(fm, ~1|sub) fm1 - lme(sqrt(mass) ~ iheight + moist*sub*pop, data=fm, random=list(grp=pdBlocked(list(pdIdent(~pl - 1), pdIdent(~pmu - 1), pdIdent(~pmu:su - 1), pdIdent(~pmu:mo - 1) Error in chol((value + t(value))/2) : non-positive definite matrix in chol I know the problem is with the random interaction terms, but I don't know how to overcome this. Any advice would be greatly appreciated. I'm new to R and analysis such as this. Thank you, Sarah Mclean [EMAIL PROTECTED] http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help
[R] crossed random effects
Hi, I have a data set on germination and plant growth with the following variables: dataset=fm mass (response) sub (fixed effect) moist (fixed effect) pop (fixed effect) mum (random effect nested within population) iheight (covariate) plot (random effect- whole plot factor for split-plot design). I want to see if moist or sub interacts with mum for any of the pops, but I am getting an error message. This is the formula I used: fm$pmu - getGroups(fm, ~1|pop/mum, level=2) fm$grp = as.factor(rep(1,nrow(fm))) fm$pl - getGroups(fm, ~1|plot) fm$mo - getGroups(fm, ~1|moist) fm$su - getGroups(fm, ~1|sub) fm1 - lme(sqrt(mass) ~ iheight + moist*sub*pop, data=fm, random=list(grp=pdBlocked(list(pdIdent(~pl - 1), pdIdent(~pmu - 1), pdIdent(~pmu:su - 1), pdIdent(~pmu:mo - 1) Error in chol((value + t(value))/2) : non-positive definite matrix in chol I know the problem is with the random interaction terms, but I don't know how to overcome this. Any advice would be greatly appreciated. I'm new to R and analysis such as this. Thank you, Sarah Mclean [EMAIL PROTECTED] http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help
[R] crossed random effects
Hi, if I have posted this twice, please ignore this. I'm not sure if I sent it to the correct e-mail address the first time. I have a data set on germination and plant growth with the following variables: dataset=fm mass (response) sub (fixed effect) moist (fixed effect) pop (fixed effect) mum (random effect nested within population) iheight (covariate) plot (random effect- whole plot factor for split-plot design). I want to see if moist or sub interacts with mum for any of the pops, but I am getting an error message. This is the formula I used: fm$pmu - getGroups(fm, ~1|pop/mum, level=2) fm$grp = as.factor(rep(1,nrow(fm))) fm$pl - getGroups(fm, ~1|plot) fm$mo - getGroups(fm, ~1|moist) fm$su - getGroups(fm, ~1|sub) fm1 - lme(sqrt(mass) ~ iheight + moist*sub*pop, data=fm, random=list(grp=pdBlocked(list(pdIdent(~pl - 1), pdIdent(~pmu - 1), pdIdent(~pmu:su - 1), pdIdent(~pmu:mo - 1) Error in chol((value + t(value))/2) : non-positive definite matrix in chol I know the problem is with the random interaction terms, but I don't know how to overcome this. Any advice would be greatly appreciated. I'm new to R and analysis such as this. Thank you, Sarah Mclean [EMAIL PROTECTED] http://mobile.yahoo.com.au - Yahoo! Mobile - Check compose your email via SMS on your Telstra or Vodafone mobile. __ [EMAIL PROTECTED] mailing list https://www.stat.math.ethz.ch/mailman/listinfo/r-help
