Dear R-helpers, i get some strange results using a linear mixed-effects model (lme), of the type: lme1 <- lme(y ~ x, random=~x|group, ...) For some datasets, i obtain very small standard deviations of the random effects. I compared these to standard deviations of the slope and intercept using a lmList approach. Of course, the SD from the lme is always smaller (shrinkage estimator), but in some cases (the problem cases) the SD from the lme seems way too small. E.g.: SD of intercept = 0.14, SD of slope = 0.0004, SD residual=0.11. An lmList gives a slope SD of 0.07. I have about n=6 observations per group, and about 20-100 groups depending on the dataset. thank you for any suggestions, Remko ^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~' Remko Duursma, Ph.D. student Forest Biometrics Lab / Idaho Stable Isotope Lab University of Idaho, Moscow, ID, U.S.A. ____________________________________________________________ Diabetics: Click here for a Free Glucose Meter from Access Diabetic. http://r.hotbot.com/r/lmt_ad/http://mocda4.com/1/c/563632/102938/302214/302214 This offer applies to U.S. Residents Only
Peter Dalgaard BSA
2003-Sep-23 07:46 UTC
[R] Very small estimated random effect variance (lme)
"Remko Duursma" <den.duurs at lycos.com> writes:> Dear R-helpers, > > i get some strange results using a linear mixed-effects model (lme), of the type: > > lme1 <- lme(y ~ x, random=~x|group, ...) > > For some datasets, i obtain very small standard deviations of the random effects. I compared these to standard deviations of the slope and intercept using a lmList approach. Of course, the SD from the lme is always smaller (shrinkage estimator), but in some cases (the problem cases) the SD from the lme seems way too small. E.g.: SD of intercept = 0.14, SD of slope = 0.0004, SD residual=0.11. An lmList gives a slope SD of 0.07. > > I have about n=6 observations per group, and about 20-100 groups depending on the dataset. > > thank you for any suggestions,It's not a shrinkage estimator it is a "subtraction estimator", measuring the excess variance of the empirical slopes over what would be expected from their s.e. if all (true) slopes were identical. This can even be negative, although the parametrizations in lme() will enforce a zero or very small variance in that case. (There are occasional cases where a negative variance can be interpreted, e.g. plants competing for the same growth medium, but you're generally in trouble if the design is unbalanced.) -- O__ ---- Peter Dalgaard Blegdamsvej 3 c/ /'_ --- Dept. of Biostatistics 2200 Cph. N (*) \(*) -- University of Copenhagen Denmark Ph: (+45) 35327918 ~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk) FAX: (+45) 35327907