Houhou Li
2012-Oct-28 23:55 UTC
[R] Why are coefficient estimates using ML and REML are different in lme?
Hi, All, My data collection is from 4 regions (a, b, c, d). Within each region, it has 2 or 3 units. Within each unit, it has measurement from about 25 sample site. I was trying to use lme function to discribe relationship between y and a few covariates. Both y and covariates were measured at the sample site level. My question is when I use exactlly the same model but choose different estimation method (ML vs REML), I got quite different coefficients esimate for fixed effect and variance estimate for random effect(see below). Can anyone here please help me to understand why? Thank you very much. 1) Using REML lme(y~ Region*(x1+x2+x3), random=~1|unit, data=temp) Linear mixed-effects model fit by REML Data: temp AIC BIC logLik 1498.871 1558.059 -731.4353 Random effects: Formula: ~1 | unit (Intercept) Residual StdDev: 5.837025 7.104742 Fixed effects: y ~ Region * (x1 + x2 + x3) Value Std.Error DF t-value p-value (Intercept) 162.28206 22.340090 193 7.264163 0.0000 Regionb -11.06624 24.582841 5 -0.450161 0.6714 Regionc 5.01670 29.177730 5 0.171936 0.8702 Regiond -36.63434 26.262448 5 -1.394932 0.2218 x1 0.04091 0.034732 193 1.177953 0.2403 x2 -0.71649 0.356771 193 -2.008252 0.0460 x3 -0.15945 0.375098 193 -0.425095 0.6712 Regionb:x1 -0.04451 0.046075 193 -0.965975 0.3353 ............. 2) using ML lme(y~ Region*(x1+x2+x3), random=~1|unit, data=temp, method="ML") Linear mixed-effects model fit by maximum likelihood Data: temp AIC BIC logLik 1478.793 1539.38 -721.3964 Random effects: Formula: ~1 | unit (Intercept) Residual StdDev: 0.0002763015 7.043271 Fixed effects: y ~ Region * (x1 + x2 + x3) Value Std.Error DF t-value p-value (Intercept) 155.05508 21.512500 193 7.207674 0.0000 Regionb 10.56095 21.981366 5 0.480450 0.6512 Regionc 9.88513 28.595621 5 0.345687 0.7436 Regiond -34.68996 24.177548 5 -1.434801 0.2108 x1 0.05274 0.033903 193 1.555701 0.1214 x2 -0.67642 0.365633 193 -1.849995 0.0658 x3 0.09977 0.293438 193 0.340007 0.7342 Regionb:x1 -0.05692 0.046042 193 -1.236259 0.2179 ........ [[alternative HTML version deleted]]
S Ellison
2012-Oct-29 10:02 UTC
[R] Why are coefficient estimates using ML and REML are different in lme?
Yi) Different criteria would be _exepcted_ to give different estimates. ii) Look at teh standard errors on the coefficients. Essentially all of them are larger than the estimates for both fitting criteria. Essentially, both models are telling you that your coefficients are not significantly different from zero. S Ellison ________________________________________ From: r-help-bounces at r-project.org [r-help-bounces at r-project.org] On Behalf Of Houhou Li [lidarfly at yahoo.com] Sent: 28 October 2012 23:55 To: r-help at r-project.org Subject: [R] Why are coefficient estimates using ML and REML are different in lme? Hi, All, My data collection is from 4 regions (a, b, c, d). Within each region, it has 2 or 3 units. Within each unit, it has measurement from about 25 sample site. I was trying to use lme function to discribe relationship between y and a few covariates. Both y and covariates were measured at the sample site level. My question is when I use exactlly the same model but choose different estimation method (ML vs REML), I got quite different coefficients esimate for fixed effect and variance estimate for random effect(see below). Can anyone here please help me to understand why? Thank you very much. 1) Using REML lme(y~ Region*(x1+x2+x3), random=~1|unit, data=temp) Linear mixed-effects model fit by REML Data: temp AIC BIC logLik 1498.871 1558.059 -731.4353 Random effects: Formula: ~1 | unit (Intercept) Residual StdDev: 5.837025 7.104742 Fixed effects: y ~ Region * (x1 + x2 + x3) Value Std.Error DF t-value p-value (Intercept) 162.28206 22.340090 193 7.264163 0.0000 Regionb -11.06624 24.582841 5 -0.450161 0.6714 Regionc 5.01670 29.177730 5 0.171936 0.8702 Regiond -36.63434 26.262448 5 -1.394932 0.2218 x1 0.04091 0.034732 193 1.177953 0.2403 x2 -0.71649 0.356771 193 -2.008252 0.0460 x3 -0.15945 0.375098 193 -0.425095 0.6712 Regionb:x1 -0.04451 0.046075 193 -0.965975 0.3353 ............. 2) using ML lme(y~ Region*(x1+x2+x3), random=~1|unit, data=temp, method="ML") Linear mixed-effects model fit by maximum likelihood Data: temp AIC BIC logLik 1478.793 1539.38 -721.3964 Random effects: Formula: ~1 | unit (Intercept) Residual StdDev: 0.0002763015 7.043271 Fixed effects: y ~ Region * (x1 + x2 + x3) Value Std.Error DF t-value p-value (Intercept) 155.05508 21.512500 193 7.207674 0.0000 Regionb 10.56095 21.981366 5 0.480450 0.6512 Regionc 9.88513 28.595621 5 0.345687 0.7436 Regiond -34.68996 24.177548 5 -1.434801 0.2108 x1 0.05274 0.033903 193 1.555701 0.1214 x2 -0.67642 0.365633 193 -1.849995 0.0658 x3 0.09977 0.293438 193 0.340007 0.7342 Regionb:x1 -0.05692 0.046042 193 -1.236259 0.2179 ........ [[alternative HTML version deleted]] ******************************************************************* This email and any attachments are confidential. Any use...{{dropped:8}}