Note: This page is designed to show the how multilevel model can be done using R and to be able to compare the results with those in the book.On this page we will use the lmer function which is found in the lme4 package. There are several other possible choices but we will go with lmer.
The data were downloaded in Stata format from here and imported into R using the foreign library from a directory called rdata on the local computer. This page is updated using R 2.11.1 in January, 2011.
library(foreign) imm10 <- read.dta(file="~/rdata/imm10.dta") attach(imm10)
Table 3.2, page 46. OLS regression lines over 10 schools.
tmp<-by(imm10, schnum, function(x) lm(math ~ homework, data=x)) tmat<-t(sapply(tmp,coef)) tmat (Intercept) homework 1 50.68354 -3.553797 2 49.01229 -2.920123 3 38.75000 7.909091 4 34.39382 5.592664 5 53.93863 -4.718412 6 49.25896 -2.486056 7 59.21022 1.094640 8 36.05535 6.496310 9 38.52000 5.860000 10 37.71392 6.335052
Two equations at the top of page 47.
# get value of public for each school
ptmp<-by(imm10, schnum, function(x) lm(public~1, data=x))
pmat<-as.matrix(sapply(ptmp,coef))
# combine public with intercept and slope
a<-cbind(tmat,pmat)
a
(Intercept) homework
1 50.68354 -3.553797 1
2 49.01229 -2.920123 1
3 38.75000 7.909091 1
4 34.39382 5.592664 1
5 53.93863 -4.718412 1
6 49.25896 -2.486056 1
7 59.21022 1.094640 0
8 36.05535 6.496310 1
9 38.52000 5.860000 1
10 37.71392 6.335052 1
summary(lm(a[,1]~a[,3]))
Call:
lm(formula = a[, 1] ~ a[, 3])
Residuals:
Min 1Q Median 3Q Max
-8.754 -5.232 -2.199 6.050 10.791
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 59.210 7.435 7.964 4.51e-05
a[, 3] -16.063 7.837 -2.050 0.0745
Residual standard error: 7.435 on 8 degrees of freedom
Multiple R-squared: 0.3443, Adjusted R-squared: 0.2624
F-statistic: 4.201 on 1 and 8 DF, p-value: 0.07455
summary(lm(a[,2]~a[,3]))
Call:
lm(formula = a[, 2] ~ a[, 3])
Residuals:
Min 1Q Median 3Q Max
-6.776 -4.869 1.768 4.159 5.852
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 1.0946 5.2680 0.208 0.841
a[, 3] 0.9626 5.5530 0.173 0.867
Residual standard error: 5.268 on 8 degrees of freedom
Multiple R-squared: 0.003742, Adjusted R-squared: -0.1208
F-statistic: 0.03005 on 1 and 8 DF, p-value: 0.8667
Equation near bottom of page 47 and Table 3.3.
library(lme4)
lmer(math ~ homework + (homework|schnum), REML=FALSE)
Linear mixed model fit by maximum likelihood
Formula: math ~ homework + (homework | schnum)
AIC BIC logLik deviance REMLdev
1781 1803 -884.7 1769 1764
Random effects:
Groups Name Variance Std.Dev. Corr
schnum (Intercept) 61.806 7.8617
homework 19.979 4.4697 -0.804
Residual 43.067 6.5625
Number of obs: 260, groups: schnum, 10
Fixed effects:
Estimate Std. Error t value
(Intercept) 44.773 2.603 17.199
homework 2.049 1.472 1.391
Correlation of Fixed Effects:
(Intr)
homework -0.803
Equation near bottom of page 49 and Table 3.4.
lmer(math ~ homework + public + (homework|schnum), REML=FALSE)
Linear mixed model fit by maximum likelihood
Formula: math ~ homework + public + (homework | schnum)
AIC BIC logLik deviance REMLdev
1765 1790 -875.4 1751 1744
Random effects:
Groups Name Variance Std.Dev. Corr
schnum (Intercept) 40.677 6.3778
homework 21.683 4.6565 -0.982
Residual 42.955 6.5540
Number of obs: 260, groups: schnum, 10
Fixed effects:
Estimate Std. Error t value
(Intercept) 58.056 2.695 21.546
homework 1.941 1.525 1.273
public -14.651 1.831 -8.000
Correlation of Fixed Effects:
(Intr) homwrk
homework -0.772
public -0.602 0.010
Equation at the bottom of page 50 and Table 3.5. The negative value for the interaction coefficient in the book is probably a typo error, it should be positive.
lmer(math ~ homework + public + homework:public + (homework|schnum), REML=FALSE)
Linear mixed model fit by maximum likelihood
Formula: math ~ homework + public + homework:public + (homework | schnum)
AIC BIC logLik deviance REMLdev
1767 1795 -875.4 1751 1739
Random effects:
Groups Name Variance Std.Dev. Corr
schnum (Intercept) 40.503 6.3642
homework 21.577 4.6451 -0.982
Residual 42.954 6.5540
Number of obs: 260, groups: schnum, 10
Fixed effects:
Estimate Std. Error t value
(Intercept) 59.2102 6.5976 8.975
homework 1.0946 4.6688 0.234
public -15.9419 6.9775 -2.285
homework:public 0.9472 4.9385 0.192
Correlation of Fixed Effects:
(Intr) homwrk public
homework -0.966
public -0.946 0.913
homwrk:pblc 0.913 -0.945 -0.965
Table 3.6, page 52.
# quietly rerun model from page 47 m1<-lmer(math ~ homework + (homework|schnum), REML=FALSE) # random intercepts and slopes for each school rcoef<-coef(m1) rcoef
$schnum (Intercept) homework 1 50.27101 -3.143116 2 48.88642 -2.753799 3 39.19512 7.566056 4 35.15670 5.393500 5 53.08124 -3.737865 6 48.58612 -1.765064 7 58.05519 1.336116 8 37.15232 6.059531 9 39.16960 5.430173 10 38.17276 6.101294
Figure 3.8, page 53.
n<-nrow(rcoef[[1]])
x<-cbind(0, 7)
y<-cbind(seq(1:n), seq(1:n))
y[,1]<-rcoef[[1]][,1]
y[,2]<-rcoef[[1]][,1]+rcoef[[1]][,2]*7
plot(x, y[1,], type="l", ylab="predicte", ylim=c(20,100))
for (i in 2:n){points(x, y[i,], type="l", lty=i)}
UCLA Researchers are invited to our Statistical Consulting Services
We recommend others to our list of Other Resources for Statistical Computing Help
These pages are Copyrighted (c) by UCLA Academic Technology Services