General Notes / Setup

The Handbook of Research Synthesis and Meta-Analysis by Cooper et al. (2019), now in its third edition, has been one of the quintessential texts on meta-analysis and the entire research synthesis process as a whole. In this document, I provide the R code to reproduce the worked examples and analyses from various chapters. Emphasis will be on using the metafor package, but several other packages will also be used. To read more about the metafor package, see the package website and the package documentation.

The package can be installed with:

install.packages("metafor")

Once the package is installed, we can load it with:

library(metafor)

A few additional notes:

  1. Results are only reproduced for chapters containing worked examples.
  2. Occasionally, there are some minor discrepancies between the results shown in the book and those obtained below. These can result from using different software packages that implement methods in slightly different ways, due to intermittent rounding or using a different rounding scheme, or due to chance when the analyses involve some stochastic process (e.g., when using MCMC sampling in the chapter on Bayesian methods). Minor discrepancies will (usually) not be commented on. However, where discrepancies are more substantial, they will be noted (and the reasons for them if they are known).
  3. The results are generally given without discussion or context. The code below is not a substitute for reading the book, but is meant to be used together with it. In other words, readers of the book interested in replicating the results with R can see here how this is possible.

1) Research Synthesis as a Scientific Process

To recreate Figure 1.1 (showing the number of citations to articles including the terms ‘research synthesis’, ‘systematic review’, or ‘meta-analysis’ in their titles), I redid the search in the Web of Science Core Collection, which yielded the following data:

dat <- read.table(header=TRUE, text = "
year hits
2020 15973
2019 32120
2018 26293
2017 23885
2016 21318
2015 18487
2014 14781
2013 12357
2012 9802
2011 7528
2010 6120
2009 5121
2008 4006
2007 3553
2006 2771
2005 2336
2004 1911
2003 1526
2002 1309
2001 1005
2000 891
1999 832
1998 729
1997 580
1996 466
1995 70
1994 25
1993 11
1992 5
1991 9
1990 20
1989 127
1988 104
")
dat <- dat[-1,]               # remove current year (not complete)
dat <- dat[dat$year >= 1995,] # keep only 1995 or later
dat <- dat[nrow(dat):1,]      # reverse order

We can then create a bar chart based on these data:

# Figure 1.1

par(mar=c(4,4,2,2))
barplot(dat$hits, names.arg=dat$year, las=2, space=0.4, col="#6c9ece", border=NA)
abline(h=seq(0, 30000, by=5000), col="gray")
barplot(dat$hits, space=0.4, col="#6c9ece", border=NA, add=TRUE, axes=FALSE)

10) Evaluating Coding Decisions

Part of the code for this chapter (by Jack Vevea, Nicole Zelinksy, and Robert Orwin) is adapted from the chapter itself (see section 10.5). Below, we will make use of several additional packages that need to be installed (if they are not already installed). So let’s do this first.

# install the 'irr' package (if it is not already installed) and load it

if (!require(irr)) {
   install.packages("irr")
   library(irr)
}
# install the 'psych' package (if it is not already installed) and load it

if (!require(psych)) {
   install.packages("psych")
   library(psych)
}
# install the 'vcd' package (if it is not already installed) and load it

if (!require(vcd)) {
   install.packages("vcd")
   library(vcd)
}
# install the 'lme4' package (if it is not already installed) and load it

if (!require(lme4)) {
   install.packages("lme4")
   library(lme4)
}
# Table 10.1

dat <- read.table(header=TRUE, text = "
study c1 c2 c3
1  3 2 3
2  3 1 1
3  2 2 2
4  3 2 3
5  1 1 1
6  3 1 3
7  2 2 1
8  1 1 1
9  2 2 1
10 2 1 3
11 2 2 2
12 3 3 3
13 3 1 2
14 2 1 1
15 1 1 1
16 1 1 2
17 3 3 1
18 2 2 2
19 2 2 2
20 3 1 1
21 2 1 2
22 1 1 3
23 3 2 2
24 3 3 3
25 2 2 3")
# put ratings for coders 1, 2, and 3 into separate vectors

c1 <- dat$c1
c2 <- dat$c2
c3 <- dat$c3

# combine ratings for coders 1 and 2 into a matrix

c1c2 <- cbind(c1, c2)

# combine ratings from all three coders into a matrix

all3 <- cbind(c1, c2, c3)
# cell counts and marginal totals for coders 1 and 2

addmargins(table(c2, c1))
##      c1
## c2     1  2  3 Sum
##   1    5  3  4  12
##   2    0  7  3  10
##   3    0  0  3   3
##   Sum  5 10 10  25

# note: first variable is for rows, second is for columns, so to reproduce
# panel A of Table 10.2, we have to use table(c2, c1)
# agreement rate for coders 1 and 2

mean(c1 == c2)
## [1] 0.6

# agreement rate for all three coders

mean(c1 == c2 & c1 == c3)
## [1] 0.36

# agreement rate (in %) between coders 1 and 2

irr::agree(c1c2)
##  Percentage agreement (Tolerance=0)
## 
##  Subjects = 25 
##    Raters = 2 
##   %-agree = 60

# note: agree(c1c2) would have been sufficient, but due to the large number of
# additional packages being used, I will make it clear by using the :: operator
# which package a function belongs to (unless this is clear from the context)
# agreement rate (in %) between all three coders

irr::agree(all3)
##  Percentage agreement (Tolerance=0)
## 
##  Subjects = 25 
##    Raters = 3 
##   %-agree = 36

# unweighted Cohen's kappa for coders 1 and 2

irr::kappa2(c1c2)
##  Cohen's Kappa for 2 Raters (Weights: unweighted)
## 
##  Subjects = 25 
##    Raters = 2 
##     Kappa = 0.425 
## 
##         z = 3.51 
##   p-value = 0.000455

# unweighted Cohen's kappa for all three coders

irr::kappam.fleiss(all3)
##  Fleiss' Kappa for m Raters
## 
##  Subjects = 25 
##    Raters = 3 
##     Kappa = 0.315 
## 
##         z = 3.85 
##   p-value = 0.00012

# weighted Cohen's kappa for coders 1 and 2

irr::kappa2(c1c2, weight=0:2)
##  Cohen's Kappa for 2 Raters (Weights: 0,1,2)
## 
##  Subjects = 25 
##    Raters = 2 
##     Kappa = 0.386 
## 
##         z = 3.22 
##   p-value = 0.00126

We can also use the psych package to compute Cohen’s kappa, which also provides corresponding confidence intervals.

# unweighted and weighted Cohen's kappa for coders 1 and 2

W <- outer(1:3, 1:3, FUN = function(x,y) abs(x-y)) # create weight matrix
W
##      [,1] [,2] [,3]
## [1,]    0    1    2
## [2,]    1    0    1
## [3,]    2    1    0

res <- psych::cohen.kappa(c1c2, w=W)
## Warning in cohen.kappa1(x, w = w, n.obs = n.obs, alpha = alpha, levels = levels): upper or
## lower confidence interval exceed abs(1) and set to +/- 1.
print(res, digits=3)
## Call: cohen.kappa1(x = x, w = w, n.obs = n.obs, alpha = alpha, levels = levels)
## 
## Cohen Kappa and Weighted Kappa correlation coefficients and confidence boundaries 
##                   lower estimate upper
## unweighted kappa  0.185    0.425 0.666
## weighted kappa   -1.000    0.386 1.000
## 
##  Number of subjects = 25

Note that the CI for weighted kappa is not correct! Using the vcd package, we can also compute Cohen’s kappa and obtain the correct CI for weighted kappa.

print(vcd::Kappa(table(c1,c2)), digits=3, CI=TRUE)
##            value   ASE    z Pr(>|z|) lower upper
## Unweighted 0.425 0.123 3.46 0.000532 0.185 0.666
## Weighted   0.386 0.128 3.02 0.002558 0.135 0.637

# note: the (default) weighting scheme used for computing weighted kappa by
# the function is the one described in the chapter
# Krippendorff's alpha for coders 1 and 2 when treating the data as ratings on
# a nominal, on an ordinal, or on a ratio scale

irr::kripp.alpha(t(c1c2))
##  Krippendorff's alpha
## 
##  Subjects = 25 
##    Raters = 2 
##     alpha = 0.403

irr::kripp.alpha(t(c1c2), method="ordinal")
##  Krippendorff's alpha
## 
##  Subjects = 25 
##    Raters = 2 
##     alpha = 0.278

irr::kripp.alpha(t(c1c2), method="ratio")
##  Krippendorff's alpha
## 
##  Subjects = 25 
##    Raters = 2 
##     alpha = 0.311

# correlation between coders 1 and 2

cor(c1, c2)
## [1] 0.4520228

# note: the cor() function is part of the 'stats' package, which comes with R
# mean correlation between all pairs of coders

irr::meancor(all3)
##  Mean of bivariate correlations R
## 
##  Subjects = 25 
##    Raters = 3 
##         R = 0.331 
## 
##         z = 1.55 
##   p-value = 0.121

# intraclass correlation coefficient for coders 1 and 2

psych::ICC(c1c2)
## Call: psych::ICC(x = c1c2)
## 
## Intraclass correlation coefficients 
##                          type  ICC   F df1 df2     p lower bound upper bound
## Single_raters_absolute   ICC1 0.28 1.8  24  25 0.082      -0.118        0.60
## Single_random_raters     ICC2 0.35 2.6  24  24 0.010      -0.025        0.65
## Single_fixed_raters      ICC3 0.45 2.6  24  24 0.010       0.075        0.71
## Average_raters_absolute ICC1k 0.43 1.8  24  25 0.082      -0.268        0.75
## Average_random_raters   ICC2k 0.52 2.6  24  24 0.010      -0.050        0.79
## Average_fixed_raters    ICC3k 0.62 2.6  24  24 0.010       0.140        0.83
## 
##  Number of subjects = 25     Number of Judges =  2
## See the help file for a discussion of the other 4 McGraw and Wong estimates,

# note: this function computes 6 different types of ICCs; the first three are
# discussed in the chapter and correspond to the three different designs
# described on page 187

Using the lmer() function from the lme4 package, we can also do these calculations manually.

# restructure data into 'long' format

dat <- data.frame(study = 1:25,
                  rater = rep(1:2, each=25),
                  rating = c(c1,c2))

# absolute agreement based on one-way random-effects model

res <- lmer(rating ~ (1 | study), data = dat)
vcs <- data.frame(VarCorr(res))
vcs$vcov[1] / (vcs$vcov[1] + vcs$vcov[2])
## [1] 0.2777018

# absolute agreement based on two-way random-effects model

res <- lmer(rating ~ (1 | study) + (1 | rater), data = dat)
vcs <- data.frame(VarCorr(res))
vcs$vcov[1] / (vcs$vcov[1] + vcs$vcov[2] + vcs$vcov[3])
## [1] 0.3545258

# absolute agreement based on two-way mixed-effects model

res <- lmer(rating ~ rater + (1 | study), data = dat)
vcs <- data.frame(VarCorr(res))
vcs$vcov[1] / (vcs$vcov[1] + vcs$vcov[2])
## [1] 0.4503106

# example data from page 199

dat <- data.frame(
   study = 1:25,
   rater = rep(1:3, each=25),
   rating = c(3,3,2,3,NA,3,2,1,2,2,NA,3,3,2,1,1,3,2,2,3,2,1,3,NA,2,
              2,1,NA,2,1,1,2,1,2,1,2,3,1,1,NA,1,3,2,2,1,1,1,2,3,2,
              3,1,2,3,1,3,1,1,NA,3,2,3,2,1,1,2,1,2,2,1,2,3,2,3,3))
dat[c(1:4, 71:75),]
##    study rater rating
## 1      1     1      3
## 2      2     1      3
## 3      3     1      2
## 4      4     1      3
## 71    21     3      2
## 72    22     3      3
## 73    23     3      2
## 74    24     3      3
## 75    25     3      3

# absolute agreement for all three raters (based on one-way random-effects model)

res <- lmer(rating ~ (1 | study), data = dat)
vcs <- data.frame(VarCorr(res))
vcs$vcov[1] / (vcs$vcov[1] + vcs$vcov[2])
## [1] 0.1794355

11) Effect Sizes for Meta-Analysis

# data for Figure 11.1

dat <- read.table(header=TRUE, text = "
study    md   n   var    se  pval
A     0.400  60 0.067 0.258 0.121
B     0.200 600 0.007 0.082 0.014
C     0.300 100 0.040 0.201 0.134
D     0.400 200 0.020 0.141 0.005
E     0.300 400 0.010 0.100 0.003
F    -0.200 200 0.020 0.141 0.157")
dat
##   study   md   n   var    se  pval
## 1     A  0.4  60 0.067 0.258 0.121
## 2     B  0.2 600 0.007 0.082 0.014
## 3     C  0.3 100 0.040 0.201 0.134
## 4     D  0.4 200 0.020 0.141 0.005
## 5     E  0.3 400 0.010 0.100 0.003
## 6     F -0.2 200 0.020 0.141 0.157

# Figure 11.1

res <- rma(md, var, data=dat, method="EE", slab=study)

tmp <- dat[-1]
tmp$se  <- fmtx(tmp$se,  digits=3)
tmp$var <- fmtx(tmp$var, digits=3)

size <- sqrt(weights(res))
size <- 2.5 * size / max(size)

par(mar=c(5,4,2,2))

forest(res, xlim=c(-6.5,1.4), psize=size, header=TRUE, mlab="Combined",
       efac=c(0,1,2), annotate=FALSE, xlab="Standardized Mean Difference",
       ilab=tmp, ilab.xpos=c(-5.0, -4.1, -3.2, -2.3, -1.4))
text(-5.0, 8, "Mean\nDifference", font=2)
text(-4.1, 8, "Sample\nSize", font=2)
text(-3.2, 8, "Variance", font=2)
text(-2.3, 8, "Standard\nError", font=2)
text(-1.4, 8, "p-Value", font=2)

Effect Sizes for a Comparison of Means

# mean difference assuming sigma^2_1 = sigma^2_1

dat <- escalc("MD", m1i=103, m2i=100, sd1i=5.5, sd2i=4.5, n1i=50, n2i=50, vtype="HO")
summary(dat) # note: summary() so we can also see the standard error (sei)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 3.0000 1.0100 1.0050 2.9851 0.0028 1.0303 4.9697

# mean difference not assuming sigma^2_1 = sigma^2_1

dat <- escalc("MD", m1i=103, m2i=100, sd1i=5.5, sd2i=4.5, n1i=50, n2i=50)
summary(dat)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 3.0000 1.0100 1.0050 2.9851 0.0028 1.0303 4.9697

# note: since n1i=n2i in this example, the results are exactly the same
# mean change

dat <- escalc("MC", m1i=105, m2i=100, sd1i=10, sd2i=10, ni=50, ri=0.5)
summary(dat)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 5.0000 2.0000 1.4142 3.5355 0.0004 2.2282 7.7718

# standardized mean difference (Hedges' g)

dat <- escalc("SMD", m1i=103, m2i=100, sd1i=5.5, sd2i=4.5, n1i=50, n2i=50)
summary(dat)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 0.5924 0.0418 0.2043 2.8993 0.0037 0.1919 0.9929

# note: by default, the sampling variance is computed in a slightly different way in the
# book compared to the metafor package; by using vtype="LS2", the exact same equation as
# given in the book is used; but the difference is usually negligible (the results below
# differ slightly from those given in the book due to intermittent rounding in the book)

dat <- escalc("SMD", m1i=103, m2i=100, sd1i=5.5, sd2i=4.5, n1i=50, n2i=50, vtype="LS2")
summary(dat)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 0.5924 0.0411 0.2028 2.9208 0.0035 0.1949 0.9900

# standardized mean change (using raw score standardization)

dat <- escalc("SMCR", m1i=103, m2i=100, sd1i=5.5/sqrt(2*(1-0.7)), ni=50, ri=0.7)
summary(dat)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 0.4160 0.0137 0.1172 3.5502 0.0004 0.1863 0.6457

# note: 5.5 is the SD of the change scores, which we can convert to the SD of
# the raw scores by dividing it by sqrt(2*(1-r)) (this assumes that the SD was
# the same at the pre- and post-test)
# note: by default, the sampling variance is computed in a slightly different way in the
# book compared to the metafor package; by using vtype="LS2", the exact same equation as
# given in the book is used; but the difference is usually negligible (the results below
# differ slightly from those given in the book due to intermittent rounding in the book
# and also because the equation given in the book contains a mistake)

dat <- escalc("SMCR", m1i=103, m2i=100, sd1i=5.5/sqrt(2*(1-0.7)), ni=50, ri=0.7, vtype="LS2")
summary(dat)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 0.4160 0.0134 0.1156 3.5986 0.0003 0.1894 0.6426

# standardized mean difference based on ANCOVA results

# note: not implemented in metafor, so we have to do the computations manually

Sw <- 5.5 / sqrt(1 - 0.7^2)
d  <- (103 - 100) / Sw
Vd <- (50 + 50) * (1 - 0.7^2) / (50 * 50) + d^2 / (2*(50 + 50 - 2 - 1))
J  <- metafor:::.cmicalc(50 + 50 - 2 - 1)
g  <- J * d
Vg <- J^2 * Vd
round(g,  digits=4)
## [1] 0.3865

round(Vg, digits=4)
## [1] 0.0209

round(sqrt(Vg), digits=4)
## [1] 0.1444

# note: the results given in the book are not quite correct

Correlations

# r-to-z transformed correlation coefficient

dat <- escalc("ZCOR", ri=0.50, ni=100)
summary(dat)
##       yi     vi    sei     zi   pval  ci.lb  ci.ub 
## 1 0.5493 0.0103 0.1015 5.4100 <.0001 0.3503 0.7483

# back-transformation

c(transf.ztor(dat$yi))
## [1] 0.5

Effect Sizes for Comparing Risks

# risk difference

dat <- escalc("RD", ai=5, n1i=100, ci=10, n2i=100)
summary(dat)
##        yi     vi    sei      zi   pval   ci.lb  ci.ub 
## 1 -0.0500 0.0014 0.0371 -1.3484 0.1775 -0.1227 0.0227

# risk ratio (log transformed)

dat <- escalc("RR", ai=5, n1i=100, ci=10, n2i=100)
summary(dat)
##        yi     vi    sei      zi   pval   ci.lb  ci.ub 
## 1 -0.6931 0.2800 0.5292 -1.3099 0.1902 -1.7303 0.3440

# odds ratio (log transformed)

dat <- escalc("OR", ai=5, n1i=100, ci=10, n2i=100)
summary(dat)
##        yi     vi    sei      zi   pval   ci.lb  ci.ub 
## 1 -0.7472 0.3216 0.5671 -1.3175 0.1877 -1.8588 0.3643

# odds ratio (log transformed) for a case-control study

dat <- escalc("OR", ai=25, bi=20, ci=75, di=80)
summary(dat)
##       yi     vi    sei     zi   pval   ci.lb  ci.ub 
## 1 0.2877 0.1158 0.3403 0.8453 0.3980 -0.3794 0.9547

12) Statistically Analyzing Effect Sizes: Equal- and Random-Effects Models

# Table 12.1: Data for the Gender Differences in Conformity Example

dat <- read.table(header=TRUE, text = "
study group stdingrp nitems pmaleauth n d v
1  1 1  2 141 25  -0.330 0.029
2  1 2  2 119 25   0.070 0.034
3  2 1  2 191 50  -0.300 0.022
4  3 1 38 254 100  0.350 0.016
5  3 2 30  64 100  0.700 0.066
6  3 3 45  20 100  0.850 0.218
7  3 4 45  90 100  0.400 0.045
8  3 5 45  60 100  0.480 0.069
9  3 6  5  80 100  0.370 0.051
10 3 7  5 125 100 -0.060 0.032")

# note: including the 'percent male authors' variable from Table 12.3

dat
##    study group stdingrp nitems pmaleauth   n     d     v
## 1      1     1        1      2       141  25 -0.33 0.029
## 2      2     1        2      2       119  25  0.07 0.034
## 3      3     2        1      2       191  50 -0.30 0.022
## 4      4     3        1     38       254 100  0.35 0.016
## 5      5     3        2     30        64 100  0.70 0.066
## 6      6     3        3     45        20 100  0.85 0.218
## 7      7     3        4     45        90 100  0.40 0.045
## 8      8     3        5     45        60 100  0.48 0.069
## 9      9     3        6      5        80 100  0.37 0.051
## 10    10     3        7      5       125 100 -0.06 0.032

# equal-effects model analysis

res <- rma(d, v, data=dat, method="EE")
print(res, digits=3)
## Equal-Effects Model (k = 10)
## 
## I^2 (total heterogeneity / total variability):   71.68%
## H^2 (total variability / sampling variability):  3.53
## 
## Test for Heterogeneity:
## Q(df = 9) = 31.776, p-val < .001
## 
## Model Results:
## 
## estimate     se   zval   pval  ci.lb  ci.ub    
##    0.124  0.060  2.074  0.038  0.007  0.241  * 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Note: The book uses the term ‘fixed-effects model’, but I will use the term ‘equal-effects model’ throughout this document. My reasons for using the latter term are explained here.

# random-effects model analysis

res <- rma(d, v, data=dat, method="DL")
print(res, digits=3)
## Random-Effects Model (k = 10; tau^2 estimator: DL)
## 
## tau^2 (estimated amount of total heterogeneity): 0.094 (SE = 0.066)
## tau (square root of estimated tau^2 value):      0.307
## I^2 (total heterogeneity / total variability):   71.68%
## H^2 (total variability / sampling variability):  3.53
## 
## Test for Heterogeneity:
## Q(df = 9) = 31.776, p-val < .001
## 
## Model Results:
## 
## estimate     se   zval   pval   ci.lb  ci.ub    
##    0.189  0.118  1.594  0.111  -0.043  0.421    
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# note: unfortunately, the estimate of tau^2 was not computed correctly in the
# book (c=242.1138, not 269.798) and hence all of the results given in the left
# column on page 251 are incorrect
# fixed-effects ANOVA-type analysis

res <- rma(d, v, mods = ~ factor(group) - 1, data=dat, method="FE")
print(res, digits=3)
## Fixed-Effects with Moderators Model (k = 10)
## 
## I^2 (residual heterogeneity / unaccounted variability): 36.85%
## H^2 (unaccounted variability / sampling variability):   1.58
## 
## Test for Residual Heterogeneity:
## QE(df = 7) = 11.084, p-val = 0.135
## 
## Test of Moderators (coefficients 1:3):
## QM(df = 3) = 24.992, p-val < .001
## 
## Model Results:
## 
##                 estimate     se    zval   pval   ci.lb   ci.ub      
## factor(group)1    -0.146  0.125  -1.166  0.244  -0.391   0.099      
## factor(group)2    -0.300  0.148  -2.023  0.043  -0.591  -0.009    * 
## factor(group)3     0.339  0.077   4.421  <.001   0.189   0.490  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# note: by removing the intercept, the three coefficients directly provide the
# estimated average effect for the three groups
# weighted grand mean effect size

rma(coef(res), diag(vcov(res)), method="FE", digits=3)
## Fixed-Effects Model (k = 3)
## 
## I^2 (total heterogeneity / total variability):   90.33%
## H^2 (total variability / sampling variability):  10.35
## 
## Test for Heterogeneity:
## Q(df = 2) = 20.692, p-val < .001
## 
## Model Results:
## 
## estimate     se   zval   pval  ci.lb  ci.ub    
##    0.124  0.060  2.074  0.038  0.007  0.241  * 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# partitioning of the Q-statistics

res <- rma(d, v, mods = ~ factor(group), data=dat, method="FE")
res
## Fixed-Effects with Moderators Model (k = 10)
## 
## I^2 (residual heterogeneity / unaccounted variability): 36.85%
## H^2 (unaccounted variability / sampling variability):   1.58
## R^2 (amount of heterogeneity accounted for):            55.15%
## 
## Test for Residual Heterogeneity:
## QE(df = 7) = 11.0843, p-val = 0.1350
## 
## Test of Moderators (coefficients 2:3):
## QM(df = 2) = 20.6916, p-val < .0001
## 
## Model Results:
## 
##                 estimate      se     zval    pval    ci.lb   ci.ub      
## intrcpt          -0.1459  0.1251  -1.1660  0.2436  -0.3911  0.0993      
## factor(group)2   -0.1541  0.1940  -0.7943  0.4270  -0.5344  0.2262      
## factor(group)3    0.4851  0.1468   3.3053  0.0009   0.1975  0.7728  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# not removing the intercept, so the QM-statistic is equal to Q-between

round(res$QM, digits=3) # Q-between
## [1] 20.692

round(res$QE, digits=3) # Q-within
## [1] 11.084

# Q-within for each group

res1 <- rma(d, v, data=dat, method="FE", subset=group==1)
res2 <- rma(d, v, data=dat, method="FE", subset=group==2)
res3 <- rma(d, v, data=dat, method="FE", subset=group==3)

round(res1$QE, digits=3)
## [1] 2.54

round(res2$QE, digits=3) # 0.0004 in the book, but must be exactly 0 since k=1
## [1] 0

round(res3$QE, digits=3)
## [1] 8.545

# these add up to Q-within above

round(res1$QE + res2$QE + res3$QE, digits=3)
## [1] 11.084

# contrast between group 1 and 3

res <- rma(d, v, mods = ~ factor(group) - 1, data=dat, method="FE")
anova(res, X=c(-1,0,1), digits=3)
## Hypothesis:                                        
## 1: -factor(group)1 + factor(group)3 = 0 
## 
## Results:
##    estimate    se  zval  pval 
## 1:    0.485 0.147 3.305 <.001 
## 
## Test of Hypothesis:
## QM(df = 1) = 10.925, p-val < .001

predict(res, newmods=c(-1,0,1), digits=3)
##   pred    se ci.lb ci.ub 
##  0.485 0.147 0.197 0.773

# note: the results given in the book are slightly off
# mixed-effects model analysis

# distribution-free (method of moments) estimate of tau^2

res <- rma(d, v, mods = ~ factor(group), data=dat, method="DL")
round(res$tau2, digits=3)
## [1] 0.025

# maximum likelihood estimate of tau^2

res <- rma(d, v, mods = ~ factor(group), data=dat, method="ML")
round(res$tau2, digits=3)
## [1] 0

# restricted maximum likelihood estimate of tau^2

res <- rma(d, v, mods = ~ factor(group), data=dat, method="REML")
round(res$tau2, digits=3)
## [1] 0.027

# note: the REML estimate is incorrectly claimed to be 0 in the book
res <- rma(d, v, mods = ~ factor(group) - 1, data=dat, method="DL")
print(res, digits=3)
## Mixed-Effects Model (k = 10; tau^2 estimator: DL)
## 
## tau^2 (estimated amount of residual heterogeneity):     0.025 (SE = 0.037)
## tau (square root of estimated tau^2 value):             0.157
## I^2 (residual heterogeneity / unaccounted variability): 36.85%
## H^2 (unaccounted variability / sampling variability):   1.58
## 
## Test for Residual Heterogeneity:
## QE(df = 7) = 11.084, p-val = 0.135
## 
## Test of Moderators (coefficients 1:3):
## QM(df = 3) = 15.062, p-val = 0.002
## 
## Model Results:
## 
##                 estimate     se    zval   pval   ci.lb  ci.ub      
## factor(group)1    -0.139  0.168  -0.829  0.407  -0.467  0.190      
## factor(group)2    -0.300  0.216  -1.387  0.165  -0.724  0.124      
## factor(group)3     0.361  0.102   3.528  <.001   0.161  0.562  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# note: by removing the intercept, the three coefficients directly provide the
# estimated average effect for the three groups

# weighted grand mean effect size

rma(coef(res), diag(vcov(res)), method="FE", digits=3)
## Fixed-Effects Model (k = 3)
## 
## I^2 (total heterogeneity / total variability):   82.69%
## H^2 (total variability / sampling variability):  5.78
## 
## Test for Heterogeneity:
## Q(df = 2) = 11.557, p-val = 0.003
## 
## Model Results:
## 
## estimate     se   zval   pval   ci.lb  ci.ub    
##    0.152  0.081  1.872  0.061  -0.007  0.310  . 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# contrast between group 1 and 3

anova(res, X=c(-1,0,1), digits=3)
## Hypothesis:                                        
## 1: -factor(group)1 + factor(group)3 = 0 
## 
## Results:
##    estimate    se  zval  pval 
## 1:    0.500 0.196 2.547 0.011 
## 
## Test of Hypothesis:
## QM(df = 1) = 6.485, p-val = 0.011

predict(res, newmods=c(-1,0,1), digits=3)
##   pred    se ci.lb ci.ub pi.lb pi.ub 
##  0.500 0.196 0.115 0.885 0.007 0.993

# meta-regression model

res <- rma(d, v, mods = ~ I(log(nitems)), data=dat, method="FE")
print(res, digits=3)
## Fixed-Effects with Moderators Model (k = 10)
## 
## I^2 (residual heterogeneity / unaccounted variability): 7.66%
## H^2 (unaccounted variability / sampling variability):   1.08
## R^2 (amount of heterogeneity accounted for):            69.33%
## 
## Test for Residual Heterogeneity:
## QE(df = 8) = 8.664, p-val = 0.371
## 
## Test of Moderators (coefficient 2):
## QM(df = 1) = 23.112, p-val < .001
## 
## Model Results:
## 
##                 estimate     se    zval   pval   ci.lb   ci.ub      
## intrcpt           -0.323  0.111  -2.922  0.003  -0.540  -0.106   ** 
## I(log(nitems))     0.210  0.044   4.808  <.001   0.125   0.296  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# note: when doing transformations on predictors (such as taking the log) in
# the model formula, then we need to wrap this inside the I() function
# mixed-effects meta-regression model

# distribution-free (method of moments) estimate of tau^2

res <- rma(d, v, mods = ~ I(log(nitems)), data=dat, method="DL")
round(res$tau2, digits=3)
## [1] 0.003

# maximum likelihood estimate of tau^2

res <- rma(d, v, mods = ~ I(log(nitems)), data=dat, method="ML")
round(res$tau2, digits=3)
## [1] 0

# restricted maximum likelihood estimate of tau^2

res <- rma(d, v, mods = ~ I(log(nitems)), data=dat, method="REML")
round(res$tau2, digits=3)
## [1] 0.004

# note: the REML estimate is incorrectly claimed to be 0 in the book
# continuing with the distribution-free (method of moments) estimate of tau^2

res <- rma(d, v, mods = ~ I(log(nitems)), data=dat, method="DL")
print(res, digits=3)
## Mixed-Effects Model (k = 10; tau^2 estimator: DL)
## 
## tau^2 (estimated amount of residual heterogeneity):     0.003 (SE = 0.021)
## tau (square root of estimated tau^2 value):             0.057
## I^2 (residual heterogeneity / unaccounted variability): 7.66%
## H^2 (unaccounted variability / sampling variability):   1.08
## R^2 (amount of heterogeneity accounted for):            96.53%
## 
## Test for Residual Heterogeneity:
## QE(df = 8) = 8.664, p-val = 0.371
## 
## Test of Moderators (coefficient 2):
## QM(df = 1) = 20.821, p-val < .001
## 
## Model Results:
## 
##                 estimate     se    zval   pval   ci.lb   ci.ub      
## intrcpt           -0.321  0.117  -2.744  0.006  -0.550  -0.092   ** 
## I(log(nitems))     0.212  0.046   4.563  <.001   0.121   0.303  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# robust variance estimation

robust(res, cluster=study, digits=3)
## Mixed-Effects Model (k = 10; tau^2 estimator: DL)
## 
## tau^2 (estimated amount of residual heterogeneity):     0.003 (SE = 0.021)
## tau (square root of estimated tau^2 value):             0.057
## I^2 (residual heterogeneity / unaccounted variability): 7.66%
## H^2 (unaccounted variability / sampling variability):   1.08
## R^2 (amount of heterogeneity accounted for):            96.53%
## 
## Test for Residual Heterogeneity:
## QE(df = 8) = 8.664, p-val = 0.371
## 
## Number of estimates:   10
## Number of clusters:    10
## Estimates per cluster: 1
## 
## Test of Moderators (coefficient 2):¹
## F(df1 = 1, df2 = 8) = 25.315, p-val = 0.001
## 
## Model Results:
## 
##                 estimate     se¹    tval¹  df¹   pval¹   ci.lb¹   ci.ub¹     
## intrcpt           -0.321  0.124   -2.586    8   0.032   -0.607   -0.035    * 
## I(log(nitems))     0.212  0.042    5.031    8   0.001    0.115    0.309   ** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## 1) results based on cluster-robust inference (var-cov estimator: CR1,
##    approx t/F-tests and confidence intervals, df: residual method)

# note: the test statistic for log(nitems) is somewhat off in the book

13) Stochastically Dependent Effect Sizes

# Table 13.1

dat <- dat.kalaian1996[c(1:8,17:29,9:16,30:31),c(2,4:5,7,6,8)]
dat$study <- rep(1:26, times=rle(dat$study)$lengths)
names(dat) <- c("study", "nt", "nc", "d", "outcome", "v")
dat
##    study  nt  nc     d outcome      v
## 1      1  28  22  0.22  verbal 0.0817
## 2      2  39  40  0.09  verbal 0.0507
## 3      3  22  17  0.14  verbal 0.1045
## 4      4  48  43  0.14  verbal 0.0442
## 5      5  25  74 -0.01  verbal 0.0535
## 6      6  37  35  0.14  verbal 0.0557
## 7      7  24  70  0.18  verbal 0.0561
## 8      8  16  19  0.01  verbal 0.1151
## 17     9  43  37  0.01  verbal 0.0503
## 18    10  19  13  0.67  verbal 0.1366
## 19    11  16  11 -0.38  verbal 0.1561
## 20    12  20  12 -0.24  verbal 0.1342
## 21    13  39  28  0.29  verbal 0.0620
## 22    14  38  25  0.26    math 0.0669
## 23    15  18  13 -0.41    math 0.1352
## 24    16  19  13  0.08    math 0.1297
## 25    17  37  22  0.30    math 0.0732
## 26    18  19  11 -0.53    math 0.1482
## 27    19  17  13  0.13    math 0.1360
## 28    20  20  12  0.26    math 0.1344
## 29    21  20  13  0.47    math 0.1303
## 9     22 145 129  0.13  verbal 0.0147
## 10    22 145 129  0.12    math 0.0147
## 11    23  72 129  0.25  verbal 0.0218
## 12    23  72 129  0.06    math 0.0216
## 13    24  71 129  0.31  verbal 0.0221
## 14    24  71 129  0.09    math 0.0219
## 15    25  13  14  0.00  verbal 0.1484
## 16    25  13  14  0.07    math 0.1484
## 30    26  16  17  0.13  verbal 0.1216
## 31    26  16  17  0.48    math 0.1248

# note: this is a subset of the studies in dat.kalaian1996; instead of using
# dummy variable 'x' (where x=0 for the verbal and x=1 for the math subtest),
# we use the 'outcome' variable that has more descriptive labels
# construct variance-covariance matrices assuming rho = 0.7

V <- vcalc(v, cluster=study, type=outcome, rho=0.7, data=dat)

# var-cov matrix for studies 22 to 26

blsplit(V, cluster=dat$study, round, 4)[22:26]
## $`22`
##        [,1]   [,2]
## [1,] 0.0147 0.0103
## [2,] 0.0103 0.0147
## 
## $`23`
##        [,1]   [,2]
## [1,] 0.0218 0.0152
## [2,] 0.0152 0.0216
## 
## $`24`
##        [,1]   [,2]
## [1,] 0.0221 0.0154
## [2,] 0.0154 0.0219
## 
## $`25`
##        [,1]   [,2]
## [1,] 0.1484 0.1039
## [2,] 0.1039 0.1484
## 
## $`26`
##        [,1]   [,2]
## [1,] 0.1216 0.0862
## [2,] 0.0862 0.1248

# multivariate model with (correlated) random effects for both outcomes

res <- rma.mv(d, V, mods = ~ outcome - 1,
              random = ~ outcome | study, struct="UN", data=dat)
print(res, digits=3)
## Multivariate Meta-Analysis Model (k = 31; method: REML)
## 
## Variance Components:
## 
## outer factor: study   (nlvls = 26)
## inner factor: outcome (nlvls = 2)
## 
##            estim   sqrt  k.lvl  fixed   level 
## tau^2.1    0.001  0.025     13     no    math 
## tau^2.2    0.000  0.021     18     no  verbal 
## 
##         rho.math  rho.vrbl    math  vrbl 
## math           1                 -     5 
## verbal    -1.000         1      no     - 
## 
## Test for Residual Heterogeneity:
## QE(df = 29) = 19.902, p-val = 0.896
## 
## Test of Moderators (coefficients 1:2):
## QM(df = 2) = 9.417, p-val = 0.009
## 
## Model Results:
## 
##                estimate     se   zval   pval   ci.lb  ci.ub     
## outcomemath       0.089  0.059  1.511  0.131  -0.026  0.204     
## outcomeverbal     0.158  0.051  3.067  0.002   0.057  0.258  ** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# note: by removing the intercept, the two coefficients directly provide the
# estimated average effect for the two outcomes

# note: both variance components are very close to 0; as in the book, we
# proceed with a model where both variances are constrained to be equal to 0
# multivariate fixed-effects model

res <- rma.mv(d, V, mods = ~ outcome - 1, data=dat)
print(res, digits=3)
## Multivariate Meta-Analysis Model (k = 31; method: REML)
## 
## Variance Components: none
## 
## Test for Residual Heterogeneity:
## QE(df = 29) = 19.902, p-val = 0.896
## 
## Test of Moderators (coefficients 1:2):
## QM(df = 2) = 9.813, p-val = 0.007
## 
## Model Results:
## 
##                estimate     se   zval   pval   ci.lb  ci.ub     
## outcomemath       0.086  0.057  1.507  0.132  -0.026  0.198     
## outcomeverbal     0.159  0.051  3.131  0.002   0.059  0.258  ** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# fit model with varying values of rho

rhos <- c(0, 0.5, 0.6, 0.7, 0.8)
res <- list()
for (i in 1:length(rhos)) {
   V <- vcalc(v, cluster=study, type=outcome, rho=rhos[i], data=dat)
   res[[i]] <- rma.mv(d, V, mods = ~ outcome - 1, data=dat)
}

# Table 13.2

tab <- data.frame(rho  = rhos,
                  b1   = sapply(res, function(x) coef(x)[1]),
                  s1   = sapply(res, function(x) x$se[1]),
                  lcl1 = sapply(res, function(x) x$ci.lb[1]),
                  ucl1 = sapply(res, function(x) x$ci.ub[1]),
                  z1   = sapply(res, function(x) x$zval[1]),
                  b2   = sapply(res, function(x) coef(x)[2]),
                  s2   = sapply(res, function(x) x$se[2]),
                  lcl2 = sapply(res, function(x) x$ci.lb[2]),
                  ucl2 = sapply(res, function(x) x$ci.ub[2]),
                  z2   = sapply(res, function(x) x$zval[2]))
dfround(tab, digits=c(1,2,3,2,2,2,2,3,2,2,2))
##   rho   b1    s1  lcl1 ucl1   z1   b2    s2 lcl2 ucl2   z2
## 1 0.0 0.11 0.064 -0.01 0.24 1.77 0.15 0.053 0.05 0.26 2.93
## 2 0.5 0.09 0.060 -0.02 0.21 1.56 0.16 0.052 0.06 0.26 3.03
## 3 0.6 0.09 0.059 -0.03 0.21 1.53 0.16 0.051 0.06 0.26 3.08
## 4 0.7 0.09 0.057 -0.03 0.20 1.51 0.16 0.051 0.06 0.26 3.13
## 5 0.8 0.08 0.055 -0.03 0.19 1.49 0.16 0.050 0.06 0.26 3.20

# note: there are some printing errors in the table in the book;
# also the values given in the Z_2 column in the book are quite off
# robust variance estimation

V <- vcalc(v, cluster=study, type=outcome, rho=0.7, data=dat)
res <- rma.mv(d, V, mods = ~ outcome - 1, data=dat)
robust(res, cluster=study)
## Multivariate Meta-Analysis Model (k = 31; method: REML)
## 
## Variance Components: none
## 
## Test for Residual Heterogeneity:
## QE(df = 29) = 19.9016, p-val = 0.8960
## 
## Number of estimates:   31
## Number of clusters:    26
## Estimates per cluster: 1-2 (mean: 1.19, median: 1)
## 
## Test of Moderators (coefficients 1:2):¹
## F(df1 = 2, df2 = 24) = 12.1231, p-val = 0.0002
## 
## Model Results:
## 
##                estimate      se¹    tval¹  df¹    pval¹   ci.lb¹   ci.ub¹      
## outcomemath      0.0861  0.0383   2.2490   24   0.0340   0.0071   0.1651     * 
## outcomeverbal    0.1589  0.0362   4.3950   24   0.0002   0.0843   0.2336   *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## 1) results based on cluster-robust inference (var-cov estimator: CR1,
##    approx t/F-tests and confidence intervals, df: residual method)

# note: this is not the exact same approach that is described in the book; it
# uses a different weighting scheme and uses k-p for the degrees of freedom
# for the t-statistics; we can make use of the 'effective degrees of freedom'
# method described in the book with the 'clubSandwich' package
# install the 'clubSandwich' package (if it is not already installed) and load it

if (!require(clubSandwich))
   install.packages("clubSandwich")
robust(res, cluster=study, clubSandwich=TRUE)
## Multivariate Meta-Analysis Model (k = 31; method: REML)
## 
## Variance Components: none
## 
## Test for Residual Heterogeneity:
## QE(df = 29) = 19.9016, p-val = 0.8960
## 
## Number of estimates:   31
## Number of clusters:    26
## Estimates per cluster: 1-2 (mean: 1.19, median: 1)
## 
## Test of Moderators (coefficients 1:2):¹
## F(df1 = 2, df2 = 5.28) = 10.5212, p-val = 0.0144
## 
## Model Results:
## 
##                estimate      se¹    tval¹     df¹    pval¹    ci.lb¹   ci.ub¹     
## outcomemath      0.0861  0.0395   2.1819    6.67   0.0674   -0.0082   0.1803    . 
## outcomeverbal    0.1589  0.0370   4.2915   11.03   0.0013    0.0774   0.2404   ** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## 1) results based on cluster-robust inference (var-cov estimator: CR2,
##    approx t/F-tests and confidence intervals, df: Satterthwaite approx)

# note: but these are still not the same results as given in the book because
# of the different weighting scheme; we can get the same results as given in
# the book with the 'robumeta' package
# install the 'robumeta' package (if it is not already installed) and load it

if (!require(robumeta)) {
   install.packages("robumeta")
   library(robumeta)
}
robu(d ~ outcome - 1, data=dat, studynum=study, var.eff.size=v, rho=0.7, small=TRUE)
## RVE: Correlated Effects Model with Small-Sample Corrections 
## 
## Model: d ~ outcome - 1 
## 
## Number of studies = 26 
## Number of outcomes = 31 (min = 1 , mean = 1.19 , median = 1 , max = 2 )
## Rho = 0.7 
## I.sq = 0 
## Tau.sq = 0 
## 
##                 Estimate StdErr t-value   dfs P(|t|>)  95% CI.L 95% CI.U Sig
## 1   outcomemath    0.114 0.0489    2.32  7.69 0.05018 -0.000115    0.227   *
## 2 outcomeverbal    0.140 0.0338    4.12 13.09 0.00119  0.066450    0.213 ***
## ---
## Signif. codes: < .01 *** < .05 ** < .10 *
## ---
## Note: If df < 4, do not trust the results

# note: the value of rho actually has no influence on these results
# reproduce robumeta results using metafor and clubSandwich

vcalcfun <- function(v, tau2)
   diag((tau2 + mean(v)) * length(v), nrow=length(v), ncol=length(v))

V <- lapply(split(dat$v, dat$study), vcalcfun, tau2=0)
res <- rma.mv(d, V, mods = ~ outcome - 1, data=dat)
robust(res, cluster=study, clubSandwich=TRUE)
## Multivariate Meta-Analysis Model (k = 31; method: REML)
## 
## Variance Components: none
## 
## Test for Residual Heterogeneity:
## QE(df = 29) = 14.6868, p-val = 0.9873
## 
## Number of estimates:   31
## Number of clusters:    26
## Estimates per cluster: 1-2 (mean: 1.19, median: 1)
## 
## Test of Moderators (coefficients 1:2):¹
## F(df1 = 2, df2 = 10.54) = 11.0146, p-val = 0.0026
## 
## Model Results:
## 
##                estimate      se¹    tval¹     df¹    pval¹    ci.lb¹   ci.ub¹     
## outcomemath      0.1135  0.0489   2.3198    7.69   0.0502   -0.0001   0.2272    . 
## outcomeverbal    0.1395  0.0338   4.1220   13.09   0.0012    0.0664   0.2126   ** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## 1) results based on cluster-robust inference (var-cov estimator: CR2,
##    approx t/F-tests and confidence intervals, df: Satterthwaite approx)

# construct dataset with synthetic effects

agg <- aggregate(dat, by=list(dat$study), function(x) {
   if (is.character(x))
      paste(unique(x), collapse="/")
   else
      mean(x)
   })
agg$Group.1 <- agg$x <- NULL
agg
##    study  nt  nc      d     outcome      v
## 1      1  28  22  0.220      verbal 0.0817
## 2      2  39  40  0.090      verbal 0.0507
## 3      3  22  17  0.140      verbal 0.1045
## 4      4  48  43  0.140      verbal 0.0442
## 5      5  25  74 -0.010      verbal 0.0535
## 6      6  37  35  0.140      verbal 0.0557
## 7      7  24  70  0.180      verbal 0.0561
## 8      8  16  19  0.010      verbal 0.1151
## 9      9  43  37  0.010      verbal 0.0503
## 10    10  19  13  0.670      verbal 0.1366
## 11    11  16  11 -0.380      verbal 0.1561
## 12    12  20  12 -0.240      verbal 0.1342
## 13    13  39  28  0.290      verbal 0.0620
## 14    14  38  25  0.260        math 0.0669
## 15    15  18  13 -0.410        math 0.1352
## 16    16  19  13  0.080        math 0.1297
## 17    17  37  22  0.300        math 0.0732
## 18    18  19  11 -0.530        math 0.1482
## 19    19  17  13  0.130        math 0.1360
## 20    20  20  12  0.260        math 0.1344
## 21    21  20  13  0.470        math 0.1303
## 22    22 145 129  0.125 verbal/math 0.0147
## 23    23  72 129  0.155 verbal/math 0.0217
## 24    24  71 129  0.200 verbal/math 0.0220
## 25    25  13  14  0.035 verbal/math 0.1484
## 26    26  16  17  0.305 verbal/math 0.1232

# fit standard random-effects model to the dataset with synthetic effects

res <- rma(d, v, data=agg, method="DL")
print(res, digits=3)
## Random-Effects Model (k = 26; tau^2 estimator: DL)
## 
## tau^2 (estimated amount of total heterogeneity): 0 (SE = 0.018)
## tau (square root of estimated tau^2 value):      0
## I^2 (total heterogeneity / total variability):   0.00%
## H^2 (total variability / sampling variability):  1.00
## 
## Test for Heterogeneity:
## Q(df = 25) = 13.530, p-val = 0.969
## 
## Model Results:
## 
## estimate     se   zval   pval  ci.lb  ci.ub     
##    0.130  0.048  2.699  0.007  0.036  0.224  ** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

14) Bayesian Meta-Analysis

For these analyses, we could do most of what is described in the chapter using the excellent brms package. However, to fully reproduce all of the analyses conducted, we either need to use WinBUGS (as was done by the chapter authors) or we can use JAGS (which has the advantage that it runs not just under Windows, but also macOS / Mac OS X and Linux). Here, we will make use of the latter, which we can interact with directly from R via the rjags package. Note that JAGS needs to be installed separately (follow the link above for installation instructions).

# install the 'rjags' package (if it is not already installed) and load it

if (!require(rjags)) {
   install.packages("rjags")
   library(rjags)
}
# Table 14.1: Respiratory Tract Infections Data

dat <- dat.damico2009[c(1:8,16,9:15),-8]
dat$conceal <- 1 - dat$conceal
rownames(dat) <- 1:nrow(dat)
dat <- escalc(measure="OR", ai=xt, n1i=nt, ci=xc, n2i=nc, data=dat, digits=2)
dat
##             study year xt  nt  xc  nc conceal    yi   vi 
## 1      Abele-Horn 1997 13  58  23  30       1 -2.43 0.29 
## 2          Aerdts 1991  1  28  29  60       0 -3.23 1.10 
## 3           Blair 1991 12 161  38 170       0 -1.27 0.12 
## 4          Boland 1991 14  32  17  32       1 -0.38 0.25 
## 5       Cockerill 1992  4  75  12  75       1 -1.22 0.36 
## 6           Finch 1991  4  20   7  24       0 -0.50 0.51 
## 7          Jacobs 1992  0  45   4  46       1 -2.27 2.27 
## 8          Kerver 1988  5  49  31  47       1 -2.84 0.32 
## 9         Krueger 2002 91 265 149 262       0 -0.92 0.03 
## 10        Palomar 1997 10  50  25  49       1 -1.43 0.21 
## 11          Rocha 1992  7  47  25  54       1 -1.59 0.24 
## 12 Sanchez-Garcia 1992 32 131  60 140       0 -0.84 0.07 
## 13    Stoutenbeek 2007 62 201 100 200       0 -0.81 0.04 
## 14         Ulrich 1989  7  55  26  57       0 -1.75 0.23 
## 15       Verwaest 1997 22 193  40 185       0 -0.76 0.08 
## 16         Winter 1992  3  91  17  92       0 -1.89 0.42

# note: including 'conceal' variable (coded 0 = allocation concealment
# adequate, 1 = allocation concealment inadequate)
# Bayesian equal-effects model

k <- length(dat$yi)

jags.data <- list(yi=dat$yi, vi=dat$vi, k=k)

fe.model <- "model {
   for (i in 1:k) {
      yi[i] ~ dnorm(theta, 1/vi[i])
   }
   theta ~ dnorm(0, 0.0001)
}"

inits <- list(theta=0, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

# note: usually 'inits <- list(theta=0)' would have been sufficient; however,
# for full reproducibility, I am setting the random number generator and
# the seed for each chain (the same will be done below for other models)

model <- jags.model(textConnection(fe.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.ee <- coda.samples(model, variable.names=c("theta"), n.iter=100000, progress.bar="none")
dic.ee <- dic.samples(model, n.iter=100000, type="pD", progress.bar="none")
bma.ee <- summary(bma.ee)$quantiles[c(3,1,5)]
bma.ee <- rbind(theta=bma.ee, exp.theta = exp(bma.ee))
round(bma.ee, digits=2)
##             50%  2.5% 97.5%
## theta     -1.09 -1.27 -0.91
## exp.theta  0.34  0.28  0.40

# compare with results from a non-Bayesian analysis

res.ee <- rma(yi, vi, data=dat, method="EE", digits=2)
predict(res.ee, transf=exp)
##  pred ci.lb ci.ub 
##  0.34  0.28  0.40

# Bayesian random-effects model with uniform(0,2) prior for tau

re.model <- "model {
   for (i in 1:k) {
      yi[i] ~ dnorm(theta[i], 1/vi[i])
      theta[i] ~ dnorm(mu, 1/tau2)
   }
   mu ~ dnorm(0, 0.0001)
   tau ~ dunif(0, 2)
   tau2 <- tau^2
   theta.2 <- theta[2]
   theta.3 <- theta[3]
   theta.new ~ dnorm(mu, 1/tau2)
}"

inits <- list(theta=rep(0,k), theta.new=0, mu=0, tau=1, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

model <- jags.model(textConnection(re.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.re <- coda.samples(model, variable.names=c("mu","tau2","theta.2","theta.3","theta.new"),
                       n.iter=100000, progress.bar="none")
dic.re <- dic.samples(model, n.iter=100000, type="pD", progress.bar="none")
bma.re <- summary(bma.re)$quantiles[,c(3,1,5)]
bma.re <- rbind(mu=bma.re[1,], exp.mu=exp(bma.re[1,]), tau2=bma.re[2,],
                bma.re[3:5,], exp.theta.new=exp(bma.re[5,]))
round(bma.re[1:3,], digits=2)
##          50%  2.5% 97.5%
## mu     -1.29 -1.73 -0.94
## exp.mu  0.28  0.18  0.39
## tau2    0.27  0.01  1.09

# compare with results from a non-Bayesian analysis

res.re <- rma(yi, vi, data=dat, method="DL", digits=2)
pred.re <- predict(res.re, transf=exp)
pred.re
##  pred ci.lb ci.ub pi.lb pi.ub 
##  0.28  0.20  0.38  0.12  0.67

conf.re <- confint(res.re)
round(conf.re$random["tau^2",], digits=2)
## estimate    ci.lb    ci.ub 
##     0.18     0.04     1.20

# Bayesian random-effects model with half-normal(0,0.5^2) prior for tau

re.model <- "model {
   for (i in 1:k) {
      yi[i] ~ dnorm(theta[i], 1/vi[i])
      theta[i] ~ dnorm(mu, 1/tau2)
   }
   mu ~ dnorm(0, 0.0001)
   tau ~ dnorm(0, 1/0.5^2) T(0,)
   tau2 <- tau^2
}"

inits <- list(theta=rep(0,k), mu=0, tau=1, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

model <- jags.model(textConnection(re.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.hn <- coda.samples(model, variable.names=c("mu","tau2"), n.iter=100000, progress.bar="none")
bma.hn <- summary(bma.hn)$quantiles[,c(3,1,5)]
bma.hn <- rbind(mu=bma.hn[1,], exp.mu=exp(bma.hn[1,]), tau2=bma.hn[2,])
# Bayesian random-effects model with gamma(0.001,0.001) prior for 1/tau^2

re.model <- "model {
   for (i in 1:k) {
      yi[i] ~ dnorm(theta[i], 1/vi[i])
      theta[i] ~ dnorm(mu, prec)
   }
   mu ~ dnorm(0, 0.0001)
   prec ~ dgamma(0.001, 0.001)
   tau2 <- 1/prec
}"

inits <- list(theta=rep(0,k), mu=0, prec=1, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

model <- jags.model(textConnection(re.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.ig <- coda.samples(model, variable.names=c("mu","tau2"), n.iter=100000, progress.bar="none")
bma.ig <- summary(bma.ig)$quantiles[,c(3,1,5)]
bma.ig <- rbind(mu=bma.ig[1,], exp.mu=exp(bma.ig[1,]), tau2=bma.ig[2,])
# Table 14.2

tab <- rbind(
   c(or = pred.re$pred, ci.lb = pred.re$ci.lb, ci.ub = pred.re$ci.ub,
     tau2 = conf.re$random[1,1], ci.lb = conf.re$random[1,2], ci.ub = conf.re$random[1,3]),
   c(bma.re[2,], bma.re[3,]), c(bma.hn[2,], bma.hn[3,]), c(bma.ig[2,], bma.ig[3,]))
rownames(tab) <- c("Frequentist RE model",
                   "Bayesian RE model, uniform(0,2) prior for tau",
                   "Bayesian RE model, half-normal(0,0.5^2) prior for tau",
                   "Bayesian RE model, gamma(0.001,0.001) prior for 1/tau^2")
round(tab, digits=2)
##                                                           or ci.lb ci.ub tau2 ci.lb ci.ub
## Frequentist RE model                                    0.28  0.20  0.38 0.18  0.04  1.20
## Bayesian RE model, uniform(0,2) prior for tau           0.28  0.18  0.39 0.27  0.01  1.09
## Bayesian RE model, half-normal(0,0.5^2) prior for tau   0.28  0.19  0.39 0.19  0.01  0.70
## Bayesian RE model, gamma(0.001,0.001) prior for 1/tau^2 0.29  0.19  0.39 0.16  0.00  0.82

# Bayesian random-effects meta-regression model with uniform(0,2) prior for tau

jags.data <- list(yi=dat$yi, vi=dat$vi, xi=dat$conceal, k=k)

me.model <- "model {
   for (i in 1:k) {
      yi[i] ~ dnorm(theta[i], 1/vi[i])
      theta[i] ~ dnorm(beta0 + beta1 * xi[i], 1/tau2)
   }
   beta0 ~ dnorm(0, 0.0001)
   beta1 ~ dnorm(0, 0.0001)
   tau ~ dunif(0, 2)
   tau2 <- tau^2
   betasum <- beta0 + beta1
}"

inits <- list(theta=rep(0,k), beta0=0, beta1=1, tau=1, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

model <- jags.model(textConnection(me.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.me <- coda.samples(model, variable.names=c("beta0","beta1","tau2","betasum"),
                       n.iter=100000, progress.bar="none")
dic.me <- dic.samples(model, n.iter=100000, type="pD", progress.bar="none")
bma.me <- summary(bma.me)$quantiles[,c(3,1,5)]
bma.me <- rbind(bma.me[c(1,2,4),], exp.beta0=exp(bma.me[1,]), exp.betasum=exp(bma.me[3,]))
round(bma.me, digits=2)
##               50%  2.5% 97.5%
## beta0       -1.05 -1.55 -0.71
## beta1       -0.58 -1.22  0.16
## tau2         0.12  0.00  0.87
## exp.beta0    0.35  0.21  0.49
## exp.betasum  0.19  0.11  0.33

# compare with results from a non-Bayesian analysis

res.me <- rma(yi, vi, mods = ~ conceal, data=dat, method="DL", digits=2)
predict(res.me, newmods=c(0,1), transf=exp)
##   pred ci.lb ci.ub pi.lb pi.ub 
## 1 0.34  0.25  0.48  0.16  0.74 
## 2 0.20  0.12  0.32  0.08  0.45

round(confint(res.me)$random["tau^2",], digits=2)
## estimate    ci.lb    ci.ub 
##     0.12     0.00     1.23

# Table 14.3

tab <- rbind(c(sum(dic.ee$deviance), sum(dic.ee[[2]]), sum(dic.ee$deviance) + sum(dic.ee[[2]])),
             c(sum(dic.re$deviance), sum(dic.re[[2]]), sum(dic.re$deviance) + sum(dic.re[[2]])),
             c(sum(dic.me$deviance), sum(dic.me[[2]]), sum(dic.me$deviance) + sum(dic.me[[2]])))
colnames(tab) <- c("mean(D)", "p_D", "DIC")
rownames(tab) <- c("Equal-effect meta-analysis",
                   "Random-effects meta-analysis",
                   "Random-effects meta-regression")
round(tab, digits=1)
##                                mean(D) p_D  DIC
## Equal-effect meta-analysis        40.3 1.0 41.3
## Random-effects meta-analysis      24.4 9.7 34.1
## Random-effects meta-regression    26.5 8.3 34.7

# shrunken estimates for Aerdts (1991) and Blair (1991)

round(bma.re[4:5,], digits=2)
##           50%  2.5% 97.5%
## theta.2 -1.62 -3.01 -0.77
## theta.3 -1.26 -1.86 -0.71

# expected results for a new study

round(bma.re[6:7,], digits=2)
##                 50%  2.5% 97.5%
## theta.new     -1.27 -2.61 -0.10
## exp.theta.new  0.28  0.07  0.91

# Bayesian random-effects model for binary data

jags.data <- list(xt=dat$xt, nt=dat$nt, xc=dat$xc, nc=dat$nc, k=k)

re.model <- "model {
   for (i in 1:k) {
      xc[i] ~ dbin(pc[i], nc[i])
      xt[i] ~ dbin(pt[i], nt[i])
      logit(pc[i]) <- alpha[i] - theta[i]/2
      logit(pt[i]) <- alpha[i] + theta[i]/2
      theta[i] ~ dnorm(mu, 1/tau2)
      alpha[i] ~ dnorm(0, 0.0001)
   }
   mu ~ dnorm(0, 0.0001)
   tau ~ dunif(0, 2)
   tau2 <- tau^2
}"

inits <- list(theta=rep(0,k), alpha=rep(0,k), mu=0, tau=1, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

model <- jags.model(textConnection(re.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.re <- coda.samples(model, variable.names=c("mu","tau2"), n.iter=100000, progress.bar="none")
bma.re <- summary(bma.re)$quantiles[,c(3,1,5)]
bma.re <- rbind(mu=bma.re[1,], exp.mu=exp(bma.re[1,]), tau2=bma.re[2,])
round(bma.re, digits=2)
##          50%  2.5% 97.5%
## mu     -1.40 -1.88 -1.01
## exp.mu  0.25  0.15  0.36
## tau2    0.40  0.08  1.44

# Bayesian random-effects model with log-normal(-2.49,1.52^2) prior for tau^2

jags.data <- list(yi=dat$yi, vi=dat$vi, k=k)

re.model <- "model {
   for (i in 1:k) {
      yi[i] ~ dnorm(theta[i], 1/vi[i])
      theta[i] ~ dnorm(mu, 1/tau2)
   }
   mu ~ dnorm(0, 1.0E-4)
   tau2 ~ dlnorm(-2.49, 1/1.52^2)
}"

inits <- list(theta=rep(0,k), mu=0, tau2=1, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

model <- jags.model(textConnection(re.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.ip <- coda.samples(model, variable.names=c("mu","tau2"), n.iter=100000, progress.bar="none")
bma.ip <- summary(bma.ip)$quantiles[,c(3,1,5)]
bma.ip <- rbind(mu=bma.ip[1,], exp.mu=exp(bma.ip[1,]), tau2=bma.ip[2,])
round(bma.ip, digits=2)
##          50%  2.5% 97.5%
## mu     -1.25 -1.63 -0.95
## exp.mu  0.29  0.20  0.39
## tau2    0.15  0.01  0.62

# Bayesian random-effects meta-regression model with log-normal(-2.49,1.52^2) prior for tau^2

jags.data <- list(yi=dat$yi, vi=dat$vi, xi=dat$conceal, k=k)

me.model <- "model {
   for (i in 1:k) {
      yi[i] ~ dnorm(theta[i], 1/vi[i])
      theta[i] ~ dnorm(beta0 + beta1 * xi[i], 1/tau2)
   }
   beta0 ~ dnorm(0, 0.0001)
   beta1 ~ dnorm(0, 0.0001)
   tau2 ~ dlnorm(-2.49, 1/1.52^2)
   betasum <- beta0 + beta1
}"

inits <- list(theta=rep(0,k), beta0=0, beta1=1, tau2=1, .RNG.name="base::Mersenne-Twister")
inits <- list(inits, inits, inits)
inits[[1]]$.RNG.seed <- 12341
inits[[2]]$.RNG.seed <- 12342
inits[[3]]$.RNG.seed <- 12343

model <- jags.model(textConnection(me.model), inits=inits, data=jags.data, n.chains=3, quiet=TRUE)
update(model, n.iter=10000, progress.bar="none")
bma.me <- coda.samples(model, variable.names=c("beta0","beta1","tau2","betasum"),
                       n.iter=100000, progress.bar="none")
bma.me <- summary(bma.me)$quantiles[,c(3,1,5)]
bma.me <- rbind(bma.me[c(1,2,4),], exp.beta0=exp(bma.me[1,]), exp.betasum=exp(bma.me[3,]))
round(bma.me, digits=2)
##               50%  2.5% 97.5%
## beta0       -1.03 -1.42 -0.74
## beta1       -0.60 -1.16  0.02
## tau2         0.07  0.00  0.44
## exp.beta0    0.36  0.24  0.48
## exp.betasum  0.20  0.12  0.32

# Table 14.4: Recurrence of Violence Data

dat <- read.table(header=TRUE, text = "
study            year  xt   nt   xc   nc
'Bronx'          2005  20  202   11  218
'Brooklyn'       2000  13  129  100  386
'Broward'        2000  52  216   45  188
'San Diego Navy' 2000  63  218   75  214")
dat <- escalc(measure="OR", ai=xt, n1i=nt, ci=xc, n2i=nc, data=dat, digits=2)
dat
##            study year xt  nt  xc  nc    yi   vi 
## 1          Bronx 2005 20 202  11 218  0.73 0.15 
## 2       Brooklyn 2000 13 129 100 386 -1.14 0.10 
## 3        Broward 2000 52 216  45 188  0.01 0.05 
## 4 San Diego Navy 2000 63 218  75 214 -0.28 0.04

Using the same code as above, one can repeat all analyses with this dataset. Doing so yields the following results (code now shown; see cooper2019.rmd file for the code):

# Table 14.5

round(tab, digits=2)
##                                                               or ci.lb ci.ub tau2 ci.lb ci.ub
## Frequentist RE model                                        0.82  0.45  1.51 0.31  0.07  8.15
## Bayesian RE model, uniform(0,2) prior for tau               0.82  0.28  2.54 0.74  0.06  3.51
## Bayesian RE model, half-normal(0,0.5^2) prior for tau       0.82  0.41  1.67 0.30  0.02  1.28
## Bayesian RE model, gamma(0.001,0.001) prior for 1/tau^2     0.82  0.29  2.36 0.42  0.01  6.05
## Bayesian RE model, log-normal(-2.01,1.64^2) prior for tau^2 0.82  0.39  1.74 0.29  0.03  1.93

# note: the prior for the last model was log-normal(-2.01,1.64^2) (see page 311)
# and not log-normal(-3.95,1.79^2) as stated in the table

17) Missing Data in Meta-Analysis

# Table 17.1: Oral Anticoagulant Therapy Data

dat <- dat.anand1999

# reorder rows
dat <- dat[c(1:21,29,22:28,30:34),]
rownames(dat) <- 1:nrow(dat)

# add 'age' variable
dat$age <- 1999 - dat$year

# compute log odds ratios and corresponding sampling variances
dat <- escalc(measure="OR", ai=ai, n1i=n1i, ci=ci, n2i=n2i, data=dat, to="all", digits=2)
dat$yi <- round(dat$yi, 2)
dat$vi <- round(dat$vi, 2)

# flip signs of some of the log odds ratios (see note below)
sel <- c(4:6,11:14,16,19,20,22,24:27,29,30,32,33)
dat$yi[sel] <- -1 * dat$yi[sel]

# code intensity for study 22 as 'moderate' (see note below)
dat$intensity[22] <- "moderate"

# turn intensity into a factor with the desired order of levels
dat$intensity <- factor(dat$intensity, levels=c("low", "high", "moderate"))

# add mcar and mar dummy variables
dat$mcar <- c(0,1,0,1,1,1,1,0,1,0,0,1,0,1,1,1,1,0,1,0,1,0,1,1,1,1,1,1,1,1,1,0,1,1)
dat$mar  <- c(0,0,1,0,0,1,0,1,0,0,0,0,1,1,1,1,1,1,1,0,1,1,1,1,1,1,1,1,1,1,1,1,1,1)

# select (order of) columns
dat <- dat[c(11,12,3,2,10,13,14)]
dat
##       yi   vi intensity year age mcar mar 
## 1   3.02 2.21      high 1960  39    0   0 
## 2  -1.84 0.80      high 1960  39    1   0 
## 3   0.24 0.16      high 1961  38    0   1 
## 4   0.15 0.07      high 1961  38    1   0 
## 5   0.47 0.07      high 1964  35    1   0 
## 6   0.38 0.28      high 1964  35    1   1 
## 7  -0.38 0.18      high 1966  33    1   0 
## 8  -0.47 0.22      high 1967  32    0   1 
## 9  -0.25 0.07      high 1967  32    1   0 
## 10  0.22 0.10      high 1969  30    0   0 
## 11  0.84 0.08      high 1969  30    0   0 
## 12  0.35 0.04      high 1980  19    1   0 
## 13  0.33 0.02      high 1990   9    0   1 
## 14  0.12 0.01      high 1994   5    1   1 
## 15 -1.81 0.82      high 1969  30    1   1 
## 16  0.43 0.02      high 1974  25    1   1 
## 17  0.18 0.06      high 1980  19    1   1 
## 18  0.39 0.07      high 1980  19    0   1 
## 19  0.90 0.41      high 1993   6    1   1 
## 20  0.65 0.29      high 1996   3    0   0 
## 21  1.42 2.74      high 1990   9    1   1 
## 22  0.14 4.06  moderate 1990   9    0   1 
## 23  0.04 1.36  moderate 1982  17    1   1 
## 24  0.35 0.35  moderate 1981  18    1   1 
## 25  0.08 0.03  moderate 1982  17    1   1 
## 26  0.06 0.03  moderate 1969  30    1   1 
## 27  0.43 0.07  moderate 1964  35    1   1 
## 28 -1.16 0.93  moderate 1986  13    1   1 
## 29  0.75 0.98  moderate 1982  17    1   1 
## 30  0.81 0.60  moderate 1998   1    1   1 
## 31  0.04 0.82  moderate 1994   5    1   1 
## 32  0.35 0.70       low 1998   1    0   1 
## 33  0.34 0.06       low 1997   2    1   1 
## 34  0.15 0.02       low 1997   2    1   1

# note: compared to the original dataset, some of the log odds ratios have
# flipped signs; also, intensity is coded as 'moderate' for study 22, but
# this was actually 'high' in the original dataset; all of these differences
# are introduced above
# fit model to full data

res.full <- rma(yi, vi, mods = ~ intensity + age, data=dat)

# fit model to MCAR data

dat$age[dat$mcar == 0] <- NA
res.mcar <- rma(yi, vi, mods = ~ intensity + age, data=dat)

# fit model to MAR data

dat$age <- 1999 - dat$year
dat$age[dat$mar == 0] <- NA
res.mar  <- rma(yi, vi, mods = ~ intensity + age, data=dat)

# Table 17.2

tab <- cbind(full = coef(summary(res.full))[1:2],
             mcar = coef(summary(res.mcar))[1:2],
             mar  = coef(summary(res.mar))[1:2])
round(tab, digits=3)
##                   full.estimate full.se mcar.estimate mcar.se mar.estimate mar.se
## intrcpt                   0.213   0.142         0.214   0.153        0.193  0.122
## intensityhigh             0.060   0.180         0.014   0.209       -0.008  0.152
## intensitymoderate        -0.033   0.216        -0.011   0.239       -0.149  0.206
## age                      -0.001   0.005        -0.002   0.006        0.004  0.006

# fit model to MCAR data with mean imputation

dat$age <- 1999 - dat$year
dat$age[dat$mcar == 0] <- NA
dat$age <- replmiss(dat$age, mean(dat$age, na.rm=TRUE))
res.mcar <- rma(yi, vi, mods = ~ intensity + age, data=dat)

# fit model to MAR data with mean imputation

dat$age <- 1999 - dat$year
dat$age[dat$mar == 0] <- NA
dat$age <- replmiss(dat$age, mean(dat$age, na.rm=TRUE))
res.mar <- rma(yi, vi, mods = ~ intensity + age, data=dat)

# Table 17.3

tab <- cbind(mcar = coef(summary(res.mcar))[1:2],
             mar  = coef(summary(res.mar))[1:2])
round(tab, digits=3)
##                   mcar.estimate mcar.se mar.estimate mar.se
## intrcpt                   0.215   0.144        0.197  0.128
## intensityhigh             0.063   0.189        0.001  0.160
## intensitymoderate        -0.029   0.226       -0.132  0.216
## age                      -0.001   0.006        0.003  0.006

# fit model to MCAR data with regression imputation

dat$age <- 1999 - dat$year
dat$age[dat$mcar == 0] <- NA
fit <- lm(age ~ intensity + yi, data=dat)
dat$age <- replmiss(dat$age, predict(fit, newdata=dat))
res.mcar <- rma(yi, vi, mods = ~ intensity + age, data=dat)

# fit model to MAR data with regression imputation

dat$age <- 1999 - dat$year
dat$age[dat$mar == 0] <- NA
fit <- lm(age ~ intensity + yi, data=dat)
dat$age <- replmiss(dat$age, predict(fit, newdata=dat))
res.mar <- rma(yi, vi, mods = ~ intensity + age, data=dat)

# Table 17.4

tab <- cbind(mcar = coef(summary(res.mcar))[1:2],
             mar  = coef(summary(res.mar))[1:2])
round(tab, digits=3)
##                   mcar.estimate mcar.se mar.estimate mar.se
## intrcpt                   0.221   0.151        0.220  0.149
## intensityhigh             0.098   0.204        0.086  0.192
## intensitymoderate         0.012   0.238        0.013  0.243
## age                      -0.003   0.006       -0.003  0.007

# note: results in the table are different; not sure why
# install the 'mice' package (if it is not already installed) and load it

if (!require(mice)) {
   install.packages("mice")
   library(mice)
}
# multiple imputation using the MAR data

dat$age <- 1999 - dat$year
dat$age[dat$mar == 0] <- NA

predMatrix <- make.predictorMatrix(dat)
predMatrix[,] <- 0
predMatrix["age", c("yi", "intensity")] <- 1
predMatrix
##           yi vi intensity year age mcar mar
## yi         0  0         0    0   0    0   0
## vi         0  0         0    0   0    0   0
## intensity  0  0         0    0   0    0   0
## year       0  0         0    0   0    0   0
## age        1  0         1    0   0    0   0
## mcar       0  0         0    0   0    0   0
## mar        0  0         0    0   0    0   0

impMethod <- make.method(dat)
impMethod["age"] <- "norm"
impMethod
##        yi        vi intensity      year       age      mcar       mar 
##        ""        ""        ""        ""    "norm"        ""        ""

imp <- mice(dat, print=FALSE, m=100, predictorMatrix=predMatrix, method=impMethod, seed=1234)
fit <- with(imp, rma(yi, vi, mods = ~ intensity + age))
pool <- summary(pool(fit))

# Table 17.5

pool[-1] <- round(pool[-1], digits=3)
pool[1:3]
##                term estimate std.error
## 1         intercept    0.214     0.144
## 2     intensityhigh    0.068     0.186
## 3 intensitymoderate   -0.018     0.243
## 4               age   -0.002     0.007

# note: results in the table are different; this could be due to several reasons:
#
# 1) multiple imputation involves an element of randomness (above the seed of the
#    random number generator is set to some arbitrary value for reproducibility,
#    but a different seed would generate (slightly) different results)
# 2) the chapter author used the Amelia package instead of the mice package for
#    the analyses and there could be differences in how these packages implement
#    the multiple imputation algorithms
#
# the coefficients are actually not all that different, but the SEs shown in the
# chapter are much larger than the ones obtained above
#
# note: above m=100 imputations were used (instead of 10 as was done in the chapter)
# to obtain more stable results
#
# let's try this with the Amelia package
# install the 'Amelia' package (if it is not already installed) and load it

if (!require(Amelia)) {
   install.packages("Amelia")
   library(Amelia)
}
# multiple imputation using the MAR data

set.seed(1234)
invisible(capture.output(imp <- amelia(dat, m=100, idvars=c(2,4,6,7), noms=3, p2s=0)))

# the invisible(capture.output()) part is used here to suppress some output that
# is generated by the amelia() function, but in general this can be left out

fit <- lapply(imp$imputations,
   function(x) if (length(x) == 1L) NULL else rma(yi, vi, mods = ~ intensity + age, data=x))
fit <- fit[!sapply(fit, is.null)]

b  <- sapply(fit, function(x) coef(x))
se <- sapply(fit, function(x) x$se)
pool2 <- mi.meld(b, se, byrow=FALSE)
pool2 <- data.frame(estimate=pool2$q.mi[1,], se=pool2$se.mi[1,])
round(pool2, digits=3)
##                   estimate    se
## intrcpt              0.214 0.144
## intensityhigh        0.068 0.190
## intensitymoderate   -0.017 0.248
## age                 -0.002 0.007

# note: these results are very similar to the ones obtained above using the mice
# package and the SEs again differ substantially from those shown in Table 17.5

18) Publication Bias

Part of the code for this chapter (by Jack Vevea, Kathleen Coburn, and Alexander Sutton) is adapted from the chapter itself (see section 18.6).

Below is the code for the analysis of the irritable bowl syndrome (IBS) data.

# compute log risk ratios and corresponding sampling variances

dat <- escalc(measure="RR", ai=x.a, n1i=n.a, ci=x.p, n2i=n.p,
              data=dat.dorn2007, slab=paste(study, ", ", year))

# fit random-effects model using ML estimation

res <- rma(yi, vi, data=dat, digits=2, method="ML")
res
## Random-Effects Model (k = 19; tau^2 estimator: ML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.15 (SE = 0.07)
## tau (square root of estimated tau^2 value):      0.39
## I^2 (total heterogeneity / total variability):   72.90%
## H^2 (total variability / sampling variability):  3.69
## 
## Test for Heterogeneity:
## Q(df = 18) = 62.40, p-val < .01
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub      
##     0.42  0.11  3.77  <.01   0.20   0.63  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# funnel plot

par(mar=c(5,4,2,2))
funnel(res, atransf=exp, at=log(c(0.25, 0.5, 1, 2, 4, 8)),
       ylim=c(0,0.6), steps=7, las=1)

# cumulative meta-analysis (in order of precision) and corresponding forest plot

sav <- cumul(res, order=vi)
par(mar=c(5,4,2,2))
forest(sav, header=TRUE, xlim=c(-1,1.5))

An alternative way of plotting the results from such a cumulative meta-analysis is to plot the pooled estimates (in this case, the estimated average log risk ratios) against the estimates of the amount of heterogeneity as studies are added. The color gradient of the points/lines indicates the order of the cumulative results (light gray at the beginning, dark gray at the end).

# plot cumulative meta-analysis results

par(mar=c(5,4,2,2))
plot(sav, xlim=c(0.1,0.5), ylim=c(0,0.16), las=1)

# trim and fill analysis

tf.l.l0 <- trimfill(res, side="left",  estimator="L0")
tf.l.r0 <- trimfill(res, side="left",  estimator="R0")
tf.r.l0 <- trimfill(res, side="right", estimator="L0")
tf.r.r0 <- trimfill(res, side="right", estimator="R0")

# show results for side="left" and estimator="R0"

tf.l.r0
## Estimated number of missing studies on the left side: 1 (SE = 2.00)
## Test of H0: no missing studies on the left side:      p-val = 0.25
## 
## Random-Effects Model (k = 20; tau^2 estimator: ML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.18 (SE = 0.08)
## tau (square root of estimated tau^2 value):      0.43
## I^2 (total heterogeneity / total variability):   75.56%
## H^2 (total variability / sampling variability):  4.09
## 
## Test for Heterogeneity:
## Q(df = 19) = 69.35, p-val < .01
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub     
##     0.38  0.12  3.26  <.01   0.15   0.61  ** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# Egger regression using a weighted regression model with the standard error as predictor

regtest(res, model="lm")
## Regression Test for Funnel Plot Asymmetry
## 
## Model:     weighted regression with multiplicative dispersion
## Predictor: standard error
## 
## Test for Funnel Plot Asymmetry: t =  2.22, df = 17, p = 0.04
## Limit Estimate (as sei -> 0):   b = -0.23 (CI: -0.78, 0.32)

# Egger regression using a meta-regression model with the standard error as predictor

regtest(res)
## Regression Test for Funnel Plot Asymmetry
## 
## Model:     mixed-effects meta-regression model
## Predictor: standard error
## 
## Test for Funnel Plot Asymmetry: z =  2.64, p < .01
## Limit Estimate (as sei -> 0):   b = -0.23 (CI: -0.72, 0.27)

# The PET-PEESE procedure is this: Start with Egger regression using a weighted regression
# model with the standard error as predictor. If the limit estimate (i.e., intercept) is not
# significant, the adjusted estimate is the limit estimate. On the other hand, if the limit
# estimate is significant, repeat the Egger regression but using the sampling variances as
# predictor. The adjusted estimate is then the limit estimate from that model.

regtest(res, model="lm", ret.fit=TRUE)
## Regression Test for Funnel Plot Asymmetry
## 
## Model:     weighted regression with multiplicative dispersion
## Predictor: standard error
## 
## Call:
## lm(formula = yi ~ X - 1, weights = 1/vi)
## 
## Weighted Residuals:
##     Min      1Q  Median      3Q     Max 
## -2.5967 -1.6355  0.6988  1.4555  1.9691 
## 
## Coefficients:
##          Estimate Std. Error t value Pr(>|t|)  
## Xintrcpt  -0.2313     0.2612  -0.885   0.3883  
## Xsei       2.4617     1.1087   2.220   0.0403 *
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 1.687 on 17 degrees of freedom
## Multiple R-squared:  0.4949, Adjusted R-squared:  0.4355 
## F-statistic: 8.329 on 2 and 17 DF,  p-value: 0.00301
## 
## Test for Funnel Plot Asymmetry: t =  2.22, df = 17, p = 0.04
## Limit Estimate (as sei -> 0):   b = -0.23 (CI: -0.78, 0.32)

# note: since the intercept is not significant (p = .39), the adjusted estimate is the limit
# estimate; but just for illustration purposes, if the intercept had been significant, then
# the adjusted estimate would be the limit estimate from the following model

regtest(res, model="lm", predictor="vi")
## Regression Test for Funnel Plot Asymmetry
## 
## Model:     weighted regression with multiplicative dispersion
## Predictor: sampling variance
## 
## Test for Funnel Plot Asymmetry: t = 2.13, df = 17, p = 0.05
## Limit Estimate (as vi -> 0):    b = 0.09 (CI: -0.20, 0.38)

# rank correlation test

ranktest(res)
## Rank Correlation Test for Funnel Plot Asymmetry
## 
## Kendall's tau = 0.26, p = 0.12

For a p-curve analysis, one needs to use the online app at http://p-curve.com/.

# install the 'puniform' package (if it is not already installed) and load it

if (!require(puniform)) {
   install.packages("puniform")
   library(puniform)
}
# p-uniform

par(mar=c(5,4,2,2))
puniform(yi=dat$yi, vi=dat$vi, side="right", plot=TRUE)
## 
## Method: P
## 
## Effect size estimation p-uniform
## 
##        est     ci.lb     ci.ub       L.0    pval      ksig
##     0.6439    0.4067    0.9005     -4.03   <.001        10
## 
## ===
## 
## Publication bias test p-uniform
## 
##       L.pb      pval
##    -2.6085    0.9955
## 
## ===
## 
## Fixed-effect meta-analysis
## 
##     est.fe     se.fe   zval.fe pval.fe  ci.lb.fe  ci.ub.fe     Qstat   Qpval
##     0.3123    0.0541     5.777   <.001    0.2063    0.4182   62.4013   <.001

# excess significance test

tes(yi, vi, data=dat)
## Test of Excess Significance
## 
## Observed Number of Significant Findings: 10 (out of 19)
## Expected Number of Significant Findings: 4.9233
## Observed Number / Expected Number:       2.0311
## 
## Estimated Power of Tests (based on theta = 0.3123)
## 
##    min      q1  median      q3     max 
## 0.0894  0.1270  0.2118  0.3537  0.5397 
## 
## Test of Excess Significance: p = 0.0078 (exact test)
## Limit Estimate (theta_lim):  0.3971 (where p = 0.1)

# install the 'selectMeta' package (if it is not already installed) and load it

if (!require(selectMeta)) {
   install.packages("selectMeta")
   library(selectMeta)
}
# Dear and Begg weight-function model

sav <- DearBegg(dat$yi, sqrt(dat$vi), trace=FALSE)

par(mar=c(5,4,2,2))
plot(0, 0, type="n", xlim=c(0, 1), ylim=c(0, 1), xlab="p-Values", ylab="Estimated Weight Function")
ps <- seq(0, 1, by = 0.01)
pval <- summary(dat)$pval
rug(pval, lwd=3)
weightLine(pval, w=sav$w, col0="black", lwd0=1, lty0="solid")

# Rufibach weight-function model

sav <- DearBeggMonotone(dat$yi, sqrt(dat$vi), trace=FALSE)

par(mar=c(5,4,2,2))
plot(0, 0, type="n", xlim=c(0, 1), ylim=c(0, 1), xlab="p-Values", ylab="Estimated Weight Function")
ps <- seq(0, 1, by = 0.01)
pval <- summary(dat)$pval
rug(pval, lwd=3)
weightLine(pval, w=sav$w, col0="black", lwd0=1, lty0="solid")

# Vevea and Hedges (1995) model (with a single cutpoint at p=0.025)

selmodel(res, type="stepfun", steps=0.025)
## Random-Effects Model (k = 19; tau^2 estimator: ML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.09 (SE = 0.07)
## tau (square root of estimated tau^2 value):      0.31
## 
## Test for Heterogeneity:
## LRT(df = 1) = 6.10, p-val = 0.01
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub    
##     0.16  0.15  1.01  0.31  -0.14   0.45    
## 
## Test for Selection Model Parameters:
## LRT(df = 1) = 3.59, p-val = 0.06
## 
## Selection Model Results:
## 
##                      k  estimate    se   zval  pval  ci.lb  ci.ub      
## 0     < p <= 0.025  10      1.00   ---    ---   ---    ---    ---      
## 0.025 < p <= 1       9      0.19  0.18  -4.57  <.01   0.00   0.54  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# Table 18.1: Sample Selection Patterns for the Vevea and Woods Method

ssp <- data.frame(
 steps = c(0.005, 0.01, 0.05, 0.10, 0.25, 0.35, 0.50, 0.65, 0.75, 0.90, 0.95, 0.99, 0.995, 1),
 weights.mod.1 = c(1, 0.99, 0.95, 0.80, 0.75, 0.65, 0.60, 0.55, 0.50, 0.50, 0.50, 0.50, 0.50, 0.50),
 weights.sev.1 = c(1, 0.99, 0.90, 0.75, 0.60, 0.50, 0.40, 0.35, 0.30, 0.25, 0.10, 0.10, 0.10, 0.10),
 weights.mod.2 = c(1, 0.99, 0.95, 0.90, 0.80, 0.75, 0.60, 0.60, 0.75, 0.80, 0.90, 0.95, 0.99, 1.00),
 weights.sev.2 = c(1, 0.99, 0.90, 0.75, 0.60, 0.50, 0.25, 0.25, 0.50, 0.60, 0.75, 0.90, 0.99, 1.00))
ssp
##    steps weights.mod.1 weights.sev.1 weights.mod.2 weights.sev.2
## 1  0.005          1.00          1.00          1.00          1.00
## 2  0.010          0.99          0.99          0.99          0.99
## 3  0.050          0.95          0.90          0.95          0.90
## 4  0.100          0.80          0.75          0.90          0.75
## 5  0.250          0.75          0.60          0.80          0.60
## 6  0.350          0.65          0.50          0.75          0.50
## 7  0.500          0.60          0.40          0.60          0.25
## 8  0.650          0.55          0.35          0.60          0.25
## 9  0.750          0.50          0.30          0.75          0.50
## 10 0.900          0.50          0.25          0.80          0.60
## 11 0.950          0.50          0.10          0.90          0.75
## 12 0.990          0.50          0.10          0.95          0.90
## 13 0.995          0.50          0.10          0.99          0.99
## 14 1.000          0.50          0.10          1.00          1.00

# Vevea and Wood (2005) model using moderate one-tailed selection

selmodel(res, type="stepfun", steps=ssp$steps, delta=ssp$weights.mod.1)
## Random-Effects Model (k = 19; tau^2 estimator: ML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.16 (SE = 0.09)
## tau (square root of estimated tau^2 value):      0.40
## 
## Test for Heterogeneity:
## LRT(df = 1) = 17.74, p-val < .01
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub     
##     0.31  0.12  2.61  <.01   0.08   0.54  ** 
## 
## Test for Selection Model Parameters:
## LRT(df = 0) = 0.00, p-val = NA
## 
## Selection Model Results:
## 
##                     k  estimate   se  zval  pval  ci.lb  ci.ub   
## 0     < p <= 0.005  7      1.00  ---   ---   ---    ---    ---   
## 0.005 < p <= 0.01   2      0.99  ---   ---   ---    ---    ---   
## 0.01  < p <= 0.05   2      0.95  ---   ---   ---    ---    ---   
## 0.05  < p <= 0.1    0      0.80  ---   ---   ---    ---    ---   
## 0.1   < p <= 0.25   2      0.75  ---   ---   ---    ---    ---   
## 0.25  < p <= 0.35   0      0.65  ---   ---   ---    ---    ---   
## 0.35  < p <= 0.5    0      0.60  ---   ---   ---    ---    ---   
## 0.5   < p <= 0.65   2      0.55  ---   ---   ---    ---    ---   
## 0.65  < p <= 0.75   1      0.50  ---   ---   ---    ---    ---   
## 0.75  < p <= 0.9    3      0.50  ---   ---   ---    ---    ---   
## 0.9   < p <= 0.95   0      0.50  ---   ---   ---    ---    ---   
## 0.95  < p <= 0.99   0      0.50  ---   ---   ---    ---    ---   
## 0.99  < p <= 0.995  0      0.50  ---   ---   ---    ---    ---   
## 0.995 < p <= 1      0      0.50  ---   ---   ---    ---    ---   
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# Vevea and Wood (2005) model using severe one-tailed selection

selmodel(res, type="stepfun", steps=ssp$steps, delta=ssp$weights.sev.1)
## Random-Effects Model (k = 19; tau^2 estimator: ML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.23 (SE = 0.14)
## tau (square root of estimated tau^2 value):      0.48
## 
## Test for Heterogeneity:
## LRT(df = 1) = 19.45, p-val < .01
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub    
##     0.11  0.18  0.57  0.57  -0.26   0.47    
## 
## Test for Selection Model Parameters:
## LRT(df = 0) = 0.66, p-val = NA
## 
## Selection Model Results:
## 
##                     k  estimate   se  zval  pval  ci.lb  ci.ub   
## 0     < p <= 0.005  7      1.00  ---   ---   ---    ---    ---   
## 0.005 < p <= 0.01   2      0.99  ---   ---   ---    ---    ---   
## 0.01  < p <= 0.05   2      0.90  ---   ---   ---    ---    ---   
## 0.05  < p <= 0.1    0      0.75  ---   ---   ---    ---    ---   
## 0.1   < p <= 0.25   2      0.60  ---   ---   ---    ---    ---   
## 0.25  < p <= 0.35   0      0.50  ---   ---   ---    ---    ---   
## 0.35  < p <= 0.5    0      0.40  ---   ---   ---    ---    ---   
## 0.5   < p <= 0.65   2      0.35  ---   ---   ---    ---    ---   
## 0.65  < p <= 0.75   1      0.30  ---   ---   ---    ---    ---   
## 0.75  < p <= 0.9    3      0.25  ---   ---   ---    ---    ---   
## 0.9   < p <= 0.95   0      0.10  ---   ---   ---    ---    ---   
## 0.95  < p <= 0.99   0      0.10  ---   ---   ---    ---    ---   
## 0.99  < p <= 0.995  0      0.10  ---   ---   ---    ---    ---   
## 0.995 < p <= 1      0      0.10  ---   ---   ---    ---    ---   
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# Vevea and Wood (2005) model using moderate two-tailed selection

selmodel(res, type="stepfun", steps=ssp$steps, delta=ssp$weights.mod.2)
## Random-Effects Model (k = 19; tau^2 estimator: ML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.14 (SE = 0.07)
## tau (square root of estimated tau^2 value):      0.37
## 
## Test for Heterogeneity:
## LRT(df = 1) = 16.73, p-val < .01
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub      
##     0.37  0.11  3.52  <.01   0.16   0.58  *** 
## 
## Test for Selection Model Parameters:
## LRT(df = 0) = 0.24, p-val = NA
## 
## Selection Model Results:
## 
##                     k  estimate   se  zval  pval  ci.lb  ci.ub   
## 0     < p <= 0.005  7      1.00  ---   ---   ---    ---    ---   
## 0.005 < p <= 0.01   2      0.99  ---   ---   ---    ---    ---   
## 0.01  < p <= 0.05   2      0.95  ---   ---   ---    ---    ---   
## 0.05  < p <= 0.1    0      0.90  ---   ---   ---    ---    ---   
## 0.1   < p <= 0.25   2      0.80  ---   ---   ---    ---    ---   
## 0.25  < p <= 0.35   0      0.75  ---   ---   ---    ---    ---   
## 0.35  < p <= 0.5    0      0.60  ---   ---   ---    ---    ---   
## 0.5   < p <= 0.65   2      0.60  ---   ---   ---    ---    ---   
## 0.65  < p <= 0.75   1      0.75  ---   ---   ---    ---    ---   
## 0.75  < p <= 0.9    3      0.80  ---   ---   ---    ---    ---   
## 0.9   < p <= 0.95   0      0.90  ---   ---   ---    ---    ---   
## 0.95  < p <= 0.99   0      0.95  ---   ---   ---    ---    ---   
## 0.99  < p <= 0.995  0      0.99  ---   ---   ---    ---    ---   
## 0.995 < p <= 1      0      1.00  ---   ---   ---    ---    ---   
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# Vevea and Wood (2005) model using severe two-tailed selection

selmodel(res, type="stepfun", steps=ssp$steps, delta=ssp$weights.sev.2)
## Random-Effects Model (k = 19; tau^2 estimator: ML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.11 (SE = 0.06)
## tau (square root of estimated tau^2 value):      0.34
## 
## Test for Heterogeneity:
## LRT(df = 1) = 14.29, p-val < .01
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub      
##     0.32  0.10  3.30  <.01   0.13   0.51  *** 
## 
## Test for Selection Model Parameters:
## LRT(df = 0) = 0.00, p-val = NA
## 
## Selection Model Results:
## 
##                     k  estimate   se  zval  pval  ci.lb  ci.ub   
## 0     < p <= 0.005  7      1.00  ---   ---   ---    ---    ---   
## 0.005 < p <= 0.01   2      0.99  ---   ---   ---    ---    ---   
## 0.01  < p <= 0.05   2      0.90  ---   ---   ---    ---    ---   
## 0.05  < p <= 0.1    0      0.75  ---   ---   ---    ---    ---   
## 0.1   < p <= 0.25   2      0.60  ---   ---   ---    ---    ---   
## 0.25  < p <= 0.35   0      0.50  ---   ---   ---    ---    ---   
## 0.35  < p <= 0.5    0      0.25  ---   ---   ---    ---    ---   
## 0.5   < p <= 0.65   2      0.25  ---   ---   ---    ---    ---   
## 0.65  < p <= 0.75   1      0.50  ---   ---   ---    ---    ---   
## 0.75  < p <= 0.9    3      0.60  ---   ---   ---    ---    ---   
## 0.9   < p <= 0.95   0      0.75  ---   ---   ---    ---    ---   
## 0.95  < p <= 0.99   0      0.90  ---   ---   ---    ---    ---   
## 0.99  < p <= 0.995  0      0.99  ---   ---   ---    ---    ---   
## 0.995 < p <= 1      0      1.00  ---   ---   ---    ---    ---   
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# install the 'metasens' package (if it is not already installed) and load it

if (!require(metasens)) {
   install.packages("metasens")
   library(metasens)
}
# Copas and Shi (2001) model

res <- metagen(yi, sqrt(vi), method.tau="ML", data=dat)
sav <- copas(res)
plot(sav)

summary(sav)
## Copas selection model analysis
## 
##               p.publ                   95%-CI  tau^2    tau  p.trt  p.rsb  N
##               1.0000 0.4176 [ 0.1966; 0.6386] 0.1534 0.3917 0.0002 0.0144  0
##               0.9612 0.3999 [ 0.1903; 0.6095] 0.1391 0.3729 0.0002 0.0225  0
##               0.8468 0.3501 [ 0.1535; 0.5466] 0.1197 0.3460 0.0005 0.0541  2
##               0.7417 0.3004 [ 0.1295; 0.4714] 0.1117 0.3342 0.0006 0.0997  3
##               0.6475 0.2506 [-0.1802; 0.6813] 0.1097 0.3313 0.2542 0.1504  6
##               0.5432 0.1925 [-0.0460; 0.4310] 0.1115 0.3339 0.1137 0.1957  9
##               0.4352 0.1511 [-0.1019; 0.4041] 0.1083 0.3291 0.2418 0.2460 15
##               0.3288 0.1026 [-0.1695; 0.3746] 0.1047 0.3236 0.4599 0.3191 24
##                                                                             
##    Adjusted estimate 0.2506 [-0.1802; 0.6813] 0.1097 0.3313 0.2542 0.1504  6
##  Unadjusted estimate 0.4176 [ 0.2003; 0.6348] 0.1534 0.3917 0.0002          
## 
## Significance level for test of residual selection bias: 0.1 
## 
##                        min    max
##  range of gamma0:  -0.7558 2.0000
##  range of gamma1:   0.0000 0.2723
## 
## Largest standard error (SE): 0.5383 
## 
## Range of probability publishing trial with largest SE:
##     min    max
##  0.2249 0.9939
## 
## Calculation of orthogonal line:
## 
##  level nobs adj.r.square     slope    se.slope
##  -0.05    5    0.9999830 -1.362614 0.002809321
##   0.00    9    0.9998969 -1.440954 0.005173595
##   0.05   14    0.9994926 -1.513392 0.009457356
##   0.10   19    0.9925764 -1.774954 0.036173183
##   0.15   24    0.9886879 -2.100479 0.046836982
##   0.20   28    0.9914366 -2.508877 0.044866233
##   0.25   26    0.9997682 -2.980742 0.009078334
##   0.30   26    0.9995475 -3.081597 0.013113046
##   0.35   26    0.9993143 -3.196848 0.016748454
##   0.40   24    0.9990672 -3.426336 0.021830322
## 
##  Legend:
##  p.publ - Probability of publishing study with largest SE
##  p.trt  - P-value for test of overall treatment effect
##  p.rsb  - P-value for test of residual selection bias
##  N      - Estimated number of unpublished studies

# Rücker limit meta-analysis

sav <- limitmeta(res)
sav
## Result of limit meta-analysis:
## 
##  Random effects model                   95%-CI    z   pval
##     Adjusted estimate 0.0854 [-0.2800; 0.4508] 0.46 0.6469
##   Unadjusted estimate 0.4176 [ 0.2003; 0.6348] 3.77 0.0002
## 
## Quantifying heterogeneity:
## tau^2 = 0.1534; I^2 = 71.2% [54.1%; 81.9%]; G^2 = 99.9%
## 
## Test of heterogeneity:
##      Q d.f.  p-value
##  62.40   18 < 0.0001
## 
## Test of small-study effects:
##   Q-Q' d.f.  p-value
##  14.03    1   0.0002
## 
## Test of residual heterogeneity beyond small-study effects:
##     Q' d.f.  p-value
##  48.37   17 < 0.0001
## 
## Details on adjustment method:
## - expectation (beta0)

19) Interpreting Effect Sizes

# converting d-values to U1, U2, and U3

transf.dtou1(1.0)
## [1] 0.5537899

transf.dtou2(1.0)
## [1] 0.6914625

transf.dtou3(1.0)
## [1] 0.8413447

# converting d-values to common language effect size values

transf.dtocles(0.3)
## [1] 0.583998

# converting d-values to binomial effect-size display values

transf.dtobesd(0.3)
## [1] 0.5741702

# note: this is the proportion in the first group scoring above the median; the proportion
# in the second group scoring above the median is simply 1 minus the computed value:

1 - transf.dtobesd(0.3)
## [1] 0.4258298

# converting a 2x2 table to a correlation (phi coefficient)

escalc(measure="PHI", ai=100, bi=50, ci=60, di=90)
##       yi     vi 
## 1 0.2673 0.0031

# converting a 2x2 table to an odds ratio

summary(escalc(measure="OR", ai=100, bi=50, ci=60, di=90), transf=exp)
##       yi  ci.lb  ci.ub 
## 1 3.0000 1.8729 4.8053

# note: escalc(measure="OR", ...) computes log odds ratios, so need to exponentiate
# converting a 2x2 table to a risk ratio

summary(escalc(measure="RR", ai=100, bi=50, ci=60, di=90), transf=exp)
##       yi  ci.lb  ci.ub 
## 1 1.6667 1.3291 2.0900

# note: escalc(measure="RR", ...) computes log risk ratios, so need to exponentiate
# converting a 2x2 table to a risk difference

escalc(measure="RD", ai=100, bi=50, ci=60, di=90)
##       yi     vi 
## 1 0.2667 0.0031

# converting a 2x2 table to a number needed to treat

summary(escalc(measure="RD", ai=100, bi=50, ci=60, di=90), transf=function(x) 1/x)
##       yi  ci.lb  ci.ub 
## 1 3.7500 2.6634 6.3344

# using effect-size translations with confidence intervals (after a meta-analysis)

# note: using some made-up data that yield the same results as shown on page 443

# example with standardized mean differences

dat <- data.frame(di = c(0.03, 0.93, 0.66),
                  vi = c(0.072, 0.063, 0.047))
res <- rma(di, vi, data=dat, digits=2)
res
## Random-Effects Model (k = 3; tau^2 estimator: REML)
## 
## tau^2 (estimated amount of total heterogeneity): 0.14 (SE = 0.20)
## tau (square root of estimated tau^2 value):      0.37
## I^2 (total heterogeneity / total variability):   69.70%
## H^2 (total variability / sampling variability):  3.30
## 
## Test for Heterogeneity:
## Q(df = 2) = 6.28, p-val = 0.04
## 
## Model Results:
## 
## estimate    se  zval  pval  ci.lb  ci.ub    
##     0.55  0.26  2.15  0.03   0.05   1.05  * 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

predict(res, transf=transf.dtou3)
##  pred ci.lb ci.ub pi.lb pi.ub 
##  0.71  0.52  0.85  0.37  0.92

# example with (log) odds ratios

dat <- data.frame(logori = c(0.40, 0.74, 0.99, 0.84, 0.75),
                  vi = c(0.111, 0.093, 0.165, 0.128, 0.114))
res <- rma(logori, vi, data=dat, digits=3)
res
## Random-Effects Model (k = 5; tau^2 estimator: REML)
## 
## tau^2 (estimated amount of total heterogeneity): 0 (SE = 0.083)
## tau (square root of estimated tau^2 value):      0
## I^2 (total heterogeneity / total variability):   0.00%
## H^2 (total variability / sampling variability):  1.00
## 
## Test for Heterogeneity:
## Q(df = 4) = 1.488, p-val = 0.829
## 
## Model Results:
## 
## estimate     se   zval   pval  ci.lb  ci.ub      
##    0.724  0.154  4.715  <.001  0.423  1.025  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

predict(res, transf=exp, digits=2)
##  pred ci.lb ci.ub pi.lb pi.ub 
##  2.06  1.53  2.79  1.53  2.79

predict(res, transf=function(x) transf.lnortord(x, pc=0.12), digits=2)
##  pred ci.lb ci.ub pi.lb pi.ub 
##  0.10  0.05  0.16  0.05  0.16

See also the R functions provided by the chapter authors (Jeffrey Valentine, Ariel Aloe, and Sandra Wilson) at the end of the chapter. The code is also available on the publisher’s website for the book (see here).

20) Heterogeneity in Meta-Analysis

# data for the illustrative example

dat <- data.frame(yi = seq(10, 90, by=10), vi = 562.5)
dat
##   yi    vi
## 1 10 562.5
## 2 20 562.5
## 3 30 562.5
## 4 40 562.5
## 5 50 562.5
## 6 60 562.5
## 7 70 562.5
## 8 80 562.5
## 9 90 562.5

# fit random-effects model

res <- rma(yi, vi, data=dat, method="DL")
res
## Random-Effects Model (k = 9; tau^2 estimator: DL)
## 
## tau^2 (estimated amount of total heterogeneity): 187.5000 (SE = 375.0000)
## tau (square root of estimated tau^2 value):      13.6931
## I^2 (total heterogeneity / total variability):   25.00%
## H^2 (total variability / sampling variability):  1.33
## 
## Test for Heterogeneity:
## Q(df = 8) = 10.6667, p-val = 0.2213
## 
## Model Results:
## 
## estimate      se    zval    pval    ci.lb    ci.ub      
##  50.0000  9.1287  5.4772  <.0001  32.1081  67.8919  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# forest plot

par(mar=c(5,4,2,2))
forest(res, header=TRUE, xlim=c(-170, 310))

# approximate 95% prediction interval (based on equation 20.6)

predict(res, pi.type="simple", digits=0)
##  pred se ci.lb ci.ub pi.lb pi.ub 
##    50  9    32    68    23    77

# approximate 95% prediction interval (based on equation 20.7)

predict(res, pi.type="riley", digits=0)
##  pred se ci.lb ci.ub pi.lb pi.ub 
##    50  9    32    68    11    89

Note: The prediction interval is computed in a slightly different way in the book compared to how this is done (by default) in the metafor package. For more details, see here. However, by using argument pi.type="riley", equation 20.7 is used.

# data for Figure 20.3

dat <- read.table(header=TRUE, text = "
study   mean1 sd1  n1 mean2 sd2  n2
Carroll    94  22  60    92  20  60
Grant      98  21  65    92  22  65
Peck       98  28  40    88  26  40
Donat      94  19 200    82  17 200
Stewart    98  21  50    88  22  45
Young      96  21  85    92  22  85")

dat <- escalc("SMD", m1i=mean1, sd1i=sd1, n1i=n1, m2i=mean2, sd2i=sd2, n2i=n2,
              slab=study, data=dat, vtype="LS2")
dat
##     study mean1 sd1  n1 mean2 sd2  n2     yi     vi 
## 1 Carroll    94  22  60    92  20  60 0.0945 0.0329 
## 2   Grant    98  21  65    92  22  65 0.2774 0.0307 
## 3    Peck    98  28  40    88  26  40 0.3665 0.0499 
## 4   Donat    94  19 200    82  17 200 0.6644 0.0105 
## 5 Stewart    98  21  50    88  22  45 0.4618 0.0427 
## 6   Young    96  21  85    92  22  85 0.1852 0.0234

# fit equal- and random-effects models

res.ee <- rma(yi, vi, data=dat, method="EE")
res.re <- rma(yi, vi, data=dat, method="DL")

# show random-effects model results
print(res.re, digits=3)
## Random-Effects Model (k = 6; tau^2 estimator: DL)
## 
## tau^2 (estimated amount of total heterogeneity): 0.037 (SE = 0.042)
## tau (square root of estimated tau^2 value):      0.193
## I^2 (total heterogeneity / total variability):   58.34%
## H^2 (total variability / sampling variability):  2.40
## 
## Test for Heterogeneity:
## Q(df = 5) = 12.003, p-val = 0.035
## 
## Model Results:
## 
## estimate     se   zval   pval  ci.lb  ci.ub      
##    0.358  0.105  3.404  <.001  0.152  0.565  *** 
## 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

# Figure 20.3: Meta-Analysis Showing Relative Weights for the Equal- and Random-Effects Models

wi <- fmtx(cbind(weights(res.ee), weights(res.re)), digits=2)

par(mar=c(5,4,2,2))
forest(res.ee, header=TRUE, xlim=c(-1.5,4), ylim=c(-2.5,9), psize=1, efac=c(0,1),
       ilab=wi, ilab.xpos=c(1.7,2.4), ilab.pos=2, digits=c(3,2))
addpoly(res.re)
text(1.45, 8, "Weight\n(Equal)",  font=2)
text(2.15, 8, "Weight\n(Random)", font=2)

License

This documented is licensed under the following license: CC Attribution-Noncommercial-Share Alike 4.0 International.