This dataset include 21 studies which examine whether the chances of receiving a grant or fellowship award differs for men and women. Source: Bornmann, L., Mutz, R., & Daniel, H. (2007). Gender differences in grant peer review: A metaanalysis. Journal of Informetrics, 1(3), 226–238. https://doi.org/10.1016/j.joi.2007.03.001.
The data consists of 13 variables:
study (character) study reference
obs (numeric) observation within study
doctype (character) document type
gender (character) gender of the study authors
year (numeric) (average) cohort year
org (character) funding organization / program
country (character) country of the funding organization / program
type (character) fellowship or grant application
discipline (character) discipline / field
waward (numeric) number of women who received a grant/fellowship award
wtotal (numeric) number of women who applied for an award
maward (numeric) number of men who received a grant/fellowship award
mtotal (numeric) number of men who applied for an award
Many studies provide multiple comparisons (e.g., for different years / cohorts / disciplines), so a multilevel meta-analytic model is be used to account for the multilevel structure in these data. Metafor pacakge was used to conduct the analysis.
The data is publically available at metafor package.
library(metafor)
dat <- dat.bornmann2007
head(dat)
## study obs doctype gender year org country type
## 1 Ackers (2000) 1 Grey M&F 1996 MSCA Europe Fellowship
## 2 Ackers (2000) 2 Grey M&F 1996 MSCA Europe Fellowship
## 3 Ackers (2000) 3 Grey M&F 1996 MSCA Europe Fellowship
## 4 Ackers (2000) 4 Grey M&F 1996 MSCA Europe Fellowship
## 5 Ackers (2000) 5 Grey M&F 1996 MSCA Europe Fellowship
## 6 Ackers (2000) 6 Grey M&F 1996 MSCA Europe Fellowship
## discipline waward wtotal maward mtotal
## 1 Physical Sciences 139 711 274 1029
## 2 Physical Sciences 45 258 166 908
## 3 Physical Sciences 44 236 219 928
## 4 Physical Sciences 63 251 96 507
## 5 Social Sciences / Biology 157 910 252 1118
## 6 Physical Sciences 114 589 460 2244
The effect size is calculated as log odds ratio.
dat <- escalc(measure="OR", ai=waward, n1i=wtotal, ci=maward, n2i=mtotal, data=dat)
dat <- as.data.frame(dat)
head(dat)
## study obs doctype gender year org country type
## 1 Ackers (2000) 1 Grey M&F 1996 MSCA Europe Fellowship
## 2 Ackers (2000) 2 Grey M&F 1996 MSCA Europe Fellowship
## 3 Ackers (2000) 3 Grey M&F 1996 MSCA Europe Fellowship
## 4 Ackers (2000) 4 Grey M&F 1996 MSCA Europe Fellowship
## 5 Ackers (2000) 5 Grey M&F 1996 MSCA Europe Fellowship
## 6 Ackers (2000) 6 Grey M&F 1996 MSCA Europe Fellowship
## discipline waward wtotal maward mtotal yi vi
## 1 Physical Sciences 139 711 274 1029 -0.40107542 0.01391663
## 2 Physical Sciences 45 258 166 908 -0.05726822 0.03428886
## 3 Physical Sciences 44 236 219 928 -0.29852194 0.03391225
## 4 Physical Sciences 63 251 96 507 0.36093779 0.03404192
## 5 Social Sciences / Biology 157 910 252 1118 -0.33336360 0.01282044
## 6 Physical Sciences 114 589 460 2244 -0.07172953 0.01361164
To fit a multi-level model, we have to specify the following parameters.
yi The name of the column in our data set which contains the calculated effect sizes.
Vi The name of the column in our data set which contains the variance of the calculated effect sizes.
slab The name of the column in our data set which contains the study labels.
data The name of the data set.
test The test we want to apply for our regression coefficients. We can choose from "z" (default) and "t" (recommended; uses a test similar to the Knapp-Hartung method).
method The method used to estimate the model parameters. Both "REML" (recommended; restricted maximum-likelihood) and "ML" (maximum likelihood) are possible.
res <- rma.mv(yi, vi, random = ~ 1 | study/obs, data=dat, test = "t")
res
##
## Multivariate Meta-Analysis Model (k = 66; method: REML)
##
## Variance Components:
##
## estim sqrt nlvls fixed factor
## sigma^2.1 0.0161 0.1268 21 no study
## sigma^2.2 0.0038 0.0613 66 no study/obs
##
## Test for Heterogeneity:
## Q(df = 65) = 221.2850, p-val < .0001
##
## Model Results:
##
## estimate se tval df pval ci.lb ci.ub
## -0.1010 0.0417 -2.4196 65 0.0183 -0.1843 -0.0176 *
##
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The model results consist of 3 parts:
First, the variance components in which the random-effects variances calculated for each level of our model. Sigma^2.1 shows variance at level 3 or betweeen-study variance while sigma^2.2 shows variance at level 2 or within-studies variance. The number of studies is 21 with 66 effect sizes.
Second, The test of heterogeneity part. This part shows the overall heterogeneity, however this is informative as we need to quantify the heterogeneity at the 3 levels. We can claculate it using dmetar package
library(dmetar)
## Extensive documentation for the dmetar package can be found at:
## www.bookdown.org/MathiasHarrer/Doing_Meta_Analysis_in_R/
i2 <- var.comp(res)
summary(i2)
## % of total variance I2
## Level 1 24.03213 ---
## Level 2 14.36985 14.37
## Level 3 61.59802 61.6
## Total I2: 75.97%
The output shows that the heterogeneity variance at level 3 = 61.6 and at level 2 = 14.37. We also can plot the distribution of the total variance.
plot(i2)
Third, the model results shows the effect size estimate with confidence intervals. The mean log odds ratio = - 0.1, P value = 0.016. That means that the odds women to receive a grant or felloship is lesser than men.
Let's test for a difference between fellowship and grant applications as well as difference between countries
res2 <- rma.mv(yi, vi, mods = ~ type + country, random = ~ 1 | study/obs, data=dat, test = "t")
res2
##
## Multivariate Meta-Analysis Model (k = 66; method: REML)
##
## Variance Components:
##
## estim sqrt nlvls fixed factor
## sigma^2.1 0.0023 0.0479 21 no study
## sigma^2.2 0.0031 0.0559 66 no study/obs
##
## Test for Residual Heterogeneity:
## QE(df = 60) = 106.6330, p-val = 0.0002
##
## Test of Moderators (coefficients 2:6):
## F(df1 = 5, df2 = 60) = 5.1969, p-val = 0.0005
##
## Model Results:
##
## estimate se tval df pval ci.lb ci.ub
## intrcpt -0.1416 0.0998 -1.4189 60 0.1611 -0.3412 0.0580
## typeGrant 0.1390 0.0541 2.5699 60 0.0127 0.0308 0.2471
## countryCanada -0.0774 0.1258 -0.6152 60 0.5408 -0.3290 0.1742
## countryEurope -0.1125 0.1001 -1.1235 60 0.2657 -0.3127 0.0878
## countryUnited Kingdom 0.0789 0.1149 0.6863 60 0.4951 -0.1510 0.3087
## countryUnited States 0.0581 0.0992 0.5859 60 0.5601 -0.1403 0.2565
##
## intrcpt
## typeGrant *
## countryCanada
## countryEurope
## countryUnited Kingdom
## countryUnited States
##
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The test of moderators shows that that F5 = 25.98 , with p < 0.001.This means that there is significant difference between the subgroups. The estimates table shows only significant with application type with the effect in grant type is -0.2806.
1- Funnel plot
The funnel plot shows that studies are slightly asymmetrically distributed, so publication bias could exist.
funnel.rma(res, label = F)
2- Egger's regression test.
There is no direct function to conduct Egger's test for multi-level model. Alternatively, we calculate it by using the standard errors of the effect size estimates as a predictor in the meta-regression.
res3 <- rma.mv(yi, vi, mods = ~ sqrt(vi) , random = ~ 1 | study/obs, data=dat, test = "t")
res3
##
## Multivariate Meta-Analysis Model (k = 66; method: REML)
##
## Variance Components:
##
## estim sqrt nlvls fixed factor
## sigma^2.1 0.0165 0.1283 21 no study
## sigma^2.2 0.0039 0.0626 66 no study/obs
##
## Test for Residual Heterogeneity:
## QE(df = 64) = 186.5724, p-val < .0001
##
## Test of Moderators (coefficient 2):
## F(df1 = 1, df2 = 64) = 0.6776, p-val = 0.4135
##
## Model Results:
##
## estimate se tval df pval ci.lb ci.ub
## intrcpt -0.0661 0.0597 -1.1087 64 0.2717 -0.1853 0.0530
## sqrt(vi) -0.2151 0.2613 -0.8232 64 0.4135 -0.7372 0.3070
##
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
The coefficient of the standard error was insignificant (r=-0.2, P= 0.41). The results mean that studies are symmetrically distributed, and publication bias are absent which contraindicates the funnel plot.