metadat (version 1.4-0)

dat.vanhowe1999: Studies on the Association between Circumcision and HIV Infection

Description

Results from 33 studies examining the association between male circumcision and HIV infection.

Usage

dat.vanhowe1999

Format

The data frame contains the following columns:

study	character	study author
category	character	study type (high-risk group, partner study, or population survey)
non.pos	numeric	number of non-circumcised HIV positive cases
non.neg	numeric	number of non-circumcised HIV negative cases
cir.pos	numeric	number of circumcised HIV positive cases
cir.neg	numeric	number of circumcised HIV negative cases

Author

Wolfgang Viechtbauer, , https://www.metafor-project.org

Concepts

medicine, epidemiology, odds ratios

Details

The 33 studies provide data in terms of 2 22x2 tables in the form:

	HIV positive	HIV negative
non-circumcised	non.pos	non.neg
circumcised	cir.pos	cir.neg

The goal of the meta-analysis was to examine if the risk of an HIV infection differs between non-circumcised versus circumcised men.

The dataset is interesting because it can be used to illustrate the difference between naively pooling results by summing up the counts across studies and then computing the odds ratio based on the aggregated table (as was done by Van Howe, 1999) and conducting a proper meta-analysis (as illustrated by O'Farrell & Egger, 2000). In fact, a proper meta-analysis shows that the HIV infection risk is on average higher in non-circumcised men, which is the opposite of what the naive pooling approach yields (which makes this an illustration of Simpson's paradox).

References

O'Farrell, N., & Egger, M. (2000). Circumcision in men and the prevention of HIV infection: A 'meta-analysis' revisited. International Journal of STD & AIDS, 11(3), 137--142. https://doi.org/10.1258/0956462001915480

Examples

### copy data into 'dat' and examine data
dat <- dat.vanhowe1999
dat

if (FALSE) {
### load metafor package
library(metafor)

### naive pooling by summing up the counts within categories and then
### computing the odds ratios and corresponding confidence intervals
cat1 <- with(dat[dat$category=="high-risk group",],
   escalc(measure="OR", ai=sum(non.pos), bi=sum(non.neg), ci=sum(cir.pos), di=sum(cir.neg)))
cat2 <- with(dat[dat$category=="partner study",],
   escalc(measure="OR", ai=sum(non.pos), bi=sum(non.neg), ci=sum(cir.pos), di=sum(cir.neg)))
cat3 <- with(dat[dat$category=="population survey",],
   escalc(measure="OR", ai=sum(non.pos), bi=sum(non.neg), ci=sum(cir.pos), di=sum(cir.neg)))
summary(cat1, transf=exp, digits=2)
summary(cat2, transf=exp, digits=2)
summary(cat3, transf=exp, digits=2)

### naive pooling across all studies
all <- escalc(measure="OR", ai=sum(dat$non.pos), bi=sum(dat$non.neg),
                            ci=sum(dat$cir.pos), di=sum(dat$cir.neg))
summary(all, transf=exp, digits=2)

### calculate log odds ratios and corresponding sampling variances
dat <- escalc(measure="OR", ai=non.pos, bi=non.neg, ci=cir.pos, di=cir.neg, data=dat)
dat

### random-effects model
res <- rma(yi, vi, data=dat, method="DL")
res
predict(res, transf=exp, digits=2)

### random-effects model within subgroups
res <- rma(yi, vi, data=dat, method="DL", subset=category=="high-risk group")
predict(res, transf=exp, digits=2)
res <- rma(yi, vi, data=dat, method="DL", subset=category=="partner study")
predict(res, transf=exp, digits=2)
res <- rma(yi, vi, data=dat, method="DL", subset=category=="population survey")
predict(res, transf=exp, digits=2)
}