brief intro on what is meta-analysis, since the same study hypothesis will be tested over and over again by different researchers over the world, meta analysis is a method to collect all those similar studies, and end up with a single results.
so these are the packages im going to load for just once,
#install.packages("remotes")
#remotes::install_github("MathiasHarrer/dmetar", force = T)
library(metafor) # install.packages("metafor")
library(dmetar) # install.packages("dmetar")
Let’s see the data.
data("DepressionMortality")
DepressionMortality
## author event.e n.e event.c n.c country
## 1 Aaroma et al., 1994 25 215 171 3088 Finland
## 2 Black et al., 1998 65 588 120 1901 USA
## 3 Bruce et al., 1989 5 46 107 2479 USA
## 4 Bruce et al., 1994 26 67 1168 3493 USA
## 5 Enzell et al., 1984 32 407 269 6256 Sweden
## 6 Fredman et al., 1989 1 44 87 1520 USA
## 7 Murphy et al., 1987 24 60 200 882 Canada
## 8 Penninx et al., 1999 15 61 437 2603 Netherlands
## 9 Pulska et al., 1998 15 29 227 853 Finland
## 10 Roberts et al., 1990 173 1015 250 3375 USA
## 11 Saz et al., 1999 37 105 66 409 Spain
## 12 Sharma et al., 1998 41 120 9 54 UK
## 13 Takeida et al., 1997 29 258 24 1490 Japan
## 14 Takeida et al., 1999 61 388 34 1257 Japan
## 15 Thomas et al., 1992 15 211 73 1366 USA
## 16 Thomas et al., 1992 21 237 67 1340 USA
## 17 Weissman et al., 1986 10 48 36 216 USA
## 18 Zheng et al., 1997 31 615 1468 57674 USA
the columns are =>event.e n.e event.c n.c it include the source to calculate Risk Ratio. => Risk Ratio(RR) So we decide to use RR as our measurement.
library(metafor)
res = rma(measure = "RR", ai = event.e, bi = n.e-event.e , ci = event.c , di =n.c-event.c ,
data = DepressionMortality, slab = paste(author ))
forest(res,header = T)
So by the forest plot, we can see there is a lot of squares in diff
sizes, bigger the size higher the weight, and the weight is related to
the research sample size, which means, if the research have a greater
study population, it will leads to more stable and precise result, and
ends up weighting higher to the final estimation for all the studies.and
the width reflects the 95% confidence interval, wider means less
precision. now we look at the last row RE model, it means over all these
study, since the log risk ratio is above 0 and the CI doesn’t cross
through 0, we can confidently says that the depression group is
associated with an increased risk of the event, which is death in this
case. since we’re not familiar with the log scale, you can take
exponential of 0.7, e^0.7 =2.014, which means exposre group comparing to
control group’s RR is 2 times.
funnel(res)
the x axis is log risk ratio, 0 means no effect, for this case the middle point is around 0.7,which means increased risk. y axis is standard error, greater the sample, more precise and stable, and lower the se, and there are a bunch of study’s se is near 0 in this case, which means most of the researches are reliable. the assumption of the funnel plot is, as a scatter plot, it should be symmetrical, if not, youmay consider publication bias. publication bias is a systematic bias due to the lazy researchers, they don’t have enough time and resources to publish every findings, which leads that the published paper we see are all “successful” publish.