CompileR.Rmd

---
title: "CompileR"
csl: https://raw.githubusercontent.com/citation-style-language/styles/master/apa.csl
nocite: |
  @*
output:
  html_document: default
  word_document: default
bibliography: R-Pckgs.bib
---

```{r settings}
## Specify the file containing your data
dat_comp <- read.csv("missing_data.csv")


### Specify the columns containing the means, SDs, and Ns for intervention and control within the file

col.int.means     <- "m1"
col.int.sds       <- "sd1"
col.int.ns        <- "n1"
col.cont.means    <- "m2"
col.cont.sds      <- "sd2"
col.cont.ns       <- "n2"

col.dval <- "dval"
col.tval <-"tval"
col.pval <- "pval"
col.sign <- "sign"

```

```{r setup, include=FALSE}
library(knitr)
knitr::opts_chunk$set(echo = TRUE)
# Set so that long lines in R will be wrapped:
opts_chunk$set(tidy.opts=list(width.cutoff=80),tidy=TRUE)
```

This script demonstrates the method for assessembling SMDs described by Wolfgang Viechtbauer at http://metafor-project.org/doku.php/tips:assembling_data_smd. The comparable method for Odds Ratios is available here: http://metafor-project.org/doku.php/tips:assembling_data_or 

If you are working with other data types or have different pieces of information about studies available you may find the package 'compute.es' helpful.

## Load packages

```{r warning=FALSE, message=FALSE}
library(metafor)
library(knitr)
```

## Show dat_comp
`r kable(dat_comp)`

## Let's compile!

### Start by calculating what we can
```{r}
dat_comp <- escalc(measure="SMD", 
                        m1i = get(col.int.means),
                        sd1i = get(col.int.sds),  
                        n1i = get(col.int.ns),  
                        m2i = get(col.cont.means),
                        sd2i = get(col.cont.sds),  
                        n2i = get(col.cont.ns),  
                        data=dat_comp)

```
`r kable(dat_comp)`

###This is our starting meta-analysis based on complete cases
```{r}
dat_MA <- rma(yi, vi, data = dat_comp)
dat_MA
forest(dat_MA)
```



### Try and calculate t-values from available exact p-values [^1]
```{r}
dat_comp$tval <- replmiss(dat_comp$tval, with(dat_comp, sign * qt(pval/2, df=get(col.int.ns)+get(col.cont.ns)-2, lower.tail=FALSE)))
```
`r kable(dat_comp)`

## Try and calculate d-values from available t-values
```{r}
dat_comp$dval <- replmiss(dat_comp$dval, 
                          with(dat_comp, 
                               tval * sqrt(1/get(col.int.ns) + 1/get(col.cont.ns))))
```
`r kable(dat_comp)`

### Convert from Cohen's d to hedges g
```{r}
dat_comp$yi <- replmiss(dat_comp$yi, 
                        with(dat_comp, 
                             (1 - 3/(4*(get(col.int.ns)+get(col.cont.ns)-2) - 1)) * dval))
```
`r kable(dat_comp)`

### Calculate missing sampling variances
```{r}
dat_comp$vi <- replmiss(dat_comp$vi, 
                        with(dat_comp, 
                             1/get(col.int.ns) + 1/get(col.cont.ns) + yi^2/(2*(get(col.int.ns)+get(col.cont.ns)))))
```
`r kable(dat_comp)`


## This is our meta-analysis based on compiled data
```{r}
dat_MA <- rma(yi, vi, data = dat_comp)
dat_MA
forest(dat_MA)
```



## Write the data to a file
```{r}
write.csv(dat_comp, "compiled_data.csv", row.names = FALSE)
```

[^1]: If exact p values are not availble, one possible approach is to use the reported p cut off for the statistic (e.g. p<0.05) would use the p value 0.05. This is very conservative but may be preferrable to excluding the study entirely

## Packages used in this document
```{r include=FALSE}
citPkgs <- names(sessionInfo()$otherPkgs)
write_bib(citPkgs, file="R-Pckgs.bib")
```