What is metaanalysis?
Metaanalysis is the use of statistical methods to combine results of individual studies. This allows us to make the best use of all the information we have gathered in our systematic review by increasing the power of the analysis. By statistically combining the results of similar studies we can improve the precision of our estimates of treatment effect, and assess whether treatment effects are similar in similar situations. The decision about whether or not the results of individual studies are similar enough to be combined in a metaanalysis is essential to the validity of the result, and will be covered in the next module on heterogeneity. In this module we will look at the process of combining studies and outline the various methods available.
There are many approaches to metaanalysis. We have discussed already that metaanalysis is not simply a matter of adding up numbers of participants across studies (although unfortunately some nonCochrane reviews do this). This is the 'pooling participants' or 'treatasonetrial' method and we will discuss it in a little more detail now.
Pooling participants (not a valid approach to metaanalysis).
This method effectively considers the participants in all the studies as if they were part of one big study. Suppose the studies are randomised controlled trials: we could look at everyone who received the experimental intervention by adding up the experimental group events and sample sizes and compare them with everyone who received the control intervention. This is a tempting way to 'pool results', but let's demonstrate how it can produce the wrong answer.
A Cochrane review of trials of daycare for preschool children included the following two trials. For this example we will focus on the outcome of whether a child was retained in the same class after a period in either a daycare treatment group or a nondaycare control group. In the first trial (Gray 1970), the risk difference is 0.16, so daycare looks promising:
Gray 1970 
Retained 
Total 
Risk 
Risk difference 
Daycare 
19 
36 
0.528 
0.16 
Control 
13 
19 
0.684 
In the second trial (Schweinhart 1993) the absolute risk of being retained in the same class is considerably lower, but the risk difference, while small, still lies on the side of a benefit of daycare:
Schweinhart 
Retained 
Total 
Risk 
Risk difference 
Daycare 
6 
58 
0.1034 
0.004 
Control 
7 
65 
0.1077 
What would happen if we pooled all the children as if they were part of a single trial?
Pooled results 
Retained 
Total 
Risk 
Risk difference 
Daycare 
25 
94 
0.266 
+0.03 WRONG! 
Control 
20 
84 
0.238 
