Let us consider the birth weight example. We can view the number of standard deviations' difference as a 'standardized', or dimensionless, form of the actual findings. The value of 0.2 is the number of SDs by which the intervention changes outcome - if it is measured in grammes (where the SD is 500g) it changes by 0.2 x 500 = 100g, if it is measured in ounces (where the SD is 15oz) it changes by 0.2 x 15 = 3oz.
In practice, of course, we would not want to use the SMD method to analyse birth weight as we are able to convert between units of measurement using an exact conversion factor. However, we commonly have to use it when different measurement tools (e.g. scales) are used to measure the same clinical outcome.
For example, suppose a potential treatment for depression in the elderly achieves an average improvement of 2 points on the Hamilton Rating Scale for depression (HAMD). And suppose that the pooled standard deviation of HAMD scores is 8. Then the standardized mean difference is 2/8 = 0.25. If a similar treatment effect was to be observed on an alternative depression scale, say the Geriatric Depression Scale (GDS) which has a standard deviation of 5 points, then a standardized mean difference of 0.25 is equivalent to an improvement of 1.25 points on the GDS.
We must be careful with using the standardized mean difference, however. First, we must be sure that the different measurement scales are indeed measuring the same clinical outcome. Second, problems can arise through the use of the pooled standard deviation for the standardizing. To illustrate the latter, let us return to our study with a 2-point improvement in HAMD score (pooled SD = 8). Imagine a second study in the same meta-analysis that also used the HAMD, but had more restrictive inclusion criteria. The tight inclusion criteria meant that participants were more similar to each other, and their pooled standard deviation in HAMD scores was only 5. Imagine further that the drug was equally effective in this study in that it also achieved a 2-point average improvement in HAMD score. The standardized mean difference for this study is 2/5 = 0.4. Therefore the same effect of treatment gives a different standardized mean difference just because of the tighter inclusion criteria. This is an unfortunate implication of using standardized mean differences. Nevertheless, if studies do use different scales, there are usually few alternatives to using the standardized mean difference to combine results in a meta-analysis.
Finally, we should point out that in RevMan and The Cochrane Library, the mean difference method is referred to as ‘MD’ and the standardized mean difference method as ‘SMD’.