I see that this thread is a little old, but was just updated with a response, so I thought I could chime in. I believe the above answers are correct, but based on some comments we have received from reviewers over the past few years we have learned a lot about how to describe our methods.
Reviewers have asked us for a subgroup analysis in some manuscripts, which involved running separate analyses (i.e. meta-regression) on one or two specific populations or subgroups. For example, typically our meta-analyses have examined the effect of exercise training on specific outcomes. One subgroup analysis that was recommended to us was to stratify our analysis by exercise type (resistance vs. aerobic) and explore potential moderators within each of those subgroups.
For example:
The subgroup analysis of ONLY the resistance training effects could examine potential moderators like sets, repetitions, rest, load etc. using one meta-regression analysis
A separate subgroup analysis of ONLY the aerobic training effects could examine potential moderators like intensity, duration, frequency, etc. using a second meta-regression analysis
In this example, different exercise training programs have different variables that could change how effective exercise is, and other variables that are not relevant to that type of training (i.e. total duration may not be important for resistance training, and "sets" are not really every used in aerobic training.)
Subgroup comparisons are typically used for categorical variables, and meta-regression is typically used for continuous variables (but can also include categorical variables that are dummy coded). I like to perform subgroup comparisons prior to meta-regression to identify any categorical variables associated with the change in my outcome, and then include the significant categorical variables in the meta-regression to determine if they remain statistically significant in the overall model.
Last but not least, as a rule of thumb, I generally try not to force my continuous variables in to categories for a subgroup comparison unless there are very easy/natural threshold values that have a widely agreed upon cut-off limit to use. Male and Female are easy groups to use because they naturally fall into one group or another. Physical activity could be easily categorized as Active or Inactive based on something like steps/day where there are easy threshold values to use. A variable like Age doesn't have easy threshold values to use for categories, since we don't really have widely used definitions of older, younger, etc.
Below are a couple links that could help you out, from our own studies to see how the methods were explained:
Article Comparison of Periodized and Non-Periodized Resistance Train...
Article Effect of exercise training on C reactive protein: A systema...
In addition, the syntax available for Wilson and Lipsey's book is really good, and very user friendly when you are getting started:
Both methods are used to explain the source of heterogeneity. For subgroup analysis, the source is categorical in nature (e.g. regions, definition of outcomes, treatment regimens), while that for meta-regression is continuous (mean age, proportion of males).
For example, I would like to explore the relationship between sunlight and skin cancer risk. The association probably differs by the regions that the studies were conducted, so subgroup analysis is suitable for separating the results by region, so as to see if the heterogeneity is reduced.
On the other hand, mean ages of different population may explain the heterogeneity since elder people are more susceptible to carcinogens. Meta-regression using mean age as independent variations can show that if the heterogeneity changes significantly with age.
Subgroup analysis is applied by basically running two meta-analysis (or more) for different subgroups of the population of studies. You can then examine whether the subgroups differ in the mean effect size (and variation). You split the population based on some categorical moderator, such as gender. With subgroup analysis you assess one moderator at a time, or two with hierarchical subgrouping.
In contrast, meta-regression is applied by regressing the effect sizes on study-level moderators, which mostly include continuous moderators, such as age. You can also include categorical moderators using dummies. With meta-regression, you can include a lot of moderators.
For more examples, try the Borenstein book or see one of our papers on how to interpret meta-analysis, including sections on subgroup and meta-regression.
A sub-group analysis is the same as a meta-regression except that a sub-group analysis uses stratification to demonstrate moderator effects while a meta-regression uses regression modelling to do the same.
Sub-group analysis can be done if there is a single categorical moderator variable and will give exactly the same result as a meta-regression using that same variable (see the MetaXL User Guide on www.epigear.com for an example of this). If however there are continuous moderators or several categorical moderators then only meta-regression is possible.
I see that this thread is a little old, but was just updated with a response, so I thought I could chime in. I believe the above answers are correct, but based on some comments we have received from reviewers over the past few years we have learned a lot about how to describe our methods.
Reviewers have asked us for a subgroup analysis in some manuscripts, which involved running separate analyses (i.e. meta-regression) on one or two specific populations or subgroups. For example, typically our meta-analyses have examined the effect of exercise training on specific outcomes. One subgroup analysis that was recommended to us was to stratify our analysis by exercise type (resistance vs. aerobic) and explore potential moderators within each of those subgroups.
For example:
The subgroup analysis of ONLY the resistance training effects could examine potential moderators like sets, repetitions, rest, load etc. using one meta-regression analysis
A separate subgroup analysis of ONLY the aerobic training effects could examine potential moderators like intensity, duration, frequency, etc. using a second meta-regression analysis
In this example, different exercise training programs have different variables that could change how effective exercise is, and other variables that are not relevant to that type of training (i.e. total duration may not be important for resistance training, and "sets" are not really every used in aerobic training.)
Subgroup comparisons are typically used for categorical variables, and meta-regression is typically used for continuous variables (but can also include categorical variables that are dummy coded). I like to perform subgroup comparisons prior to meta-regression to identify any categorical variables associated with the change in my outcome, and then include the significant categorical variables in the meta-regression to determine if they remain statistically significant in the overall model.
Last but not least, as a rule of thumb, I generally try not to force my continuous variables in to categories for a subgroup comparison unless there are very easy/natural threshold values that have a widely agreed upon cut-off limit to use. Male and Female are easy groups to use because they naturally fall into one group or another. Physical activity could be easily categorized as Active or Inactive based on something like steps/day where there are easy threshold values to use. A variable like Age doesn't have easy threshold values to use for categories, since we don't really have widely used definitions of older, younger, etc.
Below are a couple links that could help you out, from our own studies to see how the methods were explained:
Article Comparison of Periodized and Non-Periodized Resistance Train...
Article Effect of exercise training on C reactive protein: A systema...
In addition, the syntax available for Wilson and Lipsey's book is really good, and very user friendly when you are getting started: