3

I've used the confint() function of R package metafor to calculate heterogeneity in random effects models in a meta-analysis.

I get sensible values for the various parameters (tau, tau^2, I^2) for most of the calculations I have run, however am getting back 0 values for all three parameters for one particular analysis. I have checked the raw data and nothing looks suspiciously different to that used for other calculations.

The confint() output I get is as follows. 

       estimate  ci.lb   ci.ub
tau^2    0.0000 0.0000  0.1061
tau      0.0000 0.0000  0.3257
I^2(%)   0.0000 0.0000 60.4187
H^2      1.0000 1.0000  2.5264

Does anyone have any idea about why I might be getting a 0 estimate with 0 as the lower end of the confidence interval?

Any thoughts would be much appreciated.

John
  • 21,167
  • 9
  • 48
  • 84
user26050
  • 33
  • 1
  • 3

1 Answers1

3

This is telling you that the data is homogenous but the degree of precision is not reported because it's less than 0.0001. For example your uncertainty intervals around the I-squared are 0% to 60%. For I-squared, it's a percentage and you can't get a negative number (only 0 - 100%) even if it makes it look skewed (as in this case).

Hope this helps.

abousetta

abousetta
  • 1,030
  • 5
  • 9
  • 1
    Hi Abousetta, thanks for your response. That seems to be a reasonable interpretation. Is it unusual that the upper limit of the confidence interval is so high (60%) though, if the data is homogenous? Also, looking at the forest plot, the heterogeneity of effect estimates looks akin to those of other analyses I have done, where the I squared estimate is more like 70% (point estimate). – user26050 Sep 10 '13 at 04:06
  • 1
    Hi, You have to remember that it's uncertainty intervals (rather than confidence intervals). In other words, the 'true' amount of heterogeneity not due to chance lies somewhere between 0% to 60% (because we can't have a negative percent). Therefore this is a good indication of a lack of statistical heterogeneity between the data. If you look at your Q statistic, it is probably less than the degrees of freedom (df) and has an associated p value that is close to 1.00. All this suggests that your data is statistically non-heterogeneous and can be pooled to generate a pooled effect estimate. – abousetta Sep 10 '13 at 08:06
  • 1
    As for the forest plot, heterogeneity is not only present when the confidence intervals don't overlap, but can be present when they partially overlap, but the studies have small standard errors. Without any bias in the individual trials, they should all be measuring a similar (if not exactly the same) 'true' point estimate. Therefore, if each trial has a small standard error and they are different from each other, the amount of heterogeneity can be large. We see this often in reviews of observational studies were the sample size can often be very large in each study. Hope this helps. abousetta – abousetta Sep 10 '13 at 08:10