Inputting confidence interval of 0 and 1 into STATA

[Sarah Cheung]

I am currently doing a meta-analysis with dichotomous outcome. Some of my included studies contain effect size = 0 or 1. When effect size equals to zero, i.e. p=0, the upper and lower limit of the confidence interval will be equal to 0 as well, vice versa. If the effect size equals to 1, i.e. p=1, both the lower limit and upper limit of the confidence interval will be 1. However, If I input these results into the STATA, error message the error message will emerge. May I know if there is any other method to put back all those studies consisting of effect size= 0 or 1 into the STATA? So that I can conduct a more a more accurate analysis. Thank you.

[Jared Greathouse]

I want to be very precise about this: you're telling me that the OUTCOME of the meta-analysis, not the outcomes of the studies, but the outcomes of the meta-analysis, are 0 and 1? That is, your partial correlations, Cohen's d's, Glass' g's, and so on, those are all 0 or 1? In MA, the outcome is the effect size. So before we continue, I want to be specific about this point.

In fact, the best way to do this is to show me what your data look like using the dataex command, as the FAQ requests.

Oh, and hey Sarah. Welcome to Statalist.

[Leonardo Guizzetti]

In addition to what Jared mentioned, it can sometimes happen in meta-analysis that an observed effect for a study has a point estimate that lies on the boundary of possible effect sizes. For example, a proportion is constrained to be in the closed interval from 0 to 1, and it may be that a study has a proportion of exactly 0 or 1. When this happens, and let's assume that a confidence interval can be calculated, then the confidence interval in the direction that lies outside the domain must be truncated at the limit. For example, if the proportion point estimate is 0, then the lower confidence limit must also be 0 (it can't possibly be less than 0), but the upper limit is greater than zero. it is not possible (in realistic settings) to have the lower and upper limits confined to be equal to the point estimate. So here is where I believe your understanding of confidence intervals or meta-analysis has gone astray. And it would be a good idea to show us a data example using -dataex- to have a concrete discussion given your dataset.

Note: it is often possible to conduct a meta-analysis on the scale of a transformation that gets around the original scale (e.g., a logit transformation for proportions), and then transform back to the original scale for common understanding. In these cases, wrinkles may arise either due to the transformation method or analytical model (or both) that require a so-called continuity correction or other slight change in the data to allow the math to work correctly (again, logit(0) is undefined, so some small number like 0.5 may commonly be added to the numerator to find a valid estimate, if slightly biased). We can set these idiosyncrasies aside to first better understand the question at hand.

[Sarah Cheung]

Sorry that I may not explain my question clear enough. What I mean is that the result of my included studies belong to dichotomous outcome. And I am trying to calculate the CI for proportions. The following is how my data look like.

"Study Name" "Effect size" "Lower limit of CI" "Upper limit of CI"
"Study A" ".138" ".099" ".181"
"Study B" ".447" ".308" ".592"
"Study C" ".4" ".327" ".473"
"Study D" ".2273". ".0522" ".4024"
"Study E" "0" "0". ".2281"
"Study F" "0" "0" ".0941"
"Study G" ".0213" "-.0083" ".048"
"Study H" ".25" ".116" ".384"
"Study I" ".398" ".346" ".454"
"Study J" ".14" ".054" ".226"

[Jared Greathouse]

For us to really see your data, you'll need to use the dataex command, as I asked in my first post. So, use that and then post back how your data look, please. Sarah Cheung

[Leonardo Guizzetti]

To make life easier on yourself, why not work directly with the raw values (the numerator and denominator of your proportion) ? This gets around this whole issue that you are hitting against.