For example, when we want to calculate sensitivity, but we have few TP amounts (but not zero) due to small sample size (especially in meta-analysis).

I've only glanced quickly at those two resources, but are you sure they are addressing the same thing? Yates' (continuity) correction as typically described entails subtraction 0.5 from |O-E| before squaring in the usual equation for Pearson's Chi2. E.g.,

https://en.wikipedia.org/wiki/Yates%27s_correction_for_continuity

But adding 0.5 to each cell in a 2x2 table is generally done to avoid division by 0 (e.g., when computing an odds ratio), not to correct for continuity (AFAIK). This is what makes me wonder if your two resources are really addressing the same issues. But as I said, I only had time for a very quick glance at each. HTH.

Just adding the caveat that you subtract .5 if that is greater than |O-E| and otherwise just treat as zero. But there are lots of alternatives to these (e.g., https://www.jstor.org/stable/2981423).

Bruce Weaver

Thank you for your kind reply. The formula you mentioned is toward punishment (i.e. reduction of positive or negative association toward zero). I am looking for way to correct toward increasing power.

In https://www.ncbi.nlm.nih.gov/books/NBK115736/pdf/Bookshelf_NBK115736.pdf:

"We used a continuity correction of 0.5 for studies where the observed count was zero for any of the cells of the 2×2 table" "Differences between methods were larger in meta-analyses where a large proportion of the available studies required a continuity correction (for the normal approximation) and in meta-analyses where studies were generally small."

"/* logit sensitivity, with continuity correction */ generate b1 = logit(TP/(TP + FN)) replace b1 = logit((TP+0.5)/(TP+FN+2*0.5)) if TN == 0 | FP == 0 | TP == 0 | FN == 0"

As I found above, the below uses of continuity correction are accepted by scientists.

1- To reduce exaggerated difference like you mentioned for Chi2 test of association (toward punishment).

2- To remove zero cells (toward punishment or toward power increase depending on which cell is zero).

But I am looking for a third use; is it possible to add 0.5 to small (but not zero) cells in TP and TN cells to increase test accuracy of small diagnostic studies?

... continue

Bruce Weaver

InArticle Critical review and comparison of continuity correction meth...

I found that we can use continuity correction to increase np to increase the possibility of normal approximation. This, is something like I have been looking for.

Daniel Wright

Where is your profile figure?! It was nice!