Dear professor and colleagues,
Is it good practice to round numbers during statistical analysis? For example, AUROC curve can be 0.789; can we round it to 0.79? If yes, can we apply same rule on confidence interval?
this is a good question because it can be dictated by convention, eg i think jounals will indicate how many decimals they want to see for the p-value. At some point you can round without any loss of info. For your auc example i think 2 decimal places is right but maybe 3 is the convention
1 Like
I love the question … I come from a physics background where our first lesson was in measurement error/uncertainty and some of those lessons still apply. For something like the AUC the number of decimal points depends on the accuracy (where the CI is the “uncertainty”). If the CI is small, then more decimal points OK. Here are some examples
0.789 (0.781 to 0.799) - this is good because the uncertainty is in the thousandths.
0.789 (0.754 to 0.848) - this is not good - should be 0.79 (0.75 to 0.85)
0.789 (0.610 to 0.967) - this is not good - should be 0.8 (0.6 to 1)
Hope this helps
John
2 Likes
Some nice guidelines on presenting statistical results here (including precision, although AUCs not directly mentioned). Guidelines for reporting of statistics for clinical research in urology - PMC
2 Likes