0

I'm testing different methods for carrying out variant calling with HIV sequencing data and want to compare the performance of each method. I have true and false positive counts for each method but have no way of calculating the true negative or false negative counts as I don't know the truth (true and false positive counts are estimated based on some heuristics).

I was originally assessing the performance of each method by simply taking the ratio of the true positive (TP) and false positive (FP) counts but my professor suggested I instead calculate the value of (FP-TP)/(FP+TP) for each method. I've never heard of this metric before and he forgets if it has a specific name or not. Does anyone know the name for this metric, if it even has a name?

Also open to suggestions of other ways I could assess the performance of my different methods.

Nick Cox
  • 48,377
  • 8
  • 110
  • 156
nickeener
  • 1
  • In remote sensing it is called "normalized difference" e.g. https://en.wikipedia.org/wiki/Normalized_difference_vegetation_index I would be tempted to call it a relative or scaled difference if someone insisted on a name. But it seems to me that a formula is concise and explicit and sits just as easily on a graph axis or in text. – Nick Cox Dec 10 '19 at 18:40
  • I dare say it will not bite but if FP + TP is ever zero or even very near you may have a problem. – Nick Cox Dec 10 '19 at 18:41
  • This closely related thread provides some general answers: https://stats.stackexchange.com/questions/86708. They aren't specific to FP and TP, though. It also provides alternatives to your professor's metric. – whuber Dec 10 '19 at 20:12
  • This is not a *metric*, because it can be negative, and the distance to the correct solution (FP=0, only TP) is minus one. – cdalitz Dec 11 '19 at 14:07

0 Answers0