0
$\begingroup$

Say, if I make a scatter plot of truth values vs pred values, is it appropriate to think of those 'outliers' in the plot as outliers?

$\endgroup$

2 Answers 2

1
$\begingroup$

In the sense that they fit what the model predicts least well, yes. Note: some values will always be the least well fitting ones and a perhaps more interesting question is whether they fit much more badly than one would expect. That might be a reason to look into whether something interesting is going on that might be an interesting exploratory finding.

However, what you mention does certainly not make them ouliers in the sense that this would be a justification to omit them from analyses. Doing so would be considered by many as unethical conduct that just tries to make one's results look cleaner than they are.

$\endgroup$
1
$\begingroup$

If you really, really want to do this for some useful purpose then a plot of actual versus predicted is perhaps not the most useful plot. Why not try the Bland and Altman plot also known as the Tukey sum and difference plot. One advantage is that it uses more of the plotting surface so you can see more clearly discrepancies. Another advantage is that it lets you see more clearly how discrepancy varies with the value. There are a couple of relevant papers Measurement in medicine: the analysis of method comparison studies and Statistical methods for measuring agreement between two methods of clinical measurement. The last of these has been cited more than 30000 times apparently

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.