calculation of max achievable AUC

[ZuZu]

The theoretically maximum achievable AUC for a presence-only model is
1-(prevalence/2). Maxent generates this estimate, but also estimates
prevalence (in the maxentResults.csv file). The two estimates are not
consistent: the estimate of maximum achievable AUC is considerably
lower than the prevalence estimate would predict. Does anyone know why
this is so?

I am using maxent 3.3.1 with the target-sample background approach.

Thanks!

ZuZu

[Steven Phillips]

ZuZu,

The first estimate assumes that the presence of the species is binary -
the species is either present or not. In this case, the maximum AUC is
below 1 because of contamination of the background by presences.

The second estimate allows presence to be probabilistic. For example,
you might observe your species only 20% of the times that you visit a
particular site. The estimate of maximum AUC should be lower in this
case, because the AUC is lowered both by contamination of the background
by presences, and by randomness affecting whether the species is truly
present in suitable sites. As an extreme example, if the species has
50% probability of presence everywhere, then the maximum achievable AUC
is 0.5.

-- Steven

> --
>
> You received this message because you are subscribed to the Google Groups "Maxent" group.
> To post to this group, send email to max...@googlegroups.com.
> To unsubscribe from this group, send email to maxent+un...@googlegroups.com.
> For more options, visit this group at http://groups.google.com/group/maxent?hl=en.
>
>
>