Why does Kullback-Leibler divergence favor the wrong survival model?

1 reply · 3K views · last reply
nigels
Messages
36
Reaction score
0
I’m currently evaluating the "realism" of two survival models in R by comparing the respective Kullback-Leibler divergence between their simulated survival time dataset (`dat.s1` and `dat.s2`) and a “true”, observed survival time dataset (`dat.obs`). Initially, directed KLD functions show that `dat.s2` is a better match to the observation:

> library(LaplacesDemon)
> KLD(dat.s1, dat.obs)$sum.KLD.py.px
[1] 1.17196
> KLD(dat.s2, dat.obs)$sum.KLD.py.px
[1] 0.8827712​

However, when I visualize the densities of all three datasets, it seems quite clear that `dat.s1` (green) better alignes with the observation:

> plot(density(dat.obs), lwd=3, ylim=c(0,0.9))
> lines(density(dat.s1), col='green')
> lines(density(dat.s2), col='purple')​

What is the cause behind this discrepancy? Am I applying KLD incorrectly due to some conceptual misunderstanding?
 
Attachments
  • Line plot of three density curves titled density.default(x = dat.obs): a thick black true density peaking near 0.28 at x near 0.5, a green s1 curve with KLD 1.17 and a purple s2 curve with KLD 0.88 that peaks higher, near 0.76.
    KL_nonsense_sof.png
    14.2 KB · Views: 856
Keep in mind that the KL-divergence is non-commutative, and different "orders" correspond to different objective functions (and different research questions). The way you're fitting it (that is, KL(Q||P), where Q is being fit to P) is trying to match regions of high density, and it does seem to be the case that the highest probability mass in your "worse fitting" model coincides with the highest probability mass in your target better than does the "better fitting" model. There's a fairly good discussion related to the topic here:

https://stats.stackexchange.com/questions/188903/intuition-on-the-kullback-leibler-kl-divergence
and
http://timvieira.github.io/blog/post/2014/10/06/kl-divergence-as-an-objective-function/

The other direction may actually be closer to what you're interested in:
KLD(dat.obs, dat.s1)$sum.KLD.py.px
KLD(dat.obs, dat.s2)$sum.KLD.py.px
 
Last edited: