match componentwise and total likelihood

[vbartolino]

g3_fit()returns the likelihood score by data components fit$likelihood (as I understand aggregated by year-step) but I'm surprise that their sum doesn't match the overall likelihood score of the model fit$score. Am I missing something?

> fit1$likelihood %>% mutate(x=weight*value) %>% .$x %>% sum(na.rm=T)
[1] 2833.439
> fit1$likelihood %>% mutate(x=value) %>% .$x %>% sum(na.rm=T)
[1] 78.79221
> fit1$score
       nll
1 7351.098

[willbutler42]

I would expect them to. Here's our redfish model:

> fit1$likelihood %>% mutate(x=weight*value) %>% .$x %>% sum(na.rm=T)
[1] 11661.99
> fit1$likelihood %>% mutate(x=value) %>% .$x %>% sum(na.rm=T)
[1] 45.08969
> fit1$score
    nll
1 11662

Could you share/link the model and param files?

[vbartolino]

Could you share/link the model and param files?

https://github.com/vbartolino/g3_GoB/tree/ba1652cedda0e1ca69e6ecd92d612ed110c2f754/herring_g3/gadget3/models/02.5.4-hs.dlnorm_TMP

I've updated gadgetutils on my computer last end of Oct:

GithubRepo: gadgetutils
GithubUsername: gadget-framework
GithubRef: HEAD
GithubSHA1: 52721f29082b219b48ddc162cad8b1ca857baff2
NeedsCompilation: no
Packaged: 2025-10-31 15:17:55 UTC; vano0001

[bthe]

Is this related to bounds? That is, are any parameters outside of the bounds?

[vbartolino]

Is this related to bounds? That is, are any parameters outside of the bounds?

in practice, I've never been able to get rid of some of the params hitting the bounds (especially *.init.* and some *rec).
Not sure I understand why that could affect the calculation of the likelihood

[lentinj]

in practice, I've never been able to get rid of some of the params hitting the bounds (especially .init. and some *rec).

The blue ling model also suffers badly from this, there's very little data at the start of the model and it leans on the bounds to keep it on the straight and narrow. As I see it the choice is to either add extra constraints to these parameters, or accept that the bounds are a functional part of the model not just a backstop for the optimiser.

A logAR(1) or random walk on recruitment is pretty essential for bling to get through the data-sparse start with sensible values, and doesn't then require the bounds to stop the worst of the peaks.

You could do similar with init., applying a dnorm for example, but that isn't that different to what the bounds do for you anyway.

Not sure I understand why that could affect the calculation of the likelihood

I've not checked, but I'm guessing the point is that the bounds are part of the overall likelihood score, but not reported by g3_fit in the breakdown output.

[willbutler42]

The difference mostly stems from the dlnorm:

> of <- g3_tmb_adfun(tmb_model, params_final)
> reports <- of$report(of$par)
> fit$score 
      nll
1 16376.7
> (fit$likelihood %>% mutate(x=weight*value) %>% .$x %>% sum(na.rm=T)) + 
+   (reports$proj_dlnorm_her.rec__nll |> sum())
[1] 16374.83

So the dlnorm reports are not being picked out by the regex in g3_fit. I can modify g3_fit to check for random effects and collate the nlls if they exist, will add this next week.

Perhaps the small difference that remains is due to the bounds penalty? At least, I cannot see anything else in the reports that makes up that difference

[vbartolino]

The difference mostly stems from the dlnorm: ...

this explains. It would be great to add it. Intuitively I would try to have in fit$likelihood every component contributing to the overall likelihood incl. random effects, understocking (which seems already there) and contribution from bounds_penalty too