All Pareto k estimates are ok (k < 0.7).
```
`check_hmc_diagnostics()` on the `stanfit` object yields the following result:
```
Divergences:
0 of 4000 iterations ended with a divergence.
Tree depth:
0 of 4000 iterations saturated the maximum tree depth of 15.
Energy:
E-BFMI indicated no pathological behavior.
```
The wall-time on the HPC cluster I have access to is about 9h for this model with four cores and 16Gb RAM.
* * *
Now my questions:
1. Do I need to do additional diagnostics? If yes: What do I have to do?
2. How can I check the goodness of fit and prevent overfitting?
3. Do you have ideas for other model specifications?
4. My primary hypothesis is, that strabismus leads to a lower score than no disfigurement. How can I check this?
5. Does the interaction between `disfigurement` and `situation` presents difficulties in answering question three?
6. How can I plot and communicate the `face` effect?
---
### Author:  [@paul.buerkner](https://discourse.mc-stan.org/u/paul.buerkner)
#### Post date: [April 16, 2019, 11:44am UTC](https://discourse.mc-stan.org/t/ordinal-regression-model-in-brms/8446/2 "2019-04-16T11:44:38Z")
Hi, did you already take a look at this ordinal tutorial paper for brms: [https://psyarxiv.com/x8swp/](https://psyarxiv.com/x8swp/)
Perhaps that could be helpful.
Some other quick thoughts:
1. Convergence looks good I think
2. Model comparison for instance via loo and loo\_compare
3. face has just 4 levels. Modeling so many varying effects for it is probably not going to work out reasonably well.
4. Take a look at the `hypothesis` method
5. Depends on the coding of your categorical variables.
---
### Author:  [@Pascal\_J](https://discourse.mc-stan.org/u/Pascal_J)
#### Post date: [April 16, 2019, 12:55pm UTC](https://discourse.mc-stan.org/t/ordinal-regression-model-in-brms/8446/3 "2019-04-16T12:55:50Z")
Hi Paul,
Thank you for your answer!
> [@paul.buerkner](#):
>
> Hi, did you already take a look at this ordinal tutorial paper for brms: [OSF](https://psyarxiv.com/x8swp/)
I read it already many times, that’s how I managed to specify the model until now, but now I’m stuck at interpreting the results…
I wanted to check for the proportional odds assumption, so I tried to fit a `acat()` model with category-specific effects but the model failed to even start sampling with hundreds of `Rejecting initial value: Log probability evaluates to log(0), i.e. negative infinity` errors, so I gave up.
> [@paul.buerkner](#):
>
> Model comparison for instance via loo and loo\_compare
I gradually built up the final model I mentioned by adding parameters and comparing them via `loo_compare`, but where can I see precisely that a model is overfitted?
> [@paul.buerkner](#):
>
> face has just 4 levels. Modeling so many varying effects for it is probably not going to work out reasonably well.
I’m sorry, but I do not understand. You mean there are too many levels of `disfigurement` which vary for `face`?
> [@paul.buerkner](#):
>
> Take a look at the `hypothesis` method
I already did, but as I already said, I don’t understand too much of the statistical and mathematical finesse to derive the correct methods… In concrete terms, I don’t know how to formulate the method call for this particular task.
> [@paul.buerkner](#):
>
> Depends on the coding of your categorical variables.
I apologise for my naïveté, but I don’t understand, what you mean…
Kindest regards,
Pascal
---
### Author:  [@paul.buerkner](https://discourse.mc-stan.org/u/paul.buerkner)
#### Post date: [April 16, 2019, 1:00pm UTC](https://discourse.mc-stan.org/t/ordinal-regression-model-in-brms/8446/4 "2019-04-16T13:00:52Z")
I see. Apologies for being a little bit high level. I will try to provide more detailed explanation later on today or tomorrow.
---
### Author:  [@paul.buerkner](https://discourse.mc-stan.org/u/paul.buerkner)
#### Post date: [April 17, 2019, 2:13pm UTC](https://discourse.mc-stan.org/t/ordinal-regression-model-in-brms/8446/5 "2019-04-17T14:13:32Z")
> I wanted to check for the proportional odds assumption, so I tried to fit a `acat()` model with category-specific effects but the model failed to even start sampling with hundreds of `Rejecting initial value: Log probability evaluates to log(0), i.e. negative infinity` errors, so I gave up.
You could trying using `inits = 0` or `init_r = 0.1` (or smaller for the latter) to initialize the model.
> I gradually built up the final model I mentioned by adding parameters and comparing them via `loo_compare` , but where can I see precisely that a model is overfitted?
There is no threshold or something. You could start with a model that you know does not overfit (because it is too simple) and then see what actually improves predictions.
> I’m sorry, but I do not understand. You mean there are too many levels of `disfigurement` which vary for `face` ?
Quite that. What I mean is that you can model it that way, but the hierachical standard deviations and correlations won’t be estimated very precisely as face has just 4 levels.
> I apologise for my naïveté, but I don’t understand, what you mean…
Coding categorical variables is not brms specific but works the same way in many model fitting R packages. You could starting by googeling “dummy coding” and “effect coding” to see what that implies with respect to the resulting coefficients. This will also help you understanding what to specify in the `hypothesis` method.