# Using loo function to help me choose a distribution

**URL:** <https://discourse.mc-stan.org/t/using-loo-function-to-help-me-choose-a-distribution/25511>\
**Category:** brms\
**Created:** [November 30, 2021, 11:17pm UTC](https://discourse.mc-stan.org/t/using-loo-function-to-help-me-choose-a-distribution/25511 "2021-11-30T23:17:08Z")\
**Posts on this page:** 4\
**Page:** 1

<div class="post-metadata">

**Author:** ![mountaingoat19](https://yyz2.discourse-cdn.com/flex030/user_avatar/discourse.mc-stan.org/mountaingoat19/32/11677_2.png) [@mountaingoat19](https://discourse.mc-stan.org/u/mountaingoat19)\
**Post date:** [November 30, 2021, 11:17pm UTC](https://discourse.mc-stan.org/t/using-loo-function-to-help-me-choose-a-distribution/25511/1 "2021-11-30T23:17:08Z")

</div>

I have two models exactly similar, but I’m using a Poisson distribution for one and a Bernoulli distribution for the other. Can I trust the information coming out of loo?

The Bernoulli model comes on top.

```
rbpa <- brm(status ~ ... , family = poisson(), prior = set_prior('normal(0, 4)', class = 'b'), cores = 6, iter = 4000)

rbba <- brm(status ~ ... , family = bernoulli ,prior(normal(0, 4), class = b), cores = 6, iter = 4000)

loo(rbpa,rbba)

```

Status is if my individual was harvested or not .

- Operating System: Windows 10
- brms Version: 2.16.3

---

<div class="post-metadata">

**Author:** ![Max\_Mantei](https://yyz2.discourse-cdn.com/flex030/user_avatar/discourse.mc-stan.org/max_mantei/32/7912_2.png) [@Max\_Mantei](https://discourse.mc-stan.org/u/Max_Mantei)\
**Post date:** [December 1, 2021, 1:57pm UTC](https://discourse.mc-stan.org/t/using-loo-function-to-help-me-choose-a-distribution/25511/2 "2021-12-01T13:57:27Z")

</div>

Hey there!

If the outcome is binary [0,1], personally, I’d always go for the Bernoulli model. A Poisson model would not be representative of the generative process of the data, as it could result in predictions such as status=2 or status=15.

Besides that, there was some discussion about whether you could compare loo on models with different likelihoods for the observational model here – can’t find it right now. If you follow Richard McElreath’s _Statistical Rethinking_, you shouldn’t do it. But IIRC, @avehtari might have a different view?

Cheers,  
Max

---

<div class="post-metadata">

**Author:** ![avehtari](https://yyz2.discourse-cdn.com/flex030/user_avatar/discourse.mc-stan.org/avehtari/32/5935_2.png) [@avehtari](https://discourse.mc-stan.org/u/avehtari)\
**Post date:** [December 1, 2021, 2:10pm UTC](https://discourse.mc-stan.org/t/using-loo-function-to-help-me-choose-a-distribution/25511/3 "2021-12-01T14:10:07Z")

</div>

See [CV-FAQ 12 Can cross-validation be used to compare different observation models / response distributions / likelihoods?](https://avehtari.github.io/modelselection/CV-FAQ.html#12_Can_cross-validation_be_used_to_compare_different_observation_models __response_distributions__ likelihoods)

---

<div class="post-metadata">

**Author:** ![avehtari](https://yyz2.discourse-cdn.com/flex030/user_avatar/discourse.mc-stan.org/avehtari/32/5935_2.png) [@avehtari](https://discourse.mc-stan.org/u/avehtari)\
**Post date:** [December 1, 2021, 2:11pm UTC](https://discourse.mc-stan.org/t/using-loo-function-to-help-me-choose-a-distribution/25511/4 "2021-12-01T14:11:25Z")

</div>

@Max_Mantei has a good point that if the outcome is binary, then Poisson is not a good choice anyway