troubles.Rmd

---
title: 'Troubleshooting'
---

```{r, include=FALSE}
library(tidyverse)
```

# Common problems and troubleshooting {#troubles}

### Scaling inputs {- #scaling-regression-inputs}

Sometimes you might come across error messages like these:

```
Warning messages: 1: Some predictor variables are on very different scales: consider rescaling

Model is nearly unidentifiable: very large eigenvalue - Rescale variables?;

Model is nearly unidentifiable: large eigenvalue ratio - Rescale variables?
```

Other similar messages might refer to 'convergence failure', e.g. for `lmer()`
models.

You should be aware that there are limits to the precision with which R (and
computers in general) can store decimal values. This only tends to matter when
working with very large or very small numbers — but this can crop up when
estimating regression coefficients or (most often) variances that are extremely
close to zero.

If you have predictors and/or outcomes on very different scales you may also end
up with very small regression coefficients, and this can make it harder for R to
fit your model, or [for you to interpret it](#scaling-regression-inputs).

#### Troubleshooting

If you see any of these errors the first step is to
[rescale your inputs](#scaling-regression-inputs)
([see this section for how](#scaling-regression-inputs)), for example by
dividing or multiplying so that they are in more natural or convenient units.

Where variances or covariances are very close to zero, try simplifying your
model. For example, if you have multiple random effects in a multilevel model,
[try elimintating some of them, or constraining their covariances](#simplifying-mixed-models).