The current model uses a log link to model M-ratio, which excludes the possibility of a negative M-ratio. This is unlikely to happen, but can arise for two reasons:
- A participant has below-chance performance (i.e., a negative d' value) but above-chance metacognition (i.e., positive meta-d')
- A participant has above-chance performance (i.e., positive d') but below-chance metacognition (i.e., negative meta-d')
We should not allow negative M-ratios as a default, because typical experimental paradigms should not fall into either of these cases and allowing them is likely to lead to issues in model fitting. But, it would be good to have an option for use cases that do fall into one of the above cases.
To do this, the immediate solution would be to model meta-d' as unconstrained with an identity link instead of M-ratio as brms' primary variable (mu). This should be easy to implement, but my impression is that we shouldn't expect the model to converge well.
The current model uses a log link to model M-ratio, which excludes the possibility of a negative M-ratio. This is unlikely to happen, but can arise for two reasons:
We should not allow negative M-ratios as a default, because typical experimental paradigms should not fall into either of these cases and allowing them is likely to lead to issues in model fitting. But, it would be good to have an option for use cases that do fall into one of the above cases.
To do this, the immediate solution would be to model meta-d' as unconstrained with an identity link instead of M-ratio as brms' primary variable (mu). This should be easy to implement, but my impression is that we shouldn't expect the model to converge well.