LRT vs Wald tests

[tavareshugo]

In the materials, we're using the LRT model to test:

• ~ Timepoint + Status vs ~ Status
• ~ Timepoint + Status + Timepoint:Status vs ~ Timepoint + Status (in the exercise)

However, the hypothesis here can be easily tested using the default Wald test, so it may be a bit misleading to encourage learners to use LRT for these sorts of simple hypothesis.

The exception would be if the variables had more than 2 levels, because then we'd be testing for multiple levels (i.e. coefficients) all at once. But unfortunately that's not the case in our data.

We could perhaps use ~ Timepoint + Status + Timepoint:Status vs ~ Status, because then we're testing the general effect of timepoint (including its interaction with status), i.e. simultaneously testing two model coefficients at once.

Or even ~ Timepoint + Status + Timepoint:Status vs ~ 1, asking the question of whether these two variables together explain any gene expression at all (regardless of what their individual effects might be). For example, we may have wanted to do this sort of broader test for further downstream gene clustering or co-expression network analysis.

[tavareshugo]

LRT test also needs to be expanded in the presentation and be more explicit what it is about.