How does Polychoric Correlation Work? (aka Ordinal-to-Ordinal correlation)

Let's say you've got data of many paired cases of two ordinal variables, like you might when you ask a large number of people the same two Likert scale questions (e.g. "poor", "fair", "good", "very good", "excellent").

 

What could you learn from the data from those two questions?

Here's a few common approaches:

 

1) Compare the means of each variable by abusing a t-test.

2) Compare the distribution of each variable with a chi-squared goodness-of-fit test.

3) Check for a relationship between responses of each variable with a chi-squared independence test.

4) Estimate the strength of such a relationship with a Spearman correlation.