Eric J Ma's Website

Statistical tests are just canned model comparisons

written by Eric J. Ma on 2020-06-28 | tags: data science bayesian statistics hypothesis testing

I reached another statistics epiphany today.

Statistical tests? They are nothing more than a comparison between statistical models that has been made canned.

What do I mean?

Let's say we have data for Korean men from both the North and the South. We could write a generative model for men's height that looks something like the following:

$$h^n \sim N(\mu^{n}, \sigma^{n})$$

$$h^s \sim N(\mu^{s}, \sigma^{s})$$

(The superscripts don't refer to powers, but "north" and "south"; I was just having difficulty getting Markdown + LaTeX subscripts parsed correctly.)

In effect, we have one data generating model per country.

Think back to how you might analyze Korean male height data using canned statistical procedures. You might choose the t-test. In performing the t-test, we make some assumptions:

  • The two populations have equal variance, and
  • We are only interested in comparing the means.

Yes, there are variations on the t-test, such as t-test with unequal variance and the likes, but I wonder how many of us would really think deep about the assumptions underlying the procedure when faced with data and a "canned protocol"? Instead, we'd simply ask, "is there a difference in the means?" And then calculate some convoluted p-value.

How could we go beyond canned procedures and tell if the two populations have a difference? Well, we could compare their means, but we could also add another comparison of the variance too. All of this comes very naturally from thinking about the data generating story. Let's see this in action.

In performing Bayesian inference, we would obtain posterior distributions for both the $\mu$ and $\sigma$ parameters of both models. We could then compare the $\mu$ and the $\sigma$ parameters, and they would both be informative.

For example, if $\mu$s turned out to be different and their posteriors do not overlap, we would claim that the two populations differ in their average height, and explain it as being due to the nutritional difference arising from economic mismanagement.

On the other hand, if the $\sigma$ turned out to be different and their posterios didn't overlap, we would claim that the two populations differ in their variation in height. How would we explain this? For example, if the distribution of North Korean heights was tighter giving rise to a smaller $\sigma$ value, we might explain this as being due to rationing and genetic homogeneity.

In both comparisons, by thinking more precisely about the data generating process, we had access to a broader set of hypotheses. If we had defaulted to the t-test as a "canned statistical procedure", and not thought carefully about setting up a data generating model, we would be missing out on a rich world of inquiry into our data.

I send out a newsletter with tips and tools for data scientists. Come check it out at Substack.

If you would like to sponsor the coffee that goes into making my posts, please consider GitHub Sponsors!

Finally, I do free 30-minute GenAI strategy calls for organizations who are seeking guidance on how to best leverage this technology. Consider booking a call on Calendly if you're interested!