November 13, 2021
The model now has a huge test suite, all of which passes, but I worry about test coverage still.
The lack of worked examples of benefit calculations in either the CPAG Guides or the online guides from the Government, Shelter, Age UK and the rest is a real problem. I’n guessing the incentive here is to upsell training courses on this stuff.
Instead, I’ve been testing against an online calculator from Policy in Practice. As someone who used to write this sort of thing, I have to say the PiP calculator is really very well designed and easy to use.
It’s a slightly frustrating task as comparing against another calculator means you need pretty much the whole model completed before you can start, rather than test-as-you-write against each component, which I’d prefer (there’s a lot of test-as-you go, but it’s not very thorough). Testinging against the calculator, there are some minor niggles that needed sorting out, like how many weeks are there in a year (not as obvious as you might think)? But testing against PiP was a really productive experience, at first turning up some really embarrassing errors I’d made, and then eventually getting to the point where I felt I was slightly ahead of them (my income tax calculations are a bit more complete in obscure corner cases, for instance).
I still don’t think there are enough tests though. For instance, my inequality routines broke badly when confronted by real data with negative incomes in it, despite passing a pretty large test suite based on worked examples from the World Bank (a few of the the world bank ones were wrong, incidentally).
One very good testing technique for a model like this is to drive a budget constraint generator through it and see if the results can be rationalised - the generator homes right in on all the wierd stuff - net incomes that drop off a cliff with small increments to wages, marginal tax rates over 100 or less than zero. Scotben is now doing very well there, at least to the extent that I understand how all the interactions are supposed to work. I’ll write about that next.
On the process of testing, I’ve been reading the Google Testing on the Toilet Door Series - recommended. My test suite would not do well against Google’s principles but it’s the best I can do on my own to this timescale.