nunosempere.com/blog/2022/04/17/simple-squiggle/index.md
2022-05-20 21:17:49 +00:00

4.0 KiB

Simple Squiggle

Linkpost for github.com/quantified-uncertainty/simple-squiggle

"Simple Squiggle" is a simple parser that manipulates multiplications and divisions between numbers and lognormal distributions. It uses an extremely restricted subset of Squiggle's syntax, and unlike it, the underlying code is not easily extensible.

It may be useful for testing correctness of limited features of the full Squiggle, or for sanity-checking the validity of some Squiggle models.

Built with

Getting started

Prerequisites

  • npm
  • nodejs

Installation

git clone https://github.com/quantified-uncertainty/simple-squiggle.git
cd simple-squiggle
## npm install

The last line is not necessary, since I'm saving node_packages in the repository.

Usage

Consider a squiggle model which only uses lognormals:

initialPrisonPopulation = 1.8M to 2.5M # Data for 2022 prison population has not yet been published, though this estimate is perhaps too wide.
reductionInPrisonPopulation = 0.25 to 0.75
badnessOfPrisonInQALYs = 0.2 to 5 # 80% as good as being alive to 5 times worse than living is good
accelerationInYears = 5 to 50
probabilityOfSuccess = 0.01 to 0.1 # 1% to 10%.
estimateQALYs = leftTruncate(
    initialPrisonPopulation *
    reductionInPrisonPopulation *
    badnessOfPrisonInQALYs *
    accelerationInYears *
    probabilityOfSuccess
    , 0)
cost = 2B to 20B
costEffectivenessPerQALY = leftTruncate(cost / estimateQALYs, 0)
costEffectivenessPerQALY

It can be simplified to the following simple squiggle model:

( 2000000000 to 20000000000 ) / ( (1800000 to 2500000) * (0.25 to 0.75) * (0.2 to 5) * (5 to 50) * (0.01 to 0.1) )

I provide both an exportable library and a command line interface (cli). The cli can be run with npm run cli, which produces a prompt:

> npm run cli

Model:

After filling in the prompt

> npm run cli

Model: ( 2000000000 to 20000000000 ) / ( (1800000 to 2500000) * (0.25 to 0.75) * (0.2 to 5) * (5 to 50) * (0.01 to 0.1) )

the output looks as follows:

> npm run cli

Model: ( 2000000000 to 20000000000 ) / ( (1800000 to 2500000) * (0.25 to 0.75) * (0.2 to 5) * (5 to 50) * (0.01 to 0.1) )

        = (lognormal(22.57, 0.70)) / ((lognormal(14.57, 0.10)) * (lognormal(-0.84, 0.33)) * (lognormal(0.00, 0.98)) * (lognormal(2.76, 0.70)) * (lognormal(-3.45, 0.70)))
        -> lognormal(22.57, 0.70) / (lognormal(14.57, 0.10) * lognormal(-0.84, 0.33) * lognormal(0.00, 0.98) * lognormal(2.76, 0.70) * lognormal(-3.45, 0.70))
        -> lognormal(22.57, 0.70) / (lognormal(13.73, 0.35) * lognormal(0.00, 0.98) * lognormal(2.76, 0.70) * lognormal(-3.45, 0.70))
        -> lognormal(22.57, 0.70) / (lognormal(13.73, 1.04) * lognormal(2.76, 0.70) * lognormal(-3.45, 0.70))
        -> lognormal(22.57, 0.70) / (lognormal(16.49, 1.25) * lognormal(-3.45, 0.70))
        -> lognormal(22.57, 0.70) / (lognormal(13.04, 1.43))
        -> lognormal(22.57, 0.70) / lognormal(13.04, 1.43)
        -> lognormal(9.53, 1.60)

=> lognormal(9.530291704996749, 1.596443005980748)
----------------------------------------------------

For ease of representation, the intermediary outputs are printed only to two decimal points. But this is just a display decision; the innards of the program work with the full set of decimals.

You can also run tests with npm run test

Roadmap

I consider this repository to be feature complete. As such, I may tinker with the code which wraps around the core logic, but I don't really intend to add further functionality.

  • Make wrapper code less hacky
  • Display final lognormal as a 90% confidence interval as well

License

Distributed under the MIT License