Waiting for my model to break

I am deep in the throes of data modelling, trying to find a statistical model that adequately describes my data so that I can produce some results to write about in my thesis.  In practical terms, what this means is feeding the computer my data plus a long equation that contains all of the factors that I think affect the genitive (and the ways they might interact with each other), then waiting (im)patiently for a few minutes while it whirrs away to itself and either spits out a bunch of numbers, or a nasty red error message telling me the model has broken (which basically says the calculations have got too big for the computer to cope with, so it just gave up).  Sometimes when the model breaks, just rearranging a few terms in the equation (while keeping it mathematically equivalent) is enough to fool the computer into trying a bit harder, but it’s a bit of a guessing game as to which terms to swap around, so it can take a lot of tries before I get a model that doesn’t break.
Then, once I get a model that runs properly, I note down its particular bunch of numbers, one of which is a score which tells me how good the model is at describing my data.  Then I start again with a new model with a different combination of factors and interactions, and compare its score with the previous model’s, to decide which model is better, and so on until I find the best possible model.  None of the models will ever be perfect (because it’s just a model), but in theory, if you search long enough, and try every possible combination, then you can find the one that’s going to get closest to properly describing your data.  (And once I find that, I can use all the other numbers it spits out to make pretty graphs and hopefully prove something interesting!)
But that takes a long time (see taking several minutes to run each model, plus often having to try each model several times before it works – in the last hour, I’ve managed to test 5 potential models).  And there are hundreds of potential combinations.  Of course, there are ways of narrowing down some of the possible combinations, but it still means a lot of sitting around twiddling my thumbs while I wait to see if it’ll work this time.  And I’ve got to keep one eye on the window the model is running in so I know if it’s finished (or broken again), so I can’t do much constructive work while I’m waiting.  So as an experiment, I thought I’d try writing a post here while I wait (so if this seems a bit disjointed, now you know why – every few sentences I have to go back to my modelling, then try and pick up my train of thought here while my next model runs).
Not that I’ve got a lot to write about here – life is its usual round of work, study, work, study, try and find time to eat and sleep, work and study some more.  Talking of which, it’s getting late – I should go and find something for dinner.

Similar Posts

2 Comments

    1. Actually, it’s not all that different than testing a recipe – add a pinch of this, reduce the amount of that, and see how it turns out. Cookies would be a lot tastier though! 🙂

Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.