(Updated 9/2/2009, but still unfinished; see other’s work on this that I’ve collected)
I never took a statistics class, so I only know the kind of statistics you learn on the street. But now that I’m in global health research, I’ve been doing a lot of on-the-job learning. This post is about something I’ve been reading about recently, how to decide if a simple statistical model is sufficient or if the data demands a more complicated one. To keep the matter concrete (and controversial) I’ll focus on a claim from a recent paper in Nature that my colleague, Haidong Wang, choose for our IHME journal club last week: Advances in development reverse fertility declines. The title of this short letter boldly claims a causal link between total fertility rate (an instantaneous measure of how many babies a population is making) and the human development index (a composite measure of how “developed” a country is, on a scale of 0 to 1). Exhibit A in their case is the following figure:
An astute observer of this chart might ask, “what’s up with the scales on those axes?” But this post is not about the visual display of quantitative information. It is about deciding if the data has a piecewise linear relationship that Myrskyla et al claim, and doing it in a Bayesian framework with Python and PyMC. But let’s start with a figure where the axes have a familiar linear scale! Continue reading
Filed under MCMC, statistics
(Tap… tap… tap… is this thing on? Good.)
July was vacation month, where I went on a glorious bike tour of the Oregon/California coast, and learned definitively that I don’t like biking on the side of a highway all day. Don’t worry, I escaped in Coos Bay and took trains and buses between Eugene, Santa Cruz, Berkeley, and SF for a vacation more my speed.
But now that I’m back, August is turning out to be project month. I have 3 great TCS applications to global health in the pipeline, and I have big plans to tell you about them soon. But one mixed blessing about these applications is that people actually want to see the results, like, yesterday! So first I have to deal with the results, and then I can write papers and blogs about the techniques.
Since Project Month is a little over-booked with projects, I’m going to have to triage one today. You’ve heard of the NetFlix Challenge, right? Well, github.com is running a smaller scale recommendation contest, and I was messing around with personal page rank, which seems like a fine approach for recommending code repositories to hackers. I haven’t got it working very well (best results, 15% of holdout set recovered), but I was having fun with it. Maybe someone else will take it up, let me know if you get it to work; networkx + data = good times.
f = open('download/data.txt')
for l in f:
u_id, r_id = l.strip().split(':')
[get the code]