4
$\begingroup$

I'm new to Bayesian inference, and have been following the book Bayesian Methods for Hackers. In this chapter, the author provides an example of how using small sample sizes can lead to extreme results, but doesn't say how to correct for this when modelling the problem.

This is how I'm currently modeling the problem with pymc. average_across_county is an array of the average person's height per county. There is also an array called population (unused here) which contains the population of each county.

mu  = pm.Uniform('height', 100, 200)
tau = pm.Uniform('precision', 0, 50)
hgt_dist = pm.Normal('hgt_dist', mu, tau, value=average_across_county, observed=True)

mcmc = pm.MCMC([mu, tau, hgt_dist])
mcmc.sample(50000, 2000)

mcmc.trace('height')[:].mean()

The error in above estimate of the true average height is about 10 times larger than a simple weighted average because I don't make use of the population data. How should I include the population variable in this model?

$\endgroup$

1 Answer 1

3
$\begingroup$

I think I should add some explanation. This example was to show estimates with small samples sizes can have extreme values. I think if I was to "model" this example, I would model it as follows:

For each county, there is an true $mu$, but a priori we don't know this. From other information, I can suggest that the mean is between 100 and 200. I have sample data from each county, and I want to use this to derive a posterior per county, and compare these posteriors together (perhaps against the global data).

county_samples = [
 [150., 160., 170., 150.],
 [140., 160., 170., 180., 150., 130.],
 [160., 140., 170.],
]

mus = [pm.Uniform('mu_%d' % i, 100, 200) for i in range(3)]
taus = [pm.Uniform('tau_%d' % i, 0, 50) for i in range(3)]
obs = [pm.Normal('obs_%d' % i, mus[i], taus[i], observed=True, value=county_samples[i]) for i in range(3)]

mcmc = pm.MCMC([mus, taus, obs])
mcmc.sample(50000,10000)


mu_0 = mcmc.trace('mu_0')[:]
mu_1 = mcmc.trace('mu_1')[:]
mu_2 = mcmc.trace('mu_2')[:]

plt.hist(mu_0, bins=50, histtype='stepfilled', label='mu_0')
plt.hist(mu_1, bins=50, histtype='stepfilled', label='mu_1', alpha=0.5)
plt.hist(mu_2, bins=50, histtype='stepfilled', label='mu_2', alpha=0.5)

enter image description here

So, even though the data has different sample sizes, we can't rule out that they are statistically different from each other. (Try playing around with the data to try to make that untrue.)

Your example uses the precomputed averages per country. This is not usually how I model in PyMC - I prefer to keep things granular, and as you can see, it fits the framework nicely.

Does this help?

$\endgroup$
1
  • 2
    $\begingroup$ Cam, I'm terribly sorry for taking so long to reply, and want to thank you for taking the time answering this personally. However, it appears you are answering a slightly different question than what I asked. What if we don't have access to county_samples? How would we weigh samples in average_across_county if we only had access to the population variable? $\endgroup$
    – Mark Dunne
    Nov 23, 2016 at 14:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.