(c) 2018 Justin Bois. With the exception of pasted graphics, where the source is noted, this work is licensed under a Creative Commons Attribution License CC-BY 4.0. All code contained herein is licensed under an MIT license.
This document was prepared at Caltech with financial support from the Donna and Benjamin M. Rosen Bioengineering Center.
This homework was generated from an Jupyter notebook. You can download the notebook here.
import numpy as np
import numba
a) Write your own MCMC sampler that employs a Metropolis-Hastings algorithm to sample continuous parameters (i.e., it does not need to handle discrete parameters) that uses a Gaussian proposal distribution. Since you are sampling multiple parameters, your proposal distribution will be multi-dimensional. You can use a Gaussian proposal distribution with a diagonal covariance. In other words, you generate a proposal for each variable in the posterior independently.
You can organize your code how you like, but here is a suggestion.
pass
ing).def mh_step(x, logpost, logpost_current, sigma, args=()):
"""
Parameters
----------
x : ndarray, shape (n_variables,)
The present location of the walker in parameter space.
logpost : function
The function to compute the log posterior. It has call
signature `logpost(x, *args)`.
logpost_current : float
The current value of the log posterior.
sigma : ndarray, shape (n_variables, )
The standard deviations for the proposal distribution.
args : tuple
Additional arguments passed to `logpost()` function.
Returns
-------
output : ndarray, shape (n_variables,)
The position of the walker after the Metropolis-Hastings
step. If no step is taken, returns the inputted `x`.
"""
pass
n_burn
as opposed to n_warmup
, because you are just going to step as you normally would, and then "burn" the samples. This is in contrast to Stan, which tunes the stepping procedure duing the warm-up phase.)def mh_sample(logpost, x0, sigma, args=(), n_burn=1000, n_steps=1000,
variable_names=None):
"""
Parameters
----------
logpost : function
The function to compute the log posterior. It has call
signature `logpost(x, *args)`.
x0 : ndarray, shape (n_variables,)
The starting location of a walker in parameter space.
sigma : ndarray, shape (n_variables, )
The standard deviations for the proposal distribution.
args : tuple
Additional arguments passed to `logpost()` function.
n_burn : int, default 1000
Number of burn-in steps.
n_steps : int, default 1000
Number of steps to take after burn-in.
variable_names : list, length n_variables
List of names of variables. If None, then variable names
are sequential integers.
Returns
-------
output : DataFrame
The first `n_variables` columns contain the samples.
Additionally, column 'lnprob' has the log posterior value
at each sample.
"""
pass
b) To test your code, we will get samples out of a known distribution. We will use a bivariate Gaussian with a mean of $\boldsymbol{\mu} = (10, 20)$ and covariance matrix of
\begin{align} \boldsymbol{\sigma} = \begin{pmatrix} 4 & -2 \\ -2 & 6 \end{pmatrix} \end{align}
I have written the function to be unnormalized and JITted with numba for optimal speed.
Do not be confused: In this test function we are sampling $\mathbf{x}$ out of $P(\mathbf{x}\mid \boldsymbol{\mu},\boldsymbol{\sigma})$. This is not sampling a posterior; it's just a test for your code. You will pass log_test_distribution
as the log_post
argument in the above functions.
mu = np.array([10.0, 20])
cov = np.array([[4, -2],[-2, 6]])
inv_cov = np.linalg.inv(cov)
@numba.jit(nopython=True)
def log_test_distribution(x, mu, inv_cov):
"""
Unnormalized log posterior of a multivariate Gaussian.
"""
return -np.dot((x-mu), np.dot(inv_cov, (x-mu))) / 2
If you compute the means and covariance (using np.cov()
) of your samples, you should come close to the inputed means and covariance. You might also want to plot your samples using bebi103.viz.corner()
to make sure everything makes sense.
You may want to add in some logic to your Metropolis-Hastings sampler to enable tuning. This is the process of automatically adjusting the $\sigma$ in the proposal distribution such that the acceptance rate is desirable. The target acceptance rate is about 0.4. The developers of PyMC3 use the scheme below, which is reasonable.
Acceptance rate | Standard deviation adaptation |
---|---|
< 0.001 | $\times$ 0.1 |
< 0.05 | $\times$ 0.5 |
< 0.2 | $\times$ 0.9 |
> 0.5 | $\times$ 1.1 |
> 0.75 | $\times$ 2 |
> 0.95 | $\times$ 10 |
Be sure to test your code to demonstrate that it works.
In Homework 6, you investigated a data set on reversals of optogenetic worms upon exposure to blue light. As a reminder, here are the data.
Strain | Year | Trials | Reversals |
---|---|---|---|
WT | 2017 | 55 | 7 |
ASH | 2017 | 54 | 18 |
AVA | 2017 | 52 | 28 |
WT | 2016 | 36 | 6 |
ASH | 2016 | 35 | 12 |
AVA | 2016 | 36 | 30 |
WT | 2015 | 35 | 0 |
ASH | 2015 | 35 | 9 |
AVA | 2015 | 36 | 33 |
Again, for the purposes of this problem, assume that we can pool the results from the two years to have 13/126 reversals for wild type, 39/124 reversals for ASH, and 91/124 reversals for AVA.
The pertinent parameter is $\theta$, the probability of reversal of a worm upon illumination.
a) Use Stan to get samples of $\theta$ for each of the three strains. Plot either histograms or ECDFs of your samples.
b) Use your Metropolis-Hastings sampler from the previous problem to do the same.
c) The posterior plots of $\theta$ are illuminating, but suppose we want to quantify the difference in reversal probability between the two strains, say strain 1 and strain 2. That is, we want to compute $g(\delta_{12}\mid n_1, N_1, n_2, N_2)$, where $\delta_{12} \equiv \theta_2 - \theta_1$. Note that computing this distribution by hand is quite difficult.