Bayesian inference and MCMC sampling introduction

A gentle and intuitive introduction to Markov Chain Monte Carlo sampling with rejection sampling.

Hause Lin
04-09-2019

Get source code for this RMarkdown script here.

Consider being a patron and supporting my work?

Donate and become a patron: If you find value in what I do and have learned something from my site, please consider becoming a patron. It takes me many hours to research, learn, and put together tutorials. Your support really matters.

Bayesian inference, priors, likelihood, posteriors, and especially Markov Chain Monte Carlo sampling can be mysterious. I will try to explain these concepts with a simple 1-dimensional example, using something known as rejection sampling, which is good for explaining how sampling works, but is never used in practice because it is highly inefficient (we’ll see what that means later on).

Binomial distribution example (life or death, heads or tails)

In a hospital, there are a total of 100 patients suffering from different deadly diseases. Doctor X is very optimistic and thinks that 83% of the patients will survive after 2 days. That is, Doctor X is expecting 83 patients to survive and 17 to die. This situation is like typical coin-toss scenarios, where survive is heads (coded 1) and death is tails (coded 0). Coin tosses and our scenario can be described using the binomial distribution.

Let’s generate some random data assuming we have 100 patients and 83% survival probability. Because the number is randomly generated from a binomial distribution (n = 100 and success = 83), we might not get exactly 83 as the output.


n_patients <- 100 # number of patients
survive_probability <- 0.83 # probability of surviving
number_of_simulations <- 1
rbinom(n = number_of_simulations, size = n_patients, prob = survive_probability)

[1] 83

Let’s repeat the scenario/simulation 10 times. Because the numbers are randomly generated from a binomial distribution (n = 100 and success = 83), we get slightly different numbers for each simulation. But these numbers should be close to 83.


number_of_simulations <- 10
rbinom(n = number_of_simulations, size = n_patients, prob = survive_probability)

[1] 88 80 79 87 87 84 85 76 88 87

Representing beliefs about probability with the beta distribution

Doctor X believes 83% of the 100 patients will survive. In other words, he believes 83% is the most likely survival probability, whereas 82% and 84% are also likely, but less likely than 83%. We can represent Doctor X’s belief using a beta distribution, which is useful for describing probabilities because it ranges from 0 to 1, and has two parameters that define the shape of the distribution.

If you’re not familiar with the beta distribution, here are the key properties of the beta distribution:

• it ranges from 0 to 1
• shape1 and shape2 parameter can be thought of as the “success” and “failure” parameters
• shape1 and shape2 are known as the alpha and beta parameters

Below are the shapes of beta distributions with different shape1 and shape2 parameters. As shown below, the beta distribution is very flexible and changing the two parameters of the distribution can change the shape of the distribution drastically!


par(mfrow = c(2, 3)) # display subplots in a 2 x 3 grid
curve(dbeta(x, shape1 = 1.0, shape2 = 1.0), from = 0, to = 1) # uniform
curve(dbeta(x, shape1 = 1.2, shape2 = 1.2), from = 0, to = 1)
curve(dbeta(x, shape1 = 4, shape2 = 4), from = 0, to = 1) # normal looking
curve(dbeta(x, shape1 = 0.5, shape2 = 0.5), from = 0, to = 1) # u-shape
curve(dbeta(x, shape1 = 5, shape2 = 2), from = 0, to = 1) # success more likely than failure
curve(dbeta(x, shape1 = 3, shape2 = 6), from = 0, to = 1) # failure more likely than success

We can represent Doctor X’s belief with the beta distribution below (shape1 = 3.5, shape2 = 1.5). Doctor X believes that 83% survival is the most likely to occur, and 40% survival is about half as likely to occur, but what this beta distribution says is that Doctor X believes that different survival probabilities ranging from 0 to 1 are plausible, even though he believes that 83% is most likely to occur.


par(mfrow = c(1, 1)) # change back to showing plots in a 1 x 1 grid
curve(dbeta(x, shape1 = 3.5, shape2 = 1.5), from = 0, to = 1,
xlab = "Plausible probabilities (survival percent or proportion)",
ylab = "Likelihood",
main = "Doctor X's belief")
abline(v = 0.83, col = "black", lwd = 3)

In Bayesian terminology, Doctor X’s belief is his prior (prior beliefs), $$P(\theta)$$.

After 2 days, only 72 patients survived (28 died). Doctor X’s priors were not too accurate, it seems. Before we discuss how Doctor X should update/revise his beliefs, let’s return to our simulations from earlier on to see how rare/frequent it is to get 72 survivals (out of 100), assuming that we believe that 83% should survive.


n_patients <- 100 # number of patients
survive_probability <- 0.83 # probability of surviving
number_of_simulations <- 1
rbinom(n = number_of_simulations, size = n_patients, prob = survive_probability)

[1] 80

Run rbinom(n = number_of_simulations, size = n_patients, prob = survive_probability) a few more times and see if you manage to get 72? You might never see any 72s. But if you keep trying, you eventually will!

Rather than manually re-run the function, let’s run 100,000 simulations to see how many 72s we observe.


number_of_simulations <- 100000
simdata <- rbinom(n = number_of_simulations, size = n_patients, prob = survive_probability)
sum(simdata == 72) # total number of simulations that resulted in 72 survivals

[1] 201

Only 201 of 100000 simulations (assuming 83% survival probability) generated 72 survivals!

Remember, 83% is just one of the plausible parameter probability values in Doctor X’s prior beliefs distribution. See beta distribution reflecting his beliefs for the entire range of probability values in the figure above! 72% might not be as likely as 83% in his prior beliefs distribution, but it’s still fairly likely to occur!

Let’s run another 100,000 simulations to see how many 72s we can get, assuming the prior belief of survival is 72%.


n_patients <- 100 # number of patients
survive_probability <- 0.72 # probability of surviving
number_of_simulations <- 100000
simdata <- rbinom(n = number_of_simulations, size = n_patients, prob = survive_probability)
sum(simdata == 72) # total number of simulations that resulted in 72 survivals

[1] 9096

This time, 9096 of 100000 simulations generated 72 survivals!

Now, we can start to understand rejection sampling, MCMC, and how prior information and likelihood combine to form the posterior!

Generate prior distribution by randomly creating data

First, let’s generate random samples to create Doctor X’s prior distribution. Remember, we’re using the beta distribution (ranges from 0 to 1) to generate probability values reflecting Doctor X’s beliefs.


n_patients <- 100 # number of patients
survive_probability <- 0.83 # prior belief of probability of surviving
prior_nsamples <- 20000
priordist <- rbeta(n = prior_nsamples, shape1 = 3.5, shape2 = 1.5)
head(priordist) # probability values

[1] 0.4622312 0.9634323 0.8193993 0.6176060 0.9533685 0.8154673

It’s always good to plot and visualize how your distributions look like.


hist(priordist, main = paste0("Doctor X's prior", " (prior samples = ", length(priordist), ")"),
xlab = "Plausible probability values", ylab = "Count", xlim = c(0, 1))
abline(v = survive_probability, col = "black", lwd = 3)

Define likelihood (generative model) and simulate data samples

The likelihood, $$P(data|\theta)$$, is the generative model, the theoretical model we think is generating the data.

Earlier on, we have only generated data with rbinom() assuming 83% and 72% survival. The prob parameter in the rbinom() function actually allows you to provide more than 1 probability at a time, which means we can enter the entire range of probability values ranging from 0 to 1 that reflects Doctor X’s prior belief beta distribution.

First, let’s remind ourselves what are the values contained in our priordist vector.


length(priordist)

[1] 20000

head(priordist) # different probability values generated from beta distribution

[1] 0.4622312 0.9634323 0.8193993 0.6176060 0.9533685 0.8154673

summary(priordist) # beta dist ranges from 0 to 1

Min. 1st Qu.  Median    Mean 3rd Qu.    Max.
0.04919 0.57868 0.72652 0.69931 0.84684 0.99990 

# generate data from the binomial distribution
simulated_data <- rbinom(n = length(priordist), size = n_patients, prob = priordist)
length(simulated_data)

[1] 20000

head(simulated_data)

[1] 43 95 79 56 96 82

summary(simulated_data)

Min. 1st Qu.  Median    Mean 3rd Qu.    Max.
3.00   57.00   73.00   69.88   85.00  100.00 

If rbinom(n = length(priordist), size = n_patients, prob = priordist) doesn’t make sense, stop and think about what’s going on… For each value in the prior distribution priordist (continuous values 0 to 1 generated from the beta distribution reflecting Doctor X’s prior beliefs), we simulated (randomly generated) data from a binomial distribution with 100 patients (size = n_patient) and a specific survival probability (0.4622312 in the first simulation).

Identify samples that match our observed outcome


sim_equals_observed_boolean <- simulated_data == 72 # "combine" prior with data
head(sim_equals_observed_boolean)

[1] FALSE FALSE FALSE FALSE FALSE FALSE

Identify prior samples associated with samples that match observed outcome to get the posterior!


posterior <- priordist[sim_equals_observed_boolean] # "combine" prior with data
length(posterior)

[1] 407

head(posterior)

[1] 0.7316867 0.7086794 0.7528067 0.6636405 0.7322580 0.7259339

Yes, this posterior is the Bayesian posterior, $$P(\theta|data)$$, the probability of your theory given the data, or Doctor X’s updated/revised beliefs! We have combined the prior, $$P(\theta)$$ with the likelihood, $$P(data|\theta)$$ to get the posterior, $$P(\theta|data)$$. Below is Bayes’ theorem:

$P(\theta|data) = \frac{P(data|\theta)P(\theta)}{P(data)}$

Rejection sampling: Note that we have only 407 samples in our posterior distribution, even though we started off with 20000 samples in the prior! We have rejected all the prior samples that fail to generate 72 survivals (observed outcome) during our simulations. Hence the term “rejection sampling”. Rejection sampling is highly inefficient because we reject or throw out many samples! In practice, we never use rejection sampling, but it’s good to illustrate MCMC principles and how sampling works to combine the prior with the data to produce the posterior.

Summarize posterior samples


summary(posterior)

Min. 1st Qu.  Median    Mean 3rd Qu.    Max.
0.6039  0.6884  0.7199  0.7173  0.7476  0.8317 

quantile(posterior, c(0.025, 0.975))

2.5%     97.5%
0.6324368 0.7846651 

Compare prior and posterior


hist(priordist, main = paste0("Doctor X's prior", " (prior samples = ", length(priordist), ")"),
xlab = "Plausible probability values", ylab = "Count", xlim = c(0, 1))
abline(v = 0.83, col = "black", lwd = 3)


hist(posterior, main = paste0("Posterior", " (n = ", length(posterior), ")"),
xlab = "Plausible probability values", ylab = "Count", xlim = c(0, 1))
abline(v = mean(posterior), col = "blue", lwd = 5)
abline(v = quantile(posterior, c(0.025)), col = "blue", lwd = 1)
abline(v = quantile(posterior, c(0.975)), col = "blue", lwd = 1)

Create a simple function for further simulations and exploration


posterior_from_prior <- function(total_patients, total_survive,
prior_samples, prior_beta_shape1 = 1, prior_beta_shape2 = 1) {
survive_propr <- total_survive/total_patients
priordist <- rbeta(n = prior_samples, shape1 = prior_beta_shape1, shape2 = prior_beta_shape2)
simulated_data <- rbinom(n = length(priordist), size = total_patients, prob = priordist)
posterior <- priordist[simulated_data == total_survive]

par(mfrow = c(1, 2))
hist(priordist, main = paste0("Prior", " (N = ", length(priordist), ")"),
xlab = "Plausible probability values", ylab = "Count", xlim = c(0, 1))
abline(v = mean(priordist), col = "black", lwd = 3)
hist(posterior, main = paste0("Posterior", " (n = ", length(posterior), ")"),
xlab = "Plausible probability values", ylab = "Count", xlim = c(0, 1))
abline(v = survive_propr, col = "red", lwd = 3)
abline(v = mean(posterior), col = "blue", lwd = 5)
abline(v = quantile(posterior, c(0.025)), col = "blue", lwd = 1)
abline(v = quantile(posterior, c(0.975)), col = "blue", lwd = 1)
print(paste0("posterior mean: ", round(mean(posterior), 2)))
print(paste0("posterior median: ", round(median(posterior), 2)))
par(mfrow = c(1, 1))
}

Exploring different parameters (sample size, prior shapes)

Try different parameters yourself to get a better sense of how things work.

Line colors

• black: prior mean
• blue: posterior mean (thin blue lines are 2.5% and 97.5% intervals)
• red: observed value/data

posterior_from_prior(total_patients = 50, total_survive = 10,
prior_samples = 20000,
prior_beta_shape1 = 0.5, prior_beta_shape2 = 0.5)


[1] "posterior mean: 0.2"
[1] "posterior median: 0.2"

posterior_from_prior(total_patients = 100, total_survive = 70,
prior_samples = 20000,
prior_beta_shape1 = 0.1, prior_beta_shape2 = 0.9)


[1] "posterior mean: 0.69"
[1] "posterior median: 0.69"

Despite the skewed prior, the posterior distribution reflects the true value (70/100 = 0.7), but note how few samples we have in the posterior distribution!


posterior_from_prior(total_patients = 10, total_survive = 7,
prior_samples = 20000,
prior_beta_shape1 = 1.0, prior_beta_shape2 = 5.0)


[1] "posterior mean: 0.5"
[1] "posterior median: 0.5"

This time, notice how the terribly skewed prior prevents the sampler from properly “exploring” the full range of the posterior distribution, which doesn’t capture the data well (7/10 = 0.7 survival).

Support my work

Corrections

If you see mistakes or want to suggest changes, please create an issue on the source repository.

Reuse

Text and figures are licensed under Creative Commons Attribution CC BY 4.0. Source code is available at https://github.com/hauselin/rtutorialsite, unless otherwise noted. The figures that have been reused from other sources don't fall under this license and can be recognized by a note in their caption: "Figure from ...".

Citation

Lin (2019, April 9). Data science: Bayesian inference and MCMC sampling introduction. Retrieved from https://hausetutorials.netlify.com/posts/2019-04-09-gentle-and-intuitive-intro-to-markov-chain-monte-carlo-mcmc/
@misc{lin2019bayesian,
}