What is the burn-in period in MCMC and why must those samples be discarded?

Conditions

• Applies to MCMC sampling.
• Requires discarding initial samples before using the rest for inference.
• The chain starts from an arbitrary point.

Reasoning, step by step

1. Identify the initial phase of the MCMC simulation.
2. Observe that early samples are still converging to high probability regions.
3. Recognize that these samples are biased by the starting point.
4. Discard the initial samples to ensure the remaining samples accurately reflect the target distribution.

Example

The video highlights the initial red samples on the probability density curve with a red box labeled 'Burn-in', explaining that they are still finding their way to high probability regions.

Common misconceptions

• Believing that all samples generated by an MCMC chain, including the very first ones, can be used directly for statistical inference.
• Assuming the burn-in period has a fixed length that applies universally to all problems.

Watch the explanation

Explore next

• How do you read a trace plot to determine when an MCMC chain has reached stationarity?
• What is the mathematical formula for the detailed balance condition?
• Why is the memoryless property of Markov chains useful?