If the first level of Bayesian Inference introduced you to the philosophical shift from frequentist to subjective probability, Decision Theory and Bayesian Inference II is where the mathematical “gloves come off.” This unit moves beyond simple conjugate priors and into the territory of high-dimensional integration, complex loss functions, and hierarchical modeling. For students in advanced statistics or actuarial science, this represents the pinnacle of rational decision-making under extreme uncertainty.
Below is the exam paper download link
PDF Past Paper On Decision Theory And Bayesian Inference II For Revision
Above is the exam paper download link
To help you navigate the transition from analytical solutions to computational power, we have synthesized the most demanding exam themes into this structured revision guide.
What is the ‘Hierarchical Bayesian Model’?
In many real-world scenarios, data is nested. For example, you might be studying student performance across several schools. A Hierarchical Model assumes that the parameters for each school are themselves drawn from a “hyper-distribution.” This allows for “borrowing strength”—where the data from one school helps inform the estimates for another. In your revision, focus on how to define Hyperparameters and how they influence the final posterior distribution.
How do we handle ‘Non-Conjugate’ Priors?
In the first unit, we loved conjugate priors because they made the math easy. In level II, we often use Non-Informative or Subjective Priors that don’t play nicely with the likelihood function. When the posterior doesn’t have a recognizable “name” (like Normal or Gamma), we can no longer solve it with a pen and paper. This is where Markov Chain Monte Carlo (MCMC) methods, such as the Gibbs Sampler or Metropolis-Hastings Algorithm, become essential.
What is the ‘Metropolis-Hastings Algorithm’?
This is a cornerstone of computational Bayesian statistics. It is a method for generating a sequence of random samples from a probability distribution that is otherwise difficult to sample from. By using a “Proposal Distribution” and an “Acceptance Ratio,” the algorithm eventually settles into the “target” posterior distribution. Examiners often ask you to explain why a proposal might be rejected or how to identify when the chain has “converged.”
How do we evaluate ‘Bayesian Model Choice’?
In Decision Theory II, you are often asked to compare two competing models. Since we don’t use standard p-values, we rely on:
-
Bayes Factors: The ratio of the likelihood of the data under Model 1 versus Model 2.
-
Deviance Information Criterion (DIC): This is the Bayesian version of the AIC. It balances how well the model fits the data against the number of effective parameters (complexity).
-
Posterior Predictive Checks: You simulate new data from your model and see if it looks like the original “real” data.
What is ‘Utility Theory’ in a Decision Context?
Decision theory isn’t just about minimizing loss; it’s about maximizing Utility. A “Loss Function” ($L$) tells you the cost of an error, but a Utility Function ($U$) represents a decision-maker’s preference or “happiness.” In an exam, you might be asked to find the Bayes Action—the specific decision that minimizes the Expected Posterior Loss.
Why is ‘Sensitivity Analysis’ mandatory?
Critics often argue that Bayesian results are “biased” by the choice of the prior. Sensitivity Analysis is your defense. It involves changing your prior slightly to see if the final decision stays the same. If your conclusion changes wildly when you move from a “Uniform” to a “Normal” prior, your results are not robust. This is a common essay topic in advanced theory papers.
What are ‘Credible Intervals’ versus ‘Confidence Intervals’?
This is a classic distinction. A Frequentist 95% Confidence Interval means if we repeated the experiment 100 times, 95 of the intervals would contain the true fixed parameter. A Bayesian 95% Credible Interval means there is a 95% probability that the parameter lies within that specific range. It is the intuitive answer most people actually want.
Conclusion
Decision Theory and Bayesian Inference II is where statistics meets modern computing. It requires a move away from “closed-form” solutions and toward algorithmic thinking. Success in your finals comes from your ability to set up the hierarchy of a model and understand how MCMC sampling “explores” the posterior landscape.
To help you master these advanced algorithms and secure your grade, we have provided a link to the essential PDF revision resource below.
Last updated on: March 24, 2026