LAISHA
Open Laisha

Indian Statistical Institute admission tests Bayesian Inference: practice questions

Indian Statistical Institute admission tests (BStat, MStat, JRF in Statistics) · India · Indian Statistical Institute

Original practice questions written to the published syllabus · not past papers
10worked here, free
58written to this exam in this subject
11syllabus units

About Indian Statistical Institute admission tests

Indian Statistical Institute admission tests (BStat, MStat, JRF in Statistics) is a postgraduate admission exam run by Indian Statistical Institute, for candidates in India. Written tests in two parts: an objective mathematics/statistics test and a descriptive subject test, followed by interview for some programmes (reported). JRF selection by written test and interview.

Official site: isical.ac.in.

We have not been able to confirm these details against the exam body's own specification, so fees, dates and pass marks are left out on purpose. Check the current format, eligibility, fees and dates on the official site.

Bayesian Inference in Indian Statistical Institute admission tests

Inference through prior and posterior: conjugate analysis, prior choice, credible intervals, Bayes factors and decision theory, hierarchical models, and computation by Markov-chain Monte Carlo, with agricultural uses in variety trials, genomic prediction and small-area estimation.

Our 58 Indian Statistical Institute admission tests questions in this subject: 18 easy, 31 medium, 9 hard, across 11 units.

10 practice questions, worked

Choose an answer, then open the working. Each explanation says why the right answer is right and why the others are not.

These 10 were written to Indian Statistical Institute admission tests and are not shown on any other exam's page.

Question 1 · easy

How is the posterior density related to the prior π(θ) and the likelihood f(x | θ)?

  1. π(θ | x) ∝ f(x | θ)/π(θ)
  2. π(θ | x) ∝ f(x | θ)π(θ)
  3. π(θ | x) ∝ π(θ)/f(x | θ)
  4. π(θ | x) = f(x | θ) + π(θ)
Answer and working

B. π(θ | x) ∝ f(x | θ)π(θ)

Bayes' theorem: π(θ | x) = f(x | θ)π(θ)/m(x), and m(x) does not involve θ. Dividing by the prior or adding densities has no probabilistic meaning.

Question 2 · medium

A decision rule δ₁ has risk no larger than δ₂ for every θ and strictly smaller for some θ. What is δ₂ called?

  1. Minimax
  2. Bayes
  3. Inadmissible
  4. Unbiased
Answer and working

C. Inadmissible

δ₂ is dominated, hence inadmissible. Stein showed that the sample mean of three or more normal means is inadmissible under total squared-error loss.

Question 3 · medium

Before a further season of variety trials, an analyst computes the expected value of sample information (EVSI). How is it used?

  1. Reject the variety if EVSI is below zero
  2. Use EVSI as the prior mean for yield
  3. Release the variety if EVSI is positive
  4. Run them only if EVSI exceeds their cost
Answer and working

D. Run them only if EVSI exceeds their cost

EVSI is the expected improvement in the decision's value from the extra data; it is never negative. Comparing it with the trial cost decides whether the information is worth buying.

Question 4 · easy

In the normal hierarchical model, what is the shrinkage factor Bᵢ for a group with sampling variance σᵢ² and between-group variance τ²?

  1. σᵢ²/(σᵢ² + τ²)
  2. τ²/(σᵢ² + τ²)
  3. σᵢ²/τ²
  4. 1/(1 + σᵢ²)
Answer and working

A. σᵢ²/(σᵢ² + τ²)

Bᵢ = σᵢ²/(σᵢ² + τ²) is the weight on the overall mean: large for noisy groups, small for precise ones. τ²/(σᵢ² + τ²) is the weight on the group's own estimate.

Question 5 · medium

What does Hamiltonian Monte Carlo use that random-walk Metropolis does not?

  1. The normalising constant m(x)
  2. Independent draws from the prior
  3. Gradients of the log posterior
  4. Conjugate full conditionals
Answer and working

C. Gradients of the log posterior

HMC simulates Hamiltonian dynamics using ∇log π(θ | x) to make long, directed moves, coping well with correlated posteriors. No MCMC method needs m(x).

Question 6 · hard

Why does MCMC make Bayesian computation feasible when m(x) = ∫ f(x | θ)π(θ)dθ is intractable?

  1. MCMC needs only the likelihood, not the prior
  2. MCMC computes m(x) exactly as a by-product of sampling
  3. MCMC replaces the posterior by the prior when m(x) is unknown
  4. The ratios need the posterior only up to a constant
Answer and working

D. The ratios need the posterior only up to a constant

In π(θ* | x)/π(θ | x) the constant m(x) cancels, and full conditionals are proportional to the joint posterior. Estimating m(x) from MCMC output is a separate, harder problem.

Question 7 · easy

What does the Gelman-Rubin R̂ compare?

  1. Between-chain and within-chain variance
  2. Autocorrelation at lags 1 and 2
  3. Prior and posterior variance
  4. Early and late halves of one chain's mean
Answer and working

A. Between-chain and within-chain variance

R̂ near 1 means the chains, started from dispersed values, agree. Comparing early and late segments of one chain is Geweke's diagnostic.

Question 8 · medium

An HMC sampler reports many divergent transitions. What should be done?

  1. Increase the burn-in and keep the model unchanged
  2. Investigate and reparameterise, since parts of the posterior are not explored reliably
  3. Thin the chain to remove the divergent draws
  4. Ignore them if R̂ is below 1.1
Answer and working

B. Investigate and reparameterise, since parts of the posterior are not explored reliably

Divergences indicate regions of high curvature where the sampler fails, which biases estimates; reparameterisation or a smaller step size is needed. R̂ may still look fine.

Question 9 · hard

In what situation can DIC be unreliable, even giving a negative effective number of parameters p_D?

  1. When the posterior is strongly non-normal or multimodal, as in mixtures
  2. When the posterior mean equals the posterior mode
  3. When the prior is conjugate
  4. When the model has only one parameter
Answer and working

A. When the posterior is strongly non-normal or multimodal, as in mixtures

p_D = D̄ − D(θ̄) depends on the deviance at the posterior mean, which can be a poor point for skewed or multimodal posteriors; WAIC and PSIS-LOO avoid this.

Question 10 · medium

How does parametric empirical Bayes estimate the hyperparameters of the prior?

  1. From the marginal distribution of the data
  2. By fixing them at values that make the prior flat
  3. From the posterior of each unit separately
  4. From expert opinion elicited before the study
Answer and working

A. From the marginal distribution of the data

EB maximises (or matches moments of) m(x | η) = ∫ f(x | θ)π(θ | η)dθ, so the data choose the prior. Elicitation is a fully Bayesian route.

A study plan for Bayesian Inference in Indian Statistical Institute admission tests

This plan follows where our questions actually fall, so the time goes where the marks are likely to be. It is our reading of the syllabus, not the exam body's weighting.

  1. First, the heavy units (47% of the 58 questions): Convergence diagnostics (7, 12%); Posterior predictive checks and model comparison (DIC, WAIC) (7, 12%); Posterior summaries and credible intervals (7, 12%); Bayes' theorem for parameters: prior, likelihood and posterior (6, 10%).

    Read the manual chapters for these first and do every worked question you can find on them.

  2. Then the middle band: Markov chain Monte Carlo: Gibbs sampling and Metropolis-Hastings (6, 10%); Non-informative and weakly informative priors (6, 10%); Bayesian hypothesis testing and Bayes factors (5, 9%); Empirical Bayes (4, 7%).

    Aim for one sitting per unit, then mixed questions across them.

  3. Then the lighter units: Hierarchical (multilevel) models (4, 7%); Conjugate families (3, 5%); Decision theory: loss functions and Bayes risk (3, 5%).

    Learn the definitions, classifications and standard numbers; these units are more often recalled than worked.

  4. Leave time for the hard ones: Markov chain Monte Carlo: Gibbs sampling and Metropolis-Hastings (2 hard of 6), Non-informative and weakly informative priors (2 hard of 6), Posterior predictive checks and model comparison (DIC, WAIC) (2 hard of 7).

    Come back to these after a first pass, untimed, and read every explanation, including for the options you rejected.

  5. Bank the easy marks: Non-informative and weakly informative priors (3 easy of 6); Posterior summaries and credible intervals (3 easy of 7).

    Do these early in revision so they are fast on the day.

  6. Finish timed. Sit the subject as a timed set without notes, mark it, and send each wrong answer back to its unit above.

Last updated · How Laisha writes and checks its pages · Report a mistake

Ask Laisha about this

Opens Laisha's chat with your question ready to send. Sign in or start free.

Unlock every question — Indian Statistical Institute admission tests Bayesian Inference

US$10 a month · one price worldwide · cancel any time

  • Every practice question in this subject, with full working
  • Timed exam papers and progress by unit
  • Ask Laisha to explain any answer — up to 1,000 questions a month
Free or Smarter
FreeSmarter
Grow guides, pest pages, articlesRead allRead all
Ask Laisha5 a dayUp to 1,000 a month
Exam practiceThe sample on each pageEvery question, timed papers
Crop calendar and pest alerts—For your own fields
Calculators and tools—All
Uploads and reports—100 a month, real files
10 days free for every new account, no card.