Exam MAS-II: Bayesian Analysis Flashcards
Bayesian inference for CAS Exam MAS-II: Bayes' theorem and the prior-likelihood-posterior chain, the conjugate families (beta-binomial, gamma-Poisson, normal-normal, gamma-exponential) and reading off posterior means, the prior and posterior predictive distributions, Bayes estimators under squared-error, absolute-error and 0-1 loss, credible intervals and how they differ from confidence intervals, Jeffreys/noninformative priors, and MCMC (Metropolis-Hastings, Gibbs) for intractable posteriors plus hierarchical models — with fully worked conjugate updates and predictive calculations.
Every deck is built into the Willys app
All of these decks, including the full practice problem banks, come built into Willys AI Flashcards & Quizzes for iPhone, iPad & Mac, with FSRS + SM-2 spaced repetition, streaks, and exam-date cram mode. 14-day free trial, then $14.99. To load a deck in the app: Settings → Library → Browse, then pick your exam and deck.
More Exam MAS-II decks:
Browse all 44 cards as a list
- Bayes theorem & posteriorsState **Bayes' theorem** for a parameter $\theta$ given data $x$, naming each piece.$\pi(\theta\mid x)=\frac{L(x\mid\theta)\,\pi(\theta)}{\int L(x\mid\theta)\,\pi(\theta)\,d\theta}\propto L(x\mid\theta)\,\pi(\theta)$. $\pi(\theta)$ = **prior** (belief before data); $L(x\mid\theta)$ = **likelihood** (model for the data given $\theta$); the denominator is the **marginal / normalizing constant** $m(x)=\int L(x\mid\theta)\pi(\theta)\,d\theta$; and $\pi(\theta\mid x)$ is the **posterior**. The denominator does not depend on $\theta$, so up to proportionality **posterior $\propto$ likelihood $\times$ prior**.
- Bayes theorem & posteriorsWhy can you usually work with the posterior **only up to proportionality** in $\theta$?The marginal $m(x)=\int L(x\mid\theta)\pi(\theta)\,d\theta$ is a constant in $\theta$ — it only rescales the curve so it integrates to $1$. So you can drop any factor not involving $\theta$, recognize the **kernel** of a known density (e.g. a $\text{Beta}$ or $\text{Gamma}$ shape), and supply the normalizing constant from that family. This is exactly how conjugate updates are done by inspection without computing the integral.
- Bayes theorem & posteriorsA coin has unknown $\theta=P(\text{heads})$. Prior: $\theta=0.4$ with prob $0.5$ and $\theta=0.7$ with prob $0.5$ (a discrete prior). You flip once and get **heads**. Find the posterior.Likelihood of heads is just $\theta$. Compute joint $=\pi(\theta)L$: $\theta=0.4$: $0.5\times 0.4 = 0.20$. $\theta=0.7$: $0.5\times 0.7 = 0.35$. Marginal $m=0.20+0.35=0.55$. Posterior: $\pi(0.4\mid H)=\frac{0.20}{0.55}\approx 0.3636$, $\pi(0.7\mid H)=\frac{0.35}{0.55}\approx 0.6364$. The heads outcome shifts weight toward the larger $\theta$.
- Bayes theorem & posteriorsTwo urns: $A$ has $\frac{1}{3}$ defective, $B$ has $\frac{1}{6}$ defective; an urn is picked at random (prior $0.5$ each). A drawn item is **defective**. Posterior probability it came from urn $A$?$P(A\mid D)=\frac{P(D\mid A)P(A)}{P(D\mid A)P(A)+P(D\mid B)P(B)}$. Numerator $=\frac{1}{3}\cdot\frac{1}{2}=\frac{1}{6}$. Other term $=\frac{1}{6}\cdot\frac{1}{2}=\frac{1}{12}$. $P(A\mid D)=\frac{1/6}{1/6+1/12}=\frac{1/6}{1/4}=\frac{2}{3}\approx 0.667$. Observing a defect favors the higher-defect urn $A$.
- Conjugate priorsWhat does it mean for a prior to be **conjugate** to a likelihood, and why is it useful?A prior family is **conjugate** to a likelihood if the resulting posterior is in the **same family** as the prior — only the parameters change. The update is then closed-form: read off the new parameters and the posterior mean from the family's formulas, with no integration. The conjugate families on the syllabus are beta-binomial, gamma-Poisson, normal-normal (known variance), and gamma-exponential (gamma is conjugate for the rate).
- Conjugate priorsBeta-binomial: with prior $\theta\sim\text{Beta}(\alpha,\beta)$ and $x$ successes in $n$ Bernoulli trials, give the posterior and its mean.Posterior: $\theta\mid x \sim \text{Beta}(\alpha+x,\ \beta+n-x)$. Posterior mean $=\frac{\alpha+x}{\alpha+\beta+n}$. The data add $x$ to the first ('success') shape and $n-x$ to the second ('failure') shape. The prior acts like $\alpha$ pseudo-successes and $\beta$ pseudo-failures already observed.
- Conjugate priorsPrior $\theta\sim\text{Beta}(2,3)$ on a claim probability; in $n=10$ policies you observe $x=4$ claims. Find the posterior and its mean.Posterior $=\text{Beta}(\alpha+x,\ \beta+n-x)=\text{Beta}(2+4,\ 3+10-4)=\text{Beta}(6,9)$. Posterior mean $=\frac{6}{6+9}=\frac{6}{15}=0.40$. For comparison the prior mean was $\frac{2}{5}=0.40$ and the sample proportion $\frac{4}{10}=0.40$, so both agree and the posterior mean is $0.40$.
- Conjugate priorsGamma-Poisson: with prior $\lambda\sim\text{Gamma}(\alpha,\beta)$ (rate parameterization, mean $\alpha/\beta$) and $n$ Poisson counts $x_1,\dots,x_n$, give the posterior and its mean.Posterior: $\lambda\mid \mathbf{x}\sim\text{Gamma}\!\left(\alpha+\textstyle\sum_{i=1}^{n} x_i,\ \beta+n\right)$. Posterior mean $=\frac{\alpha+\sum x_i}{\beta+n}$. With the rate parameterization the data add the total count $\sum x_i$ to the shape and the exposure $n$ to the rate parameter $\beta$.
- Conjugate priorsPrior $\lambda\sim\text{Gamma}(\alpha=3,\beta=2)$ (rate form) for a Poisson claim frequency. You observe $n=4$ years with counts $1,0,2,3$. Find the posterior mean.$\sum x_i = 1+0+2+3 = 6$, $n=4$. Posterior $=\text{Gamma}(\alpha+\sum x_i,\ \beta+n)=\text{Gamma}(3+6,\ 2+4)=\text{Gamma}(9,6)$. Posterior mean $=\frac{9}{6}=1.5$. Prior mean was $\frac{3}{2}=1.5$ and sample mean $\frac{6}{4}=1.5$, so the Bayes estimate is $1.5$.
- Conjugate priorsNormal-normal (known variance $\sigma^2$): prior $\mu\sim N(\mu_0,\tau^2)$, data mean $\bar{x}$ from $n$ obs. Give the posterior.Posterior is normal: $\mu\mid\mathbf{x}\sim N(\mu_n,\ \sigma_n^2)$ with **precision** adding: $\frac{1}{\sigma_n^2}=\frac{1}{\tau^2}+\frac{n}{\sigma^2}$, and $\mu_n=\sigma_n^2\!\left(\frac{\mu_0}{\tau^2}+\frac{n\bar{x}}{\sigma^2}\right)$. The posterior mean is a precision-weighted average of the prior mean $\mu_0$ and the sample mean $\bar{x}$.
- Conjugate priorsKnown $\sigma^2=100$. Prior $\mu\sim N(\mu_0=50,\ \tau^2=25)$. You observe $n=4$ values with mean $\bar{x}=60$. Find the posterior mean and variance.Precisions: prior $\frac{1}{\tau^2}=\frac{1}{25}=0.04$; data $\frac{n}{\sigma^2}=\frac{4}{100}=0.04$. Posterior precision $=0.04+0.04=0.08$, so $\sigma_n^2=\frac{1}{0.08}=12.5$. $\mu_n=12.5\left(\frac{50}{25}+\frac{4(60)}{100}\right)=12.5(2.0+2.4)=12.5(4.4)=55$. Posterior: $N(55,\ 12.5)$ — halfway between $50$ and $60$ since the two precisions are equal.
- Conjugate priorsGamma-exponential: prior $\lambda\sim\text{Gamma}(\alpha,\beta)$ (rate form) on the rate of an exponential, data $x_1,\dots,x_n$. Give the posterior.For $X_i\sim\text{Exp}(\text{rate }\lambda)$, the likelihood is $\lambda^{n}e^{-\lambda\sum x_i}$. Posterior $\propto \lambda^{\alpha+n-1}e^{-(\beta+\sum x_i)\lambda}$, i.e. $\lambda\mid\mathbf{x}\sim\text{Gamma}\!\left(\alpha+n,\ \beta+\textstyle\sum x_i\right)$. Posterior mean of the rate $=\frac{\alpha+n}{\beta+\sum x_i}$. The data add $n$ to the shape and the total $\sum x_i$ to the rate parameter.
- Conjugate priorsPrior $\lambda\sim\text{Gamma}(\alpha=2,\beta=300)$ (rate form) for an exponential claim-severity rate. You observe $n=3$ claims of sizes $100,200,300$. Find the posterior mean of $\lambda$.$\sum x_i = 100+200+300 = 600$, $n=3$. Posterior $=\text{Gamma}(\alpha+n,\ \beta+\sum x_i)=\text{Gamma}(2+3,\ 300+600)=\text{Gamma}(5,900)$. Posterior mean of the rate $=\frac{5}{900}\approx 0.005556$. The plug-in reciprocal $1/E[\lambda\mid\text{data}]=\frac{900}{5}=180$ is **not** the posterior mean claim size, which is $E[1/\lambda\mid\text{data}]=\frac{\beta^*}{\alpha^*-1}=\frac{900}{4}=225$.
- Predictive distributionWhat is the **prior predictive (marginal) distribution**, and how does it differ from the **posterior predictive**?**Prior predictive** of a new observation $\tilde{x}$, before seeing data: $m(\tilde{x})=\int f(\tilde{x}\mid\theta)\,\pi(\theta)\,d\theta$ — average the model over the prior. **Posterior predictive**, after observing $x$: $f(\tilde{x}\mid x)=\int f(\tilde{x}\mid\theta)\,\pi(\theta\mid x)\,d\theta$ — average the model over the **posterior**. Both integrate out $\theta$; the posterior predictive uses updated beliefs and is what you use to predict the next claim.
- Predictive distributionWhy is the predictive distribution **wider** (more variable) than just plugging the posterior mean of $\theta$ into the model?The predictive distribution accounts for **two** sources of uncertainty: the natural variability of the observation given $\theta$ (process/model variance) **and** the remaining uncertainty about $\theta$ itself (parameter variance). By the law of total variance, $\text{Var}(\tilde{x}\mid x)=E[\text{Var}(\tilde{x}\mid\theta)\mid x]+\text{Var}(E[\tilde{x}\mid\theta]\mid x)$. Plugging in a single $\hat\theta$ drops the second term and understates uncertainty.
- Predictive distributionDiscrete prior: claim count $X\sim\text{Poisson}(\lambda)$ with $\lambda=1$ (prob $0.6$) or $\lambda=3$ (prob $0.4$). Find the **prior predictive** probability of exactly $0$ claims.$P(X=0\mid\lambda)=e^{-\lambda}$. Mix over the prior: $P(X=0)=0.6\,e^{-1}+0.4\,e^{-3}$. $e^{-1}\approx 0.367879$, $e^{-3}\approx 0.049787$. $P(X=0)=0.6(0.367879)+0.4(0.049787)\approx 0.220728 + 0.019915 = 0.2406$.
- Predictive distributionContinuing: with the prior $\lambda=1$ (prob $0.6$) or $\lambda=3$ (prob $0.4$), you observe **one year with $0$ claims**. Find the posterior on $\lambda$, then the **posterior predictive** probability of $0$ claims next year.Posterior $\propto$ prior $\times e^{-\lambda}$: weights $0.6(0.367879)=0.220728$ and $0.4(0.049787)=0.019915$; sum $=0.240643$. Posterior: $P(\lambda=1\mid 0)=\frac{0.220728}{0.240643}\approx 0.9173$, $P(\lambda=3\mid 0)\approx 0.0827$. Posterior predictive of $0$: $0.9173\,e^{-1}+0.0827\,e^{-3}\approx 0.9173(0.367879)+0.0827(0.049787)\approx 0.3375+0.0041=0.3416$.
- Predictive distributionGamma-Poisson posterior predictive: with posterior $\lambda\mid x\sim\text{Gamma}(\alpha^*,\beta^*)$ (rate form), what is the distribution of the next count $\tilde{x}$?Mixing a Poisson over a gamma rate gives a **negative binomial**. The next count $\tilde{x}\mid x$ has $E[\tilde{x}\mid x]=\frac{\alpha^*}{\beta^*}$ (the posterior mean of $\lambda$) and variance $\frac{\alpha^*}{\beta^*}\!\left(1+\frac{1}{\beta^*}\right)=\frac{\alpha^*}{\beta^*}\cdot\frac{\beta^*+1}{\beta^*}$. The variance exceeds the Poisson value $\frac{\alpha^*}{\beta^*}$ because parameter uncertainty in $\lambda$ adds overdispersion.
- Predictive distributionWith posterior $\lambda\mid x\sim\text{Gamma}(\alpha^*=9,\beta^*=6)$ (rate form), find the mean and variance of the **posterior predictive** count $\tilde{x}$.Mean $=\frac{\alpha^*}{\beta^*}=\frac{9}{6}=1.5$. Variance $=\frac{\alpha^*}{\beta^*}\cdot\frac{\beta^*+1}{\beta^*}=1.5\cdot\frac{7}{6}=1.75$. The predictive variance $1.75$ exceeds the mean $1.5$ (overdispersion) because of leftover uncertainty in $\lambda$, unlike a plain Poisson where variance $=$ mean.
- Predictive distributionBeta-binomial posterior predictive: posterior $\theta\mid x\sim\text{Beta}(\alpha^*,\beta^*)$. Probability that the **next single trial** is a success?$P(\tilde{x}=1\mid x)=\int_0^1 \theta\,\pi(\theta\mid x)\,d\theta = E[\theta\mid x]=\frac{\alpha^*}{\alpha^*+\beta^*}$. So the predictive success probability is simply the **posterior mean** of $\theta$. (For $m$ future trials the count follows a beta-binomial distribution.)
- Predictive distributionWith posterior $\theta\mid x\sim\text{Beta}(6,9)$, find the probability the next two independent trials are **both successes**.For two future trials, integrate $\theta^2$ against the posterior: $E[\theta^2\mid x]=\frac{\alpha^*(\alpha^*+1)}{(\alpha^*+\beta^*)(\alpha^*+\beta^*+1)}$. $=\frac{6\cdot 7}{15\cdot 16}=\frac{42}{240}=0.175$. This exceeds $(\text{posterior mean})^2=0.40^2=0.16$ because the two outcomes are positively correlated through the shared unknown $\theta$.
- Loss functions & estimatorsState the **Bayes estimator** of $\theta$ under each of squared-error, absolute-error, and $0$-$1$ loss.**Squared-error loss** $(\hat\theta-\theta)^2$ → posterior **mean** $E[\theta\mid x]$. **Absolute-error loss** $|\hat\theta-\theta|$ → posterior **median**. **$0$-$1$ loss** (penalize any miss equally) → posterior **mode** (the MAP estimate). Each is the value minimizing the expected posterior loss for its loss function.
- Loss functions & estimatorsProve the posterior **mean** minimizes expected squared-error loss.Minimize $g(\hat\theta)=E[(\theta-\hat\theta)^2\mid x]$ over $\hat\theta$. $g'(\hat\theta)=E[-2(\theta-\hat\theta)\mid x]=-2\big(E[\theta\mid x]-\hat\theta\big)=0$. Solving gives $\hat\theta=E[\theta\mid x]$, the posterior mean; $g''=2>0$ confirms a minimum. The minimized value is the posterior variance $\text{Var}(\theta\mid x)$.
- Loss functions & estimatorsPosterior $\theta\mid x\sim\text{Beta}(8,4)$. Give the Bayes estimate under (a) squared-error and (b) $0$-$1$ loss.(a) Squared-error → posterior **mean** $=\frac{\alpha^*}{\alpha^*+\beta^*}=\frac{8}{12}\approx 0.667$. (b) $0$-$1$ loss → posterior **mode** $=\frac{\alpha^*-1}{\alpha^*+\beta^*-2}=\frac{7}{10}=0.70$. The mode exceeds the mean because the $\text{Beta}(8,4)$ density is left-skewed.
- Loss functions & estimatorsDiscrete posterior on $\theta$: $P(\theta=1)=0.2$, $P(\theta=2)=0.5$, $P(\theta=4)=0.3$. Give the Bayes estimate under squared-error, absolute-error, and $0$-$1$ loss.**Squared-error** → mean $=1(0.2)+2(0.5)+4(0.3)=0.2+1.0+1.2=2.4$. **Absolute-error** → median: cumulative weight reaches $0.5$ at $\theta=2$ (the $0.2+0.5=0.7$ mass at or below $2$ first passes $0.5$), so median $=2$. **$0$-$1$ loss** → mode $=2$ (highest mass, $0.5$).
- Loss functions & estimatorsHow is the **MAP (maximum a posteriori)** estimate related to the MLE, and when do they coincide?The MAP maximizes the posterior $\pi(\theta\mid x)\propto L(x\mid\theta)\pi(\theta)$ — it is the posterior **mode**, the Bayes estimator under $0$-$1$ loss. The MLE maximizes only $L(x\mid\theta)$. They coincide when the prior is **flat (constant)** over the relevant range, since then the posterior is proportional to the likelihood. With an informative prior the MAP is pulled toward the prior's high-density region.
- Loss functions & estimatorsUnder a **weighted/asymmetric** squared-error loss $w(\theta)(\hat\theta-\theta)^2$, what is the Bayes estimator?Minimizing $E[w(\theta)(\theta-\hat\theta)^2\mid x]$ gives the **weighted posterior mean** $\hat\theta=\frac{E[w(\theta)\,\theta\mid x]}{E[w(\theta)\mid x]}$. When $w\equiv 1$ this reduces to the ordinary posterior mean. Weighting lets you penalize errors more in regions of $\theta$ that matter more.
- Loss functions & estimatorsWith posterior $\lambda\mid x\sim\text{Gamma}(\alpha^*=5,\beta^*=900)$ (rate form), give the Bayes estimate of $\lambda$ under squared-error loss and its posterior variance.Squared-error Bayes estimate $=$ posterior mean $=\frac{\alpha^*}{\beta^*}=\frac{5}{900}\approx 0.005556$. Posterior variance of a $\text{Gamma}(\alpha^*,\beta^*)$ rate $=\frac{\alpha^*}{(\beta^*)^2}=\frac{5}{900^2}=\frac{5}{810000}\approx 6.17\times 10^{-6}$. (Standard deviation $\approx 0.002484$.)
- Credible intervalsDefine a **$95\%$ credible interval** for $\theta$ and contrast it with a frequentist confidence interval.A $95\%$ **credible interval** $(a,b)$ satisfies $P(a<\theta<b\mid x)=0.95$ under the posterior — a direct probability statement about $\theta$ given the data. A frequentist **confidence interval** treats $\theta$ as fixed: $95\%$ refers to the long-run coverage of the random interval, not to a probability that this particular interval contains $\theta$. The Bayesian statement 'there is a $95\%$ probability $\theta$ lies in $(a,b)$' is valid; the same wording is incorrect for a confidence interval.
- Credible intervalsDistinguish an **equal-tailed** credible interval from a **highest posterior density (HPD)** interval.**Equal-tailed:** cut off $2.5\%$ posterior probability in each tail — bounds are the $0.025$ and $0.975$ posterior quantiles. Easy to compute from quantiles. **HPD:** the **shortest** interval containing $95\%$ posterior probability; every point inside has higher posterior density than any point outside. For a symmetric unimodal posterior the two coincide; for a skewed posterior the HPD is shorter and not equal-tailed.
- Credible intervalsPosterior $\mu\mid x\sim N(55,\ 12.5)$. Construct a $95\%$ equal-tailed credible interval for $\mu$.Posterior SD $=\sqrt{12.5}\approx 3.5355$. Use $z_{0.975}=1.96$. Interval $=55\pm 1.96(3.5355)=55\pm 6.93$. So the $95\%$ credible interval is approximately $(48.07,\ 61.93)$. Because the posterior is symmetric, this is also the HPD interval.
- Credible intervalsPosterior $\mu\mid x\sim N(120,\ 16)$. Find a **$90\%$** equal-tailed credible interval and state the probability $\mu>125$.Posterior SD $=\sqrt{16}=4$; $z_{0.95}=1.645$. $90\%$ interval $=120\pm 1.645(4)=120\pm 6.58=(113.42,\ 126.58)$. $P(\mu>125\mid x)=P\!\left(Z>\frac{125-120}{4}\right)=P(Z>1.25)\approx 1-0.8944=0.1056$.
- Credible intervalsFrom a discrete posterior $P(\theta=1)=0.10$, $P(\theta=2)=0.55$, $P(\theta=3)=0.25$, $P(\theta=4)=0.10$, give a credible set of probability **at least $0.90$**.Rank by posterior mass and accumulate until $\ge 0.90$: $\theta=2$ ($0.55$) → $\theta=3$ ($+0.25=0.80$) → $\theta=1$ ($+0.10=0.90$). The set $\{1,2,3\}$ carries $0.90$ probability, so it is a $90\%$ credible set (highest-density style). Equivalently, excluding $\theta=4$ (mass $0.10$) leaves $0.90$.
- Bayes theorem & posteriorsWhat is a **noninformative (vague/diffuse) prior**, and what risk does it carry?A noninformative prior aims to add little or no information, letting the data dominate — e.g. a flat prior $\pi(\theta)\propto 1$ or a very wide proper prior. Risk: a flat prior can be **improper** (does not integrate to a finite value); that is acceptable only if the resulting posterior is still proper. A prior that is flat in one parameterization is generally **not** flat after a nonlinear transformation, so 'noninformative' is not invariant — motivating Jeffreys priors.
- Bayes theorem & posteriorsDefine the **Jeffreys prior** and give its key property.Jeffreys prior is $\pi(\theta)\propto\sqrt{I(\theta)}$, where $I(\theta)=-E\!\left[\frac{\partial^2}{\partial\theta^2}\ln f(X\mid\theta)\right]$ is the Fisher information. Key property: it is **invariant under reparameterization** — applying Jeffreys' rule in any one-to-one transformed parameter gives the consistent transformed prior. For a binomial proportion it is $\pi(\theta)\propto\theta^{-1/2}(1-\theta)^{-1/2}$, i.e. $\text{Beta}(\tfrac12,\tfrac12)$.
- MCMC & hierarchicalWhy do we often need **MCMC**, and what does it produce?For most realistic models the posterior $\pi(\theta\mid x)\propto L\pi$ has an **intractable normalizing constant** (the integral $\int L\pi\,d\theta$ can't be done in closed form), and $\theta$ may be high-dimensional. MCMC builds a Markov chain whose stationary distribution **is** the posterior, generating dependent draws $\theta^{(1)},\theta^{(2)},\dots$. After discarding burn-in, sample averages estimate posterior means, quantiles give credible intervals, etc. — all without ever computing the normalizing constant.
- MCMC & hierarchicalState the **Metropolis-Hastings acceptance ratio** and the accept/reject rule.From current $\theta$, propose $\theta^*$ from $q(\theta^*\mid\theta)$. The acceptance probability is $\alpha=\min\!\left\{1,\ \dfrac{\pi(\theta^*\mid x)\,q(\theta\mid\theta^*)}{\pi(\theta\mid x)\,q(\theta^*\mid\theta)}\right\}$. Draw $u\sim U(0,1)$: **accept** $\theta^*$ if $u\le\alpha$, else keep $\theta$. The ratio uses the posterior only through $L\pi$, so the unknown normalizing constant cancels. For a **symmetric** proposal ($q(\theta^*\mid\theta)=q(\theta\mid\theta^*)$) it simplifies to $\alpha=\min\{1,\ \pi(\theta^*\mid x)/\pi(\theta\mid x)\}$ (Metropolis).
- MCMC & hierarchicalMetropolis step: target $\pi(\theta\mid x)\propto \theta^{8}(1-\theta)^{4}$ (a $\text{Beta}(9,5)$ kernel). Current $\theta=0.5$, symmetric proposal $\theta^*=0.7$. Compute the acceptance probability.Symmetric proposal → $\alpha=\min\{1,\ r\}$ with $r=\frac{\theta^{*8}(1-\theta^*)^4}{\theta^{8}(1-\theta)^4}$. Numerator: $0.7^{8}(0.3)^4 = 0.05764801\times 0.0081 \approx 4.6695\times 10^{-4}$. Denominator: $0.5^{8}(0.5)^4 = 0.5^{12}\approx 2.4414\times 10^{-4}$. $r\approx \frac{4.6695\times 10^{-4}}{2.4414\times 10^{-4}}\approx 1.913$. Since $r>1$, $\alpha=\min\{1,1.913\}=1$ — the move to $0.7$ is **always accepted**.
- MCMC & hierarchicalMetropolis step toward a **lower-density** point. Same target $\pi\propto\theta^{8}(1-\theta)^4$, current $\theta=0.5$, symmetric proposal $\theta^*=0.2$. Compute $\alpha$ and decide whether to accept if $u=0.10$.Numerator: $0.2^{8}(0.8)^4 = 2.56\times 10^{-6}\times 0.4096 \approx 1.0486\times 10^{-6}$. Denominator: $0.5^{12}\approx 2.4414\times 10^{-4}$. $r\approx \frac{1.0486\times 10^{-6}}{2.4414\times 10^{-4}}\approx 0.004295$, so $\alpha=\min\{1,0.004295\}\approx 0.0043$. With $u=0.10>0.0043$, **reject** — the chain stays at $\theta=0.5$. Downhill moves are accepted only rarely.
- MCMC & hierarchicalDescribe **Gibbs sampling** and when it is preferred over Metropolis-Hastings.Gibbs sampling updates one parameter (or block) at a time by drawing it from its **full conditional** distribution given the current values of all the others: cycle $\theta_1^{(t+1)}\sim\pi(\theta_1\mid\theta_2^{(t)},\dots,x)$, then $\theta_2^{(t+1)}\sim\pi(\theta_2\mid\theta_1^{(t+1)},\dots,x)$, and so on. Every draw is **always accepted** (it is a special case of M-H with acceptance $1$). It is preferred when the full conditionals are known standard distributions — common in conjugate hierarchical models.
- MCMC & hierarchicalDefine **burn-in**, **thinning**, and how convergence is assessed in MCMC.**Burn-in:** discard the first portion of the chain, which depends on the arbitrary starting value, before computing summaries. **Thinning:** keep every $k$-th draw to reduce autocorrelation/storage (optional; reduces effective info per kept draw). **Convergence diagnostics:** trace plots (should look like stationary noise), running-mean stability, autocorrelation plots, and the **Gelman-Rubin** statistic $\hat{R}$ comparing within- vs between-chain variance across multiple chains ($\hat{R}\approx 1$ suggests convergence). MCMC estimates are valid only after the chain has reached its stationary (posterior) distribution.
- MCMC & hierarchicalEstimate a posterior mean and a $95\%$ credible interval from a thinned, post-burn-in MCMC sample of $\theta$: $0.30,\ 0.42,\ 0.55,\ 0.61,\ 0.48,\ 0.37,\ 0.52,\ 0.44$ (treat as the retained draws).Posterior mean $\approx$ sample average $=\frac{0.30+0.42+0.55+0.61+0.48+0.37+0.52+0.44}{8}=\frac{3.69}{8}\approx 0.461$. For a credible interval, use the empirical quantiles of the draws: sorted $0.30,0.37,0.42,0.44,0.48,0.52,0.55,0.61$. The crude $95\%$ interval spans roughly the extreme retained draws $(0.30,\ 0.61)$ — a real run uses thousands of draws and the $2.5\%/97.5\%$ empirical quantiles for a tight interval.
- MCMC & hierarchicalWhat is a **hierarchical (multilevel) Bayesian model**, and why is it natural for actuarial credibility?A hierarchical model adds a layer: data depend on group-level parameters $\theta_i$, and those $\theta_i$ share a common **hyperprior** with hyperparameters $\phi$. Schematically $x_{ij}\mid\theta_i\sim f$, $\theta_i\mid\phi\sim g$, $\phi\sim\pi(\phi)$. This 'partial pooling' shrinks each group's estimate toward the overall mean by an amount that depends on how much data the group has — exactly the **credibility** idea. Empirical Bayes estimates $\phi$ from the data; full Bayes puts a prior on $\phi$ and integrates (typically via MCMC).
- MCMC & hierarchicalIn a hierarchical normal model, the group posterior mean is $\hat\theta_i = Z_i\,\bar{x}_i + (1-Z_i)\,\mu$ with $Z_i=\frac{n_i}{n_i+k}$, $k=\sigma^2/\tau^2$. Compute $\hat\theta_i$ for $\bar{x}_i=80$, $\mu=70$, $n_i=5$, $\sigma^2=100$, $\tau^2=25$.$k=\frac{\sigma^2}{\tau^2}=\frac{100}{25}=4$, so $Z_i=\frac{n_i}{n_i+k}=\frac{5}{5+4}=\frac{5}{9}\approx 0.5556$. $\hat\theta_i = Z_i\,\bar{x}_i+(1-Z_i)\,\mu = 0.5556(80)+0.4444(70)$ $\approx 44.44 + 31.11 = 75.56$. The group estimate is shrunk from its raw mean $80$ partway toward the global mean $70$ — more shrinkage when $n_i$ is small.