Exam map · official factspriority 0
Final examination (2 hours) · 40% · closed book, with ONE A4 sheet of HANDWRITTEN notes permitted (official course materials).
| Put on your page | Why it earns space |
|---|---|
| Target + estimand | Prevents mean/median and association/effect swaps |
| Assumptions | Decide whether a method is defensible |
| SE + reference law | Turns estimate into interval/test |
| Failure + repair | Shows diagnostic judgement |
| Reporting sentence | Stops probability reversals |
Separate pass conditions: at least 40% of possible continuous-assessment marks and at least 35% of possible final-exam marks; both apply.
1 · Estimation skeletonL1
Name the target: θ=T(F). Estimator θ̂ is a function of the sample. Bias=E(θ̂)−θ; Var=E[(θ̂−Eθ̂)²]; MSE=bias²+variance. Consistency means MSE→0; bias can lower MSE.
Mean under independenceE(X̄)=μ · Var(X̄)=σ²/n · SE=s/√n
dependence adds covariance terms ⇒ naïve SE too small
CI rule: coverage belongs to the procedure; after observation, endpoints and θ are fixed.
Model → target → reportfirst 30 seconds
| Scientific quantity | Model shorthand | Target |
|---|---|---|
| Disease prevalence | Bernoulli | p=E(X) |
| Biomarker centre | Normal / nonparametric | μ or median |
| Event count | Poisson | λ |
| Survival time | Gamma / nonparametric | mean or median |
Do not let convenience rename the target. A median question is not a mean question merely because X̄ has an easy SE. Name the sampling unit and independence assumption; repeated patients or clinics add covariance and widen uncertainty.
Model bargain: parametric precision costs a defensible family; nonparametric robustness usually costs width.
2 · Empirical distributionL2
StaircaseF̂n(x)=n⁻¹ΣI(Xi≤x)
E(F̂n)=F(x) · Var=F(x){1−F(x)}/n
Plug-in: estimate T(F) by T(F̂n). Linear functional ∫r(x)dF becomes n⁻¹Σr(Xi); medians, correlations and ratios stay computable when analytic SEs are awkward.
Simultaneous DKW bandP{supx|F̂n−F|>ε}≤2e−2nε²
εn=√{log(2/α)/(2n)}; clip F̂±ε to [0,1]
Pointwise ≠ simultaneous. Pay for a whole-curve claim only when you intend to search across x.
Bootstrap reciperesample
- Draw n observations with replacement.
- Recompute identical T*; repeat B times.
- SEboot=√{B⁻¹Σ(T*b−T̄*)²}.
| Interval | Rule / risk |
|---|---|
| Normal | T̂±1.96 SE; symmetric |
| Percentile | Use 2.5%,97.5% resample quantiles |
| Pivotal | (2T̂−q.975,2T̂−q.025) |
Check: bootstrapped mean SE should be near s/√n. Bootstrap estimates variability; it need not remove boundary bias.
Method choiceL4
Ask: target, sample size/tails, plausible family, finite-sample guarantee? Small n/heavy tails favour resampling or a bound; large n/rough symmetry supports a normal interval. Impossible probability endpoints need Wilson or Agresti–Coull.
Four failures: boundary Wald; symmetric interval for skew; bootstrap trusted to erase bias; parametric bootstrap from the wrong family.
Coverage simulation: repeatedly sample, build the interval and count captures; it checks the procedure’s promised level, not whether one realised interval is “correct”.
3 · Moments & likelihoodL3
Moments: match sample and population moments, solve. MLE: maximise L(θ)=Πf(Xi;θ), usually via ℓ(θ)=Σlog f. MLE is invariant: estimate g(θ) with g(θ̂).
Bernoulli / Poissonp̂=X̄ · SE(p̂)=√{p̂(1−p̂)/n}
λ̂=X̄ · I₁(λ)=1/λ · SE(λ̂)=√(λ̂/n)
Uniform(0,θ): MLE=X(n); E=nθ/(n+1), bias −θ/(n+1); corrected form (n+1)X(n)/n.
Log scale: products become sums, differentiation is easier, and log preserves the maximiser.
Information routescore → SE
Four reusable liness(X;θ)=∂log f/∂θ
I₁(θ)=Var{s}=−E(∂²log f/∂θ²)
In=nI₁ · SE(θ̂)≈1/√In(θ̂)
CI: θ̂±zα/2SE
Delta: SE{g(θ̂)}≈|g′(θ̂)|SE(θ̂). If g is sharply curved or g′≈0, simulate from the fitted model; efficiency then depends on correct specification.
Logit: SE(ψ̂)=√{1/[np̂(1−p̂)]}. Exponential rate: I₁=1/λ² ⇒ SE=λ̂/√n.
Assumption ledgerchoose
| Tool | Needs | Failure |
|---|---|---|
| Wald | Approx normality | Boundary/skew/small n |
| Nonparam bootstrap | Representative sample | Bias reproduced |
| Param bootstrap | Correct family | Confident misspecification |
| Hoeffding/DKW | Bounded variable / CDF | Extra width |
Information is additive: independent observations give In=nI₁, so SE falls as 1/√n. Halving an SE takes four times the data. Use observed information at θ̂ for an estimated SE; use the null value in a null-based test when that test’s derivation requires it.
Transformation checklist: state g; differentiate; evaluate |g′| at θ̂; multiply by SE(θ̂); transform the estimate; inspect whether a symmetric interval can leave the transformed parameter space. The delta method is local linearisation, not a licence to ignore curvature.
MLE diagnostics: check parameter boundaries, verify the score root is a maximum, and remember regular asymptotics can fail when the true value lies on a boundary or the model is not identifiable.
5 · SufficiencyL5
T(X) is sufficient when the conditional law of full data given T=t is free of θ. Factorisation: p(x;θ)=h(x)g(T(x);θ); θ touches data only through T.
| Family | Sufficient summary |
|---|---|
| Bernoulli / Poisson | ΣXi |
| Normal, μ and σ² unknown | (ΣXi,ΣXi²) |
| Uniform(0,θ) | X(n), not X̄ |
Rao–Blackwell: E(θ̂|T) cannot raise MSE. One-to-one functions preserve sufficiency; many-to-one functions may lose it.
6 · Bayesian update & CRBL6
Bayes for parametersposterior ∝ likelihood × prior
Beta(α,β)+Y of n ⇒ Beta(α+Y,β+n−Y)
posterior mean=(Y+α)/(n+α+β)
Read α,β as prior successes/failures. Report the prior and show sensitivity when n is small. Normal–Normal posterior precision is prior precision plus sample precision.
Cramér–RaoVar(θ̂)≥1/{nI₁(θ)} for unbiased estimators only
Bernoulli: I₁(p)=1/[p(1−p)] ⇒ bound p(1−p)/n
Attaining the bound means efficient. A biased estimator may beat this variance floor in MSE.
EM looplatent labels
E: compute responsibilities under current parameters. M: update parameters with responsibility-weighted averages. Iterate until stable; use several starts because mixture likelihoods can have several peaks.
Two equal-variance componentsτi=φ(yi−μ₀)/{φ(yi−μ₀)+φ(yi−μ₁)}
μ₀←Στiyi/Στi · μ₁←Σ(1−τi)yi/Σ(1−τi)
Trap: EM can settle at a local peak; compare several starts and likelihoods.
Posterior ≠ point estimate. The posterior is a distribution; its mean is one chosen functional. A strong prior can reduce variance by adding bias, so compare posterior mean with the MLE and describe the prior in units of observations.
7 · Tests, errors, powerL7–8
State H₀,H₁ in parameter language; choose statistic and rejection region before seeing direction. Power β(θ)=Pθ(reject H₀). Type I rejects true H₀; Type II misses H₁.
| Size α | supθ∈Θ₀β(θ)=α |
|---|---|
| Level α | supθ∈Θ₀β(θ)≤α |
| Power | 1−Type II at stated alternative |
Neyman–Pearson: simple vs simple, reject for large L(θ₁)/L(θ₀); most powerful at level α against that one alternative.
General testsWald / LR
Wald: (θ̂−θ₀)/SE under H₀; two-sided reject when |T|≥zα/2. Proportion test uses p₀ in the null spread; interval uses p̂.
LR: λ=supΘ₀L/supΘL. Under regularity, −2logλ→χ²d, d=constrained parameters, not observations.
P-value: probability under H₀ of a statistic at least as extreme as observed; neither P(H₀|data) nor importance. Large p means insufficient evidence, not proof.
Fit & permutationL8
Pearson fitχ²=Σ(Oi−np0i)²/(np0i) · df=k−1
Permutation: under exchangeable labels, recompute any statistic over relabellings; p is the share at least as extreme. Exact if all are enumerated.
Normal testsupper-tail H₁:θ>θ₀ rejects if X̄>θ₀+zα/√n
two-sided rejects if |X̄−θ₀|>zα/2/√n
Power: rises with √{nI₁} times effect size. Near H₀ it equals α; with large n, tiny effects can be detectable, so report magnitude and interval with the verdict.
Goodness-of-fit conditions: compute expected counts before the statistic; combine categories if some are too small. If parameters are estimated from the same data, the reference degrees of freedom must account for them; for the fully specified course case it is k−1.
9 · Potential outcomesL9
Observe Yi=AiYi(1)+(1−Ai)Yi(0); individual effect is never observed. Target τ=E{Y(1)−Y(0)}.
Association ≠ effect: E(Y|A=1)−E(Y|A=0)=τ only with design/assumptions. Randomisation makes A independent of potential outcomes, turning the difference in means into a causal estimator.
Experiment rulebefore launch
Pre-specify primary outcome, α, smallest worthwhile effect, sample size and stopping rule; randomise; check balance; report estimate + interval + p-value; distinguish practical from statistical significance.
Sizing rule80% power, two-sided α=.05: n≈16σ²/δ² per arm
10 · AdjustmentL10
Ignorability: A ⟂ {Y(0),Y(1)} | X. Positivity: 0<P(A=1|X)<1 where inference is claimed. Ignorability is not testable from observed data.
Propensity / IPWπi=P(Ai=1|Xi)
τ=E[AiYi/πi−(1−Ai)Yi/(1−πi)]
Model treatment chance → plot overlap → estimate → sandwich SE/resampling → inspect weights → sensitivity analysis. Near-zero scores create enormous weights; poor overlap changes the population you can study.
11 · Trustworthy testsL11
Peeking and many metrics buy repeated chances to be wrong. K independent looks: FWER=1−(1−α)K. Bonferroni tests each planned look/metric at α/K. A correction chosen after inspection cannot restore the guarantee.
Pre-specify: primary outcome and stopping rule.
Design sentence: “Random allocation, pre-specified outcome and stopping rule; report estimate, interval and practical meaning.”