BIOS6902 Cheatsheet · Statistical Inference

City University of Hong Kong · handwritten A4 planner for the final exam. Semester A examination period, 7–19 December 2026 (check your personal exam timetable). Full BIOS6902 study guide →

BIOS6902 · two-side A4 landscapePrint at 100% only for revision; examination copy must be handwritten
BIOS6902
Statistical Inference
City University of Hong Kong · Semester A 2026/27
BIOS6902 cheatsheet QR code – open the full BIOS6902 study guide
Full guide + ask Sia about BIOS6902
SIDE 1 / 2 · COPY THIS BY HANDHANDWRITTEN-A4 PLANNER · targets → assumptions → rule → interpretationprinted sheet not permitted

Exam map · official factspriority 0

Assignments 30%Midterm 30%Final 40%

Final examination (2 hours) · 40% · closed book, with ONE A4 sheet of HANDWRITTEN notes permitted (official course materials).

Put on your pageWhy it earns space
Target + estimandPrevents mean/median and association/effect swaps
AssumptionsDecide whether a method is defensible
SE + reference lawTurns estimate into interval/test
Failure + repairShows diagnostic judgement
Reporting sentenceStops probability reversals

Separate pass conditions: at least 40% of possible continuous-assessment marks and at least 35% of possible final-exam marks; both apply.

Pop.TargetEstimateError
Course spine: population target, sample estimator, then error accounting.

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 quantityModel shorthandTarget
Disease prevalenceBernoullip=E(X)
Biomarker centreNormal / nonparametricμ or median
Event countPoissonλ
Survival timeGamma / nonparametricmean 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.

SAMPLEFn valsstatRESAMPLEEDFn drawsstatSUBSTITUTEresample spread≈ sampling spreadif EDF is close
Bootstrap substitution: resampled spread stands in for repeated-sample spread.

Bootstrap reciperesample

  1. Draw n observations with replacement.
  2. Recompute identical T*; repeat B times.
  3. SEboot=√{B⁻¹Σ(T*b−T̄*)²}.
IntervalRule / risk
NormalT̂±1.96 SE; symmetric
PercentileUse 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.

nominal 95%.50.30.15.06.02promiseddelivered
Wald coverage deteriorates as a proportion approaches a boundary.

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.

same maxlarger n: sharpsmaller n: flatℓ
Information is curvature: sharper likelihood means tighter estimation.

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

ToolNeedsFailure
WaldApprox normalityBoundary/skew/small n
Nonparam bootstrapRepresentative sampleBias reproduced
Param bootstrapCorrect familyConfident misspecification
Hoeffding/DKWBounded variable / CDFExtra 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.

FamilySufficient summary
Bernoulli / PoissonΣXi
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.

targetrawconditioned
Rao–Blackwell: conditioning on a sufficient summary reduces spread.

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.

E-stepM-steprepeat?alternate E and M updates
EM: E-step responsibilities → M-step weighted means → repeat with updated parameters.

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 αsupθ∈Θ₀β(θ)=α
Level αsupθ∈Θ₀β(θ)≤α
Power1−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.

under the nullalternativeType IType II
Moving the threshold trades Type I error (α) against Type II error (β).

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.

.8012080320power
Power rises with effect size; curves are n=20, 80, 320 (red, blue, green); at the null it equals α.

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.

OBSarmoutcomecauseRAND.armoutcomecauserandomisation breaks cause → arm
Randomisation severs the common cause → assignment arrow.

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)]

controltreatednoA=1noA=0
No overlap: left tail has no treated units; right tail has no controls. Restrict the target or redesign.

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.”

https://www.asksia.ai/cheatsheet/cityu-bios6902 · Full BIOS6902 study guide →
Shared by ________ · side 1/2
Forward this to your tutorial groupv2 · 2026-10-11
Semester A examination period, 7–19 December 2026.
Your exam allows one HANDWRITTEN A4 — copy side 1 by hand; do not bring this printout.
BIOS6902
Statistical Inference · Drills & traps
City University of Hong Kong · Semester A 2026/27
BIOS6902 cheatsheet QR code – open the full BIOS6902 study guide
Full guide + ask Sia about BIOS6902
SIDE 2 / 2 · DO THESE BEFORE THE EXAMWORKED DRILLS · calculate → diagnose → write the conclusionrevision printout only

D1 · Prevalence from countsL1 worked

Given: 70/200 ⇒ p̂=.35. SE=√{.35×.65/200}=.0337. 95% CI=.35±1.96(.0337)=(.284,.416).

Write: data are consistent with prevalence about 28%–42%; the procedure captures truth in 95% of repeated studies of this size. Never: “95% probability p is in this realised interval.”

Test H₀:p=.30: z=(.35−.30)/√{.30×.70/200}=1.54; p≈.12 ⇒ do not reject at 5%.

0.200.300.400.500.60Wald SEfinite boundestimate
Same estimate, two intervals: validity without shape assumptions costs width.

D2 · Bias–variance proofshow steps

Mean: E(X̄)=μ; Var=nσ²/n²=σ²/n; MSE→0. Uniform max: E(X(n))=nθ/(n+1); bias=−θ/(n+1), yet MSE→0: biased and consistent. Poisson mean: unbiased; Var=MSE=λ/n.

Choice rule: compare estimators by MSE, not bias alone. Slight off-centring can win when it greatly reduces sampling variance; consistency is a limit statement, not finite-sample unbiasedness.

unbiasedwidebiasedtight
Equal MSE can arise from high variance or squared bias; the remedy differs.

Estimator tablerecall

ModelEstimate · SE
Bernoulli pX̄ · √{p̂(1−p̂)/n}
Poisson λX̄ · √(λ̂/n)
Normal μX̄ · s/√n
Uniform upper θX(n); corrected ×(n+1)/n

Boundary drill: for a small observed proportion close to zero, decline the ordinary normal interval because its approximation is poor and can leave [0,1]; choose a distribution-free or named boundary-respecting interval and state the reason.

Dependence drill: if several outcomes come from the same patient or clinic, Var(ΣXi) contains covariance terms. Positive within-cluster correlation makes s/√n too optimistic; naming independence is part of the answer, not decoration.

Interval denominator: for an estimate-centred Bernoulli interval use p̂(1−p̂); for a score-type null test derived under H₀ use p₀(1−p₀). Mixing the two midway produces a statistic with no stated reference law.

D3 · EDF & DKWL2

Fix x; Bi=I(Xi≤x)~Bernoulli(F(x)). F̂=n⁻¹ΣBi; E(F̂)=F and Var(F̂)=F(1−F)/n. The supremum in DKW gives whole-curve control.

At fixed x: uncertainty is widest where F(x)=.5 because F(1−F) is maximal there. The DKW half-width is constant in x, shrinks at n−1/2, and is clipped to [0,1]. Unlike pointwise intervals, it controls the whole staircase in one statement.

01observed value10 points → 10 steps
Ten observations make ten EDF steps; a simultaneous band has constant vertical width.

D4 · Three bootstrap CIscholesterol

Median gap T̂≈18.5; SE≈7.4. Normal: (3.7,33.3). Percentile: (5.0,33.3). Pivotal: (5.0,34.0). All exclude zero; lower ends expose asymmetry. If resampling is skewed, avoid symmetric normal CI; state why.

Ratio estimate −.07 is inside ±.20, but CI (−.24,.15) is not contained: equivalence is not shown. The whole interval must fit.

Failure → repairL4

CI outside [0,1]Boundary; Wilson/Agresti–Coull
Skewed resamplesPercentile/pivotal
Max bias persistsBootstrap is not automatic bias repair
Param CI far narrowerCheck family specification

Right-skewed survival: exponential fit targets mean; nonparametric median is a different functional. Report the median and resampled interval when shape is not defensible.

Course data drill: two cholesterol groups give a plug-in difference of means about 20.9 with SE about 5.55, hence an approximate 95% interval about (9.8,32.0). State that the disease group is higher on this evidence; do not convert an observational group difference into an intervention claim.

Band reading: a DKW band controls the maximum vertical gap everywhere. To estimate P(a<X≤b), read F̂(b)−F̂(a); uncertainty for a searched-over contrast must respect simultaneity. A collection of separately valid pointwise intervals does not automatically protect the whole curve.

Bootstrap coding check: every resample has size n and is drawn with replacement from the empirical distribution; recompute the entire statistic, including any fitted nuisance parameters. Resampling residual pieces while holding part of the calculation fixed changes the procedure.

D5 · Bernoulli MLEderive

ℓ(p)=Ylogp+(n−Y)log(1−p). Score Y/p−(n−Y)/(1−p)=0 ⇒ p̂=Y/n. For 35/100: p̂=.35; SE=.0477; CI≈(.256,.444). Moments agrees.

Poisson: −E(second derivative)=1/λ ⇒ SE=√(λ̂/n). Exponential rate: I₁=1/λ² ⇒ SE=λ̂/√n. Delta: for τ=eμ, τ̂=eX̄, SE≈eX̄/√n.

D6 · Two priorsY=2,n=10

MLE=.20. Beta(1,1)→Beta(3,9), mean .25. Beta(10,10)→Beta(12,18), mean .40. Strong prior contributes 20 imaginary observations—twice the real sample—and pulls toward .50. Report sensitivity.

0.250.400.00.250.50.751.0uniformBeta(10,10)means .25 → .40
Two successes in ten: prior strength moves the posterior mean from .25 to .40.

D7 · CRB checkefficient?

Bernoulli I₁(p)=1/[p(1−p)]. Bound=1/[nI₁]=p(1−p)/n=Var(X̄); sample proportion is efficient. Trap: bound restricts unbiased estimators only.

D8 · Poisson test twicen=20

X̄=1.4,H₀:λ=1. Wald z=.4/√(1/20)=1.79; p≈.073. LR=2n{(λ₀−λ̂)−λ̂log(λ₀/λ̂)}≈2.84; compare χ²1. Neither rejects at 5%.

Report: both tests point towards a higher rate but do not reach the 5% threshold; give the estimate and interval so the reader sees the remaining uncertainty. “Retain H₀” must never become “prove λ=1”.

D8b · Normal prior weightingprecision

For Xi~N(μ,σ²), σ known, and μ~N(m,τ²), posterior variance={1/τ²+n/σ²}−1. Posterior mean is a weighted average of X̄ and m; each source receives weight proportional to its precision. As n grows, σ²/n shrinks and the prior weight vanishes. More precision means more weight in the posterior mean.

EM drill: write responsibilities first, then weighted means. A mixture likelihood is a product of sums; the complete-data version would separate if labels were known. EM uses expected labels, not recovered labels. Run different initial means; agreement is a diagnostic, disagreement exposes multiple peaks.

Prior sensitivity sentence: “The posterior mean moved from ___ to ___ across the stated priors; the stronger prior contributed the equivalent of ___ observations.” This is more informative than calling a prior merely weak or strong.

D9 · Factorise & improveL5

Poisson joint mass=[Π1/xi!]e−nλλΣxi; only T=ΣX touches λ ⇒ sufficient. Given T=t, E(X1|T)=t/n=X̄; variance falls λ→λ/n.

Uniform: θ−nI(X(n)≤θ), so max sufficient, mean not. Normal: expanding squares leaves Σx and Σx².

3 trials → totalsΣ=00001 seqΣ=1one 13 seqΣ=2two 1s3 seqΣ=31111 seq
For three Bernoulli trials, the total partitions eight sequences into sufficient cells.

D10 · Goodness of fitn=200

p₀=(.40,.35,.15,.10); O=(92,64,28,16); E=(80,70,30,20). Contributions 1.800,.514,.133,.800; sum=3.248, df=3. Below upper 5% point ⇒ compatible, not proof of equality. Estimating parameters reduces the reference degrees of freedom.

observedtail .07all label permutations
Permutation p-value: share of relabellings at least as extreme as observed.

D11 · Exact permutationby hand

(X₁,X₂,Y₁)=(1,9,3); observed |X̄−Ȳ|=2. Of 3!=6 label reshuffles, four are at least as extreme; exact p=4/6≈.67. For large samples, sample many relabellings and report the count.

Exchangeability is key. Random assignment guarantees it under the sharp null; otherwise shuffling may answer an unsupported question.

D11b · Testing languagemarks in words

Tea tasting: all four correct among four-of-eight choices occurs once in 70 random attempts, about .014. The logic is: null claim, statistic, probability under the null, line fixed in advance. Do not skip which outcomes count as at least as extreme.

Size vs level: size α means the worst null rejection probability equals α; level α means it is at most α. Discrete outcomes may not allow exactly .05, so a valid test can be conservative.

Observed z=1.70: two-sided p≈.089. It does not reject at 5% (critical 1.96) but does at 10% (critical 1.645). Reporting .089 preserves both facts and shows how close the evidence lies to either chosen line.

LR degrees of freedom: count restrictions: dim(Θ)−dim(Θ₀). Do not use sample size or number of categories unless the model’s restrictions make that count correct.

D12 · Potential outcomesL9

Each unit reveals Y(1) or Y(0), never both. Individual effect is unobserved; estimate an average across randomised groups. Do not call an observational difference a treatment effect without identifying assumptions.

UAY1Y0Δ1Tseen—?2C—seen?3Tseen—?4C—seen?
Fundamental problem: one potential outcome is permanently hidden for each unit.

D13 · Binary experiment500/arm

p̂₁=.15,p̂₀=.12 ⇒ effect=.03. SE=√{.15(.85)/500+.12(.88)/500}=.0216. CI=(−.012,.072). Conclusion: compatible with small loss through substantial gain; study did not settle the question.

Sizing: σ=20,δ=4 ⇒ n≈16(400)/16=400/arm. Halve δ and n quadruples.

D14 · Confounding reversalL10

Healthy: 320 treated,80 control, rate .80 both. Unhealthy: 80 treated,320 control, rate .30 both. Pooled treated=.70; control=.40; naïve gap=.30. Within-stratum gaps=0; standardised effect=0.

The naïve number compared different people, not interventions. Adjustment changes the comparison; it cannot verify all common causes were measured.

healthyillpooled
Blue/orange = treated/control. Healthy: n=320/80, both .80; ill: n=80/320, both .30; pooled .70/.40.

Adjustment trapsdiagnose

Extreme πHuge weights; inspect/trim and report
No overlapRestrict support; target changes
Hidden causeSensitivity analysis; qualify
Post-treatment XMay remove part of effect

Five-step observational answer: state ignorability and positivity; model π=P(A=1|X); plot overlap by arm; form the weighted/standardised estimate with appropriate uncertainty; finish with weight diagnostics and sensitivity to an unmeasured cause.

Randomisation test: under the sharp null, keep outcomes fixed, rerandomise labels by the actual mechanism and count contrasts at least as extreme.

Covariates: adjust for pre-treatment common causes. A mediator removes part of the effect; a post-treatment variable is not a baseline confounder.

D15 · Peeking & metricsL11

Ten looks: 1−.9510=.401; Bonferroni .005/look. Real looks correlate, so independent arithmetic overstates inflation but shows direction.

Twelve metrics: 1−.9512=.460; threshold .05/12=.0042. p=.031 fails; call it hypothesis-generating and confirm as a new primary outcome.

5% each look0.0510.2350.40100.6420false alarms
Independent repeated looks inflate the chance of at least one false alarm.

Fast test selectorprompt → tool

Simple vs simpleNeyman–Pearson LR
General parameterWald / composite LR
Counts vs proportionsPearson χ², df k−1
Randomised odd shapePermutation
Many looks/metricsPre-specify; adjust α

Sentence trapswrite precisely

“95% chance θ inside”95% is repeated-study coverage
“p=P(H₀ true)”Tail probability under H₀
“Not significant=no effect”Insufficient evidence; inspect CI/power
“Significant=important”Judge magnitude separately
“Association=effect”Need design/identification
“CRB limits all”Unbiased only

Formula beltrebuild

MSE=bias²+Var · SE(X̄)=s/√n
εDKW=√{log(2/α)/(2n)}
I₁=Var(score)=−E(second derivative) · SE≈1/√(nI₁)
SE{g(θ̂)}≈|g′(θ̂)|SE(θ̂)
posterior ∝ likelihood×prior; Beta adds counts
Wald=(θ̂−θ₀)/SE₀ · −2logLR→χ²constraints
χ² fit=Σ(O−E)²/E · ATE=E{Y(1)−Y(0)}
n≈16σ²/δ²/arm · FWER=1−(1−α)K

Final five checksdiscipline

  1. Target and denominator named?
  2. Assumption buying result stated?
  3. p̂ for CI, p₀ for null test?
  4. Units, df, reference law given?
  5. Conclusion avoids probability reversal and causal overreach?

Answer: method → numbers → reference → scientific conclusion.

https://www.asksia.ai/cheatsheet/cityu-bios6902 · Full BIOS6902 study guide →
Shared by ________ · side 2/2
Forward this to your tutorial groupv2 · 2026-10-11
Semester A examination period, 7–19 December 2026.
Your exam allows one HANDWRITTEN A4 — copy side 1 by hand; do not bring this printout.

Want one for YOUR exact syllabus?

Sia is your free desktop study agent. Drop your City University of Hong Kong BIOS6902 slides — Sia builds a sheet tailored to YOUR exam. Better than this library because it knows YOUR materials.

↓ Download Sia · Free