1 · Start With the RowW1
| Field role | Use / trap |
|---|---|
| Identifier | duplicate checks; do not average |
| Nominal | labels, no order → count/proportion |
| Ordinal | ordered levels; spacing not guaranteed |
| Discrete | counts; numerical arithmetic |
| Continuous | measurements; units essential |
| Missing | never silently zero |
Audit sequence: unit → variable role → missing/impossible values → duplicate identifiers → target population. Preserve the untouched source; put derived fields in new columns and reconcile row counts after filters.
Exam bite A numeric code such as postcode can be nominal. Storage format does not determine statistical type.
2 · Choose the DisplayW2
| Question | Report |
|---|---|
| How many? | count: workload / volume |
| What share? | relative frequency = count ÷ relevant total |
| Unequal exposure? | normalised rate, with exposure unit |
| Audit | title, axes, unit, order, missing groups, selected range |
Bars start from a common baseline. Histograms use adjacent numerical bins. Lines preserve genuine order, especially time. A truncated axis may exaggerate a length difference; 3-D geometry obscures magnitude. Normalisation repairs the denominator, not uncontrolled product mix, hours or case complexity.
2b · Spreadsheet AuditW1–2
| Stage | Action | Evidence retained |
|---|---|---|
| land | save raw tab; one header row | untouched source |
3 · Categorical ProbabilityW3
Core rulesP(Aᶜ)=1−P(A)
P(A∪B)=P(A)+P(B)−P(A∩B)
P(A|B)=P(A∩B)/P(B)
independence: P(A|B)=P(A), hence P(A∩B)=P(A)P(B)
| Asked | Denominator | Check |
|---|---|---|
| joint A and B | grand total | cell ÷ all |
| marginal A | grand total | A total ÷ all |
| A given B | B total | B-conditioned shares sum 1 |
Association: compare outcome proportions across groups using the same conditioning logic. A large raw cell may merely reflect a large group. Report a difference between percentages as percentage points; relative percent change divides that gap by a named baseline.
Exam bite Reversing the words after “given” reverses the denominator. Write the fraction before the decimal. Association in recorded data is not causation.
3b · Denominator ChecklistW3
| Before submit | Question |
|---|---|
| Unit | ticket, visit, order or customer? |
| Eligibility | who could experience “success”? |
| Missing | excluded, shown, or separate? |
| Scope | sample description or population inference? |
Repeated visits are not automatically independent customers. A pooled difference can also hide a lurking group pattern; investigate before choosing the preferred comparison.
3c · Probability Drill GridW3
| From one two-way table | Required working |
|---|---|
| joint | interior cell ÷ grand total; name both events |
| reversed conditionals | same cell, two different group totals |
4 · Centre, Spread, ShapeW4
| Shape / purpose | Centre | Spread |
|---|---|---|
| symmetric, no strong outlier | mean | standard deviation |
| skewed / influential outlier | median | IQR |
| workload / total planning | mean may still matter | show valid extremes |
Standardise an observationz = (x − mean) / standard deviation
z > 0 above mean; z < 0 below; unit-free
Units: mean, median, range, IQR and standard deviation use the original unit; variance uses squared units; z has no unit. Standardising does not make skewed data normal. Use a normal model only after checking shape and whether the model is suitable.
Outlier rule Verify the record and business origin. Correct a confirmed error; retain and explain a valid exception; never delete merely because the answer changes.
4b · Distribution WorkflowW4
| Step | Question | Exam output |
|---|---|---|
| 1 shape | symmetric, skewed, clusters, gaps? | name visible form |
| 2 centre | typical case or resource average? | median or mean, with unit |
| 3 spread | middle half or mean-distance? | IQR or SD, with unit |
| 4 unusual | error, valid exception, second process? | investigate; sensitivity |
| 5 compare | same unit, period, inclusion rule? | difference plus context |
Resistance test: predict which summaries change if the largest value doubles. Mean, range and SD react; median and IQR are usually more resistant. Resistance is not permission to ignore operational extremes.
5 · Sampling & IntervalsW5
Interval architectureestimate ± critical value × SE
margin = critical value × SE
width = 2 × margin
Mean · large-sample practiceSE(x̄)=s/√n
95% CI: x̄ ± 1.96(s/√n)
Proportion · usual intervalp̂=x/n
SE(p̂)=√[p̂(1−p̂)/n]
CI: p̂ ± critical × SE(p̂)
| Change | Interval effect |
|---|---|
| n increases | SE falls; narrower (4×n ≈ ½SE) |
| confidence increases | critical value rises; wider |
| variability increases | SE rises; wider |
| selection bias | not repaired by a narrow interval |
Interpretation: the confidence procedure captures the true population parameter at its stated long-run rate under assumptions. For one realised interval, state plausible population parameter values, not the percentage of individual observations and not a probability attached to a fixed parameter.
Endpoint checksum: midpoint = estimate; half-width = margin; lower < upper; both carry the parameter’s unit. Confirm whether the entry asks for an endpoint, margin, width or full interval.
Exam bite Random sampling supports generalisation. Independence, selection, nonresponse and measurement quality remain separate from arithmetic precision.
5b · CI Decision ChecksW5
| Interval location | Decision reading |
|---|---|
| entirely above target | plausible means all exceed target |
| crosses target | values on both sides remain plausible |
6 · Tests: Claim → DecisionW6–7
- State population H₀ and H₁; direction comes from the claim before data.
- Check sampling/independence and approximation conditions.
- Compute SE and standardised statistic; use the matching tail(s).
- If p ≤ α, reject H₀; otherwise do not reject H₀.
- Conclude about the represented population; report effect size and practical importance.
Mean test · large-sample practiceH₀: μ=μ₀
z=(x̄−μ₀)/(s/√n)
One-proportion z testH₀: p=p₀
SE₀=√[p₀(1−p₀)/n]
z=(p̂−p₀)/SE₀
| Claim wording | Alternative / tail |
|---|---|
| exceeds / above | > · right |
| below / reduced | < · left |
| differs / changed | ≠ · two-sided |
Test SE uses p₀; usual proportion-interval SE uses p̂. For the normal approximation, check expected success and failure counts using the proportion appropriate to the method.
| Decision | H₀ true | H₀ false |
|---|---|---|
| reject | Type I error | correct |
| do not reject | correct | Type II error |
Language A small p-value is not P(H₀ true). “Do not reject” is not “prove equal”. Statistical evidence is not automatically operational importance.
6b · Test Output DecoderW6–7
| Output | Meaning | Common mix-up |
|---|---|---|
| estimate | observed sample effect | population truth |
| SE | sampling fluctuation scale | sample SD |
| z | null distance in SE units | effect magnitude |
| p-value | null-tail probability | P(H₀ true) |
| α | pre-set Type I threshold | move after seeing p |
| CI | plausible parameter range | individual-value range |
7 · Regression & ModelsW7–10
Simple regressionŷ=b₀+b₁x
residual e=y−ŷ
R² = proportion of sample y-variation explained
| Output | Interpretation / check |
|---|---|
| b₁ | predicted y change per +1 x; name both units |
| b₀ | predicted y at x=0; may lack meaning |
| e > 0 | observed lies above fitted value |
| residual plot | curve, funnel or influential case warns |
| prediction | avoid extrapolation outside observed x range |
Correlation: unit-free linear association from −1 to 1; sensitive to outliers and blind to some curved relationships. Regression/correlation alone do not establish causation.
Multiple regression
ŷ=b₀+b₁x₁+⋯+bₖxₖ
Each slope compares predicted outcomes for a one-unit predictor change while holding the other included predictors fixed. For a 0/1 indicator, coefficient = predicted difference from the omitted reference category at the same included controls. Ordinary R² cannot fall when predictors are added; adjusted R² penalises weak additions. Neither validates causation or new-data prediction.
Optimisation
Define variables, objective and every capacity/demand/policy constraint; add non-negativity and integrality where relevant. Then substitute the returned values into every constraint, compute slack and mark binding limits. Feasible satisfies constraints; optimal is best among feasible choices. Stress-test contribution, capacity and demand. A sensitive mix changes sharply after a plausible input change.