about ETC1000
0 · Exam Mapcourse.json · brief
| Exercises, completed during your allocated tutorial | 20% |
| In-Class Tests, two tests worth 10 per cent each | 20% |
| Group Project | 20% |
| Examination | 40% |
Permitted materials: not confirmed for this offering in the course materials we hold — check the exam instructions on the LMS. BRIEF.md
The unit’s repeated scoring move is quantity + sentence: compute the right object, name its denominator or scale, then state what it establishes and what it does not.
1 · Variable → Display → SummaryCh 1–2
| Type | Test / example | Use |
|---|---|---|
| Nominal | no order; payment | counts, mode, bar |
| Ordinal | ordered bands | ordered bar, median band |
| Discrete | whole counts | frequency, histogram |
| Continuous | measured refinement | histogram, boxplot, mean |
Digits do not decide type: postcode, customer ID and store number are nominal because an average is meaningless. A sorted bar chart compares lengths and usually beats a pie chart beyond about four categories.
2 · Contingency TablesCh 1
same cell, changing denominatormarginal = row/column total ÷ grand total
joint = interior cell ÷ grand total
conditional = cell ÷ its row/column total
Independence: compute the same conditional probability inside each row. Equal conditionals → independent; unequal → relationship. Report the percentage-point gap. Signal words given, among, of those who demand a row or column denominator.
3 · Centre, Spread, ShapeCh 2
| Measure | Promise | Trap |
|---|---|---|
| Mean | balance point; uses all values | pulled by extremes |
| Median | middle rank | ignores distance |
| Mode | most common | may tie / be absent |
| Range | max − min | two least typical cases |
| IQR | Q3 − Q1; middle half | tails omitted |
| SD | typical spread about mean | quote in original units |
sample summariesx̄ = Σxᵢ/n
s² = Σ(xᵢ−x̄)²/(n−1) · s = √s²
IQR = Q3−Q1 · fences = Q1−1.5IQR, Q3+1.5IQR
Skew: mean > median → long upper/right tail; mean < median → long lower/left tail. Prices, incomes and waiting times often need median + IQR. The mean remains the right centre when the question is about a total.
Why n−1? Deviations use the sample mean, which makes them unusually small; dividing sample squared deviations by n would systematically understate population variance. Use n−1 for a sample estimate.
Histogram rule: try two or three interval widths and retain one where the shape is stable. Too wide hides skew; too narrow reports sampling noise as structure. State the width in the caption.
Outlier workflow: flag → check recording and case context → report sensitivity with and without if influential. A flagged point may be the most informative case and is never deleted merely for being unusual.
4 · Standardise Before ComparingCh 2
5 · Population → IntervalCh 3
| Population language | Sample language |
|---|---|
| population: every case the claim covers | sample: observed subset |
| parameter: fixed, usually unknown | statistic: known, varies by sample |
| bias: frame/process error | sampling variability: random noise |
A larger sample shrinks noise, not bias. Selection, survivor and non-response bias remain. Ask how cases entered before asking how many there are. The law of large numbers stabilises sample summaries around population values; it cannot repair the wrong frame.
sampling distribution of the meancentre = μ · SE(x̄) = σ/√n ≈ s/√n
quadruple n → halve SE
interval for a population meanx̄ ± t* × s/√n · df = n−1
| Use | Reason |
|---|---|
| Normal critical value | population σ known |
| t critical value | σ unknown; estimate with s; heavier tails |
95% confidence: the procedure captures the parameter in about 95 of 100 repetitions. The realised interval either covers it or does not. An interval for a mean says nothing about where most individual observations lie.
| Failure | What went wrong | Does larger n fix it? |
|---|---|---|
| Selection bias | entry relates to measured quantity | No |
| Survivor bias | failed cases absent from frame | No |
| Non-response | respondents differ from non-respondents | No |
| Sampling noise | chance sample-to-sample movement | Yes, at √n rate |
6 · Precision LeversCh 3
| larger s | wider interval |
| larger n | narrower at √n rate |
| higher confidence | larger critical value → wider |
Check Write the SE first. If it did not shrink when n grew, you divided by n instead of √n. Then name the population supported by the sampling frame.
7 · Association → RegressionCh 4
Scatter first: direction, form, strength and exceptions decide whether a straight-line summary is licensed. Put the explanatory variable on x and outcome on y. Swapping axes changes the fitted line because least squares measures vertical error; it does not change correlation.
joint movementsxy = Σ(xᵢ−x̄)(yᵢ−ȳ)/(n−1)
r = sxy/(sx sy), −1 ≤ r ≤ 1
least squaresŷ = b₀+b₁x · b₁ = sxy/sx² · b₀ = ȳ−b₁x̄
eᵢ = yᵢ−ŷᵢ · choose line minimising Σeᵢ²
R² = r² in simple regression
Slope sentence: “A one-unit increase in x is associated with b₁ units of y.” Name both units. Intercept only if x=0 is observed and meaningful. Residual curvature → wrong form; fan → changing spread; isolated large residual → influential exception.
R² is not a licence: it gives the share of sample outcome variation accounted for by the line; it proves neither linearity, causation nor validity outside the observed x-range.
12 · R + Wrangling RulesCh 7–8
| Verb / object | Change | Check |
|---|---|---|
| filter | keeps rows | condition first |
| select | keeps columns | variables retained |
| mutate | adds column | units / type |
| arrange | reorders | no data change |
| group | invisible label | expected groups |
| summarise | collapses rows; changes grain | carry count |
8 · Hypothesis-Test PipelineCh 5
- Name the population parameter.
- Write H₀ as equality/no effect; H₁ from the question.
- Fix α before seeing data; choose one- or two-sided from wording.
- Compute test statistic = (estimate − null)/SE.
- Compare p with α; decide; report effect, units and limit.
common formsH₀: μ=μ₀; H₁: μ≠μ₀ (differs/changed)
H₀: β₁=0; H₁: β₁≠0
“exceeds / improved / faster” → one-sided
Correct p sentence: if H₀ were true, a result at least this extreme would occur with probability p. A large p means insufficient evidence, not that H₀ is true. Statistical significance does not certify commercial importance; read effect size and interval.
| Error | Meaning | Lever |
|---|---|---|
| Type I | reject true H₀; false alarm | rate = α |
| Type II | miss a real effect | power = 1−β |
| Trade | lower α usually raises β | larger n improves both |
9 · Multiple RegressionCh 6
coefficient sentencechange in y for +1 unit of xⱼ, holding every other included variable fixed
Adding a variable changes the question and may shrink or reverse a coefficient because shared variation is reassigned. It does not make either estimate arithmetically wrong. Avoid causation from observational data and remember that omitted variables can still contaminate the model.
| Device | Rule | Read |
|---|---|---|
| k-category factor | k−1 dummies | each vs omitted baseline |
| Interaction | x×d | difference between slopes |
| Ordinary R² | never falls when x added | not model chooser |
| Adjusted R² | penalty per variable | compare specifications |
10 · ClassificationCh 9
A model gives an estimated positive probability; a threshold turns it into a decision. Lower threshold → more positives, higher sensitivity and more false positives. Raise it when false action is costly; lower it when a missed positive is costlier. Moving the threshold trades errors; it does not improve the fitted model.
ratesaccuracy=(TP+TN)/all
sensitivity=TP/(TP+FN) · specificity=TN/(TN+FP)
precision=TP/(TP+FP) · base rate=(TP+FN)/all
| Rate | Denominator asks | Lead when… |
|---|---|---|
| Accuracy | all cases | classes not badly imbalanced |
| Sensitivity | actual positives | misses are costly |
| Specificity | actual negatives | false alarms are costly |
| Precision | predicted positives | a positive flag must be trusted |
Accuracy must beat the base-rate benchmark. At 3% defaults, predicting none scores 0.97 and catches nothing. Fit on training data; judge on held-back data.
11 · Logistic RegressionCh 10
scalesodds=p/(1−p) · log odds=ln[p/(1−p)]
log odds = b₀+b₁x+… · odds multiplier for +1 x = eᵇ¹
p = 1/(1+e⁻η), η=b₀+b₁x+…