Week 18 — Final Exam Study Guide · The Whole Course, Condensed
Course: Introduction to Statistics (18-week generic edition)
Covers: Weeks 1–17 (Objectives 1–9) — everything, with extra weight on the inference half
The final: 60 auto-graded questions × 2 points = 120 points · mostly multiple choice, a few true/false, select-all, and matching · closed to AI · one attempt · sits mid-week in Week 18 · worth 5% of the grade
This page: ungraded — it's yours to keep
How to use this guide
Read it like a map, not a novel. The final draws roughly 24 questions from Weeks 1–8 and 36 from Weeks 10–17, so the post-midterm sections below are deliberately fuller — spend your time in proportion. For each week you get: the big idea, the must-know terms, a "can you…" checklist, the classic traps, and worked mini-examples with fresh numbers (every one pre-computed and verified — none of them appears on any graded assessment).
The efficient path through Week 18: ① read this guide with paper handy and do the mini-examples before reading their solutions → ② take the practice exam (30 questions, unlimited attempts, ungraded — it shares zero items with the real final) → ③ run the exam-prep tutorial with your chatbot to diagnose and drill whatever wobbled → ④ take the final. Two focused hours across three days beats any long night.
About lookup values: any final-exam question that needs a z, t, or chi-square value prints it in the question. You memorize how to use a stated value — never the tables themselves. The three course tables are reprinted in the exam-prep tutorial.
Part 1 — The descriptive & probability half (Weeks 1–8, ≈24 questions)
Week 1 — Statistics, data & study design
Big idea. Statistics turns a measured sample into a claim about an unmeasured population; the statistic (the value we have) estimates the parameter (the value we want). How individuals were chosen decides whether any of it deserves trust — bias is baked into the method and no sample size fixes it.
Must-know terms: population · sample · census · parameter · statistic · categorical (nominal/ordinal) vs. quantitative (interval/ratio) · SRS · stratified · cluster · systematic · convenience · voluntary response · bias (undercoverage, nonresponse, response, voluntary-response) · observational study · experiment · confounding variable.
Can you… name the population and sample in a scenario? · label a number as parameter or statistic? · classify a variable with NOIR? · identify a sampling design from its description? · tell an observational study from an experiment and say which supports cause?
Classic traps: "if it's a number, it's quantitative" (route numbers and IDs label); "a bigger sample fixes bias" (the 1936 Literary Digest poll: 2.4 million ballots, wrong winner — method beats size); stratified (within every group) vs. cluster (whole groups); "strong correlation, so it causes" (only random assignment earns a causal verb).
Mini-example (do it, then check). A city audits a random sample of 350 of its 9,000 streetlights; 21 are out. Population? Sample? What kind of number is 21/350? → Population = all 9,000 streetlights; sample = the 350 audited; 21 ÷ 350 = 0.06, a statistic estimating the parameter (the true out-of-service rate among all 9,000).
Deeper review: Chapter 1, Deck 1, the Week 1 outline's library scenario.
Week 2 — Tables & graphs
Big idea. Count, then share: relative frequencies (count ÷ total) make groups comparable and must sum to 1. Match the display to the variable — separated bars for categories, touching bars (histograms) for quantitative — then read shape with the standard words. Graphs can lie about honest data; the truncated axis is the classic move.
Must-know terms: frequency · relative frequency · bar chart · pie chart (parts of one whole only) · histogram · dot plot · stem plot · distribution · shape (symmetric, skewed right/left, uniform, bimodal) · outlier (investigate, don't delete) · misleading graph.
Can you… compute a relative frequency? · pick the right display for a variable? · name a distribution's shape from a description? · say why a truncated-axis chart misleads?
Classic traps: skew is named for the tail, not the peak; pie charts require non-overlapping parts of one whole; bars start at zero.
Mini-example. A volunteer group files 60 trail reports; 18 mention fallen trees. Relative frequency → 18 ÷ 60 = 0.30, and all categories' shares must total 1 — your built-in error check.
Deeper review: Chapter 2 and Deck 2's misleading-graphs run.
Week 3 — Center & spread
Big idea. Three centers (mean, median, mode) and the deciding question: resistance. The mean chases the tail, so skewed data report the median; the SD is the typical distance from the mean and travels with the mean, while the IQR travels with the median. z-scores measure relative standing in SD units.
Must-know terms: mean · median · mode · resistant measure · range · deviation · variance · standard deviation · five-number summary · quartiles · IQR · 1.5×IQR rule · boxplot · z-score.
Can you… compute mean, median, and s for five values? · pick the right center/spread pair for a skewed variable? · apply the 1.5×IQR fences? · compute and interpret a z-score?
Classic traps: variance is in squared units (report the SD); the range is owned by two values; deviations always sum to zero (that's why we square); z-scores carry no units.
Mini-example. Five potted ferns measure 6, 6, 10, 14, 14 cm tall. Mean = 50 ÷ 5 = 10; median = 10; squared deviations 16 + 16 + 0 + 16 + 16 = 64 → s² = 64 ⁄ 4 = 16 → s = 4 cm. An 18-cm fern would sit z = (18 − 10) ⁄ 4 = 2.0 — two SDs above the mean.
Deeper review: Chapter 3's crash-day demo (one outlier moves the mean, not the median).
Week 4 — Two variables
Big idea. Scatterplots read with Direction–Form–Strength (then stragglers); r compresses a straight-line association into one unitless number in [−1, 1]. Two-way tables ask "out of which group?" — different conditional distributions = association. And association is never causation: hunt the lurking variable.
Must-know terms: explanatory/response · scatterplot · correlation r · two-way table · marginal vs. conditional distribution · association · lurking variable · confounding.
Can you… describe a scatterplot in a sentence? · rank associations by |r|? · compute a conditional percent from a two-way table? · propose a lurking variable for a headline?
Classic traps: r sees straight lines only (a strong arch can have r ≈ 0); r has no units and survives unit changes; the denominator trap ("among weekend visitors" ≠ "among everyone").
Mini-example. A museum logs 50 morning visitors (20 buy the audio guide) and 25 evening visitors (5 buy it). Conditional percents: morning 20 ⁄ 50 = 40%, evening 5 ⁄ 25 = 20%. The conditionals differ → time of visit and buying the guide are associated — and that alone says nothing about why.
Deeper review: Chapter 4; the firefighters-and-damage example in the Week 4 quiz debrief.
Week 5 — Probability foundations
Big idea. Probability is a long-run promise: the proportion of times an event would occur over endless repetitions. Four verbs run everything — not (complement), or (add, subtract the overlap), and (multiply if independent), given (shrink the world, re-count).
Must-know terms: random phenomenon · law of large numbers · sample space · event · complement · disjoint · addition rules · independence · multiplication rule · conditional probability P(A|B) · gambler's fallacy.
Can you… interpret "P = 0.05" as a long-run rate? · apply the complement and general addition rules? · compute a conditional probability from counts? · explain why streaks create no debts?
Classic traps: disjoint and independent are different facts; P(A|B) ≠ P(B|A); "due for a win" is the gambler's fallacy — the long run works by swamping, not compensating.
Mini-example. For tomorrow's forecast, P(rain) = 0.5, P(high wind) = 0.3, P(both) = 0.15. P(rain or wind) = 0.5 + 0.3 − 0.15 = 0.65. P(rain given wind) = 0.15 ⁄ 0.3 = 0.50 — inside the windy world, rain happens half the time.
Deeper review: Chapter 5's two-way-table conditionals; the Week 5 lab's long-run convergence.
Week 6 — Random variables
Big idea. A random variable attaches numbers to chance. Its distribution table is its whole personality (probabilities in [0, 1], total exactly 1); E(X) = Σ x·P(x) is what you'd average, not what you'd expect, and the SD is the typical distance from that average. Y = a + bX: adding shifts the center only; multiplying rescales center and spread.
Must-know terms: discrete vs. continuous · probability distribution · legitimate distribution · expected value · variance/SD of X · linear transformation · density curve (area, not height; single exact values have probability 0).
Can you… find a missing probability? · compute and interpret E(X)? · push a mean and SD through Y = a + bX? · say why P(X = one exact value) = 0 for continuous X?
Classic traps: E(X) needn't be a possible value; adding a constant does not change the SD; probabilities that don't sum to 1 aren't a distribution.
Mini-example. A scratch card pays $0, $10, or $20 with probabilities 0.5, 0.3, 0.2. E(X) = 0(0.5) + 10(0.3) + 20(0.2) = $7 — the long-run average per card, even though no card ever pays $7.
Deeper review: Chapter 6's claw-machine expected value; Deck 6's "average, not guarantee."
Week 7 — The binomial distribution
Big idea. When a situation passes B·I·N·S (Binary, Independent, Number fixed, Same p), the count of successes is binomial: P(X = k) = C(n, k)·p^k·(1 − p)^(n−k) — ways × wins × losses — with mean np and SD √(np(1−p)). "At least one" has an express lane: 1 − P(none).
Must-know terms: binomial setting · trial · success/failure · n, p · binomial coefficient · binomial formula · complement rule for "at least one" · mean np · SD √(np(1−p)) · cumulative probability.
Can you… check B·I·N·S? · name n and p from a story? · compute P(X = k) for small n? · compute the mean and SD of a binomial count?
Classic traps: forgetting the C(n, k) "ways" factor; "until the first success" has no fixed n (not binomial); without-replacement draws break independence; "exactly k" is one bar, "at least/at most" are tails.
Mini-examples. ① A street performer lands a coin-flip trick with p = 0.4, three independent tries. P(exactly 2) = C(3, 2)(0.4)²(0.6) = 3 × 0.16 × 0.6 = 0.288. ② Flip a fair coin 64 times: the head count has mean 64 × 0.5 = 32 and SD √(64 × 0.25) = 4 — so 40 heads (z = 2) would raise an eyebrow.
Deeper review: Chapter 7's B·I·N·S impostors; the embedded binomial mini-table in Tutorial 7.
Week 8 — The normal distribution
Big idea. Density curves carry proportion as area (total 1). The normal model N(μ, σ) is drawn by two numbers, obeys 68–95–99.7 if bell-shaped, and the z-score z = (x − μ)/σ translates any normal problem onto one shared ruler. Forward: z → table → decide below/above/between. Inverse: percent → z → x = μ + z·σ. Check normality before trusting any of it.
Must-know terms: density curve · N(μ, σ) · empirical rule · z-score · standardizing · standard normal N(0, 1) · left-tail area · percentile (a position, never a score) · inverse calculation · assessing normality.
Can you… apply 68–95–99.7? · run a forward calculation with a stated table value? · run an inverse calculation (x = μ + z·σ)? · reject the normal model when a histogram is skewed?
Classic traps: height is never proportion — area is; "above" needs 1 − left area; the empirical rule's password is IF bell-shaped; a percentile is a position.
Mini-examples. ① A machine cuts dowels with lengths ≈ N(50, 10) mm. P(length < 62.5): z = 1.25 → left area 0.8944. ② The 97.72nd percentile: left area 0.9772 → z = 2 → 50 + 2(10) = 70 mm.
Deeper review: Chapter 8's forward/inverse recipes; the friendly z-table (reprinted in the exam-prep tutorial).
Part 2 — The inference half (Weeks 10–17, ≈36 questions)
Week 10 — Sampling distributions & the CLT
Big idea. Honest samples disagree — that's sampling variability, and the sampling distribution is the catalog of all possible answers. The sample mean's catalog is centered exactly at μ (unbiased) with spread SE = σ/√n, and the Central Limit Theorem hands it a bell shape as n grows, whatever the population looks like. Two rulers, always: σ for individuals, σ/√n for averages.
Must-know terms: sampling variability · sampling distribution · unbiased · standard error · SD vs. SE · Central Limit Theorem · n ≥ 30 rule of thumb · sampling distribution of p̂ · success/failure condition (np ≥ 10, n(1−p) ≥ 10).
Can you… distinguish the population, one sample, and the sampling distribution? · compute σ/√n? · explain why quadrupling n halves the SE? · compute a probability about x̄ using the z-table? · say what the CLT does not do (fix bias; reshape individuals)?
Classic traps: the CLT is about the distribution of sample means, never the histogram of one sample's individuals; SE shrinks by √n (half the noise costs 4× the data); averages are less variable than individuals, so extreme averages are rarer than extreme individuals.
Mini-examples. ① Daily visitor counts at a small museum have σ = 12. For samples of n = 9 days, SE = 12 ⁄ √9 = 4. ② If the true mean is μ = 50, the probability a 9-day average exceeds 54 is: z = (54 − 50) ⁄ 4 = 1.0 → 1 − 0.8413 = 0.1587. (One day exceeding 54 is far more likely — different ruler.)
Deeper review: Chapter 10; the Week 10 lab's 30-sample-means histogram — the course's single most important picture.
Week 11 — Confidence intervals for a mean
Big idea. A point estimate is almost surely not exactly μ, so the honest report is x̄ ± t*·(s/√n). Because s stands in for σ, the multiplier comes from the t-distribution with df = n − 1 — heavier tails, more humility, melting into z as n grows. "95% confident" describes the method's capture rate (about 19 of every 20 such intervals contain μ), never a probability about your one interval and never a claim about individuals.
Must-know terms: point estimate · confidence interval · confidence level · margin of error (t* × SE) · standard error s/√n · t-distribution · degrees of freedom · critical value t* · conditions (random; roughly normal for small n) · capture rate · plausible values.
Can you… pick t* from a stated table row? · build the interval in four beats (conditions → SE → ME → x̄ ± ME)? · interpret it about the mean, in context? · predict what widens/narrows it (confidence level, n, s)?
Classic traps: using s where SE belongs (it quadruples a margin at n = 16); "95% of individuals are inside" (no — it's about μ); "95% probability μ is in this interval" (confidence is the method's long-run rate); df = n − 1, not n; certainty costs width.
Mini-examples. ① Sixteen jars from a honey harvest average x̄ = 24.0 oz with s = 4. SE = 4 ⁄ √16 = 1; df = 15, so t* = 2.131 (95%); ME = 2.131 × 1 = 2.131 → interval (21.87, 26.13) oz. Say it: "We are 95% confident the interval captures the mean fill of the whole harvest." ② Same data at 99% (t* = 2.947): the interval widens — surer means wider, always.
Deeper review: Chapter 11's capture-rate simulation; the friendly t-table (reprinted in the exam-prep tutorial).
Week 12 — Confidence intervals for a proportion
Big idea. A sample percent estimates p with interval p̂ ± z*·√(p̂(1−p̂)/n) — center ± reach — after the password: random sample, at least 10 successes and 10 failures, population ≥ 10n. Planning a survey runs the machine backwards: n = p*(1−p*)(z*/ME)² with the conservative p* = 0.5, always rounded up. A media poll's ± covers random sampling error only — bias rides outside it.
Must-know terms: p vs. p̂ · standard error of p̂ · z* (1.645 / 1.96 / 2.576) · margin of error · one-proportion z-interval · large-counts condition · 10% condition · conservative planning value · sample-size formula.
Can you… compute SE from p̂ and n? · build and interpret the interval? · run the sample-size formula and round correctly? · read a reported poll (point estimate ± MOE) and say what the MOE does not cover?
Classic traps: feeding the SE formula counts instead of proportions; rounding a sample size down (600.25 people means 601); "the MOE accounts for bad wording/nonresponse" (it doesn't); means ride t, proportions ride z.
Mini-examples. ① 32 of 64 surveyed hikers carry trekking poles: p̂ = 0.50, SE = √(0.25 ⁄ 64) = 0.0625, 95% ME = 1.96 × 0.0625 = 0.1225 → interval (0.3775, 0.6225). ② To pin a proportion within ±0.10 at 95% with no prior estimate: n = 0.25(1.96 ⁄ 0.10)² = 96.04 → 97 people.
Deeper review: Chapter 12's poll-anatomy walkthrough.
Week 13 — Hypothesis testing: foundations
Big idea. Every test is a courtroom. H₀ (nothing's going on — holds the equals sign, about a parameter) gets the benefit of the doubt; Hₐ (the suspicion, direction chosen before the data) must earn its verdict. The p-value = P(data at least this extreme | H₀ true); compare it to a pre-chosen α. Two verdicts only: reject, or fail to reject — and fail to reject is never accept.
Must-know terms: null and alternative hypotheses · one- vs. two-sided · test statistic · p-value · significance level α · statistically significant · Type I error (false alarm, probability α) · Type II error (miss) · practical significance.
Can you… write H₀/Hₐ from a claim and a suspicion? · state what a p-value is (and isn't)? · make the α-decision and phrase the conclusion in context? · name Type I and Type II in a scenario? · separate significant from important?
Classic traps: hypotheses about x̄ or p̂ (they're about μ and p); "the p-value is the probability H₀ is true" (it is not); "p = 0.32, so the claim is proven" (fail to reject ≠ accept); shrinking α raises the Type II risk; huge samples make trivial effects significant.
Mini-examples. ① A claim is tested and p = 0.02 comes back. At α = 0.05: 0.02 ≤ 0.05 → reject H₀ — and say it in context, with the direction the data leaned. ② Same p at α = 0.01: 0.02 > 0.01 → fail to reject — the verdict depends on the α chosen before the data.
Deeper review: Chapter 13's courtroom framing and the verdict-sentence templates.
Week 14 — Testing claims about means
Big idea. A claimed average gets the full trial: t = (x̄ − μ₀) ⁄ (s/√n) counts standard errors between data and claim; the t-table row df = n − 1 supplies the cutoff; the verdict is stated in context. Paired data (same individual, twice) means subtract first — a one-sample t on the differences. And the 95% CI is the list of every μ₀ a two-sided α = 0.05 test would let survive: inside ⟺ fail to reject.
Must-know terms: one-sample t-test · claimed value μ₀ · test statistic t · SE = s/√n · df = n − 1 · critical value · one- vs. two-sided alternatives · paired data · mean difference μd · test ↔ interval duality · conditions.
Can you… compute t from x̄, μ₀, s, n? · compare |t| to a stated cutoff and conclude? · halve a two-sided p-value for a pre-stated one-sided alternative (and only then)? · recognize paired data on sight? · use an interval to settle a two-sided test?
Classic traps: one-sided alternatives chosen after peeking; treating paired columns as two independent samples (it buries the signal); "reject because t is bigger than α" (t compares to t*, p compares to α); fail to reject ≠ the claim is true.
Mini-examples. ① A print shop claims lamination jobs average 20 minutes. From n = 16 random jobs: x̄ = 22.5, s = 4 → SE = 1 → t = (22.5 − 20) ⁄ 1 = 2.5. Against the df 15 two-sided cutoff 2.131: 2.5 > 2.131 → reject — convincing evidence the true mean isn't 20. ② Duality check: since the test rejected 20, a 95% interval from these data would not contain 20.
Deeper review: Chapter 14's claim → ruler → distance → cutoff → verdict chain.
Week 15 — Testing proportions & two-sample inference
Big idea. A claimed percentage gets the one-proportion z-test — SE built from p₀ (the null's world), z = (p̂ − p₀)/SE. Two groups' rates get the two-proportion z-test (pool, because H₀ itself claims one shared rate); two groups' means get the two-sample t-test (no pooling, df technology-reported); the same individuals twice is paired, back in Week 14. The map — proportions vs. means, then one sample / two samples / same-twice — is most of the battle.
Must-know terms: one-proportion z-test · hypothesized p₀ · large-count condition (np₀ ≥ 10, n(1−p₀) ≥ 10) · two-proportion z-test · pooled proportion · two-sample t-test · technology-reported df · independent samples vs. paired · statistical vs. practical significance · power (informal) · A/B test.
Can you… build the z statistic from p₀? · compute a pooled proportion? · read a two-sample output block (z or t, p-value) and conclude in context? · pick the right procedure from a scenario's shape? · say which designs earn causal language (randomized A/B) and which never do?
Classic traps: interval → p̂ but test → p₀ (the SE changes); averaging two rates instead of pooling (wrong unless n's are equal); "p = 0.078, so the treatments are proven equal" (fail to reject ≠ accept); significance without practical importance in huge samples.
Mini-examples. ① A hiking club claims 40% of members join the monthly group hike. A random sample: 75 of 150 joined (p̂ = 0.50). SE = √(0.40 × 0.60 ⁄ 150) = 0.04 → z = (0.50 − 0.40) ⁄ 0.04 = 2.5 → two-sided p = 2(1 − 0.9938) = 0.0124 → reject at 0.05: convincing evidence the true rate isn't 40%. ② Two branches' attendance: branch A, 36 of 90 members attend (0.40); branch B, 24 of 30 (0.80) → pooled = (36 + 24) ⁄ (90 + 30) = 0.50 — combined successes over combined sizes. (The simple average, 0.60, is wrong because the branches differ in size.)
Deeper review: Chapter 15's procedure map; the output-block reading drills in Tutorial 15.
Week 16 — Chi-square tests for categorical data
Big idea. A whole table of counts goes on trial at once: χ² = Σ(O − E)²/E — gap, squared, scaled by expectation, summed. Goodness-of-fit tests one variable against a claimed mix (E = n × claimed proportion, df = k − 1); independence tests whether two categorical variables are associated (E = row × column ⁄ total, df = (r−1)(c−1)). "Big" is defined only by the df-matched critical value, and when the alarm rings, the contributions say where.
Must-know terms: observed vs. expected counts · chi-square statistic · contribution · goodness-of-fit · test of independence · degrees of freedom · critical value · expected-count condition (every E ≥ 5) · counts, never percents.
Can you… compute expected counts from a claimed mix? · compute χ² for 3–4 categories? · pick the df? · compare to a stated critical value and phrase the verdict? · tell GoF from independence on sight? · rank trouble spots by contribution, not by count?
Classic traps: feeding χ² percentages (it eats raw counts); "expected counts must be whole numbers" (decimals are fine — they're long-run averages); df from the table's shape, never the sample size; observed zeros are legal, expected counts must be ≥ 5; fail to reject ≠ the claimed mix is proven.
Mini-examples. ① A bakery claims its four loaf types sell equally. One random sample of 100 sales: 22, 28, 25, 25 (expected 25 each). χ² = (9 + 9 + 0 + 0) ⁄ 25 = 0.72; df = 3, 5% critical value 7.815 → 0.72 < 7.815 → fail to reject: consistent with an even split (not proof of it). ② Biggest contribution: the 22-loaf and 28-loaf categories tie at 0.36 each — contributions, not raw counts, point the finger.
Deeper review: Chapter 16's candy-mix walkthrough; the chi-square mini-table (reprinted in the exam-prep tutorial).
Week 17 — Regression with inference + synthesis
Big idea. The least-squares line ŷ = a + bx earns its keep three ways: a slope read with all four parts (per one unit of x, predicted, on average, in y's units), an r² read as the share of variation explained, and residuals (actual − predicted) whose patternless plot is the model's all-clear. Inference puts the slope on trial with t = b ⁄ SE(b), df = n − 2. And the course's last skill: choose the right procedure — mean, proportion, counts, or a line?
Must-know terms: least-squares line · ŷ · slope · intercept · r² · residual · residual plot · extrapolation · population slope β · SE(b) · t-test and CI for the slope · LINE conditions · choosing the right procedure.
Can you… interpret a slope and (when sensible) an intercept? · predict from the line inside the fitted range — and refuse outside it? · compute a residual? · convert r to r² and say what it means? · run the slope test from an output block with df = n − 2? · route any scenario to its procedure?
Classic traps: r² as "percent accurate" (it's a share of variation); extrapolation (outside the data, the line is fiction); slope df = n − 2, not n − 1; a significant slope is still a handshake, not a push; interval-contains-zero ⟺ two-sided test fails to reject.
Mini-examples. ① The line ŷ = 10 + 3x predicts weekly sales (dozens) from ad spend (hundreds of dollars). At x = 4: ŷ = 10 + 12 = 22 dozen; if that week actually sold 25, the residual is 25 − 22 = +3 (above the line). With r = 0.7, r² = 0.49 — about half the variation in sales is explained by ad spend. ② Output shows slope b = 3.0 with SE(b) = 1.5 from n = 26 weeks: t = 3.0 ⁄ 1.5 = 2.0 against the df 24 cutoff 2.064 → 2.0 < 2.064 → fail to reject β = 0 — these data don't quite clear the bar, and that is not proof the slope is zero. ③ Say the four-part slope sentence for b = 3.0 out loud before moving on.
Deeper review: Chapter 17's decision map — then use it below.
The cross-week connections map
The final's synthesis questions live on these seams:
- W1 → everywhere: no formula rescues a biased sample. Conditions ("random sample…") are Week 1 refusing to retire.
- W3 → W8 → W10: the z-score ruler becomes the normal table's input, then gets re-scaled by √n for averages. Two rulers: σ for individuals, σ/√n for means.
- W4 → W16/W17: two-way tables grow into chi-square independence; scatterplots and r grow into regression with a slope on trial.
- W7 → W8: binomial histograms reaching for a bell is why a normal table can approximate count behavior.
- W11 ↔ W13/W14: the interval is the menu of plausible values; a two-sided test asks whether μ₀ is on the menu. Inside ⟺ fail to reject.
- W13 → W14/W15/W16/W17: one courtroom, four defendants — a mean, a proportion, a table of counts, a slope. Only the test statistic changes.
- The decision map (the final's favorite): name the answer's shape — mean (t), proportion (z), counts in categories (χ²), a line's slope (t with n − 2) — then count samples: one, two independent, or the same individuals twice (paired → subtract first).
How the final is graded + honest strategies
- 60 questions × 2 points = 120 points, one attempt, closed to AI, answer order shuffled, worth 5% of the course grade. Roughly 24 questions from Weeks 1–8 and 36 from Weeks 10–17. Every needed lookup value is printed in its question.
- Budget a minute per question and bank the leftovers: flag anything slow and return. All questions are visible, so harvest the easy ones first.
- Write before you look. For any computation, jot the formula and numbers before reading the options — the distractors are built from the classic errors (s for SE, missed "at least," un-halved p-values, df off by one).
- Read the last sentence twice. "Which is the population?", "at least", "NARROWER", "within their own league" — the question's final clause decides the answer.
- For interpretation items, eliminate any option that says "proven," promises something about individuals from a statement about a mean, or converts a p-value into "the probability the claim is true." Those three cuts usually leave one survivor.
- Don't cram the night before. The bundle in order — this guide, the practice exam, the exam-prep tutorial — is two focused hours. The final is 5%; your steady weekly work has already written most of the story.
Next stop: the practice exam (unlimited attempts, ungraded, zero shared items with the real final) — then the exam-prep tutorial, where your chatbot re-teaches whatever the practice exam exposed.