MaizePrepSTATS 250
STATS 250 · Cheat sheet

Exam 3 reasoning sheet

Study-only reference, not an authorized exam aid. The official help card is provided at the exam. The supplied source folder contains Exam 1 and 2 cards, but no Exam 3 card; its exact formulas are not yet verified. All examples in this course are original. Lecture 15 is written from its lecture; Lectures 16–22 remain course-pack drafts until those lectures happen.

Help card versus what you must know

  • Verified on the earlier Exam 2 card: standard-error, interval, test-statistic, and effect-size formulas for one proportion, one mean, and independent mean differences; the course evidence and effect-size scales. This does not confirm the contents of the Exam 3 card.
  • Check the released Exam 3 card for the formulas and notation actually supplied. Until then, treat the formulas below as study material, not a promise of what will be printed.
  • You must supply the reasoning: choose the procedure, identify the population parameter, check design and model conditions, select the correct output row, keep units and subtraction order, and limit causal claims.
  • The exam allows the provided official help card and no notes. No calculator is the rule in the syllabus, lab slides and lecture, but the printed Exam 1 and 2 covers said calculators were not required yet allowed: confirm with the instructors. Practice reading supplied R output and doing short arithmetic or leaving exact fractions.

Procedure map

  • Quantitative response + three or more independent groups → ANOVA; which pairs differ → Tukey intervals.
  • Two quantitative variables → scatterplot, correlation, and simple regression; population linear association → slope t-test or interval.
  • Quantitative response + several predictors → multiple regression; categories → indicators; group-specific slopes → interactions.
  • Two categorical variables → conditional proportions to describe, chi-square to test independence. Relative risk compares two conditional outcome probabilities.
  • Foundation: one binary outcome → proportion inference; one quantitative outcome → one-mean t; two independent quantitative groups → Welch t; matched observations → t on differences.

ANOVA and Tukey

  • H0:μ1=⋯=μkH_0:\mu_1=\cdots=\mu_k; alternative: at least one population mean differs (always that sentence). FF = variation between the group means ÷ variation within the groups; lecture: F is never computed by hand, and only F and its p-value matter in the ANOVA table.
  • Define F in words: "the variation among the sample means was F times as large as the variation within groups." F is never negative; its distribution is right-skewed and the p-value P(F≥Fobs∣H0)P(F\ge F_{obs}\mid H_0) is always the right tail. Larger F, smaller p.
  • ANOVA is blind to which groups differ: a small p-value only says at least one mean differs. A boxplot's central line is a median, not a mean.
  • Tukey output: read lwr and upr. For B−A, entirely positive supports B>A; entirely negative supports B<A; containing 0 gives no evidence of a difference for that pair (it does not prove the means are equal).

Simple regression and inference

  • y^=β^0+β^1x\hat y=\hat\beta_0+\hat\beta_1x; residual e=y−y^e=y-\hat y. Positive residual means the model underpredicts.
  • R2=1−SSE/SSTR^2=1-SSE/SST measures explained response variation. In simple regression with an intercept, R2=r2R^2=r^2; recover the sign of r from the slope. RSE is in response units.
  • t=(β^1−β1,0)/SE(β^1)t=(\hat\beta_1-\beta_{1,0})/SE(\hat\beta_1); interval β^1±t∗SE(β^1)\hat\beta_1\pm t^*SE(\hat\beta_1); simple-regression residual df = n−2.
  • Check linearity, independence, constant error spread, and suitable error normality for inference. Avoid extrapolation. Unusual x implies leverage; actual change to the fit implies influence.

Multiple regression, categories, interactions

  • A multiple-regression slope holds other included predictors constant. With p slope coefficients plus an intercept, residual df = n−p−1. Count indicator and interaction coefficients too.
  • Adding predictors to nested least-squares models on the same data cannot reduce ordinary R-squared. Adjusted R-squared can fall; use it with RSE and substantive relevance. Correlated predictors can inflate individual coefficient SEs.
  • A k-level category needs k−1 indicators with an intercept. The omitted category is the reference. Recoding the reference changes coefficients, not predictions.
  • For y^=b0+b1x+b2IB+b3xIB\hat y=b_0+b_1x+b_2I_B+b_3xI_B: reference slope = b1b_1; B slope = b1+b3b_1+b_3; B−reference gap at x = b2+b3xb_2+b_3x. The interaction test concerns β3=0\beta_3=0.

Categorical tables and chi-square

  • Use the conditioning group's total as the denominator. Risk difference = π^A−π^B\hat\pi_A-\hat\pi_B; relative risk = π^A/π^B\hat\pi_A/\hat\pi_B. A 0.20 difference is 20 percentage points, not necessarily 20% higher risk.
  • E=(row total)(column total)/nE=(\text{row total})(\text{column total})/n; χ2=∑(O−E)2/E\chi^2=\sum (O-E)^2/E; df = (r−1)(c−1)(r-1)(c-1).
  • Course-pack condition: at least 80% of expected counts exceed 5 and none is below 1. Also check independent cases and a suitable random sampling design.
  • The null is population independence. The p-value is a right-tail probability under that null. A small value supports association, while conditional percentages reveal the pattern. For composed 2×2 Pearson examples here, correct = FALSE explicitly omits R's continuity correction.

Part II sentence discipline

  • Parameter: name the population quantity, variables, comparison order, and units. Coefficient: specify the reference group and held-fixed variables where relevant.
  • P-value: assume the null in context, state an observed-or-more-extreme outcome in the appropriate tail(s), and report its probability. Do not state the probability the null is true.
  • Conclusion: give the course evidence strength and the population association supported by the alternative. Above 0.10: not enough evidence; (0.05,0.10]: some; (0.01,0.05]: strong; at or below 0.01: very strong.
  • An interval estimates a population parameter, not individual outcomes. Random sampling supports generalization; random assignment supports causal inference. Adjustment alone cannot guarantee removal of all confounding.

MaizePrep is an independent study resource. It is not affiliated with, sponsored by, or endorsed by the University of Michigan. Course names and numbers are used only to describe which class the material covers. Every practice question is original, written in the style of the course.