• Defining a function and calling it; the print-vs-return distinction.
• Conditionals: if (cond) { ... } else { ... }.
• Building a vector with c(); indexing and boolean masks.
• Combining masks with &, |, !.
• Subsetting a data frame by mask: df[df$x > 0, ].
• Applying a function to every element of a vector with sapply.
• Reading short pieces of code and saying what they compute.

• Memorizing function-definition or indexing syntax (it is on the reference sheet).
• Writing an algorithm in the sense of, e.g., sorting a vector by hand — you will only be asked to write "algebra in order."

• Knowing what each dplyr verb does and when to use it: filter, select, arrange, rename, mutate, group_by, summarize, distinct, n_distinct.
• Reading a pipeline written with %>% and saying what data comes out.
• Expect at least one question that uses these verbs.

• Memorizing the exact argument order of each verb — signatures are on the reference sheet.
• ggplot syntax. You will not be asked to write ggplot code.

• Conditional probability and independence.
• Bayes' rule and the law of total probability.
• Boy-and-girl-style conditional-probability puzzles.

• Anything beyond what was covered in the probability-review notes.

• Deriving the MLE for Bernoulli on paper.
• Finding the MLE numerically by grid search (writing the loop / sapply over a grid of parameters; which.max).
• Recognizing that linear regression is MLE under Gaussian noise (conceptually).

Nothing specifically excluded.

• Prior, likelihood, posterior; computing a posterior on a grid (prior × likelihood, then normalize).
• The coin example in code.
• The Bayesian comparison of two means (the finches setup).

Nothing specifically excluded.

• Stating a null hypothesis and a test statistic.
• Computing a p-value by simulation: build a null distribution with replicate(...); compute mean(sim_stat >= observed) for one-sided, the absolute-value version for two-sided.
• One-sided vs two-sided tests; what the 5% threshold means.
• The ASA statement on p-values, at the level we discussed.
• Type I / II / S / M errors — what they are, when to worry about each.
• Publication bias, p-hacking, the garden of forking paths, the dead-salmon study as worked examples.

Nothing specifically excluded.

• Writing an lm formula and saying what it fits.
• Interpreting coefficients, including the sign and magnitude.
• Indicator variables for categorical predictors and what it means to leave one category out as the baseline.
• When a continuous variable should be encoded as categorical (the gnarly question from the 2019 test).
• Reading summary(lm(...)): coefficients, standard errors, p-values, R-squared.
• Computing SSE for the model and for the baseline.

• Memorizing lm or predict syntax (on the reference sheet).

• Writing a glm(y ~ x, family = binomial, data = df) formula and interpreting its output (as in the sample test from the 2019 SML201 term test).
• Going from the linear predictor to a probability with plogis.
• Classification accuracy, the base-rate problem, FPR / FNR / PPV / NPV.
• Reading an ROC curve and picking a sensible operating point given a cost story (as in MP2's "doctor with 25 charts").
• Identifying assumptions in a model (e.g. the made-up failure-probability formula in MP2) and explaining in English why an assumption may not hold.

• Memorizing glm syntax — it is on the reference sheet.
• Plotting an ROC from scratch in ggplot.

• What a 95% CI actually means (and what it does not).
• Reading a CI off a summary(glm(...)) table using Estimate +/- 1.96 * Std. Error.
• The CI ↔ p-value duality.

• Looking up critical values for other confidence levels (use the 1.96 approximation).

• The assumptions (linearity, normality of residuals, independence).
• Reading the four plot(model) diagnostic plots and saying what each tells you.
• Identifying outliers and high-leverage points.
• Multiple comparisons and what the F-test is for.

• Producing the plots from scratch — you only need to read them.

• R-squared as a comparison to the constant-prediction baseline; what "close to 0" and "close to 1" mean.
• Reading R-squared off the Multiple R-squared line of summary(model).
• Correlation as the signed square root of R-squared (single-predictor case).
• The four alternatives to causation: reverse causation, common cause, indirect causation, coincidence.
• Anscombe's quartet: why visualization matters and how a single high-leverage point can dominate R-squared.

• Deriving the R-squared formula from first principles.

• Identifying which set is which and what each is for.
• Why training accuracy is misleading.
• Why one-hot / categorical encoding cannot hurt training fit.
• The index-sampling pattern for splitting a data frame — this is the one piece of code worth memorizing, because everything else falls out from the concept.
• Using replicate(...) to estimate validation-accuracy variability (and the variability of estimates more generally).

• k-fold cross-validation library functions (we did not cover them).

• Collinearity and what it does to coefficient estimates and their standard errors.
• Omitted-variable bias (Fisher and smoking; ice cream and drownings).
• Interpreting "the gender wage gap controlling for X" — what the control changes about the claim.
• The Harvard admissions lawsuit as a controls example.
• Errors in variables: what bias they create.
• Interactions (avocado / bacon) — what the interaction term is doing.

• Algebraic derivations of bias terms.

• The glm formula used in MP2; what each predictor means.
• Interpreting the fitted coefficients.
• Identifying training, validation, and test sets.
• Reading the ROC and reasoning about a threshold for the "doctor with 25 charts" cost model.
• Explaining in English why one of the model assumptions might not hold.

• The file-loading boilerplate from MP2 part 1 — not testable.
• Producing the cost-vs-threshold curve in ggplot.

• Nothing testable on the term tests — these support the course project.

• Arctic Shift API details, VS Code agent UI, scraping-library specifics.