about ECON31031
0 · Exam mapcourse.json · brief
“Students will be allowed to bring one hand written A4 sheet of notes into the exam room.” (ECON31031 course page, Canvas, 2026/27 1st semester)
Printed sheets are not allowed. Use side 1 as the plan for the permitted handwritten page; use both sides of your handwritten A4.
| Put on the handwritten page | Why |
|---|---|
| 1. Time series | two of three answers |
| 2. MLE + tests | shared machinery |
| 3. Binary calcs | reliable Section A work |
| 4. Matrix OLS | derivations + testing |
0b · Scientific-calculator workflowstudy strategy
- Write the model and assumptions before arithmetic.
- Keep full precision; round only the reported result.
- Mark known values, forecasts and unobserved shocks distinctly.
- End with one sentence answering the economic/statistical question.
Three answer skeletons
Derive: define the model/density and dimensions → state the assumptions → show the algebra using a named identity → box the estimator/result → say why it holds. Calculate: write the formula → substitute visible numbers → keep the denominator and units explicit → report a rounded value → interpret. Test: H₀/H₁ → statistic → reference distribution and df → critical value/decision → conclusion in the language of the question.
What earns space
| Keep | Compress |
|---|---|
| conditions changing the method | introductory prose |
| denominators + df | software commands |
| sign/scale interpretation | repeated definitions |
| one numerical check per family | derivations you can rebuild |
1 · Matrix OLS corech.1
model + estimatory = Xβ + ε
β̂ = (X′X)⁻¹X′y
With n observations and k regressors plus intercept: y,ε are n×1; X is n×(k+1); β is (k+1)×1. Gauss–Markov gives BLUE under correct linear specification, E(ε|X)=0, no perfect collinearity and homoscedastic uncorrelated errors.
mean, variance, precisionE(β̂|X)=β
Var(β̂|X)=σ²(X′X)⁻¹
s²=e′e/(n−k−1)
projection identitiesP=X(X′X)⁻¹X′; M=I−P
ŷ=Py; e=My=Mε; X′e=0
P′=P, P²=P; M′=M, M²=M; MX=0
tr(M)=n−k−1
Where it bites: derive unbiasedness/variance, prove s² unbiased, or use X′e=0 to catch hand-arithmetic errors.
1b · Fit and restrictionsch.1 FAQ
R²=1−RSS/TSS
F=[(RSSᵣ−RSSᵤ)/q] ÷ [RSSᵤ/(n−k−1)]
| Question | Statistic | Reference |
|---|---|---|
| one β restriction | t=(β̂ⱼ−βⱼ⁰)/se | tₙ₋ₖ₋₁ if normal |
| q joint restrictions | F above | Fq,n−k−1 if normal |
| large sample | qF | χ²q |
Write restrictions as Rβ=r. Ordinary R² never falls when regressors enter; adjusted R² can fall because it accounts for degrees of freedom.
Gauss–Markov line-by-line
Use a random cross-section and a model linear in β; require full column rank, E(ε|X)=0 and Var(ε|X)=σ²I. Then β̂ is unbiased and BLUE. Normal errors are an extra condition for exact finite-sample t/F inference; without normality, invoke large-sample normal/χ² results.
Restriction workflow
- Translate words into each row of Rβ=r.
- Estimate unrestricted model and record RSSᵤ.
- Impose/substitute restrictions; estimate and record RSSᵣ.
- Use q rows of R in F; RSSᵣ cannot be below RSSᵤ.
2 · Maximum likelihood recipech.2
- State one-observation density and support.
- Use independence: L(θ)=∏fᵢ(θ).
- Log: ℓ(θ)=Σ ln fᵢ(θ).
- Score: s(θ)=∂ℓ/∂θ; solve s(θ̂)=0.
- Curvature: H(θ̂)=∂²ℓ/∂θ∂θ′.
- Test restrictions and interpret.
large-sample covarianceVar̂(θ̂)=[−H(θ̂)]⁻¹
likelihood-ratio testLR=2[ℓ(θ̂ᵤ)−ℓ(θ̂ᵣ)] ~ χ²q
q = number of restrictions
Count restrictions, not parameters. A closed form exists only when the score equations solve algebraically; nonlinear index models normally require numerical optimisation.
2b · Two benchmark likelihoodsch.2
Bernoulliℓ(p)=Σ[yᵢ ln p+(1−yᵢ)ln(1−p)]
p̂=ȳ; Var̂(p̂)=p̂(1−p̂)/n
normal linear modelβ̂ML=β̂OLS
σ̂²ML=RSS/n; s²OLS=RSS/(n−k−1)
Both variance estimators are consistent; only the OLS version is unbiased in finite samples.
What each derivative supplies
| Object | Use | Exam line |
|---|---|---|
| L(θ) | sample probability/density | product by IID |
| ℓ(θ) | workable objective | sum of log densities |
| score | first-order conditions | set every component to zero |
| Hessian | curvature/precision | invert its negative |
| ℓᵤ,ℓᵣ | joint test | LR=2(ℓᵤ−ℓᵣ) |
For binary data, the intercept-only restricted model has fitted probability equal to the sample proportion. This gives ℓᵣ directly and reappears in likelihood-based pseudo-R² measures. A sharp peak means larger negative curvature and a smaller covariance; a flat likelihood means imprecision.
3 · Binary model chooserch.3
| Model | pᵢ=Pr(yᵢ=1|xᵢ) | Key exam point |
|---|---|---|
| LPM | xᵢ′β | easy OLS; invalid p possible; constant effects; robust SE |
| Logit | Λ(z)=eᶻ/(1+eᶻ) | log[p/(1−p)]=z; OR=eᵝ |
| Probit | Φ(z) | normal table; latent threshold |
LPM: Var(εᵢ|xᵢ)=pᵢ(1−pᵢ), so heteroscedasticity is built in. White robust covariance repairs inference, not out-of-range predictions or constant effects.
latent probityᵢ*=xᵢ′β+εᵢ; yᵢ=1[yᵢ*>0]
εᵢ~N(0,1) ⇒ Pr(yᵢ=1)=Φ(xᵢ′β)
Only the threshold side is observed, so error variance must be fixed. This scale normalisation explains why logit and probit coefficients differ even when fitted probabilities are close.
3b · Probability, odds, interpretationch.3 FAQ
odds=p/(1−p)=eᶻ; p=odds/(1+odds)
one-unit xⱼ rise ⇒ odds × eᵝʲ
For a dummy, compare odds at xⱼ=1 and 0 holding other regressors fixed. Never multiply the probability by eᵝ.
GLM and response type
The generalised-linear view changes the link: identity for LPM, logit for logistic probabilities, inverse-normal for probit. Binary outcomes have two unordered values; ordered responses rank categories without equal spacing; multinomial responses choose among several exclusive alternatives.
Limited dependent variables
| Observed rule | Model idea |
|---|---|
| only which side of threshold | binary response |
| threshold recorded below limit | censoring |
| observations below limit absent | truncation |
Probit by hand: calculate z=x′β, then read Φ(z). For z<0 use symmetry, Φ(−a)=1−Φ(a). Scale warning: logistic error variance is π²/3 while standard normal variance is 1; coefficients are not directly comparable. In the course example, logit coefficients are roughly 1.6–1.8 times probit coefficients, while probability effects are close.
4 · Marginal effectsch.4
continuous regressorLPM: MEⱼ=β̂ⱼ
Logit: MEⱼ=p̂(1−p̂)β̂ⱼ
Probit: MEⱼ=φ(x′β̂)β̂ⱼ
dummy regressorΔp=F(x′β̂; xⱼ=1)−F(x′β̂; xⱼ=0)
The density is positive, so effect sign matches coefficient sign. For logit, p(1−p) peaks at .25 when p=.5; effects shrink near 0 or 1. Report probability changes in percentage points and scale them to the change asked.
| Summary | Calculation | Say |
|---|---|---|
| effect at average | ME at x̄ | effect for mean profile |
| average effect | n⁻¹ΣMEᵢ | mean individual effect |
| dummy | exact p(1)−p(0) | not derivative approximation |
Use the delta method to transform Var(β̂) into an effect’s standard error. Odds ratios are logit-specific and constant across individuals; probability changes depend on the starting index.
5 · Binary-model assessmentch.5
At cut-off c: predict 1 when p̂≥c. In the hit/miss table let TP=n₁₁, FP=n₁₀, FN=n₀₁, TN=n₀₀.
| Measure | Formula |
|---|---|
| hit count | (TP+TN)/n |
| recall / TPR | TP/(TP+FN) |
| FPR | FP/(FP+TN) |
| Kuipers | TPR−FPR |
| precision | TP/(TP+FP) |
| F1 | 2·precision·recall/(precision+recall) |
Brier=n⁻¹Σ(yᵢ−p̂ᵢ)² [lower better]
McFadden R²=1−ℓᵤ/ℓᵣ
A high hit count misleads with imbalance. Prefer precision/recall/F1 and the ROC: plot TPR against FPR across cut-offs; AUC uses trapezia. Guide bands: .5–.7 poor; .7–.8 fair; .8–.9 good; >.9 excellent.
phi coefficientφ=(TP·TN−FP·FN)/√[(TP+FP)(FN+TN)(TP+FN)(FP+TN)]
In-sample fitted values describe the estimation sample; out-of-sample predictions evaluate observations not used in estimation. Do not call in-sample fit evidence of forecast performance.
6 · Stationarity + operatorsch.6
Lyₜ=yₜ₋₁; Lʰyₜ=yₜ₋ₕ
Δyₜ=(1−L)yₜ=yₜ−yₜ₋₁
seasonal s: Δₛyₜ=(1−Lˢ)yₜ
Weak stationarity: E(yₜ)=μ finite and constant; Var(yₜ)=γ₀ finite and constant; Cov(yₜ,yₜ₋ₖ)=γₖ depends only on lag k. White noise has mean 0, constant variance σ² and zero autocorrelation for k≠0; it need not be normal.
sample autocorrelationrₖ=Σₜ₌ₖ₊₁ᵀ(yₜ−ȳ)(yₜ₋ₖ−ȳ) / Σₜ₌₁ᵀ(yₜ−ȳ)²
Same denominator at every lag. Slow ACF decay warns of strong persistence or nonstationarity; do not apply stationary formulas until diagnosed.
7 · AR/MA/ARMA identificationch.7
AR(1): yₜ=α+φyₜ₋₁+εₜ|φ|<1; μ=α/(1−φ)
Var(y)=σ²/(1−φ²); ρₖ=φᵏ
MA(q): yₜ=α+εₜ+Σθⱼεₜ₋ⱼE(y)=α; Var(y)=σ²(1+Σθⱼ²)
MA(1): ρ₁=θ/(1+θ²); ρₖ=0 for k>1
AR(2) stationarity: φ₁+φ₂<1; φ₂−φ₁<1; φ₂>−1. Equivalently check all characteristic roots have absolute value below one. MA terms need invertibility.
8 · Forecast, select, differencech.8–9
AR(1): ŷT+h|T=μ+φʰ(yT−μ)
MA(1): h=1 uses θêT; h≥2 → μ
AIC=−2ℓ+2K; BIC=−2ℓ+K ln T
Feed forecasts back when extrapolating; if actual out-of-sample values arrive, use them. Minimise AIC/BIC; BIC penalises dimension more heavily once T exceeds about seven.