EKN 325 Econometrics 325 Course Notes

Econometrics is the course where economics meets statistical evidence. EKN 325 (“Econometrics 325”) typically develops the core tools needed to estimate relationships, test hypotheses, and forecast using real-world data. These notes are designed to help you master the logic of econometric modeling—from assumptions and identification to interpretation, diagnostics, and applied implementation—within a South African university context, with emphasis on the kind of problems you are likely to face in assessments.

This study guide is framed around the typical teaching outcomes of EKN 325 in the University of Pretoria (UP) BCom Economics Study Material collection. It provides conceptually rigorous coverage plus practical examples that mirror common applied topics: demand and supply, labour economics, macro relationships, and policy evaluation using observational data. Wherever feasible, worked examples and “what to check in your answer” heuristics are included to strengthen exam performance.

1. Econometrics Foundations for EKN 325 (Model Thinking, Assumptions, and Notation)

Econometrics begins before you run regressions. The exam often rewards students who can clearly translate a real question into a statistical model and justify why the assumptions are plausible.

1.1 What Econometrics Does (and What It Cannot Do)

In econometrics, we build a statistical model connecting variables through parameters and randomness. A typical starting point is:

Population (true) model:
[
Y_i = \beta_0 + \beta_1 X_{1i} + \beta_2 X_{2i} + \cdots + u_i
]
Sample model (estimation): we use data ((Y_i, X_i)) to estimate the (\beta)'s and then evaluate hypotheses using the error term behavior.

Key exam idea: econometrics helps with estimation, testing, and forecasting—but inference depends crucially on assumptions about the error term (u_i) (and about the relationship between (X) and (u)).

Estimation vs Prediction

Estimation goal: learn about parameters ((\beta)'s) as features of the population.
Prediction goal: produce good out-of-sample forecasts, which can rely less strictly on causal interpretation—though assumptions still matter.

In many South African econometrics exam questions, “interpret the coefficient” or “discuss the validity of the model” is as important as computing the standard error.

1.2 Variables, Indices, and Dimensions

You’ll often see:

(i = 1,2,\dots,n) indexing observations (households, firms, months, students).
(Y_i) as the dependent variable.
(X_i) as regressors (explanatory variables).
(\beta) parameters (unknown constants).
(u_i) unobserved influences (the error term).

A common notation pattern:

Scalar regressor: (X_{1i})
Vector of regressors: (X_i = (1, X_{1i}, X_{2i}, …, X_{ki})) where the leading 1 gives (\beta_0).

1.3 The Error Term and Why It Matters

The error term (u_i) is not “measurement noise only.” It captures every factor affecting (Y_i) that is not included in (X). This means:

If an omitted variable is correlated with included regressors, (u_i) becomes correlated with (X_i), violating assumptions required for unbiased OLS.

1.4 Classical Assumptions for OLS (Causal Language vs Statistical Language)

EKN 325 usually emphasizes the Gauss–Markov conditions and related assumptions for inference. A standard list includes:

Linearity in parameters: The model is linear in (\beta).
Random sampling / independence of observations: typically (i) are i.i.d. (or weakly dependent depending on the topic).
Zero conditional mean (exogeneity):
[
\mathbb{E}[u_i \mid X_i] = 0
]
No perfect multicollinearity: regressors are not exact linear combinations of others.
Homoskedasticity (for standard OLS inference):
[
\text{Var}(u_i \mid X_i) = \sigma^2
]
No autocorrelation (for time series contexts):
[
\text{Cov}(u_i, u_j \mid X) = 0 \quad \text{for } i\neq j
]

When these assumptions fail, OLS may still be consistent under some forms of endogeneity correction, or still unbiased but with wrong standard errors, or biased and inconsistent.

Exam phrasing that scores

“OLS remains unbiased if exogeneity holds, but standard errors are wrong if heteroskedasticity exists.”
“If (\mathbb{E}[u_i|X_i]\neq 0), then OLS estimates can be biased and inconsistent.”

1.5 Simple OLS Derivation and Interpretation

For the simple regression:
[
Y_i = \beta_0 + \beta_1 X_i + u_i
]
OLS chooses (\hat{\beta}_0, \hat{\beta}1) to minimize:
[
\sum{i=1}^n (Y_i – \beta_0 – \beta_1 X_i)^2
]

The slope estimator:
[
\hat{\beta}1 = \frac{\sum{i=1}^n (X_i-\bar{X})(Y_i-\bar{Y})}{\sum_{i=1}^n (X_i-\bar{X})^2}
]
The intercept:
[
\hat{\beta}_0 = \bar{Y} – \hat{\beta}_1 \bar{X}
]

Interpreting coefficients

Suppose in a labour economics context:

(Y_i) = monthly wage (ZAR)
(X_i) = years of experience

Then (\hat{\beta}_1) approximates “the average change in monthly wage (ZAR) associated with one additional year of experience,” if exogeneity holds and the model is correctly specified.

1.6 Functional Form: Levels, Logs, and Elasticities

A major exam skill is mapping coefficient interpretation to the chosen functional form.

Common cases:

Level–level:
[
Y = \beta_0 + \beta_1 X + u
]
(\beta_1): change in (Y) for a one-unit increase in (X).
Log–level:
[
\ln(Y) = \beta_0 + \beta_1 X + u
]
(\beta_1): approximate % change in (Y) for 1-unit increase in (X):
[
100 \times \beta_1
]
more precisely: (100 \times (e^{\beta_1}-1)).
Level–log:
[
Y = \beta_0 + \beta_1 \ln(X) + u
]
(\beta_1): change in (Y) for a 1% increase in (X), approximately:
[
\Delta Y \approx \beta_1 \times 0.01
]
Log–log:
[
\ln(Y) = \beta_0 + \beta_1 \ln(X) + u
]
(\beta_1): elasticity—% change in (Y) for 1% change in (X).

Mini-example with numbers (for exam readiness)

Assume:
[
\ln(\text{Wage}) = 1.50 + 0.08 \ln(\text{Experience}) + u
]
Then (0.08) is the elasticity:

If experience increases by 10%, predicted wage increases by about 0.08 × 10% = 0.8%.

Be careful: elasticity interpretation requires log–log.

1.7 From Theory to Testable Statements

EKN 325 usually asks you to translate a theory claim into a hypothesis:

Theory: “Education increases earnings.”
Model: ( \ln(\text{earnings}) = \beta_0 + \beta_1 \text{education} + u )
Hypothesis:
[
H_0: \beta_1 = 0 \quad \text{vs} \quad H_1: \beta_1 \neq 0
]

Often, multiple regression questions appear:
[
Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + u
]
Then you may test:

(H_0: \beta_2 = 0) controlling for (X_1)
or joint restrictions like (H_0: \beta_1=\beta_2=0)

Key Takeaway Checklist (use in exams)

Define variables and units clearly.
Specify model form (levels/logs) and interpret coefficients accordingly.
State the exogeneity assumption (\mathbb{E}[u|X]=0).
Identify whether the question asks about unbiasedness, consistency, hypothesis testing, or prediction.
Link theory claims to hypotheses about (\beta) parameters.

2. OLS Estimation, Hypothesis Testing, and Diagnostic Reasoning (The Core EKN 325 Engine)

This section develops the heart of first econometrics: OLS estimation and the logic behind tests and diagnostics. Many EKN 325 questions are “mechanical” but the marks depend on your reasoning and interpretation.

2.1 Multiple Regression Model and Matrix Form

The general multiple regression:
[
Y_i = \beta_0 + \beta_1 X_{1i} + \beta_2 X_{2i} + \dots + \beta_k X_{ki} + u_i
]
Stacking observations yields:
[
\mathbf{y} = \mathbf{X}\boldsymbol{\beta} + \mathbf{u}
]

OLS estimator:
[
\hat{\boldsymbol{\beta}} = (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{y}
]
Interpretation:

(\hat{\beta}_j) estimates the partial association between (X_j) and (Y), holding other regressors constant.

2.2 Properties of OLS Under Gauss–Markov

If exogeneity holds and assumptions are satisfied:

OLS is unbiased: (\mathbb{E}[\hat{\beta}] = \beta).
OLS is consistent as (n\to\infty).
OLS is efficient among linear unbiased estimators (Gauss–Markov theorem).

But if assumptions fail, the properties change:

Heteroskedasticity: coefficients still unbiased, but variance formula wrong.
Endogeneity: coefficients biased/inconsistent.

2.3 Standard Errors, t-Statistics, and Confidence Intervals

Given:

(\hat{\sigma}^2 = \frac{1}{n-k-1}\sum_{i=1}^n \hat{u}_i^2)
Under homoskedasticity:
[
\widehat{\text{Var}}(\hat{\boldsymbol{\beta}})=\hat{\sigma}^2(\mathbf{X}'\mathbf{X})^{-1}
]

The t-statistic for testing a single coefficient:
[
t = \frac{\hat{\beta}j – \beta{j,0}}{\text{SE}(\hat{\beta}_j)}
]
Compare with critical values or compute a p-value.

Confidence interval:
[
\hat{\beta}j \pm t{\alpha/2,,n-k-1}\cdot \text{SE}(\hat{\beta}_j)
]

What graders want

In written responses:

State (H_0) and (H_1)
Compute (t) or reference a given t-statistic
Decide at a specified significance level (e.g., 5%)
Interpret in context

2.4 Joint Hypothesis Testing: F-Tests

For multiple restrictions, e.g.:

Test whether two variables matter jointly:
[
H_0:\beta_1 = 0 \quad \text{and}\quad \beta_2=0
]
Use an F-statistic:
[
F = \frac{(SSR_r – SSR_{ur})/q}{SSR_{ur}/(n-k-1)}
]
where:
(SSR) = sum of squared residuals
(r) = restricted model
(ur) = unrestricted model
(q) = number of restrictions

Alternatively, use R-squared and sample sizes, but SSR-based computations are usually clearer.

Concrete example (labour wage model)

Suppose you estimate:
[
\ln(wage_i)=\beta_0+\beta_1 educ_i+\beta_2 exp_i+\beta_3 tenure_i+u_i
]
You test:

(H_0:\beta_2=\beta_3=0) meaning “experience and tenure have no joint effect on log wage after controlling for education.”

You compute restricted model without exp and tenure, compare SSRs, and compute F.

Interpretation requires careful language:

“Rejecting (H_0)” means evidence that at least one of the restricted coefficients is non-zero.

2.5 Interpreting (R^2) and Adjusted (R^2)

(R^2):
[
R^2 = 1-\frac{SSR}{SST}
]
It measures fraction of explained variation in (Y), but note:

Adding regressors generally increases (R^2) even if they’re irrelevant.
Adjusted (R^2) penalizes for extra regressors:
[
\bar{R}^2 = 1 – (1-R^2)\frac{n-1}{n-k-1}
]

In exams, if a question asks “how does adding variables affect model fit,” you should mention adjusted (R^2) and the danger of overfitting.

2.6 Diagnostic Tests: Heteroskedasticity, Autocorrelation, and Multicollinearity

(a) Heteroskedasticity: Breusch–Pagan and White Tests (Conceptual)

Heteroskedasticity means:
[
\text{Var}(u_i|X_i)=\sigma_i^2
]
Common consequences:

OLS coefficients remain unbiased under exogeneity.
Standard errors are wrong.
t and F tests become unreliable.

Breusch–Pagan test (conceptually):

Estimate OLS and obtain residuals (\hat{u}_i).
Regress (\hat{u}_i^2) on the regressors (or on subset).
Test whether coefficients explain (\hat{u}_i^2).

White’s test is more general, including squared terms and cross products.

(b) Remedies: Robust Standard Errors

A typical exam response:

Use heteroskedasticity-robust (White/HC) standard errors to correct inference.
If the question gives robust SE output, you must interpret significance using those.

(c) Autocorrelation: Durbin–Watson and More

Autocorrelation often arises in time series:
[
u_t = \rho u_{t-1} + \epsilon_t
]

Breaches independence assumption.
Standard errors wrong, test sizes distorted.

Durbin–Watson:
[
DW \approx 2(1-\hat{\rho})
]
Interpretation:

(DW \approx 2): no autocorrelation
(DW < 2): positive autocorrelation
(DW > 2): negative autocorrelation

(d) Multicollinearity: What It Does (and Doesn’t Do)

Multicollinearity implies high correlation among regressors.

Effects:

Inflates variances of coefficient estimates.
Leads to unstable estimates: large SEs, insignificant t-tests, even when variables may be important.

Important clarification:

Multicollinearity does not bias OLS coefficients if exogeneity holds.
It reduces precision, not consistency.

How to check

Correlation matrix of regressors
VIF (variance inflation factor)
[
VIF_j = \frac{1}{1-R_j^2}
]
Rule-of-thumb guidance often used:
VIF > 10 suggests serious multicollinearity (thresholds vary).

2.7 Residual Analysis and Model Adequacy (Qualitative Diagnostics)

Beyond formal tests, residual plots matter:

Residual vs fitted values: look for patterns (fan shape implies heteroskedasticity).
Residual vs each regressor: look for systematic structure (nonlinearity or omitted variables).
Outliers and leverage: points with large leverage can dominate fit.

An exam question may provide scatterplots or residual diagrams and ask:

“What does this indicate about assumptions?”
“What would you do next?”

Your answer should align with observed patterns:

If residual spread increases with fitted values: heteroskedasticity.
If residuals show curvature: missing nonlinear terms.
If residuals correlate with a missing variable: omitted variable bias risk.

2.8 Endogeneity, Omitted Variable Bias, and Consequences for Inference

A central econometrics concept is omitted variable bias. Consider:

True model:
[
Y = \beta_1 X_1 + \beta_2 X_2 + u
]
But you estimate:
[
Y = \alpha_1 X_1 + v
]
where (v = \beta_2 X_2 + u).

If (X_2) is correlated with (X_1), then:
[
\mathbb{E}[v|X_1] \neq 0
]
leading to biased and inconsistent (\hat{\alpha}_1).

A classic verbal example (typical exam style):

Wage model includes education but omits ability.
Ability affects wage and is correlated with education → endogeneity.

Key Takeaway Checklist for Section 2

Coefficients depend on exogeneity (\mathbb{E}[u|X]=0).
Use t-tests for single coefficients; F-tests for joint restrictions.
Use adjusted (R^2) when comparing models with different regressors.
Diagnose assumptions: heteroskedasticity, autocorrelation, multicollinearity.
Interpret results carefully (significance vs practical meaning).

3. Econometric Modeling with Qualitative Variables, Interactions, Nonlinearity, and Applied Interpretation (What Exam Questions Love)

After mastering OLS mechanics, EKN 325 typically expands into model variants: dummy variables, interaction terms, nonlinear functional forms, and careful interpretation.

3.1 Dummy Variables and the Meaning of Coefficients

Dummy variables (binary indicators) take values 0 or 1.

Example: citizenship status

(D_i=1) if citizen, (0) otherwise
Model:
[
Y_i = \beta_0 + \beta_1 D_i + u_i
]

Then:

If (D_i=0): expected (Y = \beta_0)
If (D_i=1): expected (Y = \beta_0 + \beta_1)
(\beta_1) is the difference in expected outcome between the groups.

Log outcomes with dummies

If:
[
\ln(Y_i) = \beta_0 + \beta_1 D_i + u_i
]
Then (\beta_1) approximates the percentage difference in (Y) between groups:
[
100 \times \beta_1 % \text{ (approx.)}
]
More exactly:
[
100 \times (e^{\beta_1}-1)
]

3.2 Interactions: “The Effect Depends on Another Variable”

Interactions let you model heterogeneous effects.

Suppose you suspect education benefits depend on gender:
[
Y = \beta_0 + \beta_1 educ + \beta_2 female + \beta_3 (educ\times female) + u
]
Then:

For males (female=0): effect of educ is (\beta_1)
For females (female=1): effect of educ is (\beta_1+\beta_3)

Interpreting (\beta_3)

(\beta_3) measures how education’s effect differs for females relative to males.

Exam advice: Always state group-specific marginal effects, not only coefficient meanings.

3.3 Nonlinearity and Polynomial Terms

A common modeling choice:

Include (X) and (X^2) to capture curvature:
[
Y = \beta_0 + \beta_1 X + \beta_2 X^2 + u
]

Interpretation:

If (\beta_2 < 0): inverted U-shape (increasing then decreasing)
Marginal effect:
[
\frac{\partial Y}{\partial X} = \beta_1 + 2\beta_2 X
]

You may compute the turning point:
[
X^* = -\frac{\beta_1}{2\beta_2}
]
Only meaningful within the data range.

3.4 Piecewise Linear Models (Segmented Relationships)

Sometimes exam questions provide a structural break or threshold effect: effect changes after a cutoff (c).

A piecewise specification:
[
Y = \beta_0 + \beta_1 X + \beta_2 (X-c)D + u
]
where:
[
D =
\begin{cases}
1 & X>c \
0 & X\le c
\end{cases}
]

Then:

For (X \le c): slope is (\beta_1)
For (X > c): slope becomes (\beta_1+\beta_2)

This can be discussed as “policy thresholds” or “experience above a certain level changes returns.”

3.5 Model Selection: Choosing Functional Form and Variables

A recurring exam theme is: “Is the model appropriate?” or “Suggest modifications.”

Model selection tools often include:

Compare models with different specs using adjusted (R^2) and tests.
Use information criteria (if covered): AIC/BIC.
Add theoretical variables; avoid purely mechanical significance chasing.

Counterargument (to avoid mark loss)

Just because a variable is insignificant does not mean it should be removed if theory suggests it matters and omission would bias other coefficients. Conversely, significant variables can still be misleading if endogeneity exists.

3.6 Multicollinearity in Interaction and Polynomial Models

Interactions and polynomials increase the chance of multicollinearity.

Example:

Including (X), (X^2), and (X^3) yields strong correlation among regressors.
Interpret coefficients becomes tricky; centering can help.

Centering:

Replace (X) with (X-\bar{X}) before forming powers.
This changes coefficient interpretation but not fitted values (if done consistently).

Exam tip: If asked why coefficients’ signs or magnitudes change when adding polynomial terms, mention multicollinearity and re-interpret marginal effects rather than raw coefficients.

3.7 Applied Case Study (South African Context): Education, Labour Outcomes, and Dummies

Consider a simplified labour market model using cross-sectional survey data:

(Y_i = \ln(wage_i))
(X_{1i} = years_education_i)
(X_{2i} = age_i)
(D_i = 1) if the worker is female

Model:
[
\ln(wage_i) = \beta_0 + \beta_1 educ_i + \beta_2 age_i + \beta_3 female_i + u_i
]

Interpretations:

(\beta_1): elasticity-like interpretation if educ is years (not log). Usually log–level: a 1-year increase in education increases log wage by (\beta_1), approximating (100\beta_1%) increase in wage.
(\beta_3): difference in log wage between females and males, holding education and age constant.

If an exam asks “what does (\beta_3) mean?” you must state:

expected wage difference between female and male groups.
and specify it is conditional on education and age.

Adding an interaction (heterogeneous education returns)

If you estimate:
[
\ln(wage_i) = \beta_0 + \beta_1 educ_i + \beta_2 female_i + \beta_3 (educ_i \times female_i) + u_i
]
Then:

Male return to education: (\beta_1)
Female return to education: (\beta_1+\beta_3)

This sort of problem matches how applied econometrics questions test understanding beyond computation.

3.8 Interpreting Causality vs Association

Even when dummy and interactions are correct, causal interpretation requires exogeneity:

If gender is used as a regressor, exogeneity may be plausible in conditional wage comparisons, but it still depends on omitted variables like job type, discrimination, and productivity.
If education is endogenous (e.g., ability), causality is questionable without instruments.

In exam contexts, always separate:

“The regression shows an association.”
“Causal interpretation requires further assumptions or identification strategies.”

Key Takeaway Checklist for Section 3

Dummy variables: interpret group differences.
Interactions: interpret conditional/marginal effects.
Polynomial/threshold models: interpret marginal effects and turning points.
Model selection: align with theory; avoid omission bias.
Separate causal claims from statistical association.

4. Multiple Regression Advanced Ideas: Endogeneity, Instrumental Variables Logic, and (If Included) 2SLS Intuition

Many EKN 325 offerings progress into endogeneity and instrumental variables (IV) intuition, sometimes covering 2SLS. Even when the exam focuses mainly on OLS, you must know the logic to explain what goes wrong and what remedies exist.

4.1 Why Endogeneity Happens in Econometrics

Endogeneity means:
[
\mathbb{E}[u_i \mid X_i] \neq 0
]
Common causes:

Omitted variable bias: missing relevant factors correlated with regressors.
Reverse causality: (Y) influences (X).
Measurement error: regressors observed with error leading to attenuation bias.
Simultaneity: simultaneous determination of multiple variables (e.g., supply and demand).

Exam prompts often give a scenario and ask:

“Identify the source of endogeneity.”
“What assumption is violated?”

4.2 Omitted Variable Bias Formula (Conceptual Magnitude)

For a single omitted variable (Z) correlated with included regressor (X), a common derivation yields:
[
\text{Bias}(\hat{\beta}_1) = \beta_2 \cdot \frac{\text{Cov}(X,Z)}{\text{Var}(X)}
]
where:

(\beta_2) is effect of omitted (Z) on (Y) in the true model.

The sign depends on:

(\beta_2) sign
and (\text{Cov}(X,Z)) sign.

An exam can ask you to predict the direction of bias without calculations.

Example (direction of omitted variable bias)

(Y=) wage
(X=) education
omitted (Z=) ability
Suppose ability increases wage ((\beta_2>0)) and higher ability correlates with higher education ((\text{Cov}(X,Z)>0))

Then bias is positive: OLS overstates the effect of education.

4.3 Instrumental Variables: Core Requirements

IV uses a variable (Z_i) (an instrument) to recover the causal effect when (X) is endogenous.

Assume model:
[
Y = \beta X + u
]
Endogeneity: (\text{Cov}(X,u)\neq 0)

Instrument (Z) must satisfy:

Relevance:
[
\text{Cov}(Z,X)\neq 0
]
The instrument must meaningfully affect (X).
Exogeneity (validity):
[
\mathbb{E}[u \mid Z]=0
]
The instrument affects (Y) only through (X), not directly or through omitted variables.

Interpretation: Local variation

Even if IV is valid, it estimates a local average treatment effect conceptually (depending on framework). In general, IV identifies variation in (X) induced by (Z).

4.4 Two-Stage Least Squares (2SLS) Intuition

In the endogeneity case with one endogenous regressor (X), IV/2SLS works via two stages:

Stage 1: predict (X) using (Z):
[
X_i = \pi_0 + \pi_1 Z_i + w_i
]
Obtain predicted values (\hat{X}_i).

Stage 2: regress (Y) on predicted (\hat{X}):
[
Y_i = \beta_0 + \beta_1 \hat{X}_i + e_i
]

Intuition:

(\hat{X}) contains the component of (X) explained by (Z).
If (Z) is valid, that component is uncorrelated with the error term (u).

4.5 Testing Instrument Strength: The First-Stage F

In practical econometrics, weak instruments lead to bias and large standard errors.

A common rule-of-thumb (not a strict theorem threshold):

First-stage F-statistic should be sufficiently large (often cited: >10).

Exam questions might show a first-stage output and ask:

“Are the instruments weak?”
“What is the implication?”

Weak instruments:

cause unreliable IV estimates
can produce estimates with wrong size and inferential breakdown.

4.6 Overidentification (If Multiple Instruments Are Present)

If you have more instruments than endogenous regressors, you can test overall validity using an overidentification test (often Sargan/Hansen J).

Null hypothesis: instruments are valid.
If rejected: some instrument likely violates exogeneity.

Exam question style:

“You have 2 instruments for 1 endogenous regressor. Should you test validity?”
Answer: yes—overidentification checks can assess instrument set validity (assuming appropriate assumptions).

4.7 Applied Case Study: Education and Wage with Ability Bias (Instrument Idea)

Suppose you want the causal effect of education (X=\text{education}) on wage (Y=\ln(wage)), but ability makes education endogenous.

You propose instrument (Z): an exogenous policy variation or distance to a resource that affects education but not wage directly (except through education).

A plausible (not necessarily true) example in South African contexts:

Variation in access to schooling due to the historical placement of schools or changes in funding affecting attendance.

A strong exam answer:

specify relevance: instrument affects schooling years.
specify exogeneity: instrument has no direct effect on wage once education and other controls included (and no correlation with ability).

Then show the IV logic:

OLS confounds education and ability.
IV isolates education variation due to the instrument.

4.8 Counterarguments and Risks in IV Applications

Even with two IV assumptions, they are hard to defend.

Common pitfalls:

Direct effect: instrument affects (Y) through channels other than (X).
Exclusion violation: instrument correlated with unobserved determinants (like ability, motivation).
Weak instruments: small correlation between (Z) and (X).

Exam answers should include:

statement of two IV assumptions
argument or checklist for plausibility
caution that invalid instruments make IV inconsistent.

Key Takeaway Checklist for Section 4

Endogeneity breaks OLS exogeneity.
Instrument validity requires relevance and exogeneity.
2SLS uses predicted endogenous regressor variation from instruments.
First-stage strength matters (weak instruments).
Overidentification tests can assess instrument set validity.
Provide a scenario-based argument, not only formulas.

5. Full Exam Skill: Putting It Together for EKN 325 (Problem-Solving Templates, Interpretation, and Data-Driven Reasoning)

This final section is about performance: how to structure answers, avoid common mistakes, and handle typical EKN 325 exam question types. It integrates earlier concepts into repeatable templates.

5.1 How to Structure Short Answer Questions

Many EKN 325 exams include short theoretical questions. A high-scoring structure:

State the assumption or concept (e.g., exogeneity, homoskedasticity).
Explain consequences (e.g., unbiasedness, wrong SEs, inconsistency).
Mention remedy if asked (e.g., robust SEs, IV, include omitted variable proxy, transform variables).

Example template:

Question: “What happens to OLS t-tests under heteroskedasticity?”
Answer structure:
- exogeneity holds ⇒ coefficients unbiased
- heteroskedasticity ⇒ wrong SE formula
- t-tests use wrong variance ⇒ inference unreliable
- remedy: heteroskedasticity-robust SE

5.2 How to Structure Long Computation Questions

A strong computation write-up should include:

Model statement: write the regression equation.
Given outputs: specify which numbers are provided (SEs, t-stat, R²).
Testing: state hypotheses explicitly.
Decision rule: significance level (\alpha) and conclusion.
Interpretation: connect to context of the question.

Avoid these common losses:

compute but don’t interpret
interpret coefficient in wrong units (e.g., log vs level)
confuse statistical significance with economic significance
ignore that dummy variable coefficients represent group differences.

5.3 Worked Interpretation Example (Dummy + Log Outcome)

Suppose exam provides:
[
\ln(wage_i)=2.10+0.20,female_i+0.05,educ_i+u_i
]
Interpret (\beta_1=0.20) on female dummy.

If female=1:

log wage is higher by 0.20 relative to female=0.
In percentage terms:
approx percent difference: (100\times 0.20=20%) higher
More precise:
(100\times(e^{0.20}-1)\approx 100\times(1.2214-1)=22.14%)

In an exam, either approximation or exact computation can get marks if justified.

5.4 Worked Interaction Example (Education Returns by Gender)

Suppose:
[
\ln(wage_i)=1.20+0.06,educ_i+0.30,female_i+0.02(educ_i\times female_i)+u_i
]
Effects:

Males: female=0 ⇒ education slope = 0.06
Females: female=1 ⇒ education slope = 0.06 + 0.02 = 0.08

Interpretation:

a 1-year increase in education corresponds to about 6% wage increase for males and about 8% wage increase for females (approximation since dependent variable is in logs).

If asked “is education more valuable for females?” answer:

compare slopes: yes if (\beta_3) positive and statistically significant.
but discuss that this is association unless endogeneity is addressed.

5.5 Diagnostic “Decision Tree” for Common Exam Scenarios

When you see an exam prompt like “residuals show increasing variance,” or “Durbin–Watson suggests autocorrelation,” you should respond with a clear chain:

Identify assumption likely violated
- increasing residual variance ⇒ heteroskedasticity
- residuals correlated over time ⇒ autocorrelation
State effect on OLS inference
- coefficients may remain unbiased (if exogeneity holds)
- standard errors wrong ⇒ t/F unreliable
Propose remedy
- robust SEs for heteroskedasticity
- Newey-West or model-based correction for autocorrelation
Re-check after remedy
- interpret updated SEs/significance

5.6 Endogeneity Scenario Template

Question type: “Given that schooling is endogenous, what can you do?”

Answer structure:

Identify endogeneity source (omitted ability, reverse causality, measurement error).
State which assumption is violated: (\mathbb{E}[u|X]\neq 0).
Propose solution:
- include controls (if plausible proxies for omitted variable)
- use IV/2SLS if there’s a credible instrument
- consider difference-in-differences or fixed effects if panel data exists (depending on course scope)
Discuss assumptions required:
- exogeneity of instrument
- instrument relevance
Interpret IV estimates carefully (local effect, validity concerns).

5.7 Mini Mock Exam: Integrated Problem (Conceptual, with Calculations Placeholder)

A typical exam might provide the following dataset context:

(Y): monthly wage in ZAR
(X_1): years of education
(X_2): experience
(D): female dummy (1 if female)
Model estimated:
[
\ln(wage)=\beta_0+\beta_1 educ+\beta_2 exp+\beta_3 female+u
]
You may be asked:

Interpret (\beta_1).
Conduct a t-test for (\beta_3) given SE and estimate.
Test a joint hypothesis like (H_0:\beta_2=0) and (\beta_3=0).
Discuss heteroskedasticity based on a test output.
Suggest a remedy.

Even if numbers are not provided here, your answer should follow a template:

For interpretation: log-dependent variable implies approximately percentage interpretation.
For tests: state hypotheses, use t or F, conclude at given (\alpha).
For heteroskedasticity: propose robust SEs, interpret with robust inference.

5.8 Common Mistakes and How to Avoid Them

Using coefficient meaning across wrong functional forms
- log–log vs log–level vs level–log interpretation differs.
Ignoring the baseline group for dummy variables
- dummy coefficient represents difference relative to the omitted category (baseline coded by 0).
Confusing statistical significance with causality
- significant coefficients do not guarantee causal impact.
Treating (R^2) as the main evaluation criterion
- emphasis should be on theory, assumptions, and diagnostic checks.
Assuming “no multicollinearity = no problem”
- multicollinearity affects precision, not bias; endogeneity affects bias/consistency.
Performing hypothesis testing without specifying (H_0) correctly
- always write the null value explicitly.

5.9 South African Higher Education Context: How This Shows Up in Practice

In South African university settings, EKN 325-style assessments often include:

applying econometric reasoning to labour market and policy questions relevant to the local economy,
interpreting results with sensitivity to data limitations,
using examples that reference publicly known phenomena: schooling access disparities, unemployment dynamics, and wage inequality.

Your exam advantage comes from connecting econometric concepts to plausible institutional contexts:

schooling and ability,
access and distance,
labour market returns varying by demographic groups,
time series dependence in macroeconomic variables.

Even without naming specific brands or institutions in the model itself, you should show awareness that:

data may be survey-based (measurement error),
some regressors may reflect policy or institutional variation (possible instruments),
and outcomes like wages can reflect both productivity and discrimination (omitted variables/endogeneity risk).

5.10 Final Study Plan for EKN 325 (Practical Revision Strategy)

A simple high-yield plan (adapt to your schedule):

Week 1: OLS assumptions, interpretation of coefficients (levels/logs), basic t and F tests.
Week 2: dummies, interactions, polynomials, functional form selection, residual diagnostics.
Week 3: heteroskedasticity and robust SE logic; autocorrelation reasoning.
Week 4: endogeneity sources and IV logic; first-stage relevance/exogeneity discussion.
Final days: do mixed practice:
- interpret outputs,
- test hypotheses with provided SE/t/F,
- write diagnostic remedies.

When practicing, always write your explanations the way you would in an exam:

crisp definitions
correct null/alternative hypotheses
contextual interpretation
diagnostic remedy rationale.

Key Takeaway Checklist (Section 5)

Use templates: concept → consequence → remedy → interpretation.
Ensure functional form matches interpretation (especially logs).
Separate association from causality and mention assumptions.
Diagnose assumption violations and propose correct fixes.
Structure exam answers for clarity: hypotheses, statistics, decision, context.

Consolidated EKN 325 Master Cheat Sheet (One-Page Style)

Core regression model

[
Y_i = \beta_0 + \sum_{j=1}^k \beta_j X_{ji} + u_i
]

Exogeneity

[
\mathbb{E}[u_i|X_i]=0
]

OLS slope intuition (partial association)

(\hat{\beta}_j) measures association between (X_j) and (Y) controlling for other regressors—only interpretable causally with exogeneity.

t-test (single coefficient)

[
t = \frac{\hat{\beta}j – \beta{j,0}}{\text{SE}(\hat{\beta}_j)}
]

F-test (joint)

Test (q) restrictions using restricted vs unrestricted SSR.

Interpretation rules

Level–level: one-unit change in (X) → change in (Y).
Log–level: (\beta) approx = % change in (Y) for 1-unit (X).
Level–log: (\beta) approx = change in (Y) for 1% change in (X).
Log–log: (\beta) = elasticity.

Dummy variables

In level model: difference in means.
In log model: approx % difference.

Interactions

Effect of (X) depends on the interacting variable; compute conditional marginal effects.

Endogeneity and IV

OLS fails when (\mathbb{E}[u|X]\neq 0).
IV requires relevance and exogeneity.
2SLS uses predicted (X) from instruments to estimate causal effects.

Diagnostics

Heteroskedasticity ⇒ wrong SEs ⇒ use robust SE.
Autocorrelation ⇒ wrong SEs ⇒ use time-series robust methods or model correction.
Multicollinearity ⇒ large SEs, unstable estimates; not necessarily bias.

Final Closing Summary

EKN 325 demands both econometric fundamentals and disciplined reasoning. Master the relationship between assumptions and inference, practice interpreting coefficients accurately across functional forms (especially logs and dummies), and learn to diagnose and remedy assumption violations. Finally, build exam-ready answers with clear hypothesis statements, correct test logic, and context-sensitive interpretations that reflect the econometric validity of your conclusions.