ECO 326: Econometrics Course Notes

Econometrics is the part of economics that turns economic theory into testable, data-driven statements. In a course like ECO 326 (Econometrics), the emphasis is typically on how to specify models, estimate them with data, diagnose problems, and interpret results in ways that are economically meaningful. These notes are written with South African tertiary education in mind (universities, colleges, and TVET pathways), where students frequently encounter the same core econometrics toolkit: Ordinary Least Squares, inference, assumption checking, and applied work using real datasets.

This study guide focuses on the practical workflow you’re expected to master in ECO 326: (1) model building, (2) estimation, (3) inference, (4) diagnostics and corrections, and (5) application to policy and labour/education/health contexts relevant to South Africa. You’ll find detailed examples, common exam-style questions, and step-by-step procedures you can reuse under pressure.

Foundations of Econometric Thinking (ECO 326 Core Mindset)

Econometrics starts with the question: “What is the causal or predictive relationship between X and Y, and how can we estimate it from noisy data?” The “noise” includes sampling variation, measurement error, omitted variables, and simultaneity. The econometrician’s job is to translate a relationship into an econometric model, then use statistical methods to estimate and evaluate that model.

What Makes a Model “Econometric”?

A typical econometric model has four ingredients:

1. A dependent variable (Y)
   Example: household income, employment probability, school attainment, inflation rate, housing prices.

2. Explanatory variables (X’s)
   Example: years of education, labour market experience, unemployment rate, household size, interest rates, exchange rate.

3. A structural relationship (the theory)
   Example: “More education increases earnings,” or “Higher interest rates reduce investment.”

4. A statistical error term (u)
   The error term captures influences not explicitly included in X and also measurement/sampling problems.

A basic linear model is:

[
Y_i = \beta_0 + \beta_1 X_i + u_i
]

To estimate it, we choose an estimator (often OLS) and then make assumptions about how (u_i) behaves.

The Difference Between Correlation and Econometrics

Econometrics does not stop at correlation. Correlation answers: “Do variables move together?” Econometrics asks: “If X changes, what happens to Y, holding other factors constant?” In practice, you must consider:

• Omitted variable bias: the reason both X and Y might move together.
• Measurement error: X may be observed with error.
• Reverse causality / simultaneity: Y may also affect X.
• Selection bias: only certain units enter the sample.

In South African applied studies—labour market outcomes, education returns, housing and inflation—these issues appear often.

Units, Indices, and the Meaning of i

Most ECO 326 questions use notation like:

• (i = 1,2,\dots,n) for observations (individuals, households, firms, provinces, months).
• (t = 1,2,\dots,T) for time (if using time series).
• (X_{it}) if data is panel (both individual and time).

When you interpret results, always remember: the coefficients apply to the unit of measurement in your data. For example, if education is measured in years, then a coefficient on education means “change in Y for one additional year of education.” If education is measured in log years or categories, the interpretation changes.

Assumptions: Why They Matter for Exams

For OLS inference, you often need assumptions such as:

• Linearity in parameters (model is linear in β’s, even if X enters nonlinearly)
• Random sampling / independence
• Zero conditional mean: (E(u_i \mid X_i) = 0)
• No perfect multicollinearity
• Homoskedasticity or robust variance

A key concept:

• If (E(u_i \mid X_i)\neq 0), then OLS estimates are generally biased and inconsistent.
• If the variance of errors is not constant, standard errors can be wrong (but coefficients may still be unbiased under conditional mean).

Econometrics Workflow (The Exam Answer Skeleton)

Most ECO 326 solutions score well when they follow a standard workflow:

1. State the model (define Y and X clearly).
2. Explain the intuition (why you expect a relationship).
3. Write the regression equation.
4. Specify estimation method (OLS).
5. Assumption check: at least mention what could go wrong.
6. Report coefficient interpretation.
7. Hypothesis test: interpret p-values or critical values.
8. Model diagnostics and remedies: heteroskedasticity, multicollinearity, autocorrelation, endogeneity.

That structure maps directly to exam marking rubrics because it shows both statistical reasoning and econometric judgement.

Example: A Labour Market Model with SA Context

Suppose Y is log monthly earnings for employed adults, and X includes:

• (Education_i): years of schooling
• (Experience_i): years of labour market experience
• (Union_i): union membership indicator (1 if union member, 0 otherwise)
• (Region_i): province fixed effects (if included)

A simple regression:

[
\log(Earnings_i) = \beta_0 + \beta_1 Education_i + \beta_2 Experience_i + \beta_3 Union_i + u_i
]

Interpretation under a log-linear form:

• If (\beta_1 = 0.06), then one extra year of education is associated with about 6% higher earnings, approximately (more precisely: (e^{0.06}-1)).

But econometrics demands caution: union membership may correlate with unobserved ability or firm type, which affects the conditional mean assumption. That’s where the analysis may need robust standard errors, additional controls, or instrumental variables in later chapters.

Ordinary Least Squares (OLS), Inference, and Model Building

OLS is usually the first full econometric estimation method taught in ECO 326. The focus is not only computing coefficients but understanding when OLS is valid and how inference depends on assumptions.

The OLS Estimator

In matrix form, the standard linear regression model is:

[
Y = X\beta + u
]

OLS chooses (\hat{\beta}) to minimize the sum of squared residuals:

[
\hat{\beta} = \arg\min_{\beta} (Y – X\beta)'(Y – X\beta)
]

With assumptions like (E(u|X)=0), OLS is unbiased and consistent (under standard large-sample conditions).

Residuals and Their Economic Meaning

Residuals:

[
\hat{u}_i = Y_i – \hat{Y}_i
]

• If residuals show systematic patterns with a variable not included in X, that suggests omitted variables or functional form problems.
• If residual variance changes with fitted values, that hints at heteroskedasticity.

In exams, you may be asked to “discuss diagnostics” rather than compute them. A strong answer explains what pattern indicates what problem and what correction you would apply.

Interpreting Coefficients Correctly

Common ECO 326 interpretation patterns:

1) Linear model in levels

[
Y_i = \beta_0 + \beta_1 X_i + u_i
]

• (\beta_1): change in Y for a one-unit change in X.

2) Log-level

[
Y_i = \beta_0 + \beta_1 \log(X_i) + u_i
]

• If X changes by 1%: approximate change in Y is (\beta_1/100).

3) Log-log

[
\log(Y_i) = \beta_0 + \beta_1 \log(X_i) + u_i
]

• (\beta_1) is an elasticity: 1% change in X changes Y by approximately (\beta_1)% .

4) Level-log (very common)

[
\log(Y_i) = \beta_0 + \beta_1 X_i + u_i
]

• One-unit increase in X changes Y by approximately (100\times \beta_1)% .

5) Dummy variables

If (D_i\in{0,1}):

• In level model: coefficient on D is the mean difference between group 1 and group 0.
• In log model: interpretation uses (e^{\beta}) for exact effects.

Hypothesis Testing in Regression

A core set of exam competencies:

• Testing single coefficients: (H_0:\beta_j = 0)
• Testing multiple coefficients jointly: (H_0:\beta_{j_1}=\beta_{j_2}=\dots=0)
• Testing model fit (often using F-test)
• Interpreting t-statistics and p-values

t-test for one coefficient

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

Decision: compare with critical value or use p-value.

F-test for multiple restrictions (joint significance)

If you test (q) restrictions, the F-statistic can be used. The details depend on whether you use constrained vs unrestricted models.

The Role of Standard Errors

Standard errors depend on assumptions about the error term:

• With homoskedasticity, usual OLS standard errors can be valid.
• With heteroskedasticity, the OLS coefficients may remain unbiased, but standard errors become wrong unless you use robust (heteroskedasticity-consistent) standard errors.

In South African empirical work, heteroskedasticity is extremely common because variance differs across income groups, regions, firm sizes, and survey strata.

Multicollinearity and Why It Shows Up in SA Data

Multicollinearity means explanatory variables are correlated. It leads to:

• Larger standard errors
• Coefficients becoming statistically insignificant even when the variables are individually meaningful
• Coefficient instability across samples

A classic example:

• Education and experience might be correlated.
• In labour earnings models, if you include both education and experience (and perhaps age), you may see multicollinearity.

In exams, a good answer does not just say “multicollinearity exists.” It explains the effect: it harms inference by inflating variances, but it does not bias OLS coefficients if the conditional mean assumption holds.

Possible diagnostics:

• Correlation matrix
• VIF (variance inflation factor) if discussed in your course
• “Signs change when adding variables” (practical symptom)

Functional Form and Specification Errors

Functional form errors occur when the true relationship is not captured:

• Missing nonlinear terms (e.g., earnings vs education is nonlinear)
• Missing interactions (effect differs by group)
• Missing relevant variables (omitted variable bias)

Example:
If the effect of education increases but then levels off (diminishing returns), a model using only linear education may misfit. A better model could include:

• (Education) and (Education^2)
• or categories of education

An Applied Example: Education Returns and Heteroskedasticity

Consider a regression:

[
\log(Earnings_i) = \beta_0 + \beta_1 Education_i + \beta_2 Experience_i + \beta_3 Female_i + u_i
]

Suppose the estimated coefficient on Education is (\hat{\beta}_1=0.075) with robust standard error (0.012).

• Interpretation: one additional year of education is associated with about 7.5% higher earnings (approx. (100\times 0.075%)).
• Hypothesis test: (t \approx 0.075/0.012 = 6.25), likely significant.

If standard errors were computed under homoskedasticity but the data clearly has heteroskedasticity, p-values could be misleading. Robust standard errors would provide more reliable inference.

A strong exam answer mentions:

• why heteroskedasticity is plausible,
• why robust SEs help,
• and still interpret coefficients cautiously as association unless causal identification is established.

Multiple Regression and the Meaning of “Holding Constant”

In multiple regression:

[
Y_i = \beta_0 + \beta_1 X_{1i} + \beta_2 X_{2i} + u_i
]

(\beta_1) is interpreted as the effect of (X_1) on Y conditional on (X_2). That means:

• you compare observations with similar values of (X_2),
• then assess how Y varies with (X_1).

This conditional interpretation is crucial in policy discussions—for instance, comparing earnings between educated workers with similar experience.

Common Exam Pitfalls (What Not to Do)

1. Interpreting log coefficients incorrectly
   Students sometimes treat (\log(Y)) coefficients as if they were level effects without exponentiation.

2. Ignoring omitted variable bias
   For example, if ability is unobserved and correlated with education, OLS education returns may be biased.

3. Using “significance” as “causality”
   Statistical significance does not imply causal effect.

4. Confusing multicollinearity with endogeneity
   Multicollinearity affects precision, not necessarily bias. Endogeneity affects unbiasedness/consistency.

Diagnostics, Advanced Standard Errors, and the Problem of Endogeneity

After estimating a model, ECO 326 typically expects you to diagnose and correct issues. Real data almost always violates at least one OLS assumption.

Heteroskedasticity: Detection and Consequences

Heteroskedasticity means:

[
Var(u_i \mid X_i) = \sigma_i^2
]
where (\sigma_i^2) differs by i.

Why it matters

• OLS coefficients can remain unbiased under (E(u|X)=0).
• But OLS standard errors (and therefore t-tests and F-tests) become unreliable.

What to do

• Use heteroskedasticity-robust standard errors (commonly called “robust SE”).
• Consider transformations or weighted least squares if appropriate.

Exam phrasing

A well-graded response includes:

• What is heteroskedasticity?
• How does it affect inference?
• What estimator/SE correction do you propose?

Autocorrelation (Time Series or Panel Dependence)

Autocorrelation means error terms are correlated over time:

[
Cov(u_t, u_{t-1}) \neq 0
]

Consequences:

• OLS remains unbiased under strict exogeneity, but standard errors can be wrong.
• Forecast uncertainty increases.

Corrections:

• Use Newey-West or similar approaches (if taught).
• Adjust model structure: include dynamics, lags, or use GLS-type methods.

Exam hint:
If your course emphasizes it, mention:

• detection via residual plots or tests,
• consequence on inference,
• correction with appropriate standard errors.

Multicollinearity Revisited: What It Does and Does Not Do

Students often believe multicollinearity “breaks OLS.” Actually, if assumptions hold:

• OLS coefficient estimates are still unbiased (conditional mean holds).
• The harm is primarily to precision: higher SEs.

So in practice:

• coefficients may not be significant,
• but the joint effect (tested by F-test) might still be meaningful.

An exam-friendly approach:

• interpret coefficient magnitudes with caution,
• evaluate joint significance,
• consider dropping redundant variables only with justification.

Endogeneity: The Econometric “Big Problem”

Endogeneity arises when:

[
E(u_i \mid X_i) \neq 0
]

Common sources:

1. Omitted variables
   Ability, motivation, institutional quality omitted.

2. Reverse causality
   Y affects X.

3. Measurement error
   X measured with error leads to correlation between observed X and error term.

Endogeneity implies OLS is biased and inconsistent unless corrected.

Example: Education and Earnings—Why OLS Might Fail

If earnings depend not only on education and experience but also on unobserved ability:

[
\log(Earnings_i) = \beta_0 + \beta_1 Education_i + \beta_2 Experience_i + \gamma Ability_i + u_i
]

If Ability is omitted and correlated with Education:

• Education coefficient becomes biased.
• Standard errors do not fix bias.

In exam answers, it’s essential to show you understand the difference:

• Heteroskedasticity → affects SEs (inference),
• Endogeneity → affects coefficient bias/consistency (identification).

Instrumental Variables (IV) as a Remedy (Often Covered in ECO 326)

If your course includes IV/2SLS, you typically need:

• Instrument relevance: instrument Z must be correlated with endogenous X.
• Instrument exogeneity: Z must not be correlated with the error term u.

The IV strategy:

1. First stage: regress X on Z and other exogenous variables.
2. Second stage: regress Y on predicted X from first stage.

A common structure:

• Endogenous regressor: (Education_i)
• Instrument: something plausibly affecting education but not earnings directly (e.g., distance to school at age 14, policy exposure, scholarship eligibility)

You must carefully explain:

• relevance logic,
• exogeneity plausibility,
• and what tests you would use (like overidentification tests if multiple instruments, or relevance diagnostics).

Important exam caveat

Even if IV satisfies assumptions statistically, you must justify assumptions economically and contextually. In South African policy data, this often becomes an argument about:

• institutional pathways,
• policy rules,
• and plausible exclusion restrictions.

Overfitting vs Misspecification

When students add many variables, they sometimes chase significance. Econometric judgement is to balance:

• bias from omitted variables,
• variance from overfitting,
• and interpretability.

A strong answer should describe a “model selection” logic:

• include theoretical controls,
• avoid irrelevant variables that inflate variance,
• consider interactions based on theory.

Practical Diagnostics You Can Mention in Exams

Even if you are not asked to compute them, you can discuss:

• Residual plots vs fitted values (heteroskedasticity)
• Residuals over time (autocorrelation)
• Correlation matrix (multicollinearity)
• Test for normality if residual normality is discussed for small samples
• Ramsey RESET test (if covered) for functional form
• Structural breaks in time series (if covered)

Case Illustration: Housing Prices and Inflation—Potential Endogeneity

Suppose Y is housing price growth, and X includes:

• interest rates,
• household income,
• housing supply constraints.

Potential issues:

• reverse causality: housing booms can affect prices and possibly interest rate expectations through macro channels.
• omitted variables: sentiment, construction delays.
• measurement error: using proxies for supply constraints.

An exam answer could propose:

• robust SEs for heteroskedasticity,
• adding controls like construction cost indices,
• considering instruments for interest rates or supply shocks if endogeneity is severe.

The goal is to show you can diagnose and propose remedies rather than pretending OLS always works.

Regression with Dummies, Interactions, and Nonlinearities (Policy-Relevant Applications)

Many ECO 326 papers focus on categorical comparisons: between genders, provinces, income groups, educational levels, employment types, or policy regimes. This requires careful modeling with dummy variables and interactions.

Dummy Variables: Baseline Group and Interpretation

Suppose you include a dummy for gender:

• (Female_i = 1) if female
• (Female_i = 0) if male

In:

[
Y_i = \beta_0 + \beta_1 Female_i + u_i
]

• (\beta_0): expected Y for male group (baseline)
• (\beta_1): difference between female and male means, holding other covariates constant (if included)

If you include multiple categories (e.g., provinces), you include k-1 dummies to avoid dummy variable trap (perfect multicollinearity).

Interaction Terms: When Effects Differ Across Groups

An interaction between a continuous variable and a dummy tests whether the slope differs by group:

[
Y_i = \beta_0 + \beta_1 Education_i + \beta_2 Female_i + \beta_3 (Education_i \times Female_i) + u_i
]

Interpretation:

• For males (Female=0): slope on Education is (\beta_1)
• For females (Female=1): slope on Education is (\beta_1 + \beta_3)
• (\beta_2) is the difference in intercepts at Education=0 (sometimes not meaningful if Education=0 is outside data range; explain accordingly)

Exams reward you for showing the conditional effect.

Interactions with Two Dummies: Group-Specific Means

Example with two dummies: Urban (1 if urban) and HighIncome (1 if high income). A fully saturated group mean model can represent four groups:

• urban & high income
• urban & low income
• rural & high income
• rural & low income

You include three dummies plus interactions to create group-specific intercepts depending on your course’s approach.

Nonlinearities with Quadratic Terms and Marginal Effects

For diminishing returns, you might use:

[
Y_i = \beta_0 + \beta_1 Education_i + \beta_2 Education_i^2 + u_i
]

Marginal effect:

[
\frac{\partial Y_i}{\partial Education_i} = \beta_1 + 2\beta_2 Education_i
]

If (\beta_2<0), marginal returns decrease as education increases.

In South African education contexts, diminishing returns could reflect:

• skill mismatch at higher levels,
• overqualification,
• limited labour market absorption in certain occupations.

But be careful: a statistically significant quadratic term indicates curvature in the conditional mean, not necessarily true causality without identification.

A Detailed Example: Technology Skills and Employment Probability

Let Y be an employment indicator:

• (Emp_i = 1) if employed
• (Emp_i = 0) otherwise

A simple linear probability model (LPM) might be used early:

[
Emp_i = \beta_0 + \beta_1 TechSkill_i + \beta_2 Experience_i + \beta_3 Female_i + u_i
]

Interpretation in LPM:

• (\beta_1): change in employment probability for a one-unit increase in TechSkill (in probability points)

However, the LPM can predict probabilities outside [0,1] and has heteroskedastic errors by construction. A common exam response:

• discuss the limitations,
• mention robust SEs,
• and note that logit/probit might be used in later chapters if your course covers it.

Policy Comparison: Pre- and Post- Intervention Dummies

Suppose a training programme launched in 2019. Define:

• (Post_t = 1) for years 2019 and after, 0 otherwise
• (Treat_i = 1) for participants in affected regions, 0 otherwise

A difference-in-differences (DiD) structure (often introduced around this area) uses the interaction:

[
Y_{it} = \beta_0 + \beta_1 Treat_i + \beta_2 Post_t + \beta_3 (Treat_i \times Post_t) + u_{it}
]

Interpretation:

• (\beta_3) is the DiD estimate: the additional change in treated group after the intervention relative to control group.

Exam emphasis:

• parallel trends assumption is crucial,
• you should discuss threats (differential shocks across groups, selection into treatment, policy spillovers).

Worked Interpretation Example (No Computation, But Clear Logic)

Assume estimated coefficients in a DiD regression:

• (\hat{\beta}_3 = 0.08) in a model where Y is log earnings.
  Then:
• the treatment effect is approximately (100 \times 0.08 = 8%) higher earnings for treated group after intervention, relative to controls.
  If Y is in levels, interpret as absolute change.

The key is matching interpretation to the scale.

South African Context: Provinces, Inequality, and Group Effects

Dummies for provinces are frequently used in empirical economics because:

• labour market conditions differ by geography,
• education access and school quality differ,
• infrastructure and economic base differ.

When including province fixed effects (or province dummies), you absorb time-invariant differences across provinces. That strengthens comparisons but does not solve endogeneity by itself.

Time Series and Panel Concepts (When ECO 326 Moves Beyond Cross-Section)

Many ECO 326 courses extend into time series and panel econometrics, especially because South African datasets often have monthly/quarterly indicators (inflation, unemployment proxies, retail sales) and multi-year surveys (labour force surveys).

Time Series Basics: Trend, Seasonality, and Stationarity

Time series models require thinking about the data-generating process over time. A major concept is stationarity:

• A stationary series has mean and variance that do not systematically change over time.
• Non-stationary series can lead to misleading regression results (spurious correlation).

In an exam, if asked about stationarity, your response should mention:

• reason non-stationarity is problematic,
• how differencing or detrending can help,
• and that inference becomes invalid if you ignore dynamics.

Autoregressive (AR) and Distributed Lag Ideas (General)

A model with lags:

[
Y_t = \alpha + \phi Y_{t-1} + \beta X_t + u_t
]

Captures persistence: current Y depends on past Y. If errors are autocorrelated, you must ensure consistent standard errors and model form.

Why Lags Matter for Policy: Adjustment Dynamics

In economics, effects often arrive with delays:

• education reforms affect outcomes after years,
• infrastructure investment affects employment gradually,
• inflation affects consumption with lags,
• training programmes affect earnings after job placement.

Including lags is not just technical—it aligns model timing with economic mechanisms.

Panel Data: Individual Effects and Time Effects

Panel data has observations across individuals (or firms, provinces) over time. A standard panel model:

[
Y_{it} = \beta_0 + \beta_1 X_{it} + \alpha_i + \lambda_t + u_{it}
]

• (\alpha_i): individual (unit-specific) effect
• (\lambda_t): time effect
• (u_{it}): idiosyncratic error

Two main estimation frameworks are commonly discussed:

1. Fixed effects (within estimator)
   Removes (\alpha_i) by using deviations from unit means.

2. Random effects
   Assumes (\alpha_i) is uncorrelated with X (strong assumption).

In many labour/education applications, fixed effects are preferred because unobserved unit characteristics likely correlate with regressors (e.g., local managerial quality, baseline school quality, region-specific economic structure).

Fixed Effects: Interpretation and What It Controls For

With fixed effects, the coefficient on (X_{it}) is interpreted as:

• the effect of within-unit changes over time in X on within-unit changes in Y.

So if a province improves schooling quality between 2016 and 2021, fixed effects uses that change to estimate its association with changes in outcomes, netting out province-specific time-invariant traits.

Example: Unemployment and Education in a Panel

Suppose:

• Y = unemployment rate for province i at time t
• X = average years of schooling, plus youth training participation rate

Fixed effects model:

[
Unemp_{it} = \beta_0 + \beta_1 School_{it} + \beta_2 Training_{it} + \alpha_i + u_{it}
]

Interpretation:

• (\beta_1): association between changes in education levels within province and unemployment rate changes within province, controlling for province-specific baseline unemployment tendencies.

Exam nuance:

• Even with fixed effects, endogeneity might remain if schooling improvements respond to labour market conditions.

Dynamic Panels and Caution

If a model includes lagged dependent variables:

[
Y_{it} = \rho Y_{i,t-1} + \beta X_{it} + \alpha_i + u_{it}
]

then specialized estimators may be needed because standard fixed effects can produce bias in small samples (Nickell bias). Some ECO 326 courses mention these concepts qualitatively; others provide methods.

Clustered Standard Errors in Panel Settings

Even if you correctly estimate coefficients, standard errors should reflect dependence:

• In panel data, errors within the same unit i may be correlated over time.
• Using clustered standard errors by unit is often expected when errors are correlated within clusters.

An exam-friendly statement:

• “To account for within-province (or within-individual) correlation, cluster standard errors at the unit level.”

Small Numerical Example (Illustrative, Not Dataset-Specific)

Consider a simplified two-period panel. Fixed effects essentially subtract unit means:

1. For each unit i, compute:

   • ( \bar{Y}i = \frac{Y{i1}+Y_{i2}}{2} )
   • ( \bar{X}i = \frac{X{i1}+X_{i2}}{2} )

2. Transform:
   [
   (Y_{it} – \bar{Y}i) = \beta_1 (X{it}-\bar{X}i) + (u{it}-\bar{u}_i)
   ]

This uses only within-unit change between t=1 and t=2.

Exams often ask conceptually: “What does fixed effects eliminate?” Answer:

• It eliminates all unit-invariant components (\alpha_i).

Practical SA Data Considerations

South African panel datasets (where province is the unit) may have:

• policy changes (e.g., training programmes, school infrastructure funding),
• macroeconomic shocks (oil price, exchange rate movements affecting inflation),
• uneven measurement quality across provinces.

A robust econometric practice:

• include time effects (to absorb national shocks),
• include relevant controls,
• and use appropriate standard errors.

Worked Exam-Style Approaches and Common ECO 326 Question Types

This section consolidates the skills into exam-ready patterns: how to structure answers, what to mention, and how to interpret results without losing marks. It also includes mini “case vignettes” aligned to typical South African applied topics: education, labour markets, inflation, housing, and public policy.

Question Type 1: “Explain OLS assumptions and consequences of violations”

A high-mark answer should:

1. List assumptions (at least the ones central to inference and consistency):

   • (E(u|X)=0)
   • no perfect multicollinearity
   • appropriate variance assumption (homoskedasticity for classic SE)
   • independence / exogeneity (depending on context)

2. Link each assumption to what happens if it fails:

   • conditional mean fails → bias/inconsistency
   • heteroskedasticity fails → wrong SEs, invalid t/F tests unless robust SEs used
   • multicollinearity fails → large SEs, weak inference, coefficients unstable
   • autocorrelation fails → wrong SEs unless corrected

3. Conclude with remedies:

   • robust SEs,
   • adding controls,
   • transforming variables,
   • using IV/DiD/fixed effects depending on the problem type.

Question Type 2: “Interpret coefficients from a regression output”

A strong interpretation process:

1. Identify the scale of Y:

   • levels, log(Y), log-log, etc.

2. Identify the scale of X:

   • levels, logs, dummies, interactions.

3. Translate coefficient into an economic statement:

   • percentage change for log Y or log X,
   • absolute difference for dummy variables,
   • conditional slope for interactions.

4. Tie to statistical significance:

   • interpret the direction and magnitude,
   • mention uncertainty via standard errors or CI.

5. Avoid causality claims without identification:

   • “associated with” unless endogeneity is addressed.

Question Type 3: “Propose a model to test a theory”

Example theory statement:
“Education improves employment outcomes in South Africa.”

A possible model framework could be:

• Y: employment probability (or unemployment rate)
• X: education (years or qualifications), experience, gender, region controls
• error term: captures unobserved influences
• mention potential endogeneity (ability, family background)

Your answer should include:

1. dependent variable definition
2. explanatory variables and expected signs
3. model equation
4. estimation method (OLS for continuous outcomes; LPM or logit/probit if needed)
5. identification concern (endogeneity)
6. remedy suggestion if necessary (e.g., IV, fixed effects panel, DiD)

Question Type 4: “Diagnose problems and recommend corrections”

An examiner might provide hints such as:

• residuals fan out (heteroskedasticity)
• autocorrelation suspected (time series)
• multicollinearity suspected (many correlated X’s)

You should respond with:

• symptom → likely cause → consequence → correction.

Example mapping:

• Residuals vs fitted show increasing variance → heteroskedasticity → wrong SEs → robust SEs or WLS.
• Residuals correlate across time → autocorrelation → wrong SEs → HAC/Newey-West or model with lags.
• Coefficients change dramatically when adding a variable → multicollinearity → larger SEs → use theory, avoid redundant predictors, consider principal components (if in syllabus) or interpret joint effects.

Question Type 5: “Difference-in-differences reasoning”

A typical DiD question:
“Evaluates whether a training programme improved earnings.”

A top answer includes:

1. Treatment and control group definition.
2. Pre and post periods.
3. The DiD estimator concept:
   • difference in outcomes after vs before for treated
   • minus same difference for control.
4. Parallel trends assumption:
   • treated and control would have moved similarly absent programme.
5. Threats to parallel trends:
   • differential macro shocks,
   • selection into treatment based on trends,
   • spillovers.
6. Remedy suggestions:
   • event study if available,
   • placebo tests,
   • adding time-varying controls.

Mini Case Study Vignettes (SA-Aligned, Method-Focused)

Vignette A: Education and Income with Province Differences

A regression includes education and province dummies, but education might still be endogenous. A good answer:

• explains what province dummies do (control for time-invariant provincial differences),
• notes that endogeneity may persist through unobserved individual ability,
• suggests IV if a plausible instrument exists, or uses panel fixed effects if longitudinal data exists.

Vignette B: Unemployment and Inflation with Macro Shocks

If time series errors show autocorrelation, inference using classic OLS SEs may be wrong. A strong answer:

• discusses autocorrelation,
• recommends corrected SEs,
• emphasizes that coefficient significance could change after correction.

Vignette C: Wage Gaps by Gender with Interaction

If the education return differs by gender, include Education × Female interaction. A high-mark answer:

• shows how to interpret male slope and female slope,
• interprets baseline differences carefully (intercept at Education=0 may be non-meaningful),
• discusses policy implications: training and education may narrow or widen gaps depending on interaction sign.

High-Yield Summary: What ECO 326 Expects You to Master

Econometrics is both a technical and reasoning-based course. In ECO 326, the highest marks tend to come from students who demonstrate:

1. Correct model specification

   • define variables,
   • write the regression clearly,
   • justify inclusion of controls and functional form.

2. Proper estimation understanding

   • know what OLS does (minimizes squared residuals),
   • know conditions for unbiasedness/consistency.

3. Accurate interpretation

   • handle logs and dummy variables correctly,
   • interpret interactions through group-specific slopes or effects.

4. Valid inference

   • understand how heteroskedasticity and autocorrelation affect standard errors,
   • apply robust or clustered SEs when dependence is plausible.

5. Identification and endogeneity awareness

   • distinguish heteroskedasticity from endogeneity,
   • propose appropriate remedies (IV, fixed effects, DiD) based on the underlying problem.

6. Diagnostics and econometric judgement

   • use residual patterns and specification reasoning,
   • explain consequences, not just mention tests.

Exam-Ready Checklist (Use Under Time Pressure)

When answering any ECO 326 regression question:

• Step 1: State the model and define variables.
• Step 2: Choose estimation and justify briefly (OLS/LPM/FE/IV).
• Step 3: Interpret coefficients carefully (scale, dummies, logs, interactions).
• Step 4: Address assumptions and likely violations.
• Step 5: Correct standard errors or propose identification strategy.
• Step 6: Provide hypothesis test interpretation (significance + meaning).
• Step 7: Conclude with a coherent economic/policy statement consistent with your model.

This checklist aligns with how most lecturers mark: clarity, correctness, and a logically connected argument from assumptions → estimation → inference → interpretation.

South African Learning Pathways Context (Universities, Colleges, and TVETs)

Because the keyword “ECO 326: Econometrics Course Notes” is studied across South African institutions, it’s useful to connect econometrics concepts to how students typically experience them across pathways. While the core technical content is shared, the emphasis and practice differ depending on whether your programme is more academically research-oriented (traditional universities) or more application-focused (TVET and business/college pathways).

University-Style Expectations

At many South African universities, ECO 326 (or closely related econometrics modules) typically expects:

• strong mathematical comfort with regression notation,
• careful discussion of assumptions and their implications,
• the ability to interpret empirical findings in a research style.

Exams may include:

• derivations (e.g., OLS normal equations),
• short proofs or consistency arguments,
• or interpretation tasks with log and dummy variables.

College/Work-Integrated Learning Emphasis

At certain colleges or programmes with workplace alignment, students often focus on:

• how to implement models using statistical software,
• how to justify model choices for business/public policy,
• how to communicate results clearly.

The “diagnostics and correction” portions may be explained more through interpretation and less through heavy proof, but you still need sound econometric logic.

TVET-Adjacent Econometrics Skills

For TVET learners who progress into economics-related fields (or apply econometrics skills in internships), the most transferable parts are:

• understanding regression output interpretation,
• distinguishing correlation vs causation,
• recognising model assumptions in practical terms,
• and applying robust thinking to real data (e.g., labour and education metrics).

Even if your programme is less focused on theory derivations, an econometric “mindset” remains the core: specify, estimate, diagnose, and interpret appropriately.

Consistent Core Competence Across Institutions

Regardless of institution type, the most consistent competence requirements are:

• know what OLS coefficients mean on the chosen scale,
• interpret dummy and interaction effects correctly,
• understand how assumption violations affect inference,
• recognise endogeneity and justify remedies,
• communicate findings with economic interpretation.

Final Practice: Three Full Response Templates You Can Reuse

Below are response templates (not tied to any specific dataset) that mirror the style of high-mark ECO 326 exam answers. When you practice, replace variables and context with the question’s specifics.

Template 1: Model Specification + Interpretation + Assumption Discussion

1. Model:
   [
   Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + u
   ]
   Define Y and X’s explicitly.

2. Economic intuition:
   Explain why (X_1) and (X_2) affect Y.

3. Interpretation:

   • For levels: “one unit increase in (X_1) is associated with …”
   • For log(Y): “percentage change …”
   • For dummy: “group difference …”
   • For interaction: show group-specific slope.

4. Assumptions:
   State (E(u|X)=0), multicollinearity conditions, and variance assumptions.

5. Consequences and remedy:
   If heteroskedasticity: robust SEs.
   If endogeneity: IV/FE/DiD depending on context.

6. Conclusion:
   Summarize findings economically and avoid overstating causality.

Template 2: Diagnostics + Correct SEs + Next Steps

1. Identify the reported symptom (e.g., residual fan, time dependence, multicollinearity).
2. State which assumption likely fails.
3. Explain effect on inference vs coefficients.
4. Propose correction:
   • robust SE,
   • clustered SE,
   • transformed variables,
   • or revised model structure.
5. Mention verification steps (what you would test next).

Template 3: Difference-in-Differences (DiD) Argument

1. Define treated and control groups.
2. Define pre and post periods.
3. Write DiD equation with interaction term.
4. Interpret coefficient on interaction:
   • “additional post-treatment change relative to control.”
5. Discuss parallel trends assumption.
6. Mention threats to validity and how you would test/mitigate them.
7. Conclude with policy interpretation consistent with the estimate.

If you want, I can also generate ECO 326-style practice questions with fully worked solutions (including log/dummy interpretation, robust SE discussion, and DiD specification) matched to the exact topics your lecturer covers—just share the week-by-week syllabus or textbook chapters.