Econometrics 2B builds on core econometrics by deepening your mastery of multiple regression, inference under realistic conditions, model specification, heteroskedasticity/autocorrelation, and advanced time-series thinking. In most EMT2B-type syllabi across South African universities, you are expected to connect theory to exam-ready problem solving: derive estimators, test hypotheses, interpret outputs, and diagnose assumptions. This study guide is designed to help you prepare systematically, with South African–relevant exam practice themes and institution-focused clusters that reflect how different colleges and universities commonly structure “Econometrics 2B” assessments.
EMT2BB2 Exam Prep Cluster: University of Johannesburg (UJ) — Econometrics 2B: Advanced Regression, Diagnostics & Inference
1) What “Econometrics 2B” typically emphasizes at UJ
At institutions like the University of Johannesburg (UJ), “Econometrics 2B” often functions as the course where students graduate from basic OLS mechanics into robust inference, diagnostics, and model refinement. In exam settings, you usually see three broad themes:
-
Multiple regression deepening
- Interpreting coefficients in the presence of other regressors
- Testing joint restrictions with F-tests
- Handling dummy variables and interaction terms carefully
-
Assumptions beyond “nice data”
- Detecting and addressing heteroskedasticity
- Detecting autocorrelation (especially in time-ordered data)
- Understanding the consequences for standard errors and tests
-
Specification, functional form, and missing structure
- Testing for omitted variables
- Using RESET-type logic (or equivalent) for functional form problems
- Choosing between models and explaining consequences
While the precise wording differs by lecturer, UJ-style exam papers usually reward candidates who can:
- translate regression assumptions into testable conditions,
- compute test statistics correctly, and
- interpret outcomes in plain econometric language (not only formulas).
2) Core regression structure you must be able to “re-derive fast”
A typical exam begins with the classical multiple linear regression model:
[
y_i = \beta_0 + x_i'\beta + u_i,\quad i=1,\dots,n
]
Stacked form:
[
\mathbf{y} = \mathbf{X}\beta + \mathbf{u}
]
OLS estimator:
[
\hat{\beta}=(X'X)^{-1}X'y
]
Residuals:
[
\hat{u}=\mathbf{y}-\mathbf{X}\hat{\beta}
]
Residual variance estimate (under classical assumptions, if appropriate):
[
\hat{\sigma}^2=\frac{1}{n-k}\sum_{i=1}^n \hat{u}_i^2
]
where k is number of parameters including intercept if included.
OLS inference objects
- Standard error of (\hat{\beta}_j):
[
se(\hat{\beta}_j)=\sqrt{\widehat{Var}(\hat{\beta}_j)}
] - t statistic:
[
t=\frac{\hat{\beta}j-\beta{j,0}}{se(\hat{\beta}_j)}
] - Under correct assumptions, (t) follows a t-distribution with (n-k) degrees of freedom in small samples.
Exam habit: always state the degrees of freedom
If the exam asks “Compute the critical value” or “State the df,” write:
- df (=n-k) for tests based on (\hat{\sigma}^2) with OLS assumptions,
- df (=(q, n-k)) for F-tests where (q) restrictions are tested jointly.
3) Joint hypothesis testing (F-tests) — the exam workhorse
Suppose you test:
[
H_0: R\beta=r
]
where (R) is (q\times k). The restricted model imposes those constraints, producing (\text{SSR}_R). The unrestricted model produces (\text{SSR}_U).
A standard F statistic:
[
F=\frac{(\text{SSR}_R-\text{SSR}_U)/q}{\text{SSR}_U/(n-k)}
]
Equivalent expression using R-squared is possible, but SSR form is usually safest when given raw regression summaries.
Interpretation:
- If (F) is large relative to (F_{crit}), reject (H_0).
- In words: “the restrictions collectively explain a statistically significant amount of variation.”
Concrete UJ exam-style example (numerical pipeline)
Assume (n=50), (k=4) parameters, so (n-k=46). You test (q=2) restrictions. Given:
- (\text{SSR}_U=1200)
- (\text{SSR}_R=1270)
Compute:
[
F=\frac{(1270-1200)/2}{1200/46}=\frac{70/2}{1200/46}=\frac{35}{26.087…}=1.341
]
Then:
- df1 (=q=2)
- df2 (=n-k=46)
Your decision depends on (F_{crit}). In exam conditions, you might be asked to compute (F) and compare to a provided critical value. If the paper gives (F_{0.05,(2,46)}), use it. If not, you may approximate, but usually exam papers provide critical values or you compute p-values via a table.
4) Heteroskedasticity: detection and consequences
In EMT2B contexts, heteroskedasticity is typically a major topic. Classical OLS still estimates coefficients consistently under exogeneity, but standard errors are wrong if:
[
Var(u_i|X) \neq \sigma^2
]
A common heteroskedasticity form you may be taught:
[
Var(u_i|X)=\sigma_i^2
]
Detection: residual-based logic
Typical exam questions:
- “Perform the Breusch–Pagan test”
- “Perform the White test”
- “Briefly interpret what it means if you reject homoskedasticity.”
Breusch–Pagan (BP) outline (one version):
- Fit auxiliary regression of squared residuals (\hat{u}i^2) on regressors (and sometimes fitted values):
[
\hat{u}i^2 = a_0 + a_1 x{1i} + \cdots + a_k x{ki} + e_i
] - Compute test statistic either from (nR^2) or from an F-test in certain teaching variants.
- Under (H_0): no heteroskedasticity (variance constant), so auxiliary regression explains nothing.
White test:
- Regress (\hat{u}_i^2) on all original regressors, their squares, and cross-products.
- Again use (nR^2) style decision logic.
Consequences: why you must care
If you estimate under heteroskedasticity and use conventional OLS standard errors:
- t-tests and F-tests may be invalid.
- Confidence intervals may not have correct coverage.
Remedy:
- heteroskedasticity-robust standard errors (often called HC or “robust” SE).
- In exam writing, you can say: “Use robust (heteroskedasticity-consistent) covariance matrix.”
You might be asked for the intuition: robust SE corrects the estimated variance of (\hat{\beta}) by allowing (Var(u|X)) to vary across observations.
5) Autocorrelation: when errors are not independent
Time series data or panel data ordered by time can introduce:
[
Cov(u_i,u_{i-1})\neq 0
]
A common exam case: AR(1)-type errors:
[
u_t = \rho u_{t-1}+ \epsilon_t
]
Durbin–Watson (DW) test
Often taught in Econometrics 2B as a quick diagnostic for first-order autocorrelation in regression with an intercept. DW statistic:
[
d = \frac{\sum_{t=2}^n ( \hat{u}t – \hat{u}{t-1} )^2}{\sum_{t=1}^n \hat{u}_t^2}
]
Rule of thumb:
- (d \approx 2): little evidence of autocorrelation
- (d < 2): positive autocorrelation likely
- (d > 2): negative autocorrelation likely
Exam nuance: DW has bounds and depends on sample size and number of regressors; some papers provide a decision diagram or critical bounds.
Practical writing: “what happens to OLS inference?”
If autocorrelation exists and standard errors assume independence:
- OLS coefficients remain unbiased under exogeneity (depending on model form),
- but variance estimates are wrong, leading to invalid hypothesis tests.
Remedies in exams:
- use Newey–West style HAC standard errors (if taught)
- or model autocorrelation directly (e.g., GLS, AR correction)
6) Specification issues: omitted variables and functional form
Econometrics 2B typically wants you to show understanding of what goes wrong when the model is mis-specified.
Omitted variable bias (OVB) in exam writing
If the true model includes (z_i) but you estimate without it:
True:
[
y_i=\beta_0+\beta_1 x_i+\gamma z_i+u_i
]
Estimated:
[
y_i=\beta_0+\beta_1 x_i+\tilde{u}_i
]
Then:
[
\tilde{u}_i=\gamma z_i+u_i
]
If (Cov(x_i,\tilde{u}_i)\neq 0), then (\hat{\beta}_1) is biased and inconsistent.
Exam-friendly statement:
- “OVB occurs when the omitted regressor is correlated with included regressors.”
Testing for omitted variables (conceptual)
You may be asked:
- identify a test strategy: auxiliary regression, restricted/unrestricted model comparison, or joint significance test.
Example structure:
- Fit unrestricted model including suspected omitted variable.
- Fit restricted model without it.
- Use F-test for the additional regressor(s).
If the variable significantly improves fit, you reject the restricted model.
Functional form: log transforms and interpretation
Exams love interpretation:
- If ( \ln(y)) is regressed on (x):
coefficient on (x) approximates percentage change in (y) for a one-unit change in (x) (for small changes). - If (y) regressed on ( \ln(x)):
coefficient approximates change in (y) for a 1% change in (x) (again, approximation).
If the exam includes multiple transformations, write interpretation carefully and consistently with your model equation.
7) How to structure answers efficiently in UJ-style exam conditions
A scoring rubric often rewards clear steps. When you see a computational question:
- Write the model assumptions relevant to the question (e.g., classical vs heteroskedasticity).
- State what test you’re doing and what the null/alternative is.
- Compute the test statistic step-by-step.
- State degrees of freedom.
- Compare with critical value or interpret p-value.
- Provide one-sentence econometric interpretation.
When you see conceptual questions:
- Don’t just define terms. Provide:
- what assumption fails,
- which test becomes invalid,
- and what remedy is appropriate.
EMT2BB2 Exam Prep Cluster: University of Pretoria (UP) — Econometrics 2B: Time Series Foundations, Stationarity & Practical Estimation
1) Time-series readiness: what “Econometrics 2B” expects beyond raw regression
In University of Pretoria (UP)–type econometrics sequences, time-series concepts often appear as the bridging foundation: even if the exam is still “regression-based,” you are expected to know when regression results become unreliable if the data are nonstationary or correlated over time.
A regression of (y_t) on (x_t) might “look significant,” yet be spurious if variables are nonstationary. EMT2B problems frequently test whether you can recognize that risk and propose correct diagnostics.
2) Stationarity and why it matters for inference
A process (y_t) is (weakly) stationary if:
- (E[y_t]) is constant over time
- (Var(y_t)) is constant
- (Cov(y_t,y_{t-h})) depends only on lag (h), not on (t)
If stationarity fails:
- standard OLS assumptions (especially about error properties) can fail,
- t-tests and F-tests may be misleading.
Example scenario (exam style narrative, but solvable)
Suppose you estimate:
[
\ln(\text{GDP}_t)=\alpha+\beta t + u_t
]
and you test (H_0:\beta=0). If both GDP and trend are deterministic or unit-root-like, the usual distribution may not apply.
Your answer should emphasize: “Stationarity affects the validity of classical inference.”
3) Unit roots and the intuition behind ADF-type testing
UP-style exams frequently touch unit root ideas (Augmented Dickey–Fuller family or intuition).
A simple AR(1) representation:
[
y_t=\rho y_{t-1}+\epsilon_t
]
- If (|\rho|<1): stationary (mean-reverting)
- If (\rho=1): unit root (nonstationary random walk)
ADF rewrites into:
[
\Delta y_t = \alpha + \beta y_{t-1} + \sum_{i=1}^p \gamma_i \Delta y_{t-i} + \epsilon_t
]
- (H_0: \beta=0) (unit root)
- (H_1: \beta<0) (stationary)
Even if you don’t compute a full ADF regression in the exam, you may be asked to:
- identify null/alternative,
- interpret test statistic sign/magnitude,
- explain why the distribution is nonstandard (often you use special critical values).
4) Cointegration and spurious regression avoidance
If both (y_t) and (x_t) are nonstationary but some linear combination is stationary, you may have cointegration:
[
y_t = \beta x_t + u_t
]
If (u_t) is stationary, then the relationship is economically and statistically meaningful even though individual series are not stationary.
Exam tasks might ask:
- define cointegration,
- state its consequence (avoid spurious regression),
- or interpret output from cointegration tests.
Practical interpretation you should practice writing
If the exam reports:
- “Both series are I(1)” (integrated of order 1)
- “Cointegration test rejects the null of no cointegration”
Then:
- you say: “There exists a stable long-run relationship, so regression may be meaningful.”
If no cointegration:
- you say: “Risk of spurious regression; classical inference on levels may be unreliable.”
5) ARIMA basics and how they connect to econometric modeling
Many EMT2B syllabi introduce ARIMA logic at a conceptual or semi-computational level.
ARIMA(p,d,q) indicates:
- p: autoregressive order
- d: differencing to achieve stationarity
- q: moving average order
Example:
- ARIMA(1,0,0): (y_t=\rho y_{t-1}+\epsilon_t)
- ARIMA(0,1,1): (\Delta y_t = \epsilon_t + \theta \epsilon_{t-1})
You should connect ARIMA parameters to behavior:
- high persistence vs mean reversion
- smoothing via MA component
6) Autocorrelation functions: ACF/PACF reasoning
If an exam provides qualitative ACF/PACF shapes, you may be asked to identify model orders. Practice the standard rules:
- For AR(p):
- PACF cuts off after lag p
- ACF tails off gradually
- For MA(q):
- ACF cuts off after lag q
- PACF tails off gradually
- For ARMA: both tail off.
Concrete reasoning example
If:
- ACF drops sharply at lag 2 and near zero afterwards
- PACF shows tail-off
Then you suspect MA(2) structure (or mixed, depending on detail). Many exams accept the correct identification logic even if not the unique model.
7) Estimation in time series: what changes from OLS-only mindset?
In time series econometrics:
- dependence structure affects standard errors,
- forecasting requires model validation,
- residual diagnostics matter.
Even when you still estimate coefficients using regression form, you:
- check residual autocorrelation,
- verify stationarity assumptions (after transformation),
- interpret forecasts as model-based projections.
Residual checks you should mention
- serial correlation in residuals (e.g., LM test or DW if relevant)
- heteroskedasticity (time-varying variance)
- normality sometimes (depending on course)
8) Forecasting and evaluation: exam-ready checklist
When asked to forecast:
- State the forecasting model (ARIMA or regression with time-series errors).
- Compute one-step-ahead forecast using estimated coefficients.
- Explain multi-step logic (recursive forecast).
- Discuss evaluation metrics:
- MAE, RMSE, MAPE (if percentage errors are meaningful)
- Mention diagnostic stability if asked.
Even if numbers are not given, you should know what the outputs mean.
9) How UP-style answers typically score well
For calculation-heavy questions:
- show algebraic substitutions,
- write the test statistic clearly,
- use consistent notation.
For theory questions:
- link each concept to consequences: “Why does it matter?”
- mention at least one remedy (transformations, differencing, robust SE, cointegration modeling).
EMT2BB2 Exam Prep Cluster: Stellenbosch University (SU) — Econometrics 2B: Advanced Regression with Robust Methods & Model Selection
1) Why SU focuses so much on “model credibility”
In Stellenbosch University (SU) course emphasis, exams tend to test not only whether you can compute, but whether you can justify which model is credible. That means:
- you diagnose violations (heteroskedasticity, autocorrelation, endogeneity proxies),
- you propose robust alternatives,
- you select between models logically with tests and information criteria.
2) Model selection tools you must be fluent with
Common criteria:
- Adjusted R-squared
- AIC (Akaike Information Criterion)
- BIC (Bayesian Information Criterion)
Even if formulas are provided, you need to interpret meaning:
- lower AIC/BIC indicates preferred model (balance fit vs complexity)
- BIC penalizes complexity more heavily than AIC.
AIC and BIC formulas (standard form)
If regression uses Gaussian errors:
- [
AIC = -2\ln(L) + 2k
] - [
BIC = -2\ln(L) + k\ln(n)
]
Often exams provide simplified forms in terms of SSR and (n). If so, use the given formula exactly.
3) Robust standard errors: the “what and when” exam question
SU exams often explicitly ask:
- “If heteroskedasticity is present, what should you do?”
- “How do robust SE affect inference?”
You should describe:
- OLS coefficient estimates (\hat{\beta}) do not change when using robust SE; only the estimated covariance matrix changes.
- With robust SE, t and F tests based on “wrong” variance can become more reliable.
Mental model
- Conventional SE assumes constant variance and (often) independence.
- Robust SE replaces the assumption with a general variance structure estimate.
In written answers:
- mention “HC covariance matrix” (heteroskedasticity-consistent),
- or “sandwich estimator” language if taught.
4) Heteroskedasticity-consistent covariance: what the structure means
A general robust covariance matrix can be written as:
[
\widehat{Var}(\hat{\beta})=(X'X)^{-1}(X'\hat{\Omega}X)(X'X)^{-1}
]
where (\hat{\Omega}) is a diagonal matrix with estimated error variances on the diagonal, typically based on (\hat{u}_i^2) scaled appropriately.
Exam writing doesn’t always require you to compute this; but conceptually, it says:
- each observation can contribute differently to variance.
5) Practical endogeneity awareness (conceptual, but crucial)
Although “Econometrics 2B” may not fully switch to full IV/2SLS depth (that might be 2C or a later course), SU exams often include endogeneity discussions as a specification risk:
- Omitted variables
- Measurement error
- Simultaneity
You should know:
- if endogeneity exists, OLS coefficients are biased/inconsistent,
- robust SE fixes standard errors but does not fix endogeneity.
This is a common conceptual trick in exams: students incorrectly believe robust SE solves endogeneity too.
6) Case study: comparing two models for household energy demand
Consider a stylized problem consistent with South African energy contexts:
- Dependent variable: monthly electricity expenditure (y)
- Regressors: income (x_1), household size (x_2), temperature proxy (x_3)
Model A:
[
y_i = \beta_0 + \beta_1 income_i + \beta_2 size_i + \beta_3 temp_i + u_i
]
Model B adds a policy indicator (D_i) for subsidy eligibility:
[
y_i = \beta_0 + \beta_1 income_i + \beta_2 size_i + \beta_3 temp_i + \beta_4 D_i + u_i
]
Exam question type:
- Test if subsidy indicator is significant (t-test for (\beta_4)).
- If significant, use F-test for joint restrictions if multiple subsidy-related dummies exist.
- Diagnose heteroskedasticity (BP/White).
- Choose between A and B using AIC/BIC.
How to write the decision logic cleanly
- If (D_i) is statistically significant and improves AIC/BIC, reject Model A in favor of Model B (if the model remains theoretically justified).
- If heteroskedasticity is present, re-evaluate significance using robust SE.
- If BP test suggests heteroskedasticity, conventional p-values might be misleading.
This case-study format mirrors how SU lecturers often frame applied exam tasks: start with baseline regression, then enhance with diagnostics and selection criteria.
7) Interaction terms and dummies: where students lose marks
SU exams often include interaction effects such as:
[
y_i=\beta_0+\beta_1 x_i+\beta_2 D_i+\beta_3 (x_iD_i)+u_i
]
Interpretation:
- For (D_i=0): effect of (x_i) is (\beta_1)
- For (D_i=1): effect of (x_i) is (\beta_1+\beta_3)
In exam answers, you should explicitly state both regimes.
Common error
Students interpret (\beta_3) as “the effect of x on y in the treated group.” In reality:
- (\beta_3) is the incremental change in the slope due to (D=1).
8) Model diagnostic conclusion writing
A high-scoring SU-style conclusion typically includes:
- “We tested for heteroskedasticity; evidence suggests it is present/absent.”
- “We used robust standard errors to obtain valid inference.”
- “Model selection criteria prefer model X over Y.”
- “The economic interpretation aligns with theory.”
Make sure the conclusion matches the computed results. If you test heteroskedasticity and you reject (H_0), don’t conclude homoskedasticity holds.
EMT2BB2 Exam Prep Cluster: Cape Peninsula University of Technology (CPUT) — Econometrics 2B: Computation Practice for Tests, Confidence Intervals & Applied Regression Output Interpretation
1) How CPUT-style exams often feel: “Show you can do the mechanics”
In Cape Peninsula University of Technology (CPUT)–type teaching and assessment patterns, Econometrics 2B exams commonly combine:
- calculations (t-tests, F-tests, joint significance),
- interpreting regression output (coefficients, SEs, p-values, R-squared),
- short methodological explanations (“why robust SE?” “what does a rejected null mean?”).
This cluster emphasizes exam mechanics: converting given regression output into correct statistical conclusions.
2) Reading regression output like an examiner
When you receive a regression table, it typically includes:
- coefficient estimates
- standard errors
- t-statistics
- p-values
- (R^2) and adjusted (R^2)
- number of observations (n)
Checklist for interpretation
- Identify dependent variable and units
- For each coefficient:
- sign (+/−)
- magnitude (interpret in units or percent if logs)
- statistical significance via p-value or t-statistic
- Report overall fit:
- (R^2) (note: can be misleading when adding variables)
- adjusted (R^2) (preferred for comparing models)
- Confirm sample size (n).
Example interpretation format
- “A one-unit increase in (x_1) is associated with a (\hat{\beta}_1) unit increase in (y), holding other variables constant. At the 5% level, the effect is statistically significant because the p-value is less than 0.05.”
3) Confidence intervals: the fastest way to gain marks
A two-sided ((1-\alpha)\times 100%) confidence interval for (\beta_j):
[
\hat{\beta}j \pm t{\alpha/2,,n-k}\cdot se(\hat{\beta}_j)
]
Exam tasks:
- compute CI endpoints
- interpret whether 0 lies in the CI (significance link)
Key equivalence (often explicitly rewarded)
- If a 95% CI includes 0, then a two-sided test at 5% fails to reject (H_0:\beta_j=0).
- If a 95% CI excludes 0, you reject at 5%.
4) F-tests for overall significance (and why students confuse them)
Overall significance often tests:
[
H_0: \beta_1=\beta_2=\cdots=\beta_{k-1}=0
]
(assuming intercept included and not tested).
But sometimes exam paper asks you to compare two models:
- restricted vs unrestricted.
You must identify the correct df:
- numerator df = number of restrictions q
- denominator df = (n-k) (or other given df in the paper’s convention)
5) Worked mini-practice: from output to conclusions
Assume the exam provides:
- (n=80), (k=5) (including intercept), so df (=75)
- estimated coefficient on (x_2): (\hat{\beta}_2=0.90)
- standard error: (se(\hat{\beta}_2)=0.30)
Compute t-stat:
[
t=\frac{0.90}{0.30}=3.0
]
If the paper uses 5% significance two-sided, approximate critical value for df=75 is about 1.99 (exact table might differ). Since 3.0 > 1.99:
- reject (H_0)
- conclude (x_2) is statistically significant at 5%.
If instead p-value is given as 0.003:
- also reject; p-value method is consistent.
6) Heteroskedasticity test computations (BP/White variants) — how to avoid common arithmetic errors
In BP test problems, exam papers often provide:
- auxiliary regression output including (R^2_{aux})
- you compute (LM = nR^2_{aux})
- compare to a chi-square distribution with df equal to number of restrictions excluding intercept.
BP computation template
- Compute (LM=nR^2_{aux})
- Determine df = number of regressors in auxiliary excluding the intercept (or number of slope terms, depending on instruction)
- Compare to (\chi^2_{df, \alpha})
Common error: mixing df definitions. Always match the paper’s instruction (does it say df equals number of regressors excluding intercept?).
7) Robust regression inference: what changes in your answer
A typical exam instruction:
- “Assume heteroskedasticity is detected. Recompute/interpret the significance using robust SE.”
If the exam gives robust standard errors:
- use them to compute t-statistics
- re-evaluate p-values
- state changed conclusions if relevant.
If the exam does not provide robust SE values but asks conceptually:
- you explain what would happen: coefficients unchanged; standard errors updated; tests may change.
8) Case study: labour earnings model with potential heteroskedasticity
Consider an earnings regression:
[
\ln(wage_i)=\beta_0+\beta_1 education_i+\beta_2 experience_i+\beta_3 tenure_i+u_i
]
Interpretation:
- (\beta_1) is the approximate percent change in wage for one more unit of education (depending on education scale; if education measured in years, then it’s percent per year increase).
- experience and tenure similarly.
Why heteroskedasticity plausible here
In earnings:
- variance often increases with income/education.
- groups differ (urban vs rural, sector, etc.)
- thus errors may be heteroskedastic.
Exam prompt style
- Run BP test: reject homoskedasticity.
- Conclude standard errors need correction.
- Use robust SEs.
- Reinterpret (\beta_1) and (\beta_2).
This resembles many applied econometrics tasks found in teaching practice at technical universities: labour, household survey, and policy data.
9) Autocorrelation test and interpretation with time-ordered data
Suppose the exam provides a regression with quarterly data for 40 quarters:
- (n=40)
- include an intercept and k regressors
- ask: “Use DW statistic to test for first-order autocorrelation.”
You should:
- compute or use given DW value
- state whether autocorrelation is positive/negative based on distance from 2
- if bounds are provided, decide accordingly.
Then interpret: if autocorrelation exists, conventional SE may be wrong; propose remedy.
10) Strategy for multi-part questions in CPUT papers
CPUT exams often combine several small parts. Use a stable template:
(a) Identify null hypothesis.
(b) Compute required statistic.
(c) State df.
(d) Compare to critical value.
(e) Give conclusion in words with econometric meaning.
(f) If asked, provide remedy (robust SE, re-specification).
This structure reduces lost marks due to missing df or misinterpreting the direction of rejection.
Exam-Ready Synthesis: Cross-Institution “EMT2BB2” Problem Types & How to Answer Them
1) The five recurring problem types you must master
Across UJ, UP, SU, and CPUT–style assessments, you will repeatedly meet these patterns:
-
Multiple regression coefficient interpretation
- in levels or logs
- with dummies and interactions
-
t-tests and confidence intervals
- compute t-stat, CI endpoints
- interpret p-values
-
F-tests for joint restrictions or model comparison
- restricted vs unrestricted SSR logic
- correct degrees of freedom
-
Diagnostic testing: heteroskedasticity and autocorrelation
- BP/White logic
- DW logic or conceptual autocorrelation discussion
-
Model selection and specification improvements
- AIC/BIC or adjusted R-squared
- omitted variable reasoning
- robust SE vs endogeneity caution
2) A disciplined workflow for computations
When you see a question, follow:
- Extract given numbers carefully
- (n), (k), SSR or (R^2) values
- coefficient and SE values
- Write the formula relevant to what’s asked
- Compute step-by-step
- State the correct distribution and df
- Conclude with econometric meaning
This workflow is what differentiates “partial credit” answers from full-credit ones.
3) A unified vocabulary for interpretation (use consistently)
Use these phrases (and make them consistent with results):
- “We reject (H_0)” → implies statistic exceeds critical value (or p-value < α)
- “We fail to reject (H_0)” → implies insufficient evidence
- “Significant at the 5% level” → p-value < 0.05
- “Holding other variables constant” → partial effect interpretation
- “Robust standard errors address heteroskedasticity” → changes inference, not point estimates (unless model re-estimation is explicitly requested)
- “Robust SE does not fix endogeneity” → conceptual caution
4) Consistency practice: avoiding contradictory statements
Common exam pitfalls include:
- Declaring heteroskedasticity present, then interpreting conventional p-values without mentioning robust SE (unless the question explicitly uses conventional SE).
- Rejecting homoskedasticity but then concluding “assumptions are satisfied.”
- Calculating F with wrong df or mixing df1 and df2.
To avoid contradictions:
- write the assumption status immediately after diagnostic tests.
- if the question changes the variance assumptions, update inference accordingly.
5) Time-series caution statement (short and decisive)
If the exam touches stationarity/cointegration:
- always connect to inference reliability.
- don’t over-claim. A good line is:
- “If series are nonstationary and cointegration is not established, regression in levels can be spurious; inference based on classical assumptions may be invalid.”
6) Final checklist before you start writing exam answers
- Have you stated null and alternative correctly?
- Did you compute the right statistic (t vs F vs chi-square)?
- Did you use the correct df?
- Did you interpret the sign and magnitude correctly?
- Are your conclusions consistent with the diagnostic results?
- If robust methods are needed, did you incorporate them into inference?
Suggested High-Impact Revision Plan (Last-Minute but Still Thorough)
1) Day-by-day structure (repeat as needed)
Even if your exam is soon, a structured plan improves recall:
Day 1 (Regression + interpretation)
- coefficients in levels/logs
- dummies and interactions
- t-tests + CIs
Day 2 (Hypothesis testing)
- joint F-tests using SSR logic
- overall significance
- model comparison interpretation
Day 3 (Heteroskedasticity + robust inference)
- BP/White concepts
- robust SE consequences
- written interpretation and remedy
Day 4 (Autocorrelation + time series basics)
- DW intuition
- stationarity and unit roots
- cointegration/spurious regression caution
Day 5 (Model selection + diagnostics + synthesis)
- AIC/BIC interpretation
- residual diagnostics summary
- integrated problem practice
2) How to practice so you actually improve score
Practice in “exam mode”:
- 45–60 minutes per problem set
- timed writing
- strict checking of df and formula placement
- review mistakes and write a corrected solution immediately.
When reviewing, classify errors:
- calculation error
- formula selection error
- interpretation error
- df/critical value error
Then target the specific category in the next practice session.
Closing Exam Mindset for EMT2BB2
Econometrics 2B is ultimately about credible inference: you learn to question assumptions, test them, and adjust your methodology and interpretation accordingly. The top strategy is not memorizing more formulas—it is building consistent reasoning chains: model → assumptions → diagnostics → correct test → correct conclusion. If you can do that under exam pressure, you will consistently convert partial understanding into full marks across computation-heavy and theory-heavy questions.
