Econometrics assessments vary by course, so check your syllabus for the topics and notation you are expected to know. These 10 multiple-choice questions cover common undergraduate econometrics concepts. Each explanation identifies why the answer works and why a tempting alternative does not.
Select an option in each question, then use its feedback to check both your answer and the reasoning behind it. Keep track of the topics where you are unsure: a correct guess may still point to a concept worth revisiting.
OLS and interpretation
Question 1
This question tests the direction of omitted-variable bias when the signs of the relevant relationships are known. A question on this topic looks like this:
The true model is y = β₀ + β₁x + β₂z + u, where β₂ > 0. You estimate a regression of y on x but omit z. If Cov(x,z) < 0, what is the direction of the omitted-variable bias in the estimated coefficient on x?
The bias is β₂ Cov(x,z)/Var(x). Since β₂ is positive, Cov(x,z) is negative and Var(x) is positive, the bias is negative. Option 0 reverses the sign; Var(u) does not determine the direction of this bias.
The useful habit is to write the sign formula before choosing an answer. The bias direction depends on both the effect of the omitted variable on the outcome and its association with the included regressor—not simply on whether the omitted variable is important.
Question 2
This question checks the distinction between OLS coefficient estimates and their conventional standard errors under heteroskedasticity. A question on this topic looks like this:
Assume the conditional mean is correctly specified and regressors are exogenous, but the error variance differs across observations. What is the usual consequence for OLS inference?
Under exogeneity, heteroskedasticity alone does not bias the OLS coefficients, but the usual homoskedastic variance formula is generally inappropriate. Heteroskedasticity-robust standard errors address that inference problem. Option 2 incorrectly claims robust standard errors fix inconsistency from other problems, such as endogeneity.
Robust standard errors change the estimated uncertainty, not the fitted coefficients or the assumptions that make those coefficients meaningful. They do not repair omitted-variable bias, reverse causality or a wrongly specified conditional mean.
Question 3
This question tests interpretation of a coefficient in a log-level model. A question on this topic looks like this:
A regression estimates ln(wage) = β₀ + 0.08 education + u. What is the approximate interpretation of one additional year of education?
In a log-level model, a one-unit increase in education changes log wage by 0.08, approximately an 8% increase in wage. The exact percentage change is 100(exp(0.08) − 1), about 8.3%. Option 0 confuses a log outcome with a level outcome.
For larger coefficients, the approximation can be less accurate, so calculate 100(exp(β) − 1) when an exact percentage interpretation is needed. Also state what is held fixed: the coefficient is a conditional association unless the design supports a causal interpretation.
Instruments and panel data
Question 4
This question checks the two core conditions for an instrument in an instrumental-variables analysis. A question on this topic looks like this:
A researcher uses z as an instrument for an endogenous regressor x in an outcome equation. Which pair of conditions is required for z to be a valid instrument?
The instrument must be relevant (associated with x) and valid: it must satisfy the exogeneity and exclusion requirements, so it is uncorrelated with the structural error and affects the outcome through x rather than a separate pathway. Option 0 confuses correlation with the outcome with the required conditions and adds an unnecessary normality claim.
A first-stage association supports relevance, but by itself it cannot establish the exclusion restriction. In an applied answer, explain why the instrument should not share unobserved causes with the outcome or affect the outcome through another route.
Question 5
This question tests what individual fixed effects do to regressors that do not change over time. A question on this topic looks like this:
A panel regression includes individual fixed effects. Gender is recorded for each person and does not change over the observed period. What happens to the separate coefficient on gender?
Individual fixed effects absorb all characteristics that are constant within a person over time, including gender in this example. Its separate coefficient is therefore not identified in that specification. Option 0 describes variation that is absent: gender does not vary within a person.
Fixed effects can still estimate coefficients on variables that change over time within individuals. Be precise about the source of identification: the model uses within-person changes, rather than the between-person differences absorbed by the fixed effects.
Difference-in-differences and time series
Question 6
This question tests the condition that gives a basic difference-in-differences estimate a causal interpretation. A question on this topic looks like this:
A difference-in-differences model compares treated and control groups before and after a policy. Under which key condition can the interaction coefficient be interpreted as the policy's average treatment effect on the treated?
The parallel-trends condition says that, without treatment, the treated and control groups would have experienced the same trend. Equal starting levels are not required, so option 0 is a common but incorrect substitute. The causal interpretation also relies on other design conditions, such as no anticipation or relevant spillovers.
A good answer names the counterfactual trend, not merely the observed similarity of the groups. In applied work, pre-treatment trends can help assess plausibility, but cannot prove that the unobserved post-treatment counterfactual trends would have remained parallel.
Question 7
This question tests the stationarity condition for a first-order autoregressive process. A question on this topic looks like this:
For the AR(1) process yₜ = c + ρyₜ₋₁ + εₜ, with a white-noise error, which condition gives covariance stationarity?
An AR(1) process is covariance stationary when the autoregressive coefficient has absolute value below one. Option 0 is too restrictive: many nonzero values satisfy the condition. Values with |ρ| ≥ 1 do not give the usual stationary AR(1) process.
The absolute value matters: a negative coefficient can still produce stationarity if it is greater than −1. When a time-series question asks about long-run behaviour, check the root or coefficient condition before using stationary-process formulas.
Question 8
This question checks whether you know the null hypothesis in the commonly used augmented Dickey–Fuller test. A question on this topic looks like this:
In the standard augmented Dickey–Fuller test for a unit root, what is the null hypothesis?
The standard ADF null is a unit root; the alternative is stationarity, with the precise form depending on whether the test includes a constant or trend. Option 2 confuses the test with a test of residual autocorrelation. Failure to reject the null is not proof that a unit root is present.
Check the version of the test being used, including whether a constant or trend is included, because that affects the alternative hypothesis and critical values. Do not describe a failure to reject as confirmation: it means the test did not provide sufficient evidence against its null at the chosen significance level.
Regression limits and nonlinear models
Question 9
This question tests the usual consequence of strong multicollinearity when the other OLS assumptions hold. A question on this topic looks like this:
Two regressors are highly correlated but not perfectly collinear. Assume the regression is otherwise correctly specified and regressors are exogenous. What is the usual effect?
High, imperfect multicollinearity does not by itself bias OLS under exogeneity, but it can inflate coefficient variances and make separate effects difficult to estimate precisely. Option 0 confuses multicollinearity with endogeneity. Removing a regressor requires a substantive reason, not correlation alone.
Look at the question being asked: imprecise individual estimates can coexist with a model that predicts the outcome reasonably well. Perfect collinearity is different; it prevents separate estimation of the affected coefficients.
Question 10
This question tests the interpretation of a coefficient in a logit model. A question on this topic looks like this:
In a binary logit model, the coefficient on x is 0.70. Holding other variables fixed, what does this coefficient imply for a one-unit increase in x?
A logit coefficient is a change in log odds. Exponentiating 0.70 gives an odds ratio of about 2.01, so the odds are multiplied by about two. Option 2 incorrectly treats an odds ratio as a constant change in probability; the probability change depends on the starting probability and other covariates.
To report probability changes, calculate a marginal effect at specified covariate values or an average marginal effect. Do not translate a logit coefficient directly into percentage points: odds and probabilities are related, but they are not the same quantity.
Review the errors
Sort missed questions by the underlying concept rather than by question number. For example, an error on the ADF null and an error on AR(1) stationarity both concern time series, but they require different corrections: one is a hypothesis-test definition and the other is a process condition. Practise explaining the distinction in one or two sentences, then check it against your course notes.
The practice record below shows how errors can be grouped by topic. Its frequencies represent a sample practice history, not a claim about how often any topic appears in a particular course or examination.
If you lose the same conceptual mark repeatedly, turn the correction into a short prompt with a precise answer. For instance, distinguish instrument relevance from the exclusion restriction rather than memorising only the word “valid”.
You lost this mark twice: in a standard augmented Dickey–Fuller test, what is the null hypothesis, and what does failure to reject it mean?
How MySummaries helps
MySummaries can turn your own econometrics notes into revision boards, flashcards and written practice marked against those notes. Use it to organise the concepts behind questions you miss, then revisit them with spaced repetition. Open MySummaries.