top of page

ISS 2017 Statistics Paper-2 Solution: Question 7 (Gauss-Markov Model Assumptions)

6 hours ago
6 min read

We're continuing our question-by-question walk through the ISS Statistics Paper-2 previous year papers, and today we land on Question 7 from the 2017 paper. This one sits right in the middle of a small cluster of questions on the Gauss-Markov model, so if you've been following along with Questions 5 and 6 from the same paper, this one will feel like a natural next step.

Quick Summary

  • Topic:Assumptions of the Gauss-Markov linear model

  • Question Reference:ISS 2017, Statistics Paper-2, Question 7

  • Correct Answer:(b) 2 and 3 only

The Question, As Asked

In the Gauss-Markov model Y = Xβ + ε, which of the following assumptions are made?


  1. X is a matrix of full rank

  2. E(ε) = 0

  3. Var(ε) = σ²I


Select the correct answer using the code given below.


  • (a) 1 and 2 only

  • (b) 2 and 3 only

  • (c) 1 and 3 only

  • (d) 1, 2 and 3

Step 1: Write Down What the Gauss-Markov Model Actually Says

Before picking apart the statements, let's be completely clear about the setup. We have


Y = Xβ + ε


where Y is an n × 1 vector of observations, X is an n × p matrix of known constants (the "design matrix"), β is a p × 1 vector of unknown parameters we want to estimate, and ε is an n × 1 vector of random errors. This is the general linear model, and the Gauss-Markov theorem is the result that tells us when the least squares estimator of an estimable function of β is the Best Linear Unbiased Estimator (BLUE).


The whole question boils down to one thing: what do we actually have to assume about ε and X for the Gauss-Markov theorem to hold?

Step 2: Check Statement 2 — E(ε) = 0

This says the errors have zero mean, i.e., on average, the random disturbance doesn't push Y systematically above or below Xβ. This is absolutely essential. If E(ε) ≠ 0, then E(Y) ≠ Xβ, and the entire model is misspecified before we even start estimating anything — β would no longer represent the "true" linear relationship. So Statement 2 is a core assumption.Without E(ε) = 0, unbiasedness of any estimator of β can't even be discussed meaningfully, because the model itself would be biased from the start.

Step 3: Check Statement 3 — Var(ε) = σ²I

This says the errors are homoscedastic (same variance σ² for every observation) and uncorrelated with each other (the off-diagonal entries of the covariance matrix are zero, which is what the identity matrix I encodes). This is also a standard Gauss-Markov assumption. It's precisely this condition — constant variance, zero covariance — that lets the Gauss-Markov theorem prove the least squares estimator has the smallest variance among all linear unbiased estimators. If Var(ε) were something other than σ²I (say, heteroscedastic or correlated errors), ordinary least squares would still be unbiased, but it would no longer be the best (minimum variance) linear unbiased estimator — you'd need Generalized Least Squares instead. So Statement 3 is also a core assumption.

Step 4: Check Statement 1 — X Is a Matrix of Full Rank

This is where most students trip up, and it's exactly why the question is testing this particular point. It feels natural to assume X must be full rank, because in an introductory regression course, full rank is often assumed so that (X′X) is invertible and β has a unique least squares solution β̂ = (X′X)⁻¹X′Y.


But the Gauss-Markov model, as a general framework, does not require X to be of full rank. In fact, this exact 2017 paper tests this idea directly in the questions right before this one: Question 6 explicitly discusses the generalized inverse X⁻ of X, used precisely because X may not be of full column rank (for example, in ANOVA-type designs with redundant dummy variables). When X is not full rank, (X′X) is singular, and we can't invert it directly — but we can still estimate any estimable linear function of β using a generalized inverse, and the Gauss-Markov theorem still applies to those estimable functions.


So full rank of X is a convenience assumption for simplifying the algebra in some textbook presentations, not a requirement of the Gauss-Markov model itself. Statement 1 is therefore not one of the defining assumptions.


Pro Tip: Whenever a question lists "X is full rank" alongside E(ε) = 0 and Var(ε) = σ²I, treat it as a trap option unless the question explicitly restricts itself to the full-rank (regression) case. The two genuinely defining assumptions of the Gauss-Markov setup are always about the error vector ε — its mean and its covariance structure — never about the rank of X. This is one of those distinctions that's usually taught by first solving the non-full-rank case (g-inverse, estimable functions) before the full-rank case, precisely so this kind of mix-up doesn't happen on exam day.

Step 5: Put It Together

We've established:


  • Statement 1 (full rank of X) —not a Gauss-Markov assumption

  • Statement 2 (E(ε) = 0) —is a Gauss-Markov assumption

  • Statement 3 (Var(ε) = σ²I) —is a Gauss-Markov assumption


So the correct combination is 2 and 3 only, which corresponds to option (b).


This matches the official ISS 2017 Statistics Paper-2 answer key, which also records option (b) as correct. Our independent derivation and the official key agree.

Why This Question Matters

This question is a great test of whether you understand the Gauss-Markov theorem conceptually, rather than just having memorized the OLS formula β̂ = (X′X)⁻¹X′Y. The ISS exam frequently builds a short run of 3-5 questions around one theme — here, Questions 5 through 8 all revolve around the general linear model, ANOVA, and the Gauss-Markov setup — and examiners deliberately plant a full-rank vs. non-full-rank distinction to separate students who understand estimability theory from those who only know the textbook regression case. Getting comfortable with generalized inverses and estimable functions pays off across several questions at once, not just this one.

Frequently Asked Questions

Is X being full rank ever required in the Gauss-Markov framework?

Not for the Gauss-Markov theorem to hold in general. Full rank is only needed if you want a unique, directly invertible least squares estimator β̂ = (X′X)⁻¹X′Y for every individual parameter in β. When X is not full rank, you instead work with estimable linear functions of β and use a generalized inverse.

What does Var(ε) = σ²I actually mean in plain words?

It means two things combined: every error term has exactly the same variance σ² (homoscedasticity), and no two error terms are correlated with each other (the covariance between any pair of errors is zero). Together these give the identity matrix I when you scale by σ².

What happens to OLS if Var(ε) ≠ σ²I?

OLS remains unbiased, but it's no longer the minimum-variance linear unbiased estimator. In that situation, Generalized Least Squares (GLS), which accounts for the actual covariance structure, becomes the efficient choice instead.

How is this question connected to Questions 5 and 6 in the same paper?

Questions 5 and 6 build up the machinery of the general (possibly non-full-rank) linear model, including the generalized inverse X⁻. Question 7 then directly tests whether you internalized that full rank was never a baseline assumption — it's a nice example of how the ISS paper sequences related concepts together.

Why does E(ε) = 0 matter so much for unbiasedness?

Because E(Y) = E(Xβ + ε) = Xβ + E(ε). If E(ε) ≠ 0, then E(Y) ≠ Xβ, which means even the true, correctly specified parameter β would not match what the model predicts on average — unbiasedness of any estimator becomes impossible to define sensibly.

Is this type of "select the correct assumptions" question common in ISS Paper-2?

Yes, this multi-statement, code-based format (where you pick which numbered statements are true) appears repeatedly throughout ISS Statistics Paper-2, especially around linear models, ANOVA, and hypothesis testing topics. Practicing enough of these builds the pattern recognition needed to quickly rule out plausible-sounding but incorrect statements like Statement 1 here.


Have a doubt about this derivation, or want to see a related concept explained differently? Drop a comment below — and if this helped clear things up, consider sharing it with a fellow ISS aspirant who's working through the same paper.

Recent Posts

See All

Comments


bottom of page