ISS 2016 Statistics Paper-2 Solution: Question 18 (MSE Criterion for Selecting Estimators)

Continuing our question-by-question walk through the ISS Statistics Paper-2 previous year papers, today we pick up right where Question 17 left off and tackle Question 18 from the 2016 paper. This one looks deceptively short, but it hides a conceptual trap that catches a lot of aspirants who have only memorised formulas without understanding what they are for.

Quick Summary

• Topic: Criteria used to compare and select statistical estimators

• Question reference: ISS 2016, Statistics Paper-2, Question 18

• Correct answer: (c) Estimates

Here is the question exactly as it appears in the test booklet.

Question 18.Mean square error criteria are used for selection of which one of the following?

• (a) Explanatory variables

• (b) Model

• (c) Estimates

• (d) Response variables

Step 1: Understand What a "Selection Criterion" Is Being Asked About

Before jumping to the answer, slow down and ask a basic question: in statistics, what kinds of things do we actually "select" or "choose between"? There are really three different situations where the word "selection" comes up, and this question is testing whether you can tell them apart:

• Selecting which explanatory variables (regressors) to include in a regression equation — this is called model selection, and it typically uses tools like R², adjusted R², Mallows' Cp, AIC, or BIC.

• Selecting which estimator to use for a given unknown parameter, when more than one candidate estimator is available — this uses criteria like unbiasedness, efficiency, consistency, and importantly, the Mean Square Error (MSE).

• Checking properties of the response variable itself, such as its distribution or variance — this is a data property, not a selection criterion at all.

Once you separate these three ideas clearly, the question becomes much easier, because "Mean Square Error criterion" is a phrase with one very specific, textbook meaning in estimation theory.

Step 2: Define the Mean Square Error of an Estimator

Let θ be an unknown population parameter (for example, a regression coefficient β), and let θ̂ (read "theta hat") be an estimator of θ built from sample data. The Mean Square Error of θ̂ is defined as:

MSE(θ̂) = E[(θ̂ − θ)²]

This is literally "the average squared distance between the estimator's value and the true parameter." Now here is the key algebraic step that every serious ISS aspirant must be able to derive from scratch, not just recall:

MSE(θ̂) = E[(θ̂ − E(θ̂) + E(θ̂) − θ)²]

Add and subtract E(θ̂) inside the bracket, then expand the square:

MSE(θ̂) = E[(θ̂ − E(θ̂))²] + 2·E[(θ̂ − E(θ̂))(E(θ̂) − θ)] + [E(θ̂) − θ]²

The middle cross-term vanishes because E(θ̂) − θ is a constant (not random) and E[θ̂ − E(θ̂)] = 0 by definition of expectation. So we are left with the famous decomposition:

MSE(θ̂) = Var(θ̂) + [Bias(θ̂)]²

where Bias(θ̂) = E(θ̂) − θ. This single formula is the entire reason the MSE criterion exists as a tool for comparing estimators.

Step 3: Why Isn't Unbiasedness Enough by Itself?

A natural first instinct is to say "just pick the unbiased estimator, since Bias = 0 automatically minimises the second term." But that argument breaks down the moment there are two or more unbiased estimators to choose from, or when a slightly biased estimator has a dramatically smaller variance.

Here is a small worked illustration. Suppose two estimators θ̂₁ and θ̂₂ both exist for the same parameter θ:

• θ̂₁ is unbiased, so Bias(θ̂₁) = 0, but Var(θ̂₁) = 25

• θ̂₂ has a small bias, Bias(θ̂₂) = 2, but Var(θ̂₂) = 4

Compute the MSE of each directly from the formula in Step 2:

MSE(θ̂₁) = 25 + 0² = 25

MSE(θ̂₂) = 4 + 2² = 4 + 4 = 8

Even though θ̂₂ is biased, its MSE (8) is far lower than the unbiased θ̂₁'s MSE (25). If your goal is to get estimates that are, on average, closer to the true value, θ̂₂ is the better choice. This exact trade-off is why techniques like ridge regression (which deliberately introduces a small bias into the OLS coefficient estimator in exchange for a large reduction in variance) are justified using the MSE criterion, not the unbiasedness criterion alone.

Pro Tip: The trickiest part of this question is resisting the urge to associate "Mean Square Error" with R² and model-fit statistics just because both involve squared terms and both show up in regression chapters. Keep a strict mental separation: R², adjusted R², AIC and Cp answer the question "which set of explanatory variables should be in my model?", while the MSE criterion answers the completely different question "which estimator of a given parameter should I trust more?" The way this distinction is usually drilled into students preparing for ISS — including at a place like Sunrise Classes — is to always ask first "am I comparing variables, or am I comparing estimators of the same quantity?" before touching any formula.

Step 4: Eliminate the Other Three Options

With the definition and the bias-variance decomposition in hand, we can now rule out the remaining choices with concrete reasons rather than guesswork:

• (a) Explanatory variables — deciding which regressors to keep is a model-building decision, driven by R², adjusted R², AIC/BIC, or significance tests on coefficients, not by an MSE comparison of estimators.

• (b) Model — comparing entire competing models (say, a linear model versus a quadratic one) again falls under model-selection criteria such as AIC, BIC, or cross-validated prediction error, which are conceptually distinct from the estimator-level MSE defined in Step 2.

• (d) Response variables — the response variable y is simply the data you are trying to explain; you don't "select" it using an estimation-theory criterion at all.

Only option (c) Estimates (i.e., estimators) matches the actual technical definition of the MSE criterion: it is a tool for comparing candidate estimators of the same unknown parameter and picking the one that minimises Var(θ̂) + [Bias(θ̂)]².

Final Answer

The correct response is (c) Estimates. This matches the official answer key published alongside the 2016 ISS Statistics Paper-2, so our independent derivation and the official key are in full agreement — no discrepancy to flag for this question.

Why This Question Matters

This question is a favourite type in ISS Paper-2 because it tests conceptual clarity rather than calculation speed. The bias-variance decomposition of MSE is not just an exam formula — it is the backbone of modern estimation theory, showing up again in shrinkage estimators, ridge and Lasso regression, empirical Bayes methods, and machine learning model evaluation. Students who take the time to derive MSE(θ̂) = Var(θ̂) + [Bias(θ̂)]² from first principles, rather than memorising it, find that an entire family of "which estimator is better" questions across Paper-1 and Paper-2 becomes almost automatic to answer.

Frequently Asked Questions

Is the Mean Square Error the same as the variance of an estimator?

No. Variance is only one part of the MSE. The MSE equals variance plus the square of the bias, so the two are equal only in the special case where the estimator is unbiased. Whenever an estimator has non-zero bias, its MSE will always be strictly larger than its variance.

Can an estimator with lower variance ever be a worse choice than one with higher variance?

Yes, if the low-variance estimator carries a large enough bias. Because MSE combines both terms, a small reduction in variance is not automatically a win — you have to check whether the added bias outweighs the variance saved. This is exactly why the MSE criterion, rather than variance alone, is used to rank estimators.

Why don't we just always use the unbiased estimator with the smallest variance?

Among all unbiased estimators, picking the one with the smallest variance is indeed a well-known rule, called the Minimum Variance Unbiased Estimator (MVUE) criterion. But this rule only applies within the pool of unbiased estimators. The MSE criterion is more general because it also allows biased estimators to compete, which is essential in situations like ridge regression or shrinkage estimation.

Is Mean Square Error used anywhere outside estimation theory?

Yes, the same squared-error idea reappears in forecasting and prediction accuracy (where it is often called Mean Squared Prediction Error) and in machine learning as a loss function for regression models. The core arithmetic is identical, even though the object being evaluated (a predicted value instead of a parameter estimate) is different.

How is this different from R-squared used in model selection?

R² measures how much variation in the response variable is explained by a fitted model as a whole, and it is used to compare different sets of explanatory variables. The MSE criterion instead compares two or more candidate formulas (estimators) for the same single parameter. They answer different questions and should never be treated as interchangeable in an exam setting.

What is the safest way to revise this topic before the ISS exam?

Write out the MSE decomposition from scratch — MSE(θ̂) = E[(θ̂ − θ)²] expanded step by step into Var(θ̂) + [Bias(θ̂)]² — at least a few times without looking at notes, then apply it to two or three toy numerical examples like the one in Step 3 above. Being able to reproduce the derivation under exam pressure is far more reliable than trying to memorise the final formula alone.

If anything in this derivation felt unclear, or if you spotted a different way to think about the bias-variance trade-off, drop a comment below — doubts like this are exactly what this series exists to clear up. And if a fellow ISS aspirant is stuck on the same question, feel free to share this post with them.