ISS 2017 Statistics Paper-2 Solution: Question 9 (R² in Nested Regression Models)
This post continues our ongoing question-by-question walkthrough of the ISS Statistics Paper-2 previous year papers. Today we move to Question 9 of the 2017 paper, which tests a very fundamental but frequently misunderstood fact about regression: what happens to R² when you add more explanatory variables to a model. We will derive the answer from scratch rather than just quoting a rule.
Quick Summary
Topic:Nested linear regression models and monotonicity of R²
Question Reference:ISS 2017, Statistics Paper-2, Question 9
Correct Answer:Option (b) — R²A ≤ R²B
The Question
Consider the following two models with E(ε) =0 and Var(ε) = σ²I:
Model A :Y =X 1β1 + ε
Model B :Y =X 1β1 +X 2β2 + ε
Which one of the following is correct?
Step 1: Understand What Model A and Model B Actually Are
Look carefully at the two models. Model B has two blocks of regressors,X 1 with coefficient vector β1, and X 2 with coefficient vector β2. Model A has only the first block.
Here is the key observation a beginner often misses: Model A is not some unrelated, separate model. It is exactly what you get from Model B if you force β2 =0. In other words, every possible fitted value that Model A can produce is also a possible fitted value that Model B can produce (just set the β2 part to zero). But Model B can additionally produce many fitted values that Model A cannot, because it has the freedom to choose β2 to be non-zero.
So Model A's set of possible fits is a subset of Model B's set of possible fits. This single fact is what drives the entire answer, so hold onto it.
Step 2: Write R² in Terms of Sums of Squares
For any linear regression fitted by ordinary least squares (OLS), we define:
Total Sum of Squares, TSS = Σ(yi − ȳ)² — this measures the total variation in the observed Y values around their mean. It depends only on the data Y, not on which model (A or B) we use. So TSS is the same number for both models.
Residual Sum of Squares, RSS = Σ(yi − ŷi)² — this measures how much is left unexplained after fitting the model. This number is different for Model A and Model B, because the fitted values ŷi are different.
The coefficient of determination is defined as:
R² = 1 − (RSS / TSS)
So to compare R²A and R²B, since TSS is common to both, we only need to compare RSSA and RSSB.
Step 3: Use What OLS Actually Minimizes
OLS estimation, by definition, finds the coefficient values that minimize RSS over all coefficients available in that model.
For Model A, OLS searches over all possible β1 values only, and finds the β̂1 that gives the smallest possible RSS achievable using X 1 alone. Call this minimum value RSSA.
For Model B, OLS searches over all possible (β1, β2) pairs together, and finds the pair that gives the smallest possible RSS achievable using both X 1 and X 2. Call this minimum value RSSB.
Now here is the crucial logical step. The search space for Model B (all pairs (β1, β2))contains every point of the form (β1,0), which is exactly Model A's search space in disguise. So whatever smallest RSS Model A managed to achieve by picking its best β̂1, Model B could achieve that exact same RSS value too, simply by choosing β2 =0 and the same β1 = β̂1. That means RSSA is one of the candidate values available to Model B's minimization.
Since Model B's OLS procedure picks the minimum RSS over a set of candidates that includes RSSA, the true minimum RSSB can only be less than or equal to RSSA. It cannot possibly be larger, because RSSA itself was already an available option.
Pro Tip: Whenever you see two models where one is a special case of the other (obtained by setting some coefficients to zero or adding a linear restriction), always ask: "which model's search space contains the other's?" The bigger search space can never do worse at minimizing RSS than the smaller one — at absolute worst it ties. This one comparison (RSSrestricted vs RSSfull) is also exactly the logic behind the F-test for testing linear restrictions in a general linear model, so understanding it here pays off again later in the paper. This is usually taught by working through the restricted-vs-unrestricted RSS comparison on a small numerical example first, and only writing down the general argument once the picture is intuitive.
So we have established: RSSB ≤ RSSA.
Step 4: Translate the RSS Inequality into an R² Inequality
We now have RSSB ≤ RSSA, with the same TSS for both models. Since R² = 1 − RSS/TSS, a smaller RSS corresponds to a larger R² (the subtraction flips the direction of the inequality). So:
RSSB ≤ RSSA ⇒ RSSB/TSS ≤ RSSA/TSS ⇒ 1 − RSSB/TSS ≥ 1 − RSSA/TSS ⇒ R²B ≥ R²A
Equivalently, R²A ≤ R²B, which is exactly option (b).
Step 5: Why Not the Other Options?
Option (a), R²A = R²B, would only be true in the special case where the extra regressors in X 2 happen to add absolutely nothing useful (their OLS coefficients come out to exactly zero on this particular data). That is a possible special case, not a general rule, so it cannot be the universally correct statement.
Option (c), R²A ≥ R²B, claims the opposite direction of what we proved and would mean that adding more information to a model makes it fit worse under OLS, which contradicts the minimization argument in Step 3.
Option (d), R²A ≠ R²B, rules out equality altogether, but as noted in option (a)'s discussion, equality is a perfectly valid corner case (when the added regressors are totally useless for this sample), so a statement that always forbids equality is also not universally true.
Only option (b) holds in every case, including the corner case of equality, which is why it is the correct choice.
Final Answer
The correct answer is (b) R²A ≤ R²B. This matches the official answer key for Series A of the ISS 2017 Statistics Paper-2 exam, which lists the answer to Question 9 as B.
Why This Question Matters
This question is really testing whether a candidate understands R² as a byproduct of an optimization problem (RSS minimization) rather than as some abstract "goodness" score. That understanding is what lets you immediately see why R² only climbing when you add variables does not mean those variables are actually useful — which is precisely why practitioners prefer adjusted R² or an F-test for model comparison instead of raw R². Questions built around nested models (one model being a restricted version of another) show up repeatedly across ANOVA, general linear hypothesis testing, and model selection topics in this paper, so this single idea has a lot of leverage once it clicks.
Frequently Asked Questions
Does R² always strictly increase when you add a new variable to a regression model?
No, it never decreases, but it does not have to strictly increase either. If the new variable is completely useless for predicting Y in that particular sample, OLS can set its estimated coefficient to exactly zero, leaving RSS and therefore R² unchanged. In virtually all real datasets with continuous variables, however, R² increases by at least a tiny amount just by chance.
If R² always goes up (or stays the same) when adding variables, why not just add every variable possible?
Because a higher raw R² does not mean a better or more meaningful model — it can simply reflect overfitting to noise in the sample. This is exactly why adjusted R², which includes a penalty for the number of regressors, or a formal F-test for the significance of the added variables, is used instead of comparing raw R² values directly.
How is this question related to the F-test for linear restrictions?
The F-test for testing whether β2 = 0 in Model B is built directly from the difference (RSSA − RSSB). Since we proved this difference is always ≥ 0, the F-statistic used in that test is guaranteed to be non-negative, which is a nice consistency check on the whole framework.
What does "nested models" mean in this context?
Two models are nested when one model can be obtained from the other by imposing restrictions on its parameters — here, Model A is Model B with the restriction β2 =0. Whenever models are nested like this, their minimized RSS values (and hence R² values) are always ordered in a predictable direction, exactly as derived above.
Does the same R² ≤ R² logic apply if we instead compare adjusted R² values?
No, and this is a common trap. Adjusted R² includes a correction for the number of parameters in the model, and it can actually decrease when you add a regressor that does not improve the fit enough to offset that penalty. The monotonic, never-decreasing property proved in this question applies only to the plain (unadjusted) R².
Is this result specific to OLS, or does it hold for any estimation method?
It specifically relies on OLS being a minimization procedure over a search space that grows as you add regressors. If a different estimation method were used that does not minimize RSS (for example, some constrained or biased estimator), this particular guarantee would not automatically carry over.
If anything in this derivation felt unclear, drop a comment with the exact step you got stuck on, and feel free to share this with a fellow ISS aspirant who is working through the same paper — explaining the RSS argument out loud to someone else is often the fastest way to make it truly stick.

Comments