ISS 2017 Statistics Paper-2 Solution: Question 8 (Interaction Variance Component in Two-Way ANOVA)
Continuing our question-by-question walk through the ISS Statistics Paper-2 previous year papers, today we pick up right where the 2017 paper left off. This is Question 8 of 2017, and it takes us into two-way random effects ANOVA — a topic that looks intimidating the first time you see it, but is really just careful bookkeeping once you know the pattern.
Quick Summary
Topic: Two-way random effects ANOVA — estimating the variance component due to interaction
Question reference: ISS 2017, Statistics Paper-2, Question 8
Correct answer: (a) 1
The Question
An experiment was conducted to study the variation in fuel consumption of diesel vehicles. For this purpose, different types of fuel injectors (A) and different qualities of fuel (B) were assigned randomly to the same type of diesel vehicles. The analysis of variance table of the random effect model is given below:
ANOVA Table Source — Degrees of freedom — Sum of squares A — 4 — 48 B — 3 — 36 A × B — x — 96 Error — y — z Total — 59 — 380
The estimate of component of variance due to interaction (σ̂²AB) is:
(a) 1
(b) 1.6
(c) 5
(d) 8
Step 1: Recover the Number of Levels of A and B
In a two-way ANOVA with factor A having a levels and factor B having b levels, the degrees of freedom for A is a − 1, and for B is b − 1. The table tells us directly:
df(A) = 4, so a − 1 = 4, which gives a = 5. df(B) = 3, so b − 1 = 3, which gives b = 4.
So there are 5 types of fuel injectors and 4 qualities of fuel — 5 × 4 = 20 combinations (cells) in all.
Step 2: Fill in the Missing Degrees of Freedom (x and y)
For a balanced two-way layout with n observations per cell, the degrees of freedom split up as:
df(A × B) = (a − 1)(b − 1) = 4 × 3 =12, so x = 12.
The total degrees of freedom is always one less than the total number of observations: df(Total) =abn − 1 = 59, which means abn = 60. Since ab = 20, we get n = 60 / 20 =3 observations per cell.
Now the error degrees of freedom is df(Error) =ab (n − 1) = 20 × (3 − 1) =40, so y = 40.
Check: 4 + 3 + 12 + 40 = 59. That matches the given total, so our values are consistent.
Step 3: Fill in the Missing Sum of Squares (z)
The sums of squares for all sources must add up to the total sum of squares:
SS(A) + SS(B) + SS(A × B) + SS(Error) = SS(Total) 48 + 36 + 96 + z = 380 180 + z = 380 z =200
Step 4: Compute the Mean Squares
A mean square is just a sum of squares divided by its own degrees of freedom:
MS(A × B) = SS(A × B) / df(A × B) = 96 / 12 =8 MS(Error) = SS(Error) / df(Error) = 200 / 40 =5
Step 5: Use the Expected Mean Square for the Random Effects Model
This is the part that actually decides the question, so it's worth slowing down here. In the two-way random effects model (both A and B are random factors, not fixed ones), the model is written as:
Yijk = μ + Ai + Bj + (AB)ij + εijk
where Ai, Bj, (AB)ij and εijk are all independent random variables with mean zero and variances σ²A, σ²B, σ²AB and σ²e respectively. For this model, standard ANOVA theory gives us the expected mean squares:
E[MS(A × B)] = σ²e +n σ²AB E[MS(Error)] = σ²e
Subtracting the second from the first and solving for σ²AB gives the estimator:
σ̂²AB = (MS(A × B) − MS(Error)) /n
Pro Tip: The single most common slip on this question is stopping at MS(A × B) − MS(Error) = 8 − 5 = 3 and marking option nowhere close to the key. The interaction mean square doesn't estimate σ²AB directly — it estimates σ²e + nσ²AB, because every interaction cell actually contains n replicate observations contributing to that mean square. You must divide the difference by n to isolate σ²AB itself. This is exactly the kind of detail that's usually drilled into students through repeated expected-mean-square derivations until identifying the right "n" to divide by becomes automatic rather than a guess under exam pressure.
Step 6: Final Calculation and Answer
Plugging in the numbers:
σ̂²AB = (MS(A × B) − MS(Error)) /n = (8 − 5) / 3 = 3 / 3 =1
This matches option (a) 1, which is also confirmed by the official ISS 2017 Paper-2 answer key (Series A) for Question 8.
Why This Question Matters
This question bundles together three separate skills examiners love to test at once: reconstructing a partially filled ANOVA table using the relationships between degrees of freedom and sums of squares, recognizing the structure of a balanced two-way random effects design, and correctly applying expected mean square formulas to build variance component estimators. Any one of these skills could be tested alone, but seeing them combined is a strong signal that the ISS Paper-2 syllabus expects you to treat ANOVA not as a memorized formula sheet but as a connected system you can reconstruct from first principles, even when parts of the table are deliberately hidden.
Frequently Asked Questions
Why is this called a "random effects" model instead of a "fixed effects" model?
In a fixed effects model, the specific levels of A and B chosen for the experiment are the only ones of interest, and inferences apply only to those levels. In a random effects model, the levels used (here, the specific fuel injector types and fuel qualities) are treated as a random sample from a larger population of possible levels, so the goal shifts from estimating individual effects to estimating variance components like σ²A, σ²B and σ²AB.
How would the expected mean square formulas change if this were a mixed model instead?
If one factor were fixed and the other random (a mixed effects model), the expected mean square for the fixed factor would no longer have a clean σ² term but would instead involve a sum of squared fixed effects, while the random factor and interaction terms would still involve variance components, often with a correction factor depending on which convention (restricted or unrestricted) is used. It's important to check which model the question specifies before writing down expected mean square formulas.
What if the variance component estimate comes out negative?
Since MS(A × B) and MS(Error) are both sample quantities, it is entirely possible for MS(A × B) to be smaller than MS(Error) purely due to sampling variation, which would make the estimator negative — an impossible value for a true variance. In practice, such negative estimates are usually truncated to zero and reported as evidence that the interaction variance component may genuinely be negligible.
Why do we divide by n and not by ab or some other quantity?
The "n" in the formula is the number of replicate observations within each A × B cell, because that is exactly the quantity that appears as the coefficient of σ²AB in the expected mean square derivation. If the design had unequal replications per cell, the formulas would need suitable modification since the clean balanced-design result would no longer hold directly.
Is there a quicker way to solve this without deriving the expected mean square formula from scratch?
For a standard balanced two-way random effects design, the result σ̂²AB = (MS(A × B) − MS(Error)) / n is standard enough to memorize directly, alongside the analogous main-effect formulas σ̂²A = (MS(A) − MS(A × B)) / (bn) and σ̂²B = (MS(B) − MS(A × B)) / (an). Knowing all three together, rather than just the interaction one, protects you against questions that ask for σ²A or σ²B instead.
Why does the question specify "random effect model" explicitly in the problem statement?
Because the formula for estimating variance components depends entirely on which model (fixed, random, or mixed) generated the ANOVA table, examiners are careful to state this explicitly so there is no ambiguity about which set of expected mean square relationships you should use.
If anything in this derivation felt unclear, or if you solved it a different way, drop a comment below — and if you know a fellow ISS aspirant working through these same papers, consider sharing this post with them.

Comments