Why an Importance Ranking Is Not a Unique Explanation
A model may use several predictors that carry overlapping information. For example, a hypothetical insurer might record both a vehicle’s age and its mileage; older vehicles may tend to have higher mileage. If a model predicts claim severity using both, its reported importance scores do not necessarily reveal how much of the prediction is uniquely due to either variable.
An importance score is a summary defined by a particular method, model, and performance measure. It is not automatically a share of the prediction, a causal effect, or a measure of information that belongs exclusively to one feature. With correlated predictors, a score may be divided between features, assigned mainly to one, or change substantially across model fits.
The earlier material on variable importance distinguishes common measures such as coefficient-based and permutation-based scores. Here the key issue is how to interpret those measures when predictors overlap. An importance result describes the fitted model and the chosen calculation; it does not by itself show that one feature explains a prediction on its own.
Exact Redundancy: The Prediction Can Be Identified While Coefficients Are Not
Consider a linear model with two identical predictors. The model can determine their combined contribution, but the data cannot determine how to divide that contribution between the two coefficients. This is more than a ranking issue: different coefficient assignments can produce exactly the same fitted values.
Worked Example: Two Identical Predictors in a Linear Model
A hypothetical portfolio study has three observations. For each, two recorded measures, \(X_1\) and \(X_2\), are identical. The response \(Y\) is loss cost in thousands of dollars. Fit the specified regression through the origin, \(\hat{Y}=b_1X_1+b_2X_2\), to the data below, and assess whether either coefficient gives a unique allocation.
| Observation | \(X_1\) | \(X_2\) | \(Y\) |
|---|---|---|---|
| 1 | 1 | 1 | 2 |
| 2 | 2 | 2 | 4 |
| 3 | 3 | 3 | 6 |
Method. Substitute the data into the squared-error objective. For each observation \(j\), \(X_{1j}=X_{2j}=x_j\) and \(Y_j=2x_j\). Therefore,
The objective depends on \(b_1\) and \(b_2\) only through their sum. It reaches its minimum, zero, whenever \(b_1+b_2=2\). Thus \((b_1,b_2)=(2,0)\), \((1,1)\), and \((0,2)\) all fit the observations exactly.
At \(X_1=X_2=4\), all three coefficient pairs predict \(4b_1+4b_2=8\) thousand dollars. Yet their separate feature contributions are, respectively, \((8,0)\), \((4,4)\), and \((0,8)\) thousand dollars. The prediction is the same, but the allocation between features is not identified by these data.
Check. Each fitted value is \(2x_j\), so the residuals are \(0,0,0\) and SSE is \(0\) for every coefficient pair satisfying \(b_1+b_2=2\). The common mistake is to interpret one possible coefficient allocation as the unique importance of each predictor.
This example gives a direct justification for caution: because the predictors are identical, exchanging some coefficient from one to the other leaves every model prediction unchanged. If predictors are strongly but not perfectly correlated, the same issue is usually less absolute but can still make coefficient estimates sensitive to the sample. A coefficient’s sign or size is a conditional model quantity, not a stand-alone measure of a feature’s unique predictive value.
Permutation Importance: Individual and Grouped Questions Differ
For a loss measure such as mean squared error, one-at-a-time permutation importance compares the model’s original evaluation loss with the loss after one feature is shuffled across observations. If \(L_{\text{base}}\) is the original loss and \(L_{\text{perm},k}\) is the loss after shuffling feature \(k\), the increase is
A larger increase means that, for this fitted model, evaluation set, and loss, shuffling that feature damaged performance more. It does not mean that the feature owns that amount of predictive information. When another feature carries similar information, the model may still predict well after one is shuffled. Conversely, shuffling just one feature can create combinations of feature values that were rare or absent in the observed data.
Worked Example: Individual Versus Grouped Permutation
A hypothetical model predicts a binary outcome \(Y\) using two perfectly correlated binary predictors \(A\) and \(B\). The four evaluation rows are \((A,B,Y)=(0,0,0),(0,0,0),(1,1,1),(1,1,1)\). The fixed model is \(\hat{Y}=(A+B)/2\), and performance is measured by mean squared error (MSE). Calculate the expected increase in MSE from permuting \(A\) alone, \(B\) alone, and both together using the same random row permutation.
Baseline. In every row, \(A=B=Y\), so \(\hat{Y}=Y\). The squared error in each row is \(0\), and \(L_{\text{base}}=0\).
Permute \(A\) alone. The original \(A\) values contain two zeros and two ones. Under a uniformly random permutation of the four entries, any row has probability \(2/4=1/2\) of receiving a value different from its original value. Since \(B=Y\), the prediction error in a mismatched row is \((A_{\text{perm}}-Y)/2\), which is either \(1/2\) or \(-1/2\); its square is \(1/4\). The expected MSE is therefore
By symmetry, permuting \(B\) alone also gives expected MSE \(0.125\). The expected individual permutation importance of each feature is \(0.125-0=0.125\).
Permute both together. Apply the same random row permutation to \(A\) and \(B\). They remain equal after the shuffle, so the prediction is the permuted binary value. A row has probability \(1/2\) of receiving the opposite value from its original \(Y\). In that case the squared error is \(1\); otherwise it is \(0\). Thus,
The grouped increase is \(0.5-0=0.5\), while the two individual increases are each \(0.125\). Their sum, \(0.25\), does not equal the grouped increase. These are different perturbation questions, not pieces of a uniquely divisible total.
Check. The mismatch probability is \(1/2\) for each stated shuffle, and the squared error is \(1/4\) for an individual shuffle versus \(1\) for the joint shuffle. The common mistake is to add individual importance scores and call the result the group’s unique contribution.
Grouped permutation can be useful when the reporting question concerns a related set of features rather than one member of that set. The grouping must be stated, and the shuffle rule matters: shuffling a group together preserves its internal relationships, while shuffling features separately does not. Neither approach turns an importance score into a causal or exclusive attribution.
Tree Importance Can Be Assigned to One Substitute
A tree-based model illustrates another form of instability. A split-based importance measure may accumulate the reduction in impurity from splits that use a feature. When two correlated predictors offer nearly the same split, the fitted tree may choose one first. The other may then receive little or no split importance, even though it could have served a similar predictive role.
Worked Example: A Tree Chooses One of Two Equivalent Features
A hypothetical training sample has four records with binary outcome \(Y=(0,0,1,1)\). Two binary features \(A\) and \(B\) both equal \(Y\) on every record. A classification tree uses Gini impurity and may split on either feature. Calculate the impurity reduction if it splits on \(A\), and describe what happens to split-based importance if a tie-breaking rule selects \(A\).
Method. For a node with class proportions \(p_0\) and \(p_1\), Gini impurity is \(1-p_0^2-p_1^2\). The root has two records in each class, so \(p_0=p_1=1/2\) and its impurity is \(1-(1/2)^2-(1/2)^2=1/2\).
Splitting on \(A\) creates one child containing the two \(Y=0\) records and one containing the two \(Y=1\) records. Each child is pure and has Gini impurity \(0\). The weighted child impurity is \((2/4)(0)+(2/4)(0)=0\), so the split’s impurity reduction is \(1/2-0=1/2\).
Because \(B\) gives exactly the same partition, splitting on \(B\) would also reduce impurity by \(1/2\). If the fitted tree chooses \(A\) and stops, its split-based importance assigns the reduction to \(A\); \(B\) has no split and receives zero from this tree. A tree that instead chooses \(B\) can reverse those scores without changing the predictions.
Check. Both possible root splits create pure children and therefore give the same reduction, \(0.5\). The ranking reflects the chosen tree structure and tie resolution, not evidence that \(A\) uniquely explains the outcome.
With real samples, the features may not be exactly equivalent. Small changes in the data, candidate splits, or model settings can then change which feature is selected and how importance is distributed. A ranking that moves between correlated features should not be described as a stable discovery about which feature matters most.
Why Importance Scores Can Be Distorted
Correlation is one reason scores can mislead, but the method and evaluation setup also matter. Interpret an importance score only after naming what was calculated and what its scale means.
- Coefficient size depends on scale and model specification. In a regression, a raw coefficient describes the modeled change in the response per unit change in that predictor, holding included predictors fixed. Predictors measured in different units cannot be ranked fairly by raw coefficient size alone. Correlation can also make the conditional coefficients unstable, even when fitted predictions are similar.
- One-at-a-time permutation measures reliance in the fitted model. If a substitute feature remains available, shuffling one predictor may have a modest effect. If the shuffle breaks relationships among predictors, the resulting rows can also represent combinations that are uncommon in the evaluation data.
- Split-based importance reflects the fitted tree structure. A correlated feature that is not selected for a split can receive little score, despite being a plausible substitute. The selection can depend on the particular sample and model-fitting choices.
- Ranking depends on the target metric. A feature’s importance under squared error need not match its importance under another loss or evaluation measure. Always identify the metric and the evaluation data used.
These points do not imply that importance measures are useless. They define the limits of the claim. For example, a report can state that a feature had a large permutation-related increase in validation MSE for a specified model. It should not silently expand that finding into “this feature alone explains the prediction” or “changing this feature would cause the outcome to change.”
Reporting Shared Importance Carefully
When predictors are correlated, reporting should make the uncertainty in allocation visible without obscuring the useful predictive result. A clear account identifies the fitted model, importance method, evaluation basis, and the relevant relationships among features. Where appropriate, show both individual results and a justified grouped result, while explaining that the group score answers a different question.
State whether the reported ranking is based on coefficients, permutation, impurity reduction, or another specified method.
Identify the fitted model, evaluation data, and performance measure. A score is not detached from these choices.
When predictors carry similar information, say that individual rankings may be sensitive to how the model allocates or uses that information.
Describe predictive association or model reliance, not unique explanation or causation. If a group score is shown, name the group and how it was evaluated.
A suitable statement might be: “Under the stated validation-set permutation procedure, vehicle age and mileage each showed a modest individual increase in prediction error when shuffled. Because these predictors are correlated, the individual scores do not establish that either feature uniquely accounts for the model’s predictions.” If a grouped result is also presented, specify that both features were perturbed together and report its result separately.
This distinction is useful to decision-makers. They can learn which inputs the fitted model uses, while also being told when the available evidence does not support a unique allocation among related predictors. The careful wording is aligned with the SOA Exam PA (Predictive Analytics) syllabus: communicate what the analysis supports, and do not claim more than the method establishes.
Check Your Understanding
For each problem, identify what the importance result supports and what it does not establish.
- A regression uses two identical predictors \(X_1=X_2=x\), with no intercept, and the observed response is \(Y=3x\). Give two coefficient pairs that fit perfectly and state whether their separate contributions are uniquely determined. (Answer: For example, \((b_1,b_2)=(3,0)\) and \((0,3)\); both fit perfectly, and separate contributions are not uniquely determined.)
- A classifier’s baseline validation error is \(0.08\). After permuting feature \(C\) alone, its error is \(0.13\). Calculate the increase and state one limitation if a correlated feature remains in the model. (Answer: The increase is \(0.13-0.08=0.05\); the other feature may preserve similar information, so this does not establish \(C\)'s unique contribution.)
- Two features give the same perfect tree split. The tree selects the first feature and assigns it split importance \(0.4\); the second receives \(0\). What does this result indicate about the fitted tree, and what does it not prove? (Answer: The fitted tree used the first feature for that split; the zero score does not prove that the second feature lacks predictive information or could not substitute.)
- Two correlated features have individual permutation importance increases of \(0.03\) and \(0.04\), while jointly permuting them gives an increase of \(0.12\). Should the individual scores be added to estimate a unique shared contribution? Explain briefly. (Answer: No. The individual and joint perturbations define different quantities; the joint increase is \(0.12\), and none of these values is a unique causal allocation.)
- A report ranks a raw regression coefficient for annual income above a coefficient for age, although income is measured in thousands of dollars and age in years. Name two reasons this ranking alone is insufficient as a comparison of predictive importance. (Answer: The predictors use different units, so raw coefficient magnitudes are not directly comparable; also, correlated predictors can make conditional coefficients unstable and importance depends on the chosen measure.)