Why Choosing a Representative Takes More Than Ranking Correlations
When several predictors carry similar information, keeping all of them can make a model harder to interpret and may create instability. But choosing the predictor with the largest pairwise correlation is not automatically a sound solution. Correlation describes an association between two variables; it does not tell us which variable is more reliable, available when a prediction is needed, or meaningful for the decision at hand.
The aim here is narrower than general feature selection: once a group of predictors has been identified as strongly related, decide whether one can represent the group and, if so, which one is most suitable. This is a judgment informed by data and the intended use of the model, not a universal rule that the highest-ranked variable must win. This approach is aligned with the SOA Exam PA (Predictive Analytics) syllabus.
Earlier material on predictor correlations and correlation cutoffs helps identify possible groups, but no cutoff can establish that two variables have the same meaning. For example, an account balance and the number of transactions might move together, yet one measures an amount held and the other measures activity. The selection question is therefore two-part: does one predictor adequately represent the shared information, and which candidate best serves the intended use?
A Structured Comparison of Candidate Representatives
A practical comparison can examine four criteria. Define the criteria before scoring, use consistent scales, and assess candidates in the context of how the model will be used.
- Data quality: How accurate, consistently defined, complete, and timely is the measurement? Consider known recording errors, measurement processes, and the amount of missing data.
- Availability: Will the predictor be present for the population and at the time a prediction is required? A variable collected only after an event cannot serve as a legitimate input to a prediction made before that event.
- Interpretability: Can users understand what the variable measures, its units, and the practical meaning of a change in its value?
- Subject-matter relevance: Does the variable represent a concept that is appropriate to the business or policy question? Is its definition consistent with how the decision will be made?
A scoring table can make these judgments explicit. Suppose \(s_{ij}\) is the score for candidate \(j\) on criterion \(i\), and \(w_i\) is the chosen nonnegative weight for that criterion. If all criteria use the same direction, such as 1 for least suitable and 5 for most suitable, and the weights sum to 1, a weighted score is
This calculation assumes that the criteria have been defined consistently, that the scores are meaningfully comparable, and that the weighted additive trade-off is appropriate for the decision. The score is a transparent decision aid, not a probability, a statistical test, or proof that the highest-scoring variable will produce the best predictive performance. A critical requirement, such as being available at prediction time, may be a pass-or-fail condition rather than a criterion that can be offset by high scores elsewhere.
Justifying the Weighted Comparison
The weighted sum is useful because it makes the trade-offs visible. For two candidates, \(A\) and \(B\), subtracting their scores gives
This identity follows by subtracting the two sums term by term. Each difference \(s_{iA}-s_{iB}\) is the advantage of \(A\) over \(B\) on one criterion; multiplying by \(w_i\) expresses that advantage at the chosen importance of the criterion. Adding the weighted differences gives the total score difference. Thus, the calculation explains which trade-offs led to the ranking, but its result depends on the chosen weights and scores. If those inputs are poorly justified, the arithmetic cannot make the decision reliable.
Worked Example: Choosing Between Two Measures of Building Size
A hypothetical facilities team has two strongly related predictors for forecasting maintenance demand: floor area in square metres and the number of rooms. Both are recorded for most buildings. The team assigns weights of 0.35 to data quality, 0.25 to availability, 0.20 to interpretability, and 0.20 to subject-matter relevance. Each candidate is scored from 1 (least suitable) to 5 (most suitable).
Method. Compare the candidates with a weighted score, then review whether the ranking is consistent with the forecasting use. The main mistake this avoids is choosing solely by whichever variable has the slightly larger correlation with another member of the group.
The team scores floor area 4, 4, 5, and 4 across the four criteria. Its score is
The team scores room count 3, 5, 4, and 3. Its score is
The difference is \(4.20-3.70=0.50\). Floor area has higher data-quality, interpretability, and relevance scores; room count has a higher availability score. Given the stated criteria and weights, floor area is the preferred representative. The team should still document the choice and check that floor area is available with a consistent definition for the buildings to which the model will be applied.
Check: The weights sum to \(0.35+0.25+0.20+0.20=1.00\), and each score is a weighted average on the 1-to-5 scale. Neither score exceeds 5 or falls below 1.
Worked Decisions in Different Contexts
Candidate choice is context-sensitive. A measure that is clear and reliable in one setting may be unavailable, delayed, or poorly defined in another. The following examples show why criteria should be considered directly rather than inferred from correlation rank.
Worked Example: Comparing Two Hypothetical Income Measures
A hypothetical lender is considering two strongly related predictors for a model used during an initial application review: applicant-reported annual income and annual income estimated from bank deposits. The team uses weights of 0.35 for availability, 0.25 for measurement quality, 0.20 for interpretability, and 0.20 for subject-matter relevance. Scores range from 1 to 5.
Method. Apply the agreed weighted scoring rule, while separately checking whether a predictor is available at the time of the application decision. The mistake to avoid is assuming that a measure with a stronger correlation or a more precise-looking value is automatically the best representative.
Applicant-reported income receives scores of 4, 2, 5, and 4. Therefore,
The deposit-based measure receives scores of 2, 3, 2, and 2.5. Therefore,
The bank-deposit measure scores lower on availability and interpretability, but higher on measurement quality, than applicant-reported income. Its total is 2.35, compared with 3.70 for reported income. Under these stated weights, reported income is preferred. This does not establish that self-reported values are more accurate; the table explicitly gives the deposit measure the higher measurement-quality score. It shows that the other criteria outweigh that advantage in this particular decision.
Before using either variable, the team should confirm that the information is permitted and actually available for the intended application process. If deposit records are collected only after a decision, they fail the timing requirement and cannot be used for that prediction, regardless of their score.
Check: The deposit-based total is \(0.70+0.75+0.40+0.50=2.35\); the reported-income total is \(1.40+0.50+1.00+0.80=3.70\). The stated criterion-by-criterion comparison agrees with the scores.
Worked Example: Selecting a Telematics Measure With Exposure in Mind
A hypothetical fleet analyst considers three strongly related predictors for a model used to compare vehicle risk: the count of harsh-braking events, the harsh-braking rate per 1,000 kilometres, and a proprietary driving-risk index. Assume the rate is calculated as \(1{,}000\) times the event count divided by kilometres travelled. The index combines braking with other driving signals and is less transparent. The team assigns weights of 0.25 to data quality, 0.20 to availability, 0.20 to interpretability, and 0.35 to subject-matter relevance.
Method. Compare the candidates using the weighted table and ask whether each measure matches the intended comparison. The common mistake avoided is selecting a raw event count without considering that vehicles may have travelled different distances.
The event count receives scores of 4, 4, 5, and 3. Its score is
The rate receives scores of 4, 3, 4, and 5. Its score is
The risk index receives scores of 3, 5, 2, and 3. Its score is
The rate scores highest, at 4.15. It is less available than the other measures in this hypothetical setting, but it better matches the goal of comparing risk across vehicles with different distances travelled. The team selects the rate as the representative, subject to confirming consistent distance measurement and availability at prediction time.
Check: The scores rank \(4.15>3.85>3.20\), and the weights sum to 1.00. The choice is justified by the stated relevance of exposure-adjusted comparisons, not by a claim that the rate is universally superior.
When One Representative Is Not Enough
A correlated group is a warning to investigate redundancy, not a command to retain exactly one variable. Two predictors may be strongly associated but still represent distinct concepts, capture different patterns, or be important for different decisions. For example, total payments and payment frequency might move together while offering separate operational interpretations. If retaining both is useful, the decision should be supported by the modeling goal and checked in the intended validation process.
Conversely, a single representative should not be selected just to reduce the column count when it is poorly measured or unavailable for part of the target population. A defensible process distinguishes three questions:
Use association evidence and subject knowledge to identify predictors that may carry overlapping information. Treat the group as a candidate for review, not a proven set of interchangeable variables.
Check definitions, data quality, population coverage, and whether each candidate exists when and where predictions must be made. Rule out candidates that fail a non-negotiable condition.
Use clearly defined criteria and, if helpful, a weighted score. Record the reasoning, assumptions, and any trade-offs; reconsider the choice if the intended population or use changes.
Common Mistakes and Exam Traps
Suppose a candidate calculates scores of 3.70 for applicant-reported income and 2.35 for bank-deposit income, then says, “Bank-deposit income is lower on every criterion, so it is clearly inferior.” That conclusion is incorrect: the example gives the bank-deposit measure a measurement-quality score of 3, compared with 2 for reported income. The bank-deposit measure is lower on availability and interpretability, but higher on measurement quality; its total is 2.35. The weighted total and the criterion-by-criterion explanation must both agree with the scoring table.
- Treating correlation rank as a selection rule. The strongest pairwise association does not establish that a variable is most accurate, available, or useful.
- Changing the scoring scale mid-comparison. If high scores mean “more suitable,” apply that direction to every criterion and candidate.
- Letting a weighted total hide a fatal limitation. High interpretability cannot compensate for a predictor that is unknown at prediction time if availability is required.
- Presenting judgment scores as objective statistics. State who set the criteria and weights, and explain why they fit the intended use.
- Assuming correlation means interchangeability. Confirm that the chosen variable represents the concept needed; retain more than one when the variables serve distinct purposes.
Choosing representatives is a disciplined screening decision, not a mechanical consequence of a correlation matrix. Clear criteria help make the decision reviewable, while checks on timing, data quality, and meaning prevent an appealing numerical ranking from obscuring practical problems.
Check Your Understanding
For each situation, use the stated criteria and constraints rather than selecting by correlation rank alone.
- A team gives data quality, availability, interpretability, and relevance weights of 0.40, 0.30, 0.20, and 0.10. Candidate A scores 5, 2, 4, and 3; candidate B scores 3, 4, 3, and 4. Calculate both totals and select the higher-scoring candidate. (Answer: A scores 3.70, B scores 3.40, so choose A under these weights.)
- Two strongly related predictors are available at prediction time. One is a raw count of customer complaints; the other is complaints per 100 customer-months. The intended comparison is across customers with different observation periods. Which measure is more directly aligned with that comparison, and what data issue should be checked? (Answer: The rate is more directly aligned; check that customer-month exposure is measured consistently and is available.)
- Weights for quality, availability, interpretability, and relevance are 0.30, 0.25, 0.20, and 0.25. Candidate X scores 4, 3, 5, and 2; candidate Y scores 3, 5, 3, and 4. Calculate the totals and choose by the score. (Answer: X scores 3.45; Y scores 3.75; choose Y under these weights.)
- Two predictors have a correlation of 0.92, but one measures total claim dollars and the other measures claim count. A pricing team needs both severity and frequency interpretations. Must it keep only one? Explain. (Answer: No. Their strong association does not make their meanings identical; retaining both may be justified by the separate interpretations and modeling purpose.)
- A highly interpretable predictor is recorded only after the outcome occurs, while a less interpretable alternative is available before the prediction. Can a weighted score alone justify using the first predictor? (Answer: No. If prediction occurs before the outcome, the first predictor fails the availability-at-prediction-time constraint and should not be used.)