Why Proxy Predictors Need Deliberate Review
Two columns in a data set can look different and still tell a model much the same thing. A postal-area deprivation score and a measure of local housing costs, for example, might both reflect aspects of neighborhood resources. If both are supplied to a model, their overlap can make the model harder to explain without necessarily adding much new information. But overlap alone does not establish that either variable should be removed.
A proxy predictor is a measured variable that indirectly represents, or is strongly associated with, another characteristic of interest. That characteristic might be difficult to observe directly, such as access to transportation or financial resources. A proxy can be useful when it is available at the time a prediction is made, but it can also be a poor substitute, obscure the reason for a prediction, or encode sensitive group differences.
The key phrase is for a particular modeling purpose. A pair of predictors may overlap in a general description of the data but differ in how clearly they measure a construct, whether they are available at deployment, or how they affect decisions for different groups. A predictor screen is evidence for review, not an automatic feature-selection rule.
How to Investigate Possible Proxies
Begin by naming the underlying characteristic that a predictor might represent and explaining why that interpretation is plausible. Then inspect the data for evidence of overlap. For two numeric variables, a scatterplot and a correlation can reveal a strong linear relationship. For categorical variables, a cross-tabulation shows how often their categories occur together. A numeric and categorical pair can be examined with grouped summaries or plots.
These checks have limits. A near-zero correlation rules out neither a nonlinear relationship nor a proxy relationship based on categories. A high correlation does not show that the variables have the same meaning, nor does it show that either predicts the outcome well. Likewise, a strong association with an outcome does not establish that the predictor is appropriate, fair, or available when needed.
When a supervised screen is part of the assessment, use only the data permitted for that stage of the modeling workflow. In particular, information that would not be known at the prediction date cannot become a legitimate predictor merely because it is present in a historical data set. For an initial proxy review, record the feature’s definition, likely underlying construct, observed overlap, outcome relationship, timing, and potential consequences of using it.
Worked Example: Two Numeric Measures of Neighborhood Access
A hypothetical service-planning team records an access score \(A\), where larger values mean easier access to a service, and distance \(B\), in kilometers, from a service center for five areas. It also records a numeric service-use outcome \(Y\). The values are \(A=(1,2,3,4,5)\), \(B=(10,8,6,4,2)\), and \(Y=(12,10,9,7,6)\). Assess whether \(A\) and \(B\) look like overlapping predictors and what that implies.
Method. Compute the sample Pearson correlations. For paired observations, the sample correlation is the sum of cross-products of deviations from the means divided by the square root of the product of the two sums of squared deviations.
The means are \(\bar A=3\), \(\bar B=6\), and \(\bar Y=8.8\). The deviations for \(A\) are \((-2,-1,0,1,2)\); for \(B\), they are \((4,2,0,-2,-4)\); and for \(Y\), they are \((3.2,1.2,0.2,-1.8,-2.8)\).
For the pair \(A,B\), the sum of cross-products is \((-2)(4)+(-1)(2)+(0)(0)+(1)(-2)+(2)(-4)=-20\). The sums of squared deviations are \(10\) for \(A\) and \(40\) for \(B\). Therefore,
The observed values have a perfect negative linear relationship: in these five records, \(B=12-2A\). Now compare each predictor with \(Y\). The \(A,Y\) cross-product sum is \(-6.4-1.2+0-1.8-5.6=-15\), and the sum of squared \(Y\) deviations is \(22.8\). Thus \(r_{A,Y}=-15/\sqrt{10\cdot22.8}\), approximately \(-0.9934\). For \(B,Y\), the cross-product sum is \(12.8+2.4+0+3.6+11.2=30\), so \(r_{B,Y}=30/\sqrt{40\cdot22.8}\), approximately \(0.9934\).
Conclusion. In this small hypothetical data set, both measures have a strong linear association with service use, and they encode the same ordering in opposite directions. They are strong candidates for a proxy review. This five-area pattern does not establish that either will remain useful in a larger population, and it does not establish which measure is preferable. The team should also check their definitions, availability, and meaning to users.
Common mistake avoided: Treating the opposite signs of \(r_{A,Y}\) and \(r_{B,Y}\) as evidence that one predictor is unrelated to the outcome. Their signs differ because higher access and greater distance have opposite directions in this example.
What a Correlation Does—and Does Not—Establish
A useful justification for examining \(r^2\) comes from simple least-squares regression with an intercept. Let \(X\) and \(Y\) be observed numeric variables, and let \(S_{XX}=\sum_i(X_i-\bar X)^2\), \(S_{YY}=\sum_i(Y_i-\bar Y)^2\), and \(S_{XY}=\sum_i(X_i-\bar X)(Y_i-\bar Y)\). The fitted slope is \(b=S_{XY}/S_{XX}\). The fitted value for observation \(i\) is \(\bar Y+b(X_i-\bar X)\), so its fitted deviation from \(\bar Y\) is \(b(X_i-\bar X)\).
The sum of squared fitted deviations is therefore \(b^2S_{XX}=S_{XY}^2/S_{XX}\). Dividing this by \(S_{YY}\), the total sum of squared deviations in \(Y\), gives
Thus, in a simple linear regression with an intercept, the squared sample correlation is the fraction of the sample variation in \(Y\) accounted for by that one-variable linear fit. This identity explains why a large absolute correlation between two predictors is evidence that they share a linear pattern. It does not show that they have identical meaning, that they have the same relationship with an outcome, or that one adds no useful information when the other is considered. Those are distinct questions.
Worked Example: A Categorical Feature as a Potential Proxy
A hypothetical outreach program examines 100 applications. A neighborhood flag is “high deprivation” or “lower deprivation,” and an applicant group is recorded as Group P or Group Q. The cross-tabulation is: among 30 Group P applicants, 24 have a high-deprivation flag and 6 have a lower-deprivation flag; among 70 Group Q applicants, 21 have a high-deprivation flag and 49 have a lower-deprivation flag. There are 18 recorded outcomes among the 45 high-flag applications and 11 among the 55 lower-flag applications. Assess the proxy evidence and its limits.
Step 1: Check how the flag varies by group. The high-flag rate is \(24/30=0.8000\) for Group P and \(21/70=0.3000\) for Group Q. The rates differ by \(0.8000-0.3000=0.5000\), or 50 percentage points. This is evidence that the flag is associated with group membership in this sample.
Step 2: Check how well the flag identifies Group P. If “high deprivation” were used to label an applicant as Group P, while “lower deprivation” labelled an applicant as Group Q, the correct classifications would be 24 Group P applicants with a high flag plus 49 Group Q applicants with a lower flag. That is \(24+49=73\) correct out of 100, or \(0.7300\). Always labelling everyone Group Q would give \(70/100=0.7000\) correct. The flag is associated with group membership, but this simple calculation is not evidence that it should be used to infer an individual’s group.
Step 3: Check the outcome rates. The high-flag outcome rate is \(18/45=0.4000\). The lower-flag rate is \(11/55=0.2000\). The difference is \(0.2000\), so the flag is also associated with the outcome in these hypothetical data.
Conclusion. The flag is a possible proxy for group membership and is associated with the outcome. These counts do not show why the outcome rates differ, whether the flag would perform similarly in new data, or whether using it is appropriate for the program’s purpose. The team should review the decision being made, the possible effects on groups, and whether the feature measures a relevant characteristic or merely carries a sensitive pattern.
Common mistake avoided: Concluding that an outcome association makes a feature automatically acceptable. Prediction, explanation, and fairness are different considerations; this table alone settles none of them.
Usefulness, Interpretability, Availability, and Ethics
A proxy review should lead to a documented decision, not just a list of correlations. Ask four questions before retaining a candidate feature:
- Usefulness: Is there evidence that the feature is relevant to the stated prediction task? Does it provide a measurement that is meaningful for the intended use, rather than merely a convenient association in one sample?
- Interpretability: Can a reader explain what the feature measures and what a change in its value means? A direct measurement may be easier to communicate than a proxy, although direct measurements can also be noisy, incomplete, or inappropriate.
- Availability: Is the feature known at the actual prediction time and available in the population where the model will be used? A strong historical association is not operationally useful if the value is collected only after a decision has been made.
- Ethical implications: Could the feature encode a sensitive characteristic or contribute to unequal decisions? A feature’s apparent neutrality does not prove that it has no such effect. Review should reflect the application, affected groups, and applicable requirements; a statistical screen cannot certify fairness on its own.
A proxy can be retained when it is useful, available, and appropriate for the task, with limitations clearly recorded. It can be excluded when it is unavailable at deployment, difficult to justify, or unsuitable for the decision. In some settings, retaining a transparent measurement while excluding a less interpretable duplicate may simplify explanation. In others, neither feature is an adequate measure of the underlying construct. The choice should follow the use case and evidence rather than a universal rule such as “always drop the more correlated variable.”
Worked Example: Timing and Missingness Change the Decision
A hypothetical outreach team makes an initial decision when an application is received. It has a neighborhood index for all 60 historical applications and a household expenditure measure for 45; the other 15 expenditure values were not collected. The neighborhood index is recorded at application, while expenditure is collected only after an applicant completes a later assessment. Among the 60 records, 30 are in the high-index category, with 20 outcomes; the other 30 are in the lower-index category, with 6 outcomes. Decide what can be concluded for the initial decision.
Step 1: Summarize the observed outcome association. The high-index outcome rate is \(20/30=0.6667\). The lower-index rate is \(6/30=0.2000\). The difference is \(0.6667-0.2000=0.4667\), approximately 46.67 percentage points. This is a descriptive association in the historical records.
Step 2: Quantify expenditure missingness. The expenditure value is missing for \(60-45=15\) applications. Its missing proportion is \(15/60=0.2500\), or 25%. This also means that the measure is not present for every historical case in the stated data.
Step 3: Apply the prediction-time condition. Expenditure is collected only after the initial decision. It is therefore unavailable for that decision, even when it is recorded later in the historical file. The neighborhood index is available at the stated time, but its association does not establish that it is a fair or suitable basis for action.
Conclusion. For an initial-decision model, the later expenditure measure cannot be used as an input because its value is not yet known. The neighborhood index is operationally available and shows an outcome association in these data, but it needs review for relevance, interpretation, and potential group effects before use. The outcome rates alone do not justify retaining it.
Common mistake avoided: Treating a value present somewhere in a retrospective file as available to a model at every decision time. Availability is determined by the information known when the prediction is made.
A Practical Review Sequence
State what underlying characteristic a predictor may represent and why that interpretation makes sense in context.
Use numeric comparisons, plots, grouped summaries, or cross-tabulations appropriate to the variable types. Record what the data show and what they cannot establish.
Ask whether the feature is relevant to the intended outcome and known at the prediction date. Do not confuse retrospective availability with operational availability.
Consider whether users can understand the measurement and whether it may encode sensitive group differences or affect decisions unevenly.
Retain, exclude, or defer the feature for further review, and document the evidence, limitations, and intended use that support the decision.
The most common exam trap is to treat high predictor correlation as a complete decision rule. High correlation is evidence of overlapping linear variation, not proof that one variable is useless or ethically problematic. Conversely, low correlation does not rule out a nonlinear proxy relationship or a fairness concern. A sound answer identifies what the diagnostic shows, states what remains unknown, and connects the feature decision to availability and the use case.
Check Your Understanding
For each problem, distinguish the evidence from the decision it does not settle.
- A hypothetical study has two centered numeric predictors with sums of squared deviations 16 and 25, and a cross-product sum of \(-18\). Compute their sample correlation and state one conclusion it supports and one it does not support. (Answer: \(r=-18/\sqrt{16\cdot25}=-0.9000\); it supports strong negative linear overlap in the sample, but does not prove either predictor is dispensable.)
- In a sample of 80 records, a location flag is high for 18 of 24 members of Group R and for 14 of 56 members of Group S. Compute the two high-flag rates and their difference. (Answer: \(18/24=0.7500\), \(14/56=0.2500\), difference \(0.5000\), or 50 percentage points.)
- A predictor is observed for 92 of 100 historical cases. What is its missing proportion, and does the proportion alone establish that it can be used at deployment? (Answer: \(8/100=0.0800\), or 8%; no, prediction-time availability and the reasons for missingness still need review.)
- A later-collected financial measure is associated with an outcome and is present in the final historical file. The decision must be made before that measure is collected. Should it be an input to that initial decision? Explain. (Answer: No; it is unavailable at the prediction time, regardless of its retrospective association.)
- A neighborhood score is strongly associated with both an outcome and a sensitive group in a hypothetical sample. Name two further considerations before deciding whether to retain it. (Answer: Assess whether it is available at prediction time and appropriate for the intended task; also review interpretability and potential effects on affected groups. The associations alone do not settle the decision.)