Why a Predictor May Represent More Than Its Recorded Value
A research question may concern something that is important but not directly recorded: access to care, educational opportunity, or medication adherence. A data set, however, contains observed variables such as travel distance, school resources, or prescription refills. Those variables may provide useful information about the underlying idea, but they are not identical to it. Distinguishing the two helps explain why a predictor belongs in a model and what its coefficient can—and cannot—mean.
The distinction matters in both explanatory and predictive specifications. For explanation, an indirect measure may stand in for a relevant aspect of the subject matter, but its coefficient describes the recorded measure, not automatically the broader idea. For prediction, a proxy may be useful if it is available at the intended prediction time and carries relevant information. In either case, the rationale should identify what the measure represents and what it leaves out.
State the Model in Terms of What Was Measured
Suppose \(Y_i\) is an outcome for observational unit \(i\), \(W_i\) is an observed proxy for a construct, and \(Z_i\) represents other included predictors. A linear conditional-mean model can be written as:
Here, \(\beta_1\) describes the difference in the modelled mean outcome associated with a one-unit difference in the recorded proxy, comparing units with the same values of the other included predictors. It does not, by itself, describe a one-unit difference in the underlying construct. That interpretation would require a clear, justified link between changes in the proxy and changes in the construct.
The conditional-mean form assumes that the stated predictors and functional form adequately represent the mean outcome for the intended analysis. For usual regression interpretations, the observational units should also be correctly defined, the data should be measured and coded consistently, and the predictors should not be perfectly redundant. Independence of units is needed for standard uncertainty calculations when the design does not otherwise account for dependence. These conditions do not make a proxy a direct measurement: they concern the model and data, not the meaning of the proxy.
A coefficient can be calculated even when a proxy is imperfect. The key issue is what claim is being made from that coefficient. For a descriptive or predictive purpose, it may be enough that the proxy is measured reliably and carries useful information. For a scientific explanation about the construct itself, it matters whether the proxy represents the relevant aspects of that construct and whether other differences could explain its association with the outcome. A regression association alone does not establish a causal effect.
A Practical Rationale for a Proxy Term
Before including a proxy, state the intended role of the term in ordinary language. A sound rationale connects the research question to the construct, then explains why the recorded variable is informative about that construct. It should also name the main mismatch between the two. This makes the specification more transparent than describing a predictor only as “important” or “available.”
Specify the broader concept that matters to the question, without calling it the recorded variable.
State exactly what was recorded, for which unit, in what units, and over what time period.
Say why the recorded value may carry information about the construct, and identify important dimensions it does not capture.
Describe the coefficient as an association with the recorded measure, conditional on the other included terms, unless a separate justification supports a stronger claim.
This reasoning also clarifies whether a proxy suits the model’s purpose. A measure may be substantively relevant but unavailable at the time a prediction must be made. Another may be available but too loosely connected to the construct to support the intended explanation. Relevance, measurement quality, timing, and interpretation are related questions, but they are not interchangeable.
Worked Example: Travel Distance as a Proxy for Access to Care
Suppose a hypothetical data set records each patient’s distance to a clinic in kilometres and the number of preventive visits during a year. The underlying construct of interest is access to care. Distance is observable and may represent one barrier to access, but it does not record transport availability, clinic opening hours, appointment supply, or other barriers.
For five hypothetical patients, suppose the recorded pairs \((W_i,Y_i)\) are \((2,8)\), \((4,7)\), \((6,6)\), \((8,5)\), and \((10,4)\), where \(W_i\) is distance in kilometres and \(Y_i\) is annual visits. In a simple linear fit, the slope is \(b_1=S_{WY}/S_{WW}\), where \(S_{WY}\) is the sum of cross-products of deviations from the means and \(S_{WW}\) is the sum of squared deviations in distance.
The means are \(\bar{x}=6\) km and \(\bar{y}=6\) visits. The distance deviations are \(-4,-2,0,2,4\); the visit deviations are \(2,1,0,-1,-2\). Thus:
The fitted association is half a fewer preventive visit per year for each additional kilometre of distance in this simple model. The predictor in the calculation is distance, not access as a whole. The five points do not establish that distance causes fewer visits, nor do they show that every aspect of access has been measured. If these rows were treated as independent observations for formal inference, that independence would need to be justified by the data collection; the small illustrative set cannot establish it. The example supports a clear description of the proxy and its limits, not a broad causal conclusion.
Interpret the Recorded Scale, Not the Construct’s Label
A proxy’s units are part of its interpretation. If a model uses a rate, count, distance, or indicator, the coefficient is on that recorded scale. It is not enough to name the construct in the model description and then interpret the coefficient as though the construct itself had been directly measured.
Worked Example: Computers per Student as a Proxy for Educational Resources
Imagine a hypothetical analysis of school-level mean test scores. The construct is educational resources, while the recorded predictor \(R\) is the number of working computers per 100 students. Suppose a fitted conditional-mean model, including other specified school predictors, has the relevant part:
Here, \(Y\) is the school’s mean test score in points and \(R\) is computers per 100 students. Holding the included \(Z\) values fixed, a difference of three computers per 100 students corresponds to a modelled mean difference of:
The coefficient is an association with the recorded computer rate. It does not say that every additional type of educational resource raises a school’s mean score by 1.4 points, or that supplying computers would cause the stated difference. The proxy does not directly measure computer condition, student access, teacher support, or other resources. Further, the observational units are schools: the interpretation concerns school mean scores, not an individual student’s score. Any explanatory claim would also need to consider whether other school differences account for the association.
A defensible rationale for including \(R\) would say that working computers per 100 students provide a measurable, relevant indicator of one resource dimension. It would also say that this term does not represent educational resources in full. That rationale is more precise than simply labelling \(R\) “resources.”
Measurement Quality and Proxy Limitations
A useful proxy needs a credible connection to the construct for the intended question. Ask what is actually observed, how consistently it is recorded, and whether its time window matches the outcome or prediction task. A measure recorded with error or inconsistently across units may be a weak basis for explanation. A measure that captures only one narrow aspect may still be useful, but its scope should be stated rather than silently expanded.
A mismatch can matter even when the recorded values are precise. For example, a carefully measured count of clinic visits is still a count of visits, not a complete measure of health-care access. Conversely, a broad construct may be represented by a proxy that is noisy or measured at a different level from the outcome. The interpretation should preserve those distinctions: say what the proxy records and why it may be relevant, rather than claiming that it fully measures the construct.
For a predictive specification, the question is whether the proxy is measured before prediction and contains information useful for predicting the outcome in the intended setting. For an explanatory specification, the question is whether the proxy meaningfully represents the concept under study and whether the resulting conditional association answers the stated question. Neither purpose justifies treating proxy status as proof of causal influence.
Worked Example: Prescription Refills as a Proxy for Medication Adherence
Consider a hypothetical analysis relating medication adherence to a symptom score, where higher scores indicate more severe symptoms. The construct is adherence, meaning whether a patient takes medication as directed. A possible proxy is the fraction of days in a 90-day period for which prescription records show medication available. For one hypothetical patient, records indicate 72 covered days:
Suppose a fitted model includes the proxy \(A\) and other predictors \(Z\), with the relevant part written as:
At the same \(Z\) values, compare a coverage fraction of 0.80 with 0.60. The modelled difference in mean symptom score is:
Thus, a 0.20 higher recorded coverage fraction is associated with a one-point lower modelled mean symptom score, conditional on the included terms. Refill availability does not confirm that medication was taken; it can miss stockpiling, early refills, or other departures between having medication and taking it. The term may be useful as a measure of recorded medication availability, but the coefficient should not be described as the effect of increasing adherence by a specified amount. For prediction, the records would also need to be available by the prediction time. Whether that condition holds depends on the intended use and the measurement timeline.
A Short Justification for the Coefficient’s Meaning
The interpretation follows directly from the model’s predictor. Holding \(Z\) fixed, the fitted conditional mean at proxy value \(w\) is \(\beta_0+\beta_1w+\boldsymbol{\gamma}^{\mathsf T}Z\). At \(w+\Delta\), it is \(\beta_0+\beta_1(w+\Delta)+\boldsymbol{\gamma}^{\mathsf T}Z\). Subtracting the first expression from the second gives:
This is a modelled mean difference for a change in \(W\), the recorded proxy. The algebra contains no term for an unobserved construct. To interpret the difference as a change in the construct, one would need additional information or assumptions linking a change in \(W\) to a specific change in that construct. The regression equation alone supplies no such link.
Common Mistake: Renaming a Proxy as the Construct
A common mistake is to write, “The coefficient for access to care is \(-0.5\),” when the model actually contains distance to a clinic. This wording turns an indirect measure into a direct one and can imply that all dimensions of access have been measured. The correction is to name both elements: “The coefficient for distance to a clinic is \(-0.5\) visits per kilometre; distance is being used as one indicator of access to care.”
A related mistake is to treat a plausible proxy as if it established a causal effect. A proxy’s relevance supports a reason for considering it in the specification; it does not, by itself, rule out confounding, measurement limitations, or other explanations for an association. State the model’s actual predictor, its units, its role, and the limits of the claim.
Check Your Understanding
For each question, keep the underlying construct distinct from the recorded variable.
- A model uses kilometres from a patient’s home to the nearest clinic as a predictor. What construct might this proxy represent, and name two aspects it may fail to measure.
- A coefficient relates school mean test scores to computers per 100 students. What does the predictor measure, and why is it inaccurate to call the coefficient an effect of educational resources generally?
- Prescription records show medication available for 45 of 60 days. Calculate the recorded coverage fraction and state one reason it may not equal medication adherence.
- In a model with a proxy \(W\), explain what a coefficient \(\beta_1\) describes when the other included predictors are held fixed.
- Why does a plausible connection between a proxy and a construct not, on its own, establish a causal effect?