When Is a Proxy a Reasonable Substitute for What You Actually Want to Study?

When Is It Acceptable to Measure Something Else Instead?

Sometimes the variable you actually want is unavailable. Household income may be too sensitive or unreliable to collect. Historical exposure may be impossible to reconstruct directly. Societal research impact may unfold over decades. A psychological construct may not be directly observable at all.

Researchers often respond by using a proxy: an observable variable that stands in for the target of interest. That can be methodologically defensible, but availability alone is not enough. The central question is whether the substitute preserves enough information about the target to support the inference you intend to make.

A Proxy Is Reasonable When the Substitution Has a Defensible Basis

How to Decide Whether a Proxy Is Good Enough

What Makes Something a Proxy?

A proxy stands in for a target that the researcher would ideally like to observe more directly. The distinction is relational rather than intrinsic. The same variable can be a direct measure in one study and a proxy in another.

For example, if the research question concerns household income, reported income is the variable of interest. If the research question concerns broader socioeconomic circumstances, income might instead function as an indicator. If income cannot be collected and an asset index is substituted for it, the asset index functions as a proxy.

This distinction is easier to see once you separate an indicator, measure, and proxy according to the role each plays relative to the research target.

Start by Asking Why the Target Cannot Be Measured

A proxy should solve an identifiable measurement problem. Direct measurement may be impossible because the target is latent, historical, inaccessible, prohibitively expensive, ethically difficult to observe, or unavailable in an existing dataset.

In epidemiology, for example, exposure proxies are frequently used when the exposure of interest cannot be observed directly. Formal risk-of-bias guidance therefore asks whether the proxy adequately represents the exposure of interest during the time window relevant to the outcome.

If the target can be measured directly with reasonable validity, cost, and feasibility, substituting a more distant variable requires a stronger rationale.

The Proxy Needs a Plausible Connection to the Target

A useful proxy should have a substantive reason for changing when the target changes. This relationship might arise from theory, a known mechanism, previous empirical research, validation data, or some combination of these.

Suppose night-time satellite illumination is used as a proxy for local economic activity where reliable economic statistics are unavailable. The argument is not simply that both variables happen to vary geographically. Researchers would need a defensible reason to expect economic activity to influence observable night-time illumination and evidence that the relationship is useful for the setting and inference at hand.

Construct-validation work on secondary-data proxies similarly emphasizes careful specification of the target construct and evaluation of whether the proxy behaves as theory predicts in relation to relevant constructs.

Correlation Helps, but Correlation Alone Is Not Enough

A proxy may correlate strongly with the target and still behave badly under conditions that matter to your study. Correlation can arise because both variables respond to a third factor, because the relationship holds only in a particular population, or because the proxy captures only one portion of the target's variation.

For example, number of citations may correlate with some dimensions of scholarly influence. That does not make citation count a direct measure of research quality, societal benefit, methodological rigor, or every other phenomenon with which influential scholarship might be associated.

The relevant question is therefore not simply, “Are X and Y correlated?” It is, “Why should X provide information about Y in the circumstances in which I intend to use it?”

Ask How Much of the Target the Proxy Actually Represents

A proxy can be related to the target while representing it poorly. Methodological research has shown that regression estimates can change substantially as the degree to which a proxy represents its intended latent construct changes.

This matters because proxy quality is not merely a descriptive issue. Imperfect representation can alter substantive estimates and conclusions.

If possible, examine evidence comparing the proxy with a stronger reference measure. Depending on the field and variable, relevant evidence might include agreement, sensitivity and specificity, predictive performance, convergent evidence, measurement-error analysis, or relationships with variables that theory says should and should not be associated with the target.

Check What Else Influences the Proxy

A proxy is especially vulnerable when factors other than the target strongly determine its values.

Imagine using number of online discussion posts as a proxy for student engagement. Posting frequency may partly reflect engagement, but it can also depend on instructor requirements, assessment incentives, internet access, course design, communication preferences, and whether students are permitted to participate through other channels.

The proxy can therefore vary even when the target does not, or remain unchanged when important aspects of the target vary.

Computational social-science research has similarly warned that imperfect proxies can contain both noise and contamination from other concepts, potentially biasing estimates and their uncertainty.

The Time Window Must Match the Research Question

A proxy may be reasonable in general but inappropriate for the period relevant to the study.

Suppose residential address is used to estimate environmental exposure. A participant's current address may be a reasonable proxy for current residential exposure but a poor proxy for exposure ten years earlier if the participant has moved repeatedly.

Risk-of-bias guidance for exposure assessment explicitly considers whether a proxy approximates the exposure during the time window relevant to the outcome.

The Proxy May Work Better for Some Populations Than Others

Proxy-target relationships can be context dependent. An asset that reliably differentiates household economic circumstances in one country may be nearly universal in another and therefore provide little useful discrimination. Digital activity may be a reasonable indicator of participation in a fully online course but much less informative in a course where most learning occurs offline.

This means evidence supporting a proxy should be evaluated in relation to the population, setting, period, and purpose for which it will be used.

Consider the Consequences of Proxy Error

Not all proxy error behaves in the same way. Measurement-error research shows that error and misclassification can bias effect estimates, reduce statistical power, leave residual confounding, or otherwise alter results depending on the variable's role and the structure of the error.

This is an important reason to avoid the comforting but unreliable assumption that an imperfect proxy merely adds harmless “noise.” The direction and magnitude of the resulting distortion can depend on how the proxy departs from the target.

A Proxy Can Be Useful Without Being Perfect

Demanding perfect correspondence would make many legitimate forms of research impossible. A proxy can still provide useful evidence when its limitations are understood and proportionate to the inference being made.

The appropriate standard is therefore not “Does the proxy equal the target?” If it did, it would hardly be functioning as a proxy. More useful questions are:

• Does it track the relevant variation in the target?
• Is the relationship sufficiently stable in the population and context?
• Are major alternative influences understood?
• Is the measurement error tolerable for the intended analysis?
• Would a different feasible proxy provide a closer representation?
• Can the resulting claim be stated at a level that the proxy actually supports?

Triangulation Can Strengthen Proxy-Based Inference

If no single proxy provides a compelling representation, several independent forms of evidence may help. Secondary-data proxy research has noted that proxies can provide valuable corroborating or triangulating evidence when used carefully.

For example, researchers studying neighborhood socioeconomic conditions might combine information from housing, education, employment, and asset indicators rather than relying on one distant substitute.

Multiple proxies are not automatically better, however. If they all share the same bias or measure the same narrow manifestation, multiplication does not solve the underlying validity problem.

Be Explicit That You Used a Proxy

Transparency matters because readers need to distinguish what was actually observed from what was inferred. Methodological discussions of proxy exposure measurement recommend explaining the rationale for the chosen proxy and how it may differ from the target variable.

Sometimes the Better Decision Is to Narrow the Question

If the only available proxy is weak, another option is to ask a question about what you can actually measure well.

Suppose institutional records provide course completion but you originally hoped to study “student success.” Rather than treating completion as a comprehensive proxy for success, you might narrow the outcome to course completion. The revised study asks a smaller question but makes a more defensible claim.

Proxy use should therefore be compared not only with direct measurement, but also with changing the scope of the construct or research question.

Using LMS Activity as a Proxy for Student Engagement

The proxy becomes more defensible because the target has been specified narrowly enough to correspond to the available evidence. Had the researcher defined engagement as behavioral, cognitive, and emotional involvement, the same digital traces would provide a much weaker representation of the complete construct.

Common Mistakes When Using Proxies

Use a Proxy Only When You Can Defend the Substitution

The practical decision begins with the target, not the available dataset. Identify exactly what you wish you could observe before deciding whether another variable can stand in for it.

Evaluate a Proxy Before Building Your Study Around It

Questions About Using Proxy Variables

A Proxy Is Useful Only to the Extent That the Substitution Holds

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.