011 min
Why a proxy matters
Most of what a team wants to know has no meter. Customer satisfaction, patient health, learning, user intent and the climate of the year 1400 cannot be read from an instrument. To get an answer, someone measures something else and treats it as the answer. That something is a proxy, and choosing it is a decision, even when nobody calls it one.
The cost of a poor choice is a confident wrong answer. The proxy itself is measured correctly, so the number looks clean and nobody questions it. The error is in the step from the number to the conclusion. A product team that reads a rising metric as rising customer value, or a hospital that reads low spending as good health, can act on data that is accurate and still reach the wrong decision.
Almost every measurement outside of physics, money and simple counting involves a proxy. A key performance indicator such as "weekly active users" stands in for "people who find the product useful". The question is never whether you use proxies. It is whether you know which ones you use, and what each one leaves out.
022 min
What makes a proxy work
A proxy has three parts. The target is the quantity you want to know but cannot observe, such as past temperature. The proxy is the quantity you can observe, such as the width of a tree ring. The link is the reason the proxy changes when the target changes. Social scientists call the act of choosing a measurable stand-in for a hard-to-measure idea operationalization.
In statistics, a proxy is defined narrowly. It is a variable that serves in place of an unobservable or immeasurable variable, while not being otherwise relevant in its own right. The last clause matters. You do not care about tree-ring width for its own sake. You care about it only because it carries information about something else.
The link must be a close correlation, which is a tendency for two quantities to move together. It does not have to be a straight-line relationship, and it can be negative, so that the proxy falls when the target rises. Reference works describe it exactly this way.
A strong correlation on one dataset is not enough, though. In practice a useful proxy needs the link to hold in three ways:
- Strength. The proxy tracks the target closely enough for the decision at hand.
- Equality. The link is the same for every group and setting you apply it to.
- Stability. The link is the same next year as it was when you checked it.
The table shows common proxies and the way each one breaks.
| Proxy | Target | Why they move together | How the link can fail |
|---|---|---|---|
| Tree-ring width | Past temperature or moisture | Trees grow wider rings in good growing seasons | A tree limited by water records rainfall more than temperature |
| GDP per person | Standard of living | Output per person and goods per person tend to rise together | An average hides how income is shared |
| Body mass index | Body fat | Weight for a given height rises with fat | A person with much muscle scores high with little fat |
| Search queries about flu symptoms | Number of people with flu | Ill people search for their symptoms | Healthy people search too when the news reports an outbreak |
| Health-care spending | Health need | Sicker patients receive more care | Unequal access to care changes who is treated |
031 min
Where the term comes from
No single person coined the term, and a verified first use is not known. The idea is older than the word. People used proxies long before statistics gave them a name, and the word entered statistics and econometrics as a label for a variable that serves in place of an unobservable one. The Oxford Dictionary of Statistics (Upton and Cook, 2002) gives that definition.
The need for proxies is easiest to see in climate science. Reliable global records of climate only began in the 1880s, and proxies provide the only means for scientists to determine climatic patterns before record-keeping began. Without them, nothing could be said about earlier centuries at all.
In the social sciences the same need appears for a different reason. The thing to be measured exists today, but no one can read it directly. Income, intelligence and quality of life are examples. Researchers choose a measurable quantity and accept its limits.
042 min
A worked case: climate proxies
Climate science is the clearest example, because the method is written down step by step. The US National Oceanic and Atmospheric Administration (NOAA) describes proxy data as preserved physical characteristics of the environment that can stand in for direct measurements. They come from corals, pollen, ice cores, tree rings, caves, sediments and historical documents.
Tree rings show the logic. In temperate regions where there is a distinct growing season, trees generally produce one ring a year. A tree that depends on warmth or moisture grows a wider ring when those conditions are present and a narrower one when they are not. A tree can live for centuries, so one trunk holds a year-by-year record.
Historical documents work the same way. Scientists used historical grape harvest dates to reconstruct summer temperatures, from April to September, in Paris from 1370 to 1879. A harvest date is not a temperature. It is a record of when grapes ripened, and ripening depends on warmth.
The step that makes this science rather than guesswork is calibration. Most proxy records are calibrated against independent temperature measurements during the period when both exist. Researchers do this in three steps:
- Find the years in which both the proxy and a direct measurement exist.
- Estimate the relationship between them over those years.
- Apply that relationship to the older years, where only the proxy exists.
The method has one built-in assumption. Calibration proves the link held in the overlap years. It cannot prove the link held in 1400. Climate scientists reduce this risk by cross-checking several different proxies against each other, because a tree and a coral are unlikely to fail in the same way.

052 min
When the proxy was wrong: health-care spending
In 2019, Ziad Obermeyer and colleagues published a study in the journal Science of an algorithm that US health systems use to decide which patients get extra care. The algorithm needed a measure of how much care a patient needs. Health need cannot be read directly, so the designers used a proxy that was easy to obtain: how much each patient would cost.
The result was a measurable bias. At a given risk score, Black patients were considerably sicker than White patients, as evidenced by signs of uncontrolled illnesses. The authors traced the cause to the proxy. The algorithm predicted health care costs rather than illness, and unequal access to care meant that less money was spent on Black patients than on White patients. The same illness produced lower spending for one group, so the link between cost and need was not equal across groups.
The size of the effect was large. Fixing the disparity would raise the share of Black patients who receive extra help from 17.7% to 46.5%.
Two lessons follow. First, the proxy looked good. The authors note that health care cost appeared to be an effective proxy for health by some measures of predictive accuracy. A proxy can predict well overall and still be wrong for a subgroup. Second, nobody made an error in arithmetic. The algorithm did exactly what it was asked, and the problem was the question. This is a source of bias that no amount of data cleaning removes.
062 min
When the link drifted: Google Flu Trends
A second failure shows a different weakness. In 2009, Jeremy Ginsberg and colleagues reported in the journal Nature a method for tracking influenza-like illness using Google search queries. The target was the percentage of doctor visits for flu-like symptoms, which the US Centers for Disease Control and Prevention (CDC) reports with a delay. The proxy was the frequency of certain search queries. The authors found that the relative frequency of certain queries is highly correlated with the percentage of physician visits in which a patient presents with influenza-like symptoms. That finding let them estimate flu activity with a reporting lag of about one day.
The proxy worked at first, and then it did not. Over 2011 to 2013 the service consistently overestimated flu. In one interval of the 2012 flu season, it predicted twice as many doctor visits as the CDC recorded. In 2014, David Lazer, Ryan Kennedy, Gary King and Alessandro Vespignani analysed the episode in Science under the title "The Parable of Google Flu: Traps in Big Data Analysis".
The general mechanism is drift, which means the link changes over time. People may search for different reasons after a news story, and the search page itself can change what people type. Any of these changes the link between queries and illness. Nothing in the query counts shows that the link has changed. Only the direct measurement, the CDC's count, can reveal it.
This is why a proxy needs a direct measurement kept alive on a schedule, even a slow and costly one. A proxy checked once, years ago, tells you about the past link only.
072 min
How a proxy shows up in product, design, and engineering work
Each scenario below has a role, a decision, and a point where the proxy and the target separate.
- A product manager reading time on page. The team wants to know whether readers find an article useful. Time on page is the proxy, because interested readers stay longer. After a redesign with smaller text, time on page rises. The change made the text harder to read, so people stayed longer because they were slower. The decision to keep the redesign would have been wrong.
- A support lead reading ticket counts. Fewer tickets are read as fewer product problems. Then the team moves the contact link deeper into the help centre. Tickets fall, and the problems stay.
- A growth engineer reading an [A/B test](/concept/ab-testing). A test reports the effect on the metric it measures. If the metric is a proxy, such as clicks on a button, the result says the proxy moved, not that the goal moved. The shorter the test, the more likely the proxy is the only thing it can measure.
- A machine learning engineer choosing a label. The model should predict a real outcome, but the real outcome is not in the data. The engineer picks the nearest column that is, such as spending in place of illness. This is the Obermeyer pattern, and it happens whenever training data is chosen for convenience.
In each case, the proxy was a reasonable first choice. The failure was not asking what else could move the proxy.
081 min
How to choose and check a proxy
Treat a proxy as a claim that can be tested, not as a fact. A short routine helps:
Write the target in one sentence with no metric in it.
"Readers understand the article" is a target. "Time on page" is not.
List the ways the proxy could move without the target.
For time on page: slower reading, a confusing layout, a page left open in another tab. If the list is long, look for another proxy.
Find data where both are measured, and compare them.
Take a sample of cases where you have the direct measure, such as a short survey of readers, and check that the proxy tracks it.
Repeat the check inside each group the decision affects.
A link that holds on average can fail for a subgroup, as it did for health-care spending.
Re-check on a schedule.
Choose a date for the next direct measurement before you start relying on the proxy.
Use more than one proxy when the decision is large.
Two proxies that fail for different reasons are harder to fool at the same time.
The routine costs little. The direct measurement in step 3 can be small, because its job is to test the link, not to replace the proxy.
091 min
When a proxy becomes a target
A proxy is reliable only while nobody pushes on it directly. The economist Charles Goodhart expressed the core idea in a 1975 article on United Kingdom monetary policy: any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. In other words, a link that was found by observation may not survive once people try to use it.
Later writers turned this into a general adage, now called Goodhart's law:
When a measure becomes a target, it ceases to be a good measure.
The mechanism is simple. A proxy and its target move together because both respond to the same underlying cause. When people are rewarded for the proxy, they find the cheapest way to raise it. That way is rarely the way that also raises the target. The proxy rises and the link breaks.
A team scored on "tickets closed" can close tickets faster by closing them early. A school scored on exam results can teach to the exam. In both, the measured number improves and the thing it was meant to track does not. The pattern is a form of incentive-caused bias.
The practical rule is to keep the proxy for learning and a separate, harder-to-influence measure for rewards.
101 min
When a proxy fails
Proxy failures fall into a small number of named types. Each has a tell.
- Partial coverage. The proxy captures only part of the target. Montgomery and colleagues, studying living standards, wrote that in poor countries, no single empirical measure can be expected to display all of the facets of the concept of income. Tell: two reasonable proxies for the same target disagree.
- Unequal link. The link differs between groups. Health-care spending tracked illness poorly for the group with less access to care. Tell: the proxy gives different answers for people with the same real outcome.
- Drift. The link changes over time, as with the flu estimates. Tell: the proxy and the direct measurement slowly separate.
- Gaming. People push the proxy directly, as in Goodhart's law. Tell: the proxy improves quickly, while no other signal does.
- Mistaking proxy for cause. A proxy only has to correlate with the target. Acting on it also requires that changing the proxy changes the target. Wider tree rings do not cause warm summers, and padding a metric does not cause value. Tell: an action aimed at the proxy moves the proxy and nothing else.
The last failure is the one the concept of correlation and causation warns about, and it is the easiest to commit, because the proxy is the number the team sees every day.
112 min
Proxy vs. nearby concepts
The deciding fact is what the quantity is for. A proxy is used instead of a target that cannot be measured. The neighbouring ideas below are used for something else.
| Concept | What it is | How it differs from a proxy |
|---|---|---|
| Direct measure | The target itself, read by an instrument | There is no stand-in, so no link to check |
| Key performance indicator | A number a team tracks and reviews | Often a proxy, but it can also be a direct count such as revenue |
| Confounder | A variable that affects both the cause and the effect in a study | It distorts a result. A proxy for an unmeasured confounder can reduce that distortion |
| Correlation | A tendency for two quantities to move together | It is the evidence a proxy relies on, not the proxy itself |
The word also has unrelated meanings. A proxy server passes network traffic, and a proxy vote is cast on someone's behalf. This entry covers only the research sense.
People often mix up proxies and confounders because one study can use both. When a study cannot measure something that affects its result, a researcher may measure a proxy for it and include that proxy in the analysis. The proxy is then a tool for handling the confounder, and the two remain different things.
?7 questions
Questions people ask
What is a proxy in research?
What is a good example of a proxy variable?
How do you know whether a proxy is good?
Can a proxy predict well and still be wrong?
Is a KPI the same as a proxy?
What is the difference between a proxy and a confounder?
What does Goodhart's law have to do with proxies?
§7 sources
Sources
Show all 7 sourcesHide sources
Obermeyer, Z., Powers, B., Vogeli, C. and Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science.
Ginsberg, J. et al. (2009). Detecting influenza epidemics using search engine query data. Nature.
NOAA National Centers for Environmental Information (2018). What Are Proxy Data?
Wikipedia. Proxy (statistics), a summary that cites Upton and Cook (2002) and Montgomery et al. (2000). (statistics)
Wikipedia. Proxy (climate), a summary of calibration practice. (climate)
Wikipedia. Google Flu Trends, a summary that cites Lazer et al. (2014).
Wikipedia. Goodhart's law, a summary that cites Goodhart (1975).





