Econometrics is the use of statistical methods on economic data to estimate relationships between variables and test whether those relationships are causal. It asks how much revenue a pound of television spend produced, or how far unit sales fall when price rises by 1%. The discipline exists because economic data is almost never generated by a controlled experiment. Budgets, prices and demand move together, and econometrics is the toolkit built to separate one influence from another when nobody randomised anything.

How an econometric model works

The workhorse is regression. A modeller writes an equation in which an outcome, such as weekly sales, is explained by inputs such as media spend by channel, price, promotions and seasonality. The method finds the coefficients that best fit the historical data. Each coefficient is an estimate of how much the outcome moves when one input changes and the others are held constant, accompanied by a measure of uncertainty such as a standard error or a confidence interval.

A hypothetical example shows the logic. A retailer collects 156 weeks of national data. Its model regresses the logarithm of unit sales on the logarithm of average price, transformed television and search spend, and dummy variables for holidays. The price coefficient comes back at -1.8, which reads as a price elasticity: a 1% price increase is associated with a 1.8% fall in volume, other things equal. The television coefficient, converted to money, implies that 1 million pounds of spend generated 2.1 million pounds of incremental revenue, with a 90% interval of 1.2 million to 3.0 million pounds. The interval matters as much as the point estimate.

Marketing applications add two transformations before fitting. Adstock, a decay function named by the British media researcher Simon Broadbent in 1979, spreads each week's advertising over following weeks, because ads keep working after they run. Adstock has roots in Leendert Koyck's 1954 distributed-lag work. A saturation curve, often a Hill function, imposes diminishing returns so that the tenth million spent on a channel yields less than the first.

Beyond ordinary regression, the toolkit includes several families of method:

  • Time-series models handle data ordered in time, where this week's sales depend on last week's. Robert Engle and Clive Granger shared the 2003 Nobel prize for work on volatility and cointegration in such series.
  • Panel data follows many units, such as 200 regions, over many periods.
  • Instrumental variables use a factor that shifts the input but affects the outcome only through that input.
  • Difference-in-differences compares the change in a treated group before and after an intervention with the change in an untreated group over the same period.
  • Synthetic control, introduced by Alberto Abadie and Javier Gardeazabal in a 2003 study of the Basque Country, builds a weighted blend of untreated units to stand in for the treated one.

The central concern across all of them is identification: whether the data contains variation clean enough to give a coefficient a causal reading rather than a merely correlational one.

From Frisch to the credibility revolution

The Norwegian economist Ragnar Frisch coined the word econometrics in 1926. Four years later, in December 1930, 16 people met in Cleveland to found the Econometric Society, which describes itself as "an international society for the advancement of economic theory in its relation to statistics and mathematics." Its journal, Econometrica, first appeared in 1933 with Frisch as editor, a post he held for 21 years.

The statistical foundations were laid in the 1940s. Trygve Haavelmo's 1944 monograph, The Probability Approach in Econometrics, argued that economic relationships should be treated as probabilistic and studied simultaneously. He received the Nobel prize in 1989. Frisch and Jan Tinbergen had shared the first economics Nobel in 1969 "for having developed and applied dynamic models for the analysis of economic processes."

Two later critiques reshaped practice. In 1976 Robert Lucas argued that models estimated on past behaviour break down when policy changes, because people adjust to the new rules. In 1983 Edward Leamer published "Let's Take the Con Out of Econometrics" in the American Economic Review, attacking the fragility of results that depend on which variables a researcher chooses to include. The response, labelled the credibility revolution by Joshua Angrist and Jorn-Steffen Pischke in a March 2010 paper, favoured research designs built around natural or real experiments. The Royal Swedish Academy of Sciences recognised that shift on October 11, 2021, awarding half the prize to David Card and half jointly to Angrist and Guido Imbens "for their methodological contributions to the analysis of causal relationships."

Why marketers keep running into it

Most advertisers meet econometrics through marketing mix modelling (MMM). A 2025 vendor landscape from IAB Australia, the Interactive Advertising Bureau's Australian arm, described MMM as using "econometric methods to quantify marketing channel effectiveness" and profiled twelve providers, from Analytic Partners and Gain Theory to Google's Meridian and Meta's Robyn.

The open-source models have changed who runs the regressions. Google released Meridian globally on January 29, 2025, a Bayesian model in which prior beliefs, such as lift estimates from past experiments, are combined with observed data. On September 30, 2025 it added pricing and promotion variables and new prior options. Meta's Robyn fits ridge regression and tunes its settings with Nevergrad.

Privacy rules are a large part of the renewed interest. Econometric models work on aggregate data and need no cookies or device identifiers. The French trade body Alliance Digitale made that case in a January 2026 white paper that recommended pairing MMM with attribution and incrementality tests.

Econometric thinking also frames the experimental side of measurement. Incrementality studies, geo tests and sales lift studies estimate a counterfactual, and their results feed metrics such as iROAS, which differs from platform-reported ROAS precisely because it attempts a causal estimate.

The discipline also reaches courtrooms. In the multidistrict litigation against Google's advertising technology in New York, Judge P. Kevin Castel ruled on September 30, 2026 that economists Ali Hortacsu and Shengwu Li could testify, and sent damages claims including a $1.72 billion publisher class overcharge to a jury. In the UK, a roughly 5 billion pound claim on behalf of search advertisers was certified on August 5, 2026, with the figure estimated by the claimant's expert economist and resting partly on an assumed 50% pass-on rate.

Where it fails

The most serious failure is endogeneity: an input that is correlated with unobserved drivers of the outcome. Marketers raise budgets when they expect demand to rise, and targeting systems serve ads to people already likely to buy. Endogeneity usually inflates estimated returns. The eBay study published in Econometrica in 2015 by Thomas Blake, Chris Nosko and Steven Tadelis is the standard illustration: observational estimates put brand-search ROI at 1,632% even with controls, while a controlled shutdown found -63%.

Simulation evidence points the same way. A preprint posted to arXiv on August 21, 2026 by Niklas Heusch found that a standard mix model reported a paid search ROAS of 10.61 against a true value of 4.20, and still reported 8.41 when given the true confounders.

Other problems are familiar to any econometrician. Omitted variable bias arises when a driver of sales is left out and its effect is absorbed by something included. Multicollinearity occurs when channels move together, as television and search often do during a launch, so the model cannot tell them apart. Overfitting produces models that explain the past well and forecast poorly, a risk that rises when dozens of adstock and saturation parameters are tuned on 104 or 156 data points. The Lucas critique applies when a model trained under one bidding regime is used to plan spend under an automated one.

Governance and incentives draw criticism too. Practitioners have questioned whether platform-built models favour the platforms that build them. The Coalition for Innovative Media Measurement (CIMM) argued in July 2026 that default priors and embedded assumptions can shift data asymmetry rather than remove it, and proposed a standard model factsheet. Proprietary vendors face the opposite charge of opacity.

Not the same as

  • Statistics is the broader mathematical discipline of inference from data. Econometrics borrows its tools but adds economic theory about how the data was generated, and an emphasis on causal identification in non-experimental settings.
  • Data science is an applied field combining programming, data engineering and analysis, often aimed at prediction or product features.
  • Machine learning optimises predictive accuracy. A model can forecast sales well without saying what happens if one input changes. Econometrics prioritises interpretable, unbiased coefficients, sometimes at the cost of fit.
  • Marketing mix modelling is one application of econometrics, not a synonym. Lift tests, price elasticity studies, demand forecasting and antitrust damages models all use the same methods.

Recent developments

Google has moved econometric tooling closer to the media buying interface. On May 20, 2026, at Google Marketing Live, it announced Meridian inside Google Analytics 360. Meridian version 2.0.0 followed on September 2, 2026, and on September 9 Google declared its geo-experiment tool GeoX generally available, claiming budget savings of more than 31% for large advertisers. That figure had not been independently replicated as of October 2026. Google has also argued that open-sourcing lets teams audit the method.

The direction of travel is toward models calibrated by experiments, with tests used to anchor coefficients that observational data alone cannot pin down. Whether that combination resolves the identification problem, or merely relocates its assumptions into priors, is the field's open question.

Timeline

  • 1926: Ragnar Frisch coins the term econometrics.
  • December 1930: The Econometric Society is founded in Cleveland at a meeting of 16 people.
  • 1933: Econometrica publishes its first issue, edited by Frisch.
  • 1944: Trygve Haavelmo publishes The Probability Approach in Econometrics.
  • 1954: Leendert Koyck publishes his distributed-lag model.
  • 1969: Frisch and Jan Tinbergen share the first Nobel prize in economic sciences.
  • 1976: Robert Lucas publishes "Econometric Policy Evaluation: A Critique".
  • 1979: Simon Broadbent introduces adstock in the Journal of the Market Research Society.
  • 1983: Edward Leamer publishes "Let's Take the Con Out of Econometrics".
  • 1989: Haavelmo receives the Nobel prize.
  • 2000: James Heckman and Daniel McFadden share the Nobel prize for microeconometric methods.
  • 2003: Engle and Granger share the Nobel prize for time-series methods; Abadie and Gardeazabal publish the synthetic control study of the Basque Country.
  • March 2010: Angrist and Pischke publish "The Credibility Revolution in Empirical Economics".
  • 2015: Blake, Nosko and Tadelis publish the eBay search advertising experiment in Econometrica.
  • October 11, 2021: Card, Angrist and Imbens are awarded the Nobel prize.
  • January 29, 2025: Google releases Meridian globally.
  • September 18, 2025: IAB Australia publishes its MMM vendor landscape.
  • September 30, 2025: Meridian adds pricing and promotion variables.
  • January 29, 2026: Alliance Digitale publishes its measurement white paper.
  • May 20, 2026: Google announces Meridian in Analytics 360.
  • July 31, 2026: CIMM circulates its paper on MMM governance.
  • August 5, 2026: The UK Competition Appeal Tribunal certifies the search advertiser claim against Google.
  • August 21, 2026: Heusch's structural estimation preprint appears on arXiv.
  • September 9, 2026: Meridian GeoX becomes generally available.
  • September 30, 2026: Judge Castel sends ad tech damages claims to a jury.

Summary

  • Who: Academic economists, from Ragnar Frisch and Trygve Haavelmo to Joshua Angrist and Guido Imbens, built the discipline; today agency econometricians, measurement vendors, platforms such as Google and Meta, and expert witnesses in antitrust cases apply it.
  • What: The application of statistical methods, chiefly regression, time-series, panel, instrumental variable and quasi-experimental designs, to economic data in order to estimate relationships and test causal claims.
  • When: Named in 1926 and institutionalised with the Econometric Society in 1930; reshaped by the Lucas and Leamer critiques of 1976 and 1983 and the design-based shift recognised by the 2021 Nobel prize; packaged for advertisers through open-source models such as Meridian from 2025.
  • Where: In marketing mix models, incrementality and geo tests, pricing studies, central banks and government forecasting, and damages calculations in courts and tribunals including the US ad tech litigation and the UK Competition Appeal Tribunal.
  • Why: Economic data is rarely experimental, so separating the effect of advertising, price or market conduct from everything else moving at the same time requires methods designed for that purpose, and those methods remain vulnerable to endogeneity, omitted variables and the assumptions modellers choose.