The Ten Commandments of Econometrics
Guiding principles for turning data into meaningful economic conclusions.
“There are two things you are better off not watching in the making: sausages and econometric estimates.”
Edward E. Leamer (1944-2025), American econometrician
In public discourse, economics is often seen as a field of big ideas: markets and incentives, inflation and inequality, growth and trade. But behind those ideas lies a quieter, more technical discipline — econometrics — that gives empirical substance to economic theory. If economics is the study of how people make decisions under constraints, then econometrics is how we test whether our theories about those decisions actually hold up in data.
At its core, econometrics is the application of statistical methods to economic questions. It is how we move from verbal reasoning to quantitative evidence. How we distinguish patterns from noise, causation from correlation, and insight from illusion. Whether we are evaluating the impact of education on earnings, measuring the effect of a policy intervention, or estimating the price elasticity of demand, we are likely using econometric tools.
But econometrics is not merely a set of techniques. It is a way of thinking rigorously about identification, uncertainty, and model validity. It demands clarity of purpose and a deep respect for the limitations of data. Done well, it disciplines our narratives with evidence. Done poorly, it gives statistical gloss to unfounded claims.
What follows is a compact articulation of the principles that govern credible empirical work. These are not technical instructions, but foundational commitments — what might be considered the ten commandments of econometrics. They are not taught all at once, and not always stated explicitly, but they form the unwritten constitution of sound empirical practice.
1. Thou shalt begin with a theoretical model.
Every econometric analysis should begin with a question grounded in economic reasoning. Before estimating anything, you must ask what causal mechanism you are testing and what assumptions justify the specification. Without a model, regression becomes a mechanical exercise with no interpretive discipline.
Theoretical models need not be complex. They may be informal or stylized. What matters is that they provide structure: they define the outcome, clarify the role of covariates, and specify what is held constant. A theory tells you which variables belong in the model, which do not, and what sign and magnitude you might expect.
Consider the canonical Mincer equation:
Here, log wagei is the natural logarithm of individual i's wage, capturing proportional differences rather than absolute amounts. The constant 𝛽0 represents the baseline wage level for someone with zero education and experience. The coefficient 𝛽1 measures the average percentage increase in wages associated with one additional year of education. The terms 𝛽2 ⋅ expi and 𝛽3 ⋅ expi2 capture the effect of labor market experience on wages, allowing for nonlinear returns that typically diminish at higher experience levels. Finally, 𝜀i is the error term, accounting for unobserved factors affecting wages.
This equation does not arise arbitrarily. It reflects a human capital model in which schooling and experience raise productivity, with diminishing returns to experience. Each coefficient has a theoretical interpretation. Estimating this model without understanding that structure would reduce it to curve fitting.
A regression without theory risks becoming a search for statistical patterns rather than economic understanding. When models are built without a guiding framework, interpretation drifts. Significance may be misread as meaning. Causal claims may emerge where only associations exist.
Start with theory. It defines the question, guides the design, and ensures that your empirical work speaks to something beyond the data at hand. Without it, your model may be technically correct but conceptually hollow.
2. Thou Shalt Know Thy Data.
Data are the foundation of any econometric analysis. Understanding their origins, limitations, and construction is essential before proceeding with estimation. Data do not arrive as perfect representations of economic reality; they reflect measurement decisions, sampling procedures, and institutional contexts.
Knowing your data means being familiar with how variables are defined and collected. Are key variables self-reported or administrative? Is the sample representative of the population of interest, or does it suffer from selection bias? Are there missing values or outliers that require attention? Ignoring these details can lead to misleading results or invalid inference.
Moreover, understanding the timing and frequency of the data is crucial. Cross-sectional, panel, and time-series data require different approaches to modeling dependence, dynamics, and heterogeneity. Misapplying methods without regard to the data structure can invalidate standard assumptions.
Finally, exploring the data through summary statistics, visualizations, and simple cross-tabulations provides insight into distributional features, correlations, and potential anomalies. This exploratory phase is not mere routine; it informs model specification and highlights data quality issues early.
In short, econometrics is not a black box. Deep familiarity with the data grounds empirical work in reality and protects against common pitfalls that arise from treating data as abstract numbers rather than imperfect reflections of complex social phenomena.
3. Thou Shalt Make Thy Identification Strategy Explicit.
The credibility of any causal claim rests on the identification strategy. Identification explains how the variation used to estimate the parameter of interest can be interpreted as exogenous or as-if randomized. Without this clarity, estimates reflect correlations that may be confounded by omitted variables or reverse causality.
Articulating the identification strategy involves stating clearly what source of variation isolates the causal effect. This could be a natural experiment, an instrumental variable, a difference-in-differences setup, or a randomized controlled trial. Each design comes with assumptions that must be transparent and testable when possible.
Failure to make the identification strategy explicit leads to ambiguity. Researchers, policymakers, and readers cannot assess the validity of the inference without understanding the source of exogenous variation. Moreover, poor or absent identification may generate biased and inconsistent estimates that misinform decisions.
In sum, identification is not a technical afterthought but the backbone of econometric inference. It bridges the gap between statistical associations and meaningful causal interpretation. Every empirical paper must state its identification strategy clearly and justify why it is plausible.
4. Thou Shalt Respect the Error Term.
The error term is not a nuisance; it is a signal. It captures everything unobserved, unmeasured, or misspecified in the model. Whether it behaves well or misbehaves determines whether your estimates can be trusted.
In the classical linear regression model,
the error term 𝜀i carries the residual variation in yi after conditioning on the covariates x1i, … , xki. For the OLS estimator to be unbiased and consistent, the key assumption is that the regressors are exogenous:
This assumption implies that the error term is uncorrelated with the explanatory variables, and thus that the regressors capture all systematic variation in the outcome. If this fails — due to omitted variables, simultaneity, or measurement error — then OLS becomes biased and inconsistent.
Further, the assumption of homoskedasticity,
ensures that OLS is the best linear unbiased estimator (BLUE). If the variance of the error term depends on the level of some regressor, then standard errors are incorrect, and inference becomes unreliable. Similarly, if observations are not independent (as in panel or time series data), and errors are serially correlated, then the usual OLS standard errors underestimate true sampling variability.
Violations of these conditions do not just affect statistical properties — they signal economic structure. If residuals are systematically large for a certain group, you may have omitted a relevant variable. If errors are correlated with a covariate, you may be misattributing variation to the wrong source. These are not technicalities; they are substantive.
Respecting the error term means checking residual plots, testing for heteroskedasticity, and diagnosing misspecification. It means thinking carefully about what the error term contains and whether its properties are consistent with your identification strategy.
Treat it with care. It is not just random noise. It is the boundary between what your model explains and what it ignores.
5. Thou Shalt Be Wary of Functional Form.
All econometric models impose structure, and nowhere is this more obvious — and more dangerous — than in the choice of functional form. Every time you specify a linear relationship, include a quadratic term, or log-transform a variable, you are making a substantive assumption about how the world works. These choices affect not just estimation, but also interpretation and policy relevance.
A standard linear model assumes that the marginal effect of a regressor is constant:
Here, 𝛽1 implies that a one-unit increase in xi is associated with the same change in yi, regardless of the level of xi. This may be a useful approximation, but often it is not realistic. Economic behavior frequently involves diminishing returns, thresholds, or nonlinearities.
Adding a quadratic term, as in
introduces curvature, but even this remains a strong assumption. It implies a smooth, symmetric parabola, which may not match the actual relationship. Worse, if the true relation is piecewise linear, discontinuous, or involves saturation points, the quadratic fit may mislead more than it clarifies.
Functional form choices also affect identification. For example, including log-terms changes the interpretation: in
𝛽1 represents a semi-elasticity, meaning a one-unit increase in xi is associated with a 𝛽1 ⋅ 100 percent change in yi. In contrast, in
𝛽1 measures the change in yi associated with a 1 percent increase in xi. These are not cosmetic differences. They imply distinct theories of behavior.
Whenever possible, let theory and data guide your functional form. Use plots, local polynomial regressions, or binning to explore nonlinearity before settling on a parametric structure. Rely on robustness checks to verify that your conclusions do not hinge on a particular specification.
A model must simplify, but it should not distort. Functional form is a modeling choice that embeds assumptions. Make those assumptions explicit, test them when you can, and never mistake a convenient simplification for a structural truth.
6. Thou Shalt Not Confuse Statistical Significance with Economic Significance.
A small p-value is not a license to declare victory. Statistical significance only tells you that an estimated effect is unlikely to be zero, given a set of assumptions. It says nothing about the magnitude, importance, or relevance of that effect in the real world.
In large samples, even minuscule differences can appear highly significant. For instance, consider the model:
with an estimate of ^𝛽1 = 0.02 and a standard error of 0.005. The t-statistic is 4, the p-value is below 0.001, and you may be tempted to call this a strong result. But if yi is income and xi is years of education, then a 0.02 increase means just 2 percent higher earnings per year of schooling. Whether that is economically meaningful depends on the context, not on the t-statistic.
Conversely, effects that are economically large may fail to be statistically significant in small samples. This does not mean the effect is absent; it may mean that the data is too noisy or too sparse to detect it. Treating significance as a binary criterion obscures this nuance.
Always report effect sizes and confidence intervals. Ask whether the magnitude is substantively meaningful, not just statistically distinct from zero. Focus on estimation, not just hypothesis testing. The point of econometrics is to learn something about economic relationships, not merely to pass a significance threshold.
Statistical significance is a tool, not a conclusion. Use it wisely, interpret it cautiously, and never confuse precision with importance.
7. Thou Shalt Guard Against Overfitting.
A model that explains everything in your sample may explain nothing outside it. Overfitting occurs when you include too many variables, too many interactions, or too much flexibility relative to the amount of information in the data. The model captures noise as if it were signal, and the result is poor out-of-sample performance.
Overfitting is a problem of illusion. Within-sample fit looks impressive — R2 is high, residuals are small — but the model has learned the idiosyncrasies of the sample, not the underlying structure. It performs well on the data it has seen, but fails to generalize.
In a simple linear regression,
adding more covariates will always improve the fit mechanically. But unless those covariates are motivated by theory or improve prediction on new data, they may do more harm than good. The adjusted R2 penalizes extra variables, but it is no substitute for careful specification.
The risk is greater in small samples or when covariates are highly collinear. In those cases, the model may fit the sample closely while producing unstable estimates that vary wildly with slight changes in the data. This is especially dangerous when results are used for policy, forecasting, or counterfactual inference.
Guarding against overfitting requires discipline. Use theory to justify included variables. Use cross-validation or holdout samples to evaluate predictive performance. Test robustness by resampling or applying the model to new data. Resist the temptation to chase significance by adding regressors that have no grounding in the research question.
Fit is not the goal. Understanding is. A parsimonious model that captures the essential structure is more reliable than an overfit model that captures every quirk. Simplicity is not weakness; it is strength under constraint.
8. Thou Shalt Test Robustness Relentlessly.
No single specification proves a result. Econometric findings must survive stress. Robustness checks are not optional; they are essential. They test whether your conclusions depend on one modeling choice, one sample definition, or one estimator.
Start by varying the specification. Does the estimated effect hold when you drop or add controls? When you transform variables? When you use a different functional form or exclude influential observations? If results vanish when you tweak the model, the result was never solid to begin with.
Next, check the sample. Are results driven by outliers or a particular subgroup? What happens if you exclude one year, one region, or one tail of the distribution? If an effect only appears in one slice of the data, that limits its generality.
Then try alternative estimators. OLS is sensitive to assumptions; it may be biased if errors are heteroskedastic, or if regressors are endogenous. Re-estimating with robust standard errors, fixed effects, or instrumental variables can reveal whether the result is methodological or real.
Robustness is not just about mechanical checks. It is about credibility. Each test you run gives the reader more reason to believe that the result reflects a real economic relationship, not a fluke of the data or a quirk of the model.
A fragile result is not a result. If your conclusion changes under minimal pressure, it was never convincing. Robustness reveals strength. Without it, inference is built on sand.
9. Thou Shalt Not Worship the R2.
The coefficient of determination, R2, is widely reported and often misunderstood. It measures the proportion of variance in the dependent variable explained by the regressors. But a high R2 does not imply causality, relevance, or validity. It simply reflects how well the model fits the sample data, which is not the same as how well it explains the economic mechanism at hand.
In the linear model
R2 is calculated as:
where SSR is the sum of squared residuals and TSS is the total sum of squares.
A higher R2 means smaller residual variance relative to the total variance. But you can increase R2 arbitrarily by adding regressors — even irrelevant ones. This makes R2 easy to inflate and poor as a measure of model quality.
Moreover, a high R2 can coexist with severe omitted variable bias, endogeneity, or poor out-of-sample performance. Conversely, a low R2 may be entirely appropriate in models where outcomes are inherently noisy or where explanatory variables are weakly predictive but still causally important.
In causal inference, what matters is the validity of the identification strategy, the plausibility of the assumptions, and the robustness of the effect — not how much of the outcome variation is explained. A valid instrumental variable estimator may have a low R2 but still yield an unbiased estimate of a causal parameter. That makes it more valuable than a high-R2 model built on invalid assumptions.
Do not judge a model by its R2. Judge it by its ability to answer the question you are asking. The point is not to fit the data perfectly, but to understand the underlying process.
10. Thou Shalt Always Distinguish Correlation from Causation.
Nothing in econometrics is more misunderstood in public discourse, or more fundamental to empirical credibility, than the distinction between correlation and causation. Correlation measures co-movement. Causation explains why that movement occurs. The two are not interchangeable.
In formal terms, let the covariance between X and Y be nonzero. This tells us that changes in X are associated with changes in Y, but it does not tell us whether:
X causes Y,
Y causes X,
A third variable Z causes both,
The association is spurious due to data quirks or model misspecification.
For example, countries with more cellphone subscriptions per capita also tend to have higher GDP. But cellphones do not cause prosperity. Wealthier countries adopt new technology more readily. Confusing correlation for causation here would lead to misleading policy conclusions.
Causal inference requires an identification strategy — an argument for why the variation in X can be interpreted as exogenous. This may arise from random assignment, a valid instrument, or structural assumptions supported by theory. Without it, no matter how strong or statistically significant the correlation, it cannot justify a causal claim.
The danger is not just academic. Policy built on spurious correlations can misallocate resources, reinforce bias, or produce unintended harm. Econometrics offers tools to move beyond correlation, but only if we use them carefully and transparently.
Therefore, every empirical economist must internalize the most important statistical warning ever issued: correlation is not causation. Recognize it. Respect it. Explain it.
In sum, these ten principles are the most important guidelines to keep in mind for credible and rigorous econometric practice. They serve as the foundation of modern empirical economics, guiding how we turn data into meaningful evidence grounded in theory and careful reasoning.
While there are certainly more detailed rules and nuances to explore (not even including time-series econometrics, which is a whooole new can of worms), these core commitments form the essential framework for sound analysis. Keeping them in mind will help you avoid common pitfalls and maintain clarity, transparency, and integrity in your work.
I may share additional “commandments” another time, but for now, these represent the critical pillars every good economist should internalize to produce trustworthy and relevant results.
Very good (and fun) write-up! Love that Edward Leamer’s quote from the beginning of your article lol.
Can’t stress #7 and #10 enough.
Wish I could have read this years ago