Granger Causality: A Step-by-Step Guide

Key takeaways

Granger causality requires sufficiently long, regularly sampled data and is best used to generate hypotheses, not to prove definitive cause-effect relationships.
Proper data preprocessing, including stationarity testing and lag order selection, is critical to avoid spurious results and misinterpretation of predictive relationships.
Conditioning on potential confounders in multivariate models prevents false causality signals caused by third variables driving both series.
Nonlinear, time-varying, and high-dimensional extensions address the limits of classical linear Granger tests but demand more data and careful tuning.
Common pitfalls include omitted variables, sampling mismatches, structural breaks, and multiple testing, which can produce misleading causality conclusions if not properly checked.

What Is Granger Causality and When Should You Use It?

Granger causality asks a narrow, testable question: does the past of one time series improve your forecast of another series, beyond what that series’ own history already tells you? If yes, the first series “Granger-causes” the second in a predictive sense, not necessarily a mechanistic one. Analysts use it to screen for likely causal links in forecasting, economics, neuroscience, and genomics, where controlled experiments are impossible.

Granger causality is fundamentally about predictive precedence. The test, formalized by econometrician Clive Granger in 1969, asks whether including past values of a series Y reduces the prediction error of series X, compared to a forecast built only from X’s own history. If knowing Y’s past shrinks that error, Y “Granger-causes” X. That is the whole idea, and it is worth sitting with before you touch any math.

Two examples make the intuition concrete. In economics, does a rise in short-term interest rates help predict next quarter’s industrial output, once you already account for output’s own recent trend? If adding rate data to the forecasting equation meaningfully sharpens the prediction, rates Granger-cause output. In neuroscience, does the electrical activity in one brain region, recorded a few milliseconds earlier, help predict activity in a second region beyond what that second region’s own recent signal already explains? Researchers use exactly this logic to map information flow between cortical areas.

Notice what the test does not claim. It never asserts that X mechanistically causes Y in the everyday sense of one thing physically producing another. A rooster’s crow reliably precedes sunrise, and a naive time-series model might even find the crow “helps predict” daylight if the data were coarse enough. Granger causality is temporal precedence plus predictive value, not physics. That distinction is why the field usually says a series “Granger-causes” another rather than “causes” it outright.

Granger causality earns its place in your toolkit under specific conditions:

You have two or more time series measured at regular intervals, with enough observations to estimate a model reliably.
A randomized controlled experiment is not feasible, either for ethical, practical, or historical-data reasons.
Your goal is to generate or rank candidate causal hypotheses for further investigation, not to deliver final proof.
You care about forecasting performance and want to know whether adding a variable actually helps.

It is a poor fit when you can run an actual experiment. If you can randomize treatment and control groups, an A/B test or randomized design will give you a far stronger causal claim than any observational time-series method, Granger’s included. Quasi-experimental designs like interrupted time series or regression discontinuity also tend to produce sturdier causal evidence when their assumptions hold, because they exploit a known intervention point rather than relying purely on lagged correlation.

Across domains, the applications share a common thread. Economists use Granger tests to check whether money supply predicts inflation or whether one country’s stock index leads another’s. Neuroscientists use it, often under the label Granger causality analysis, to trace directional influence between brain signals. Genomics researchers apply it to gene-expression time series to hypothesize regulatory relationships between genes. Finance analysts use it to test lead-lag relationships between related assets. In every case, the output is a statistical signal worth investigating further, not a verdict.

How Does the Math Behind a Granger Test Actually Work?

The formal test rests on comparing two versions of a forecasting model called a vector autoregression, or VAR. Start with the simplest case: two stationary series, Xₜ and Yₜ, each measured over time t.

The restricted model forecasts Xₜ using only its own past values:

Xₜ = α₀ + ∑_i=1^p αᵢ Xₜ₋ᵢ + εₜ

Here, α₀ is a constant, αᵢ are coefficients on X’s own lags up to lag order p, and εₜ is the error term this model cannot explain.

The unrestricted model adds Y’s lagged values to the mix:

Xₜ = β₀ + ∑_i=1^p βᵢ Xₜ₋ᵢ + ∑_j=1^p γⱼ Yₜ₋ⱼ + ηₜ

The new terms, γⱼ, capture how much each lag of Y contributes to predicting X once X’s own history is already in the model. The null hypothesis of no Granger causality states that all the γⱼ coefficients equal zero. In plain language, that means Y’s past adds nothing once you already know X’s own past.

Testing that null typically uses an F-test comparing the residual sum of squares from the restricted model against the unrestricted one. If adding Y’s lags shrinks the unexplained error by more than chance would predict, the F-statistic crosses a critical threshold and you reject the null. Some software implementations instead report a likelihood-ratio test, which asks the same question through a slightly different statistical lens, comparing the log-likelihoods of the two model fits. Both routes should give you a consistent conclusion when the sample is reasonably large.

Choosing the lag order p is not a formality. Too few lags and the model misses genuine predictive relationships that unfold slowly; too many lags and you start fitting noise, inflating standard errors and eroding statistical power. Analysts typically choose p using information criteria such as the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC), fitting the VAR at several candidate lag lengths and picking the one that minimizes the chosen criterion. AIC tends to favor slightly longer lag structures and can overfit in smaller samples, while BIC penalizes complexity more heavily and often lands on a more parsimonious model.

Beyond the yes/no verdict from an F-test, researchers sometimes want to know how strong the predictive relationship is, and at which frequencies it operates. That is where the Geweke measure comes in: a decomposition of Granger causality in the frequency domain, letting you see whether Y’s influence on X concentrates at particular cycles (say, seasonal frequencies) rather than uniformly across time. It is a useful refinement when a simple significant or not-significant answer feels too coarse for the question you are asking, and it appears frequently in the applied literature summarized by the Annual Review of Statistics.

How Do You Run a Granger Causality Test on Real Data?

A reliable Granger test follows a sequence, not a single button press. Skipping steps is the most common way analysts generate a result that looks clean and turns out to be noise.

Check your data’s structure. Confirm both series are sampled at the same, regular interval, and that you have enough observations, generally several dozen at minimum, relative to your candidate lag order. Look for missing values and decide whether to interpolate, drop, or flag them; irregular gaps distort lag relationships badly.
Test for stationarity. Run an Augmented Dickey-Fuller (ADF) test and a KPSS test on each series. These two tests have opposite null hypotheses, ADF assumes non-stationarity until proven otherwise, KPSS assumes stationarity, so running both gives you a more honest picture than either alone.
Address non-stationarity properly. If a series carries a trend or unit root, first-differencing often fixes it. But if two series are individually non-stationary and share a long-run equilibrium relationship, differencing can destroy real information. Test for cointegration using the Johansen procedure, and if it is present, model the relationship with a vector error-correction model (VECM) instead of blindly differencing both series, a nuance the Annual Review of Statistics flags as a common source of misleading results.
Select the lag order. Fit the VAR at several lag lengths and compare AIC and BIC scores. Favor BIC in small to moderate samples to avoid overfitting, and check whether your causality conclusion holds steady across a couple of reasonable lag choices rather than pivoting entirely on one.
Fit the VAR and check residual diagnostics. Examine the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the model residuals. Leftover autocorrelation there signals your lag order is too short or your model is misspecified. Also confirm the fitted VAR is stable, meaning its characteristic roots lie inside the unit circle, or your forecasts will diverge nonsensically.
Run the Granger test itself and interpret the F-statistic or likelihood-ratio result against your chosen significance level, typically 0.05.
Add relevant covariates for a conditional test. If a third variable plausibly drives both series, include it in both the restricted and unrestricted models to guard against a spurious finding driven by a common cause rather than a genuine link between X and Y.

A few checks are easy to skip and costly to skip:

A few checks are easy to skip and costly to skip

Re-run the test with at least two lag specifications to confirm the finding is not an artifact of one arbitrary choice.
Report which stationarity transformation you applied and why.
Note whether cointegration testing changed your modeling approach.

For software, R’s vars package (functions like VARselect and causality) and base stats functions handle most of this workflow cleanly. Python’s statsmodels library offers grangercausalitytests along with ADF and KPSS implementations in the same ecosystem. MATLAB’s Econometrics Toolbox covers the same ground for users already working in that environment. None of these tools substitutes for the diagnostic steps above; they just execute the arithmetic once you have made the modeling decisions correctly. Statohub’s calculators can help you sanity-check preliminary descriptive statistics before you move into dedicated time-series software.

What Does Multivariate Granger Causality Tell You That a Simple Pair Can’t?

Bivariate Granger tests have a well-known blind spot: they cannot distinguish a genuine X-to-Y relationship from a case where a third variable Z drives both X and Y with a lag, creating an illusion of direct influence between them. Conditional, or multivariate, Granger causality fixes this by including Z explicitly in both the restricted and unrestricted models, testing whether X still helps predict Y after Z’s influence is already accounted for.

Picture three economic indicators: energy prices, transportation costs, and consumer food prices. A naive bivariate test might suggest energy prices Granger-cause food prices directly. Once you condition on transportation costs, which respond quickly to energy prices and then feed into food distribution costs, the direct energy-to-food link often weakens substantially, revealing transportation costs as the more immediate mediator. That is the value of conditioning: it separates direct predictive influence from influence that is really routed through something else.

Bivariate	Multivariate
Energy → Food (apparent link)	Energy → Transport (direct)
Transport not included	Transport → Food (direct)
Cannot separate mediation	No direct Energy → Food (weakened)

Figure 1. Energy–Food causal paths

Extending this logic to many series at once produces what researchers call network Granger causality, a directed graph where an arrow from node A to node B means A Granger-causes B, conditional on all the other nodes in the network. This approach shows up constantly in neuroscience, mapping influence between dozens of recorded brain regions, and in genomics, mapping regulatory relationships across gene-expression panels with many time points.

The catch is dimensionality. Once you have dozens or hundreds of series, fitting a full VAR with every series conditioning on every other one requires estimating an enormous number of coefficients, often more than your sample can support. Modern approaches address this with sparsity-inducing penalties, methods that assume most pairs of series have no direct causal link and shrink the corresponding coefficients toward zero, similar in spirit to lasso regression. This makes network estimation feasible, but it introduces a tuning problem: how aggressively should you penalize, and how do you know you have not shrunk away a real signal? The Annual Review of Statistics notes this trade-off between estimation feasibility and sample complexity as one of the active tension points in high-dimensional causal network research.

Reading a directed network responsibly means watching for two additional traps. First, feedback loops are common and legitimate: A Granger-causing B does not preclude B Granger-causing A too, and many real systems show bidirectional influence. Second, instantaneous causality, correlation between X and Y at the same time point rather than across a lag, sits outside what a standard Granger test can resolve. If your sampling interval is coarser than the true speed of the underlying process, a genuinely lagged relationship can appear instantaneous in your data, and the directional arrow you would want from the network simply is not recoverable at that resolution.

What Modern Extensions Go Beyond the Classic Granger Test?

The 1969 formulation assumes a linear relationship between stationary series, and real systems frequently violate both conditions. The methodological literature has spent the decades since building around those limits.

Nonlinear extensions use kernel methods or neural network architectures to capture predictive relationships that a linear VAR cannot represent, such as threshold effects or relationships that only appear during specific regimes. These methods can detect genuine nonlinear coupling that linear Granger tests miss entirely, but they trade away interpretability. A kernel-based causality score does not decompose into a clean coefficient you can report in a table, and these approaches typically demand more data to estimate reliably than a comparable linear model, a trade-off the Phys. Rev. E comparison study documented directly: standard Granger tests performed well on linear autoregressive systems but lost ground to information-theoretic and state-space methods once the underlying dynamics turned chaotic.

Time-varying Granger causality addresses a different problem: real relationships are rarely fixed forever. A monetary policy channel that predicted output strongly in one decade may weaken or reverse in the next. State-space formulations let the causal coefficients drift smoothly over time rather than assuming one fixed relationship across the entire sample, useful whenever structural change is plausible, which in economic and biological data is nearly always.

Mixed-frequency and subsampled data create a subtler headache. Economic series often mix monthly, quarterly, and annual measurements; neural recordings might be downsampled to reduce storage. Naively aggregating everything to the coarsest frequency throws away information and can distort or entirely reverse an apparent causal direction, an effect closely related to temporal aggregation bias. Recommended correction approaches include mixed-frequency VAR models that keep each series at its native sampling rate rather than forcing a lowest-common-denominator alignment, and careful attention to whether the aggregation itself, rather than the underlying process, is generating an apparent relationship.

Point-process adaptations handle event data directly rather than forcing it into evenly spaced bins. Neural spike trains and event logs (server failures, transaction timestamps) are fundamentally lists of event times, not continuous measurements. Discretizing them into fixed time bins to run a standard Granger test risks aliasing, where the true event-to-event timing gets distorted by an arbitrary bin width. Hawkes process models extend Granger-style influence measures into continuous time, letting one event stream’s history increase or decrease the momentary likelihood of events in another stream without ever forcing a bin-width choice, a technique the PMC review highlights as particularly well suited to spike-train and event-log analysis where discretization would otherwise distort the timing structure researchers actually care about.

Figure 2. Modern Granger Extensions

Where Do Granger Causality Tests Go Wrong?

Most misleading Granger results trace back to one of a handful of well-documented failure modes, and every one of them is checkable before you trust your output.

Omitted-variable confounding tops the list. If a third variable drives both series with different lags, a bivariate test will happily report a direct link that does not exist in any meaningful sense. The fix is conditioning on plausible confounders, not just running the pairwise test and calling it done.

Sampling-rate mismatch distorts timing in ways that are easy to overlook. If the true causal lag between two processes is faster than your measurement interval, the relationship can look instantaneous, weak, or backwards, depending on exactly how the aliasing falls. This is a persistent concern in neuroscience applications, where signal acquisition rates directly shape which lags are even detectable, a caveat the Brainstorm methods documentation addresses explicitly for EEG and MEG data.

Structural breaks and nonstationarity generate spurious detections at an alarming rate. Two unrelated series that both happen to trend upward over the same historical period will often produce a “significant” Granger result purely from the shared drift, not from any real predictive link. Stationarity testing is not a bureaucratic checkbox; it is the difference between a real finding and a shared-trend artifact.

Overfitting and multiple testing creep in when you run the same test across many lag specifications, many variable pairs, or many subsamples and report only the significant result. Testing a network of 20 series pairwise means running nearly 400 individual tests, and at a 0.05 significance threshold, roughly 20 of those will look significant by chance alone.

A short list of red flags worth checking before you report anything:

Does the finding survive at more than one lag specification?
Did you test and address stationarity before fitting the model?
Could a plausible third variable explain the relationship instead?
Are you correcting for multiple comparisons if you tested many pairs?
Does the residual diagnostics show any leftover autocorrelation?

Pro Tip: Run your Granger test on a randomly shuffled version of one of your series before trusting the real result. If the shuffled data still produces a “significant” causality finding at your usual threshold, something in your pipeline, likely stationarity or lag misspecification, is generating false positives regardless of what the real data show.

How Do You Run a Granger Test From Start to Finish?

Consider a realistic applied question: does a region’s weekly retail foot-traffic index help predict its weekly online search volume for local store hours, or does the relationship run the other way? Both series are observational, weekly, and span two years, roughly 104 data points each, a modest but workable sample for a low-order VAR.

Visual inspection and stationarity testing. Plot both series first. Suppose foot traffic shows a mild upward trend, alongside visible seasonal spikes around holidays. Run ADF and KPSS on both series; the raw foot-traffic series fails the ADF test (cannot reject a unit root) while search volume passes. First-differencing foot traffic resolves the issue, confirmed by re-running ADF on the differenced series.
Lag selection and VAR fit. Fit the VAR jointly on both series (differenced foot traffic and search volume) across lags 1 through 8, since weekly seasonal effects might take up to two months to show. BIC favors lag 2, while AIC favors lag 5. Fit both specifications and check whether your eventual conclusion holds at both, rather than picking whichever supports your hypothesis.
Run the Granger test and check diagnostics. At lag 2, test whether foot traffic’s past helps predict search volume beyond search volume’s own history, and vice versa in the reverse direction. Suppose the F-test in the foot-traffic-to-search-volume direction returns a p-value of 0.01, while the reverse direction returns 0.42. Check the residual ACF and PACF for both equations; if no meaningful autocorrelation remains, lag 2 is adequately specified. Confirm the fitted VAR’s stability condition holds.
Conditional check and robustness. A plausible confounder here is a regional promotional calendar (sales events) that could drive both foot traffic and search interest independently. Add a promotional-event indicator series to both the restricted and unrestricted models and re-run the test. If the foot-traffic-to-search-volume result stays significant after conditioning, that is meaningfully stronger evidence than the unconditional result alone. Re-run at lag 5 as well to confirm the direction of the finding does not flip.

To present this transparently, report the lag order chosen and why, the F-statistic (or likelihood-ratio statistic) and its p-value in both directions, the stationarity transformation applied to each series, whether a conditional variable was included, and whether the conclusion held across your two lag specifications. A reader should be able to reconstruct your reasoning from those details alone, without needing to re-run your code to trust the result.

How Should You Word and Report Granger Causality Findings?

The single most important habit when writing up Granger results is precise language. Write that “X Granger-causes Y” or “X predicts Y beyond Y’s own history,” never simply “X causes Y.” That phrasing distinction is not academic pedantry; it accurately reflects what the test can and cannot support, and reviewers in fields that use this method regularly will notice the difference.

Report the essential numbers rather than a bare significant/not-significant verdict: the lag order used and the criterion that selected it, the F-statistic or likelihood-ratio statistic with its degrees of freedom, the p-value, and whether the result held across at least one alternative lag specification. If you used a conditional model, state explicitly which covariates you controlled for and why you chose them.

For visualizations, a simple table listing each direction tested (X to Y, Y to X) alongside its statistic and p-value communicates more than a chart in most cases. When you are presenting a full network across many series, a directed graph diagram with arrow thickness scaled to effect size, paired with the underlying statistics in an appendix table, lets readers evaluate both the big picture and the details.

Recommended robustness checks to mention or include: a shuffled-data control test, sensitivity to lag choice, and, where feasible, a comparison against a nonlinear method if you suspect the true relationship might not be linear.

Where Does Granger Causality Fit in an Analyst’s Toolbox?

Granger causality is most useful as a screening tool rather than giving a definitive causal verdict. It earns its keep when you are exploring observational time-series data and need a defensible, quantitative way to rank which variables deserve deeper investigation. When a randomized experiment or a solid quasi-experimental design, like interrupted time series or regression discontinuity, is available instead, favor that route; it will carry more causal weight than any lagged-correlation method.

The learning path we suggest starts with the fundamentals: solid grounding in stationarity and regression diagnostics, then a clear look at why correlation is not causation, before tackling VAR models and hypothesis testing directly. Once the mechanics feel familiar, apply them to a real dataset you understand well enough to sanity-check the output. That is where Granger causality actually earns its reputation, or reveals its limits.

— Statohub

Where to Learn, Calculate, and Apply Granger Causality Next

Everything in this guide, stationarity checks, lag selection, VAR diagnostics, works better once the underlying concepts feel solid rather than memorized. Statohub’s Learn hub covers the statistical foundations, from probability to regression, that make a Granger workflow click instead of feeling like a checklist you are copying blindly. If time series and forecasting are your focus, the Forecasting & Time Series category walks through stationarity, autocorrelation, and model evaluation in more depth than any single article can.

For the applied side, the Applied Statistics hub is where theory meets real datasets, economic indicators, sales figures, experimental results, worked the way this article worked through the retail foot-traffic example. And when you need to double-check a descriptive statistic or run a quick calculation while you prep data for a VAR model, Statohub’s calculators handle the fast arithmetic so you can focus on the modeling decisions that actually matter. Start with the fundamentals if VAR notation still feels unfamiliar, then come back and run the workflow on a dataset of your own.

Sources

For the formal proof and original notation, read Granger’s 1969 paper directly. For a broad methodological survey covering nonlinear and high-dimensional extensions, see the Annual Review of Statistics overview. For applications across economics, neuroscience, and genomics alongside interpretation caveats, consult the PMC review. For empirical performance comparisons against nonlinear alternatives, the Phys. Rev. E study is essential. For hands-on workflow guidance, particularly in neuroscience contexts, the Brainstorm methods page is a practical companion.

Sources

Investigating Causal Relations by Econometric Models and Cross-spectral Methods — C. W. J. Granger (1969) Econometrica / JSTOR
Granger Causality: A Review and Recent Advances Annual Review of Statistics and Its Application / PMC
Comparison of six methods for the detection of causality in a bivariate time series — Phys. Rev. E (2018) Physical Review E / PubMed
Granger causality — Brainstorm / NeuroImage (methods page) University of Southern California / Brainstorm
Granger causality tests — null hypothesis, F-tests, and likelihood-ratio tests statsmodels documentation
Stationarity and detrending (ADF/KPSS) statsmodels documentation

FAQ

Frequently asked questions

What is Granger causality?: Granger causality asks a narrow, testable question: does the past of one time series improve your forecast of another series, beyond what that series' own history already tells you? If yes, the first series "Granger-causes" the second in a predictive sense, not necessarily a mechanistic one. Analysts use it to screen for likely causal links in forecasting, economics, neuroscience, and genomics, where controlled experiments are impossible.
What does multivariate Granger causality tell you that a simple pair can't?: Bivariate Granger tests have a well-known blind spot: they cannot distinguish a genuine X-to-Y relationship from a case where a third variable Z drives both X and Y with a lag, creating an illusion of direct influence between them. Conditional, or multivariate, Granger causality fixes this by including Z explicitly in both the restricted and unrestricted models, testing whether X still helps predict Y after Z's influence is already accounted for.
How should you word Granger causality findings?: The single most important habit when writing up Granger results is precise language. Write that "X Granger-causes Y" or "X predicts Y beyond Y's own history," never simply "X causes Y." That phrasing distinction is not academic pedantry; it accurately reflects what the test can and cannot support, and reviewers in fields that use this method regularly will notice the difference.