Actiknow
Business Intelligence & Analytics

Causal Analytics in BigQuery: When Correlation Is Not Enough for a Business Decision

Learn when BigQuery causal inference can estimate business intervention effects, where its assumptions fail, and how leaders can validate results before acting.

Business analytics team reviewing performance dashboards and discussing data driven decisions

Search intent: Informational; executive and analytics-leader evaluation

Contents hide

The executive question is not “what moved?” but “what changed because we acted?”

Revenue rose after a new campaign. Customer churn fell after a service redesign. A new pricing rule coincided with a jump in average order value. These are useful observations, but none tells a leadership team whether the action caused the change. The difference matters because budgets, staffing and product decisions are often made as if every movement after a launch belongs to the launch.

Correlation identifies variables that move together. Causal analysis asks what would probably have happened to an outcome if an intervention had not occurred. That missing alternative is called the counterfactual. Because a business cannot observe the same customer, store or market both receiving and not receiving the intervention at the same time, the counterfactual must be estimated or created through an experiment.

BigQuery now includes AI.CAUSAL_EFFECT, a Preview function announced by Google Cloud on September 10, 2026, for estimating intervention effects in time-series data. It is useful for a particular class of business questions, but it does not turn observational data into experimental proof. Leaders should treat the output as decision evidence with assumptions, not a certified return-on-investment number.

Business executives reviewing marketing performance charts during an office meeting

Three questions that require three different methods

1. What happened?

Descriptive reporting compares outcomes across time, channels, products or regions. It can establish that conversion rose from 3.0% to 3.5%, or that weekly revenue exceeded a target. This is a measurement question. The key controls are accurate definitions, consistent reporting periods, trustworthy source data and clear ownership of KPIs.

2. What is associated with the outcome?

Correlation, segmentation and predictive modeling can reveal that repeat purchasers tend to use a particular feature, or that customers with longer support waits are more likely to churn. These findings help prioritize investigation and forecast risk. They cannot, on their own, tell you whether changing the feature or reducing the wait would create the predicted improvement. Confounding factors and reverse causality can explain an association.

3. What effect did our action produce?

Causal analysis tries to isolate the effect of a specific intervention from the outcome that would have occurred without it. A randomized controlled experiment can do this by assigning treatment independently of other drivers. When randomization is impractical, observational approaches such as interrupted time series, difference-in-differences, matching or synthetic control may be considered. Each method has a different identification assumption, data requirement and failure mode.

An executive team should first decide which question it needs answered. A dashboard can reliably answer the first. A prediction model can help with the second. The third needs a credible comparison, a defensible intervention date and explicit tests of alternative explanations.

What BigQuery AI.CAUSAL_EFFECT actually does

According to Google Cloud’s AI.CAUSAL_EFFECT documentation, the function takes a time-series measure, its timestamp and an intervention timestamp. It splits observations into pre-intervention and post-intervention periods, uses an ARIMA_PLUS time-series forecast to estimate a no-intervention baseline, and compares observed post-intervention values with that baseline. It supports analyzing multiple series through identifier columns and can return either aggregate results or point-by-point predictions.

In the aggregate output, absolute_effect is the sum of actual minus expected values after the intervention; relative_effect divides that difference by the summed expected values. The output also includes a p_value, a field called prob_causal_effect calculated as 1 minus the p-value, and a status field. The detailed view includes the actual series, estimated counterfactual, prediction bounds and an indicator for post-intervention observations.

Two distinctions are essential. First, the counterfactual is a model-based forecast, not an observed control group. Second, the field name prob_causal_effect should not be interpreted as a Bayesian posterior probability that an intervention caused the outcome: the documentation defines it arithmetically as 1 minus a frequentist p-value. Statistical significance can indicate an unusual deviation under a model, but it does not eliminate omitted events, data errors or selection bias.

Google Cloud lists the function as Preview under its pre-general-availability terms. Teams should confirm current regional availability, permissions, query costs and feature behavior before using it in a production decision workflow. The function’s launch and status are recorded in the BigQuery release notes.

Data engineer working at a computer in an office environment

Illustrative SQL: evaluate a pricing change

The following is a template, not a query against a real client dataset. Replace the example project, dataset, table, metric, date range and intervention timestamp with validated values. Aggregate the outcome to a consistent grain before running it.

SELECT *

FROM AI.CAUSAL_EFFECT(

  (

    SELECT sales_date, net_revenue

    FROM `your_project.analytics.daily_revenue`

    WHERE sales_date BETWEEN DATE ‘2026-01-01’ AND DATE ‘2026-05-31’

  ),

  data_col => ‘net_revenue’,

  timestamp_col => ‘sales_date’,

  intervention_timestamp => TIMESTAMP ‘2026-04-01 00:00:00+00’,

  output_time_series => TRUE

);

The example assumes exactly one net-revenue observation per day, a known pricing-change date and no other transformations hidden inside the input. In real work, the first query should usually be a tested daily metric table, not raw transactions. Use identifier columns only when the series and intervention definitions are valid for each group. A query that runs successfully can still answer the wrong business question.

The business case: campaign lift, not just campaign timing

Consider a hypothetical company that launches a promotion on April 1. Its April sales are $1.12 million, while the model estimates $1.00 million without the promotion. The modeled difference is $120,000, or 12% above the expected baseline. These numbers are illustrative only, not Actiknow results and not a BigQuery benchmark.

That $120,000 is not automatically incremental profit. The business may have discounted products it would have sold at full price, shifted purchases forward from May, raised fulfillment costs or cannibalized other channels. Even the incremental revenue estimate can be biased if a competitor exited the market, a holiday changed demand or tracking was modified near April 1. The analysis needs a clearly defined measurement window, an economic bridge from revenue to contribution margin and sensitivity checks.

A useful decision memo would state: observed change; estimated no-intervention outcome; estimated difference and uncertainty; major competing explanations; expected cost and margin impact; and the action leadership will take if the result remains credible. This is much more useful than a chart titled “campaign caused 12% growth.”

When a time-series counterfactual is a reasonable starting point

A forecast-based intervention analysis can be helpful when the business has a reasonably long and stable history, a clearly dated action, consistent measurement, and no obvious simultaneous shock. Examples include a single operational policy change, a feature rollout with a documented start date, or a channel initiative that did not coincide with other major changes.

It becomes less convincing when a company launches several initiatives together, changes data collection at the same time, selects the intervention date after looking at the outcome, or operates in a volatile environment where historical patterns are poor guides. A statistically polished model cannot repair a fundamentally ambiguous treatment definition.

Before you run the function, answer these six questions

  • What exactly changed, for whom, and on what date?
  • What outcome and time grain were defined before examining results?
  • How much pre-intervention history covers relevant seasonality?
  • Were any promotions, pricing changes, outages, product releases or market events concurrent?
  • Did tracking, attribution or accounting definitions change?
  • Is there a plausible comparison group or holdout that could challenge the forecast?

If the answers are weak, treat the exercise as exploratory rather than as an ROI estimate. Sometimes the correct next step is to improve event instrumentation or plan a randomized test, not to select a more sophisticated model.

Choosing the right causal design

Randomized A/B test: strongest option when practical

Randomly assign eligible units to treatment and control, specify the primary outcome and guardrails in advance, and analyze the intention-to-treat effect. This is often the preferred approach for digital product changes, messaging, offers and other interventions where randomization is feasible. Watch for sample-ratio mismatch, spillover between groups, premature stopping and metrics that do not reflect business value.

Difference-in-differences: when comparable groups exist

Compare outcome changes in treated and untreated groups before and after the intervention. The crucial assumption is that, absent treatment, the groups would have followed sufficiently similar trends. Pre-trend checks, common shocks and changes in group composition matter. A geographic rollout may offer a candidate design, but geography alone does not make groups comparable.

Synthetic control: when one treated unit has several possible donors

Construct a weighted comparison from untreated markets, branches or business units whose historical patterns approximate the treated unit. This can be attractive for a large regional initiative with a limited number of treated units. It requires a defensible donor pool, pre-period fit and careful assessment of spillovers.

Interrupted time series: when no credible control exists

Estimate a pre-intervention trajectory and examine whether the post-intervention level or trend departs from it. BigQuery’s AI.CAUSAL_EFFECT belongs in this broad forecast-based family. It can provide a useful first estimate, but causal interpretation depends heavily on whether the intervention is the most credible explanation for the deviation. The method is especially vulnerable to concurrent shocks and unstable trends.

Regression discontinuity and matching: specialized alternatives

If eligibility changes sharply at a known threshold, regression discontinuity may estimate a local effect near that cutoff. Matching or weighting can make observed groups more comparable, but cannot remove bias from important unmeasured confounders. Neither method should be chosen merely because it is available in a tool. The assignment mechanism determines whether it fits.

Data analytics professionals collaborating while reviewing charts on computer screens

Seven failure modes that executives should challenge

1. The intervention date is chosen after inspecting the spike

Selecting the most dramatic date retrospectively exaggerates apparent evidence. Record the intervention date from deployment logs, campaign calendars or policy approvals before looking at outcomes. If several candidate dates are tested, disclose that search.

2. A tracking change looks like a business effect

Cookie consent, attribution rules, app instrumentation, payment status definitions and warehouse transformation changes can create discontinuities. Reconcile independent signals where possible. A sales lift in analytics that is absent from the order ledger should trigger a measurement investigation, not a budget increase.

3. Seasonality is mistaken for lift

Weekly cycles, holidays, paydays, academic calendars and promotional seasons affect outcomes. Evaluate whether the history covers those patterns and whether the post-period is comparable. Short histories can generate unstable counterfactuals even if the function returns an output.

4. Multiple interventions overlap

A website redesign, pricing adjustment and media campaign launched in the same week cannot usually be separated by a single-series forecast. Consider staggered rollouts, holdouts or a different design. Report the combined-period change if attribution to individual initiatives is not defensible.

5. The metric hides substitution and cannibalization

Higher online orders might be displaced store orders; lower support tickets might reflect a broken contact form; higher average order value might result from fewer small purchases. Include guardrail metrics, contribution margin, customer outcomes and adjacent channels.

6. Statistical significance is treated as commercial significance

A small effect can be statistically detectable but economically irrelevant. Conversely, a material-looking estimate may have uncertainty too wide to justify a major decision. Evaluate effect size, uncertainty, implementation cost, downside risk and reversibility together.

7. The result cannot be reproduced

If analysts cannot recover the source data snapshot, metric definition, intervention timestamp, SQL version and model output, the result is not auditable. A decision-quality analysis needs a reproducible evidence trail, not a screenshot from a one-off notebook.

How to interpret the output without overstating certainty

Start with absolute_effect and relative_effect as model-estimated differences over a specified post-intervention period. Then inspect the detailed series. Does the gap begin near the intervention? Does it persist? Are prediction bounds broad? Are there unusual missing dates or outliers? Is the model reporting a status error? Would changing the pre-period or excluding a known anomalous week materially alter the estimate?

Do not report p_value as the probability that the null hypothesis is true. Do not report 1 minus p_value as the probability the intervention caused the change. Prediction bounds around a forecast are not automatically a complete uncertainty interval for causal identification. Separate model uncertainty from uncertainty about whether the underlying assumptions are valid.

A strong executive summary might say: “Under a forecast-based counterfactual and assuming no material concurrent changes, the outcome was estimated to be above baseline over the four-week period. The result is sensitive to the pre-period selected and should be validated with a holdout in the next rollout.” This communicates both insight and limits.

Build a trustworthy measurement foundation first

Causal analysis is downstream of reliable data engineering. Define the business metric once, including exclusions, refunds, late-arriving transactions and currency treatment. Record intervention metadata such as deployment date, eligible population, geography, channel, targeting rule and known concurrent activity. Use consistent time zones and calendar logic. Preserve raw inputs, transformation versions and validated daily aggregates.

Organizations that are still reconciling dashboards should first invest in dependable business intelligence and data modeling services, including source integration and metric governance. An effect estimate built on conflicting revenue definitions will create more argument than clarity.

The dashboard layer should distinguish observed performance from estimated counterfactuals. In executive dashboard design, show actual and expected lines together, a clearly labeled intervention marker, a selected measurement window, uncertainty, guardrails and a concise assumptions panel. Avoid a single green “incremental ROI” tile with no method disclosure.

For operational data scattered across CRM, billing, web analytics and fulfillment systems, data integration and custom solutions may be necessary to align identifiers and event timestamps before any meaningful evaluation. The engineering deliverable is not only a pipeline: it is an auditable definition of who was exposed to what, when, and what outcome was subsequently measured.

Data engineer examining technical dashboards at a multi monitor office workstation

A practical 30-day pilot for a data leader

Week 1: define the decision and pre-register the analysis

Choose one consequential but bounded intervention, such as a regional operational change or a staged campaign. Write a one-page decision brief covering the action, treatment population, primary outcome, guardrails, intervention date, analysis window, alternative explanations and decision threshold. Agree who owns the business interpretation.

Week 2: validate the underlying series

Reconcile daily or weekly outcomes to source systems. Inspect missing periods, outliers, backfills, duplicate transactions and tracking changes. Confirm that enough historical observations exist to characterize relevant patterns. Build a documented metric table and preserve the SQL and source snapshot.

Week 3: run competing analyses

Use AI.CAUSAL_EFFECT as a forecast-based baseline where appropriate. Run sensitivity checks with different defensible pre-periods and post-period lengths. Examine placebo dates before the actual intervention, inspect pre-period forecast behavior and, if possible, compare with an untreated group. A fragile estimate should be reported as fragile.

Week 4: make a bounded decision

Present observed change, modeled effect, uncertainty, sensitivity results, competing explanations and estimated economic value. Decide whether to scale, run a better experiment, gather more data or stop. Capture the decision and revisit it when later evidence becomes available. The pilot succeeds if it improves decision discipline, even when the answer is “we cannot attribute this lift reliably.”

An executive decision checklist

Before approving an intervention-effect claim, ask:

  • Is the treatment unambiguous?
  • Was the outcome selected before inspecting the result?
  • Is the comparison credible?
  • Could another event explain the difference?
  • Has measurement remained stable?
  • Are uncertainty and sensitivity visible?
  • Does the estimate connect to contribution margin or another business objective?
  • Can the analysis be reproduced?
  • Is the recommended action reversible if the estimate is wrong?

A useful governance rule is to classify results as descriptive, associational, observational causal estimate or randomized experimental evidence. Do not collapse those categories into a single “data-driven” label. That distinction helps executives compare claims from different teams on equal terms.

Business professionals discussing kpi dashboards and analytics results in a meeting

Frequently asked questions

Does correlation ever prove causation?

No. Correlation alone does not establish that changing one variable will change another. Causal interpretation requires a credible design, assumptions about alternative explanations and, ideally, experimental or quasi-experimental evidence.

Is AI.CAUSAL_EFFECT generally available in BigQuery?

As of November 2026, Google Cloud documents AI.CAUSAL_EFFECT as a Preview feature. Verify current status and applicable pre-GA terms before depending on it for production workflows.

Does the BigQuery function require a control group?

No. Its documented approach uses an ARIMA_PLUS forecast to estimate a counterfactual from the time series itself. That makes the function convenient, but also means it cannot independently rule out other events that occurred near the intervention.

Can the function measure marketing incrementality?

It can estimate a time-series deviation after a marketing intervention. Calling that deviation incremental impact is defensible only to the extent that the counterfactual assumptions hold. For major spend decisions, randomized holdouts or credible comparison groups are preferable when feasible.

What does a low p-value tell us?

Under the statistical model and its assumptions, the observed deviation is less compatible with a no-effect null hypothesis. It does not tell you the probability the campaign worked, guarantee the assumptions, or establish financial return.

How much historical data is needed?

Google’s documentation notes that fewer than three observations can cause a short-series error, but that is a technical minimum, not a practical standard for credible causal inference. The required history depends on seasonality, volatility, frequency and the length of the post-period. In many business settings, substantially more data is needed.

Can we run this for multiple stores or regions?

Yes. The function supports identifier columns for multiple time series. However, different regions may have different launch dates, market conditions and data quality. Segment-level results require consistent intervention definitions and safeguards against multiple-comparison overinterpretation.

What if several initiatives launch on the same date?

A single time-series counterfactual generally cannot isolate each initiative’s contribution. Consider a staggered rollout, controlled experiment, additional comparison data or a clearly labeled estimate of the combined change.

Should executives put causal estimates directly into KPI dashboards?

Only with clear labels, visible assumptions, documented method and links to reproducible analysis. Keep actual observed metrics separate from modeled counterfactuals, and avoid displaying a modeled effect as booked revenue or realized profit.

The bottom line

BigQuery makes forecast-based intervention analysis easier to run, but ease of execution is not the same as strength of causal evidence. The value comes from matching the business question to the right design, measuring consistently, challenging alternative explanations and turning uncertainty into a proportionate decision. A leader who asks “what else could have caused this?” is doing more for analytical quality than one who asks for a more impressive chart.

Discuss an analytics decision with Actiknow

If your team is trying to connect operational changes to business outcomes, Actiknow can help with the BI, data integration and reporting foundations needed for reliable measurement. Share the decision you are trying to make and the systems involved when you contact the Actiknow team. The first step is to assess whether the available data and study design can support the claim you want to make.

Editorial references: Google Cloud BigQuery release notes (September 10 and October 6, 2026) and the AI.CAUSAL_EFFECT function reference. Feature status and technical behavior should be rechecked before publication.