Correlation vs causation is the difference between two things moving together and one thing making the other happen. On the IIBA Certification in Business Data Analytics (CBDA) exam, the safe conclusion from "sales rose after the campaign" is always correlation; causation needs a test with a comparison group, such as a holdout or an A/B test.
Why correlation vs causation matters on the CBDA exam
The CBDA exam punishes overclaiming, and causal language is its most common form. Causal claims show up in the three largest domains: you frame a diagnostic question, interpret results, and recommend action based on what the analysis can defend.
| CBDA fact | Value |
|---|---|
| Questions | 75, multiple-choice, scenario-based |
| Time | 120 minutes |
| Pass mark | Not published; pass or fail shown on screen |
| Interpret and Report Results | 20% of questions |
| Use Results to Influence Business Decision Making | 20% of questions |
Last verified: October 2026 on the IIBA CBDA page and the CBDA Certification Handbook (May 2026).
How the IIBA frames it
The IIBA treats "why did it happen?" as its own type of analytics. IIBA's public Introduction to Business Data Analytics: A Practitioner View (2019, §1.2) lists four types: descriptive (what happened), diagnostic (why it happened), predictive (what is likely to happen) and prescriptive (what should happen if we take a path). A causal claim is a diagnostic answer, so it needs diagnostic evidence, not a timeline.
The same document's guidance on deriving insights is to let the data drive the conclusions (§2.4.5). An analyst who writes "the campaign caused the lift" from a before-and-after chart lets the stakeholder's story drive it instead.
Foster et al. in An Introduction to Psychological Statistics put it bluntly: "Correlations DO NOT prove causation." A third variable can drive both measures, and randomized experiments are the most reliable way to balance such hidden influences.
Worked example: the 18% quarter
A retailer's sales rise 18% in the quarter after an ad campaign, and the marketing director asks the analyst to confirm it worked. First, the analyst lists what else changed:
- Seasonality. Was the previous quarter a seasonal low? Compare with the same quarter last year.
- Price and promotion. Did prices drop or did a discount run at the same time?
- Competition. Did a competitor run out of stock or exit a market?
- Distribution. Did new stores or regions open?
Then the analyst proposes a test that creates a counterfactual, an estimate of what sales would have been without the campaign:
- Holdout test. Withhold the ads from a random group of customers and compare.
- Geo test. Run the campaign in some regions and keep matched regions dark.
- Time-series counterfactual. When no experiment is possible, model the expected sales from control series. Google's CausalImpact package does this, and its documentation warns that "valid conclusions require strong assumptions," such as controls that the campaign did not touch.
The report says: sales and the campaign are correlated in time; causation is not established; here is the test that would establish it.
Why each wrong option fails
The original question asks what the analyst can conclude after sales rose 18% following the campaign.
| Option | Why it is tempting | Why it fails |
|---|---|---|
| A. The campaign caused the increase | Stakeholders talk like this | Infers cause from sequence; ignores rival explanations |
| B. Correlated; causation needs testing | Sounds cautious | Correct: reports what the data shows and names the next step |
| C. The campaign had no effect | Sounds skeptical | Also a causal claim, made without evidence |
| D. Proves positive ROI | Sounds business-minded | Needs causal lift and costs; neither is known |
A, C and D all sound like decisions. B sounds like hedging, but hedging that names a test is the professional answer.
What practitioners say
A 2025 Haus article on incrementality testing describes how attribution and media mix models show spend and sales moving together, but cannot tell you what would have happened anyway. Its fix is the holdout: the gap between exposed and unexposed groups is the incremental lift.
A digital analytics practitioner on Substack makes the same point about web metrics: two metrics that move together may share a hidden driver such as seasonality, so test alternative explanations before claiming cause.
CBDA candidates report a similar pattern. LN Mishra of Adaptive US, who holds the CBDA, ranks data interpretation as a high-priority study topic in his 2024 exam guide, above statistical methods. The skill tested is judging what a result means, not computing it.
Common misconceptions
- "A strong correlation is close to proof." Strength does not fix confounding. A strong seasonal driver produces strong correlations.
- "If B happened after A, A caused B." Time order is necessary for causation but not sufficient.
- "No correlation means no effect." An effect can be masked by an opposite trend, such as a price rise during the campaign.
Practice Questions
Question 1. Sales rose 18% in the quarter after a new ad campaign launched. What can the analyst conclude? A) The campaign caused the increase B) Sales and the campaign are correlated; causation needs further testing C) The campaign had no effect D) The increase proves positive ROI
Answer: B. Only B separates what the data shows (co-movement in time) from what needs testing (cause).
Question 2. An analyst finds that customers who use the mobile app spend 40% more than those who do not. A product manager wants to force all customers onto the app. What should the analyst do first? A) Support the plan; the difference is large B) Point out that heavy spenders may choose the app, and propose a randomized test of app prompts C) Run a regression on spend with app usage as the only predictor D) Report that the app has no effect on spend
Answer: B. Self-selection is a classic third-variable problem. A randomized prompt creates comparable groups. C still measures association.
Question 3. No randomized test is possible for a national TV campaign. Which approach is most defensible? A) Compare the campaign month with the previous month B) Build a counterfactual from unaffected control series and state the assumptions C) Ask customers whether they saw the ad D) Treat reach figures as the effect
Answer: B. A modelled counterfactual is weaker than an experiment but far stronger than a before-and-after comparison.
Related topics
- CBDA statistics questions on hypothesis tests and regression
- MVP and experiment design: setting up the test that settles cause
- Root cause analysis and the 5 Whys
- CBDA Interpret and Report Results quiz
- CBDA exam guide: format, domains and cost
To drill this reflex on full scenarios, use the CBDA practice exam.
FAQ
Is correlation vs causation tested on the CBDA exam?
Yes, though rarely by name. IIBA lists diagnostic analytics and interpreting results as core skills, and scenario questions often offer a causal conclusion as a tempting wrong answer. Pick the option that reports the relationship and names a test.
What is the best way to prove a campaign caused a sales increase?
A randomized experiment, such as a customer holdout or a geo test, is the strongest design. It creates a comparison group that did not see the campaign, so the difference between groups estimates the causal lift.
Why is "the campaign had no effect" also wrong?
Because it is a causal claim too. Without a comparison group you cannot show that the campaign added nothing, just as you cannot show it added everything.
Sources
- IIBA: Certification in Business Data Analytics (CBDA) — exam format and domain weights, checked October 2026
- IIBA: CBDA Certification Handbook (May 2026) — delivery and results, checked October 2026
- IIBA: Introduction to Business Data Analytics, A Practitioner View (2019) — analytics types, §1.2 and §2.4.5
- Foster et al.: Correlation versus Causation, LibreTexts
- Google: CausalImpact documentation — Brodersen and Hauser
- Haus: Understanding incrementality testing (2025)
- Correlation and causality in digital analytics, Substack
- Adaptive US: Mastering the CBDA exam, LN Mishra (2024)