Correlation versus causation in a marketing dashboard

A spend-sales correlation describes co-movement in the observed data. It does not by itself estimate how sales would change if the business increased spend. Shared trends, targeting decisions, seasonality and reverse causation can create the same visual pattern.

Separate the descriptive dashboard question from the causal decision before recommending a budget change.

Create a strong correlation without a spend effect

The following synthetic process makes both spend and sales rise with week. Sales are generated from week and noise; spend does not enter the sales-generating equation.

import numpy as np
from scipy import stats

week = np.arange(1, 13, dtype=float)
noise = np.array([0, 10, -10, 5, -5, 0, 10, -10, 5, -5, 0, 5], dtype=float)
spend = 100 + 10 * week
sales = 1000 + 50 * week + noise
correlation = stats.pearsonr(spend, sales)
assert correlation.statistic > .99
counterfactual_spend = spend + 100
counterfactual_sales = 1000 + 50 * week + noise
assert np.array_equal(counterfactual_sales, sales)
assert np.all(counterfactual_spend > spend)
print({'observed_correlation': float(correlation.statistic),
       'sales_change_under_authored_spend_intervention': float((counterfactual_sales-sales).sum())})

The correlation exceeds 0.99, yet the authored structural process gives zero sales change when spend alone increases. This is possible because week drives both series. The example proves that a high correlation is compatible with no causal spend effect; it does not claim that real advertising has no effect.

The SciPy pearsonr reference describes the correlation calculation. The code is reproducible in the analyst statistics lab.

Do not interpret the ordinary p-value as causal evidence

Time-series observations can be dependent, and shared trends can make ordinary correlation inference inappropriate. The example uses the coefficient descriptively and does not use its default p-value to make a causal claim.

NIST's autocorrelation guide explains dependence across time. More weeks of the same confounded pattern do not automatically identify the effect of changing spend.

Ask what determines the marketing allocation

Businesses may spend more during expected high-demand periods, target people already likely to convert or increase budgets after sales improve. In each case, spend is related to other causes of sales.

Adding control variables can help under a justified model, but it is not a mechanical cure. In this fixture, spend is a deterministic function of week, so there is no independent spend variation after accounting for week. The data cannot separately identify a spend effect through that simple regression.

Design evidence for the intervention question

An appropriately randomized experiment can create variation in exposure or budget that supports a causal comparison, subject to implementation, interference and measurement assumptions. When randomization is unavailable, a credible quasi-experimental design needs explicit identification assumptions and diagnostic checks.

The relevant outcome might be incremental conversions, revenue or contribution margin over a specified horizon. Define that estimand before selecting a model. Platform-attributed conversions and incremental conversions are not automatically the same quantity.

Communicate what the dashboard establishes

A defensible description is: spend and sales rose together over the observed period. A stronger statement such as increasing spend by INR 100 will produce a particular sales increase requires additional evidence.

Show timing, segment composition and known operational changes alongside the trend. Use the dashboard to generate questions and monitor an intervention, while keeping causal estimates tied to the design that supports them.

Exercise: add a genuine spend term to the synthetic sales equation and introduce spend variation independent of week. Compare the observed relationship under both processes, then explain why real data do not reveal their generating equation as conveniently as this simulation does.

NeuraPath's Data Analytics with Generative AI course connects analytical dashboards with sound decision reasoning. A useful marketing analysis distinguishes an observed association from evidence about what a budget intervention would change.

Continue learning

This article is part of the Statistics for analytical decisions sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in ANOVA: what a significant result does not tell you.
• Continue with Simpson's paradox in channel conversion rates.