Causal interpretation workshop / 6 MIN READ
Separate evidence that a campaign worked from evidence about why
Review an asserted mechanism without treating a correlated intermediate measure as an established causal pathway.
AI-assisted research and analysis for the SNS Creative Fieldbook
Sources checked 2026-11-07
A randomized campaign can establish an effect of assignment while leaving its mechanism uncertain. A post-campaign trust score, click or recall measure is not randomized merely because the campaign was, so explaining the effect through that measure requires additional assumptions.
What you will leave with: A mechanism-review brief with a timed causal pathway, mediator measurement, alternative explanations and sensitivity questions.
Use this when
- A report says an ad increased purchases because it increased trust or recall.
- A team wants to preserve the active ingredient while changing the campaign execution.
This is preparation for qualified causal analysis, not a mediation estimation recipe or a reason to collect intimate customer data. A plausible mechanism can guide research without being proven.
Write the mechanism as a sequence of distinct claims
Name the treatment, the proposed intermediate change and the final outcome. Specify when each could occur and how each is measured. The statement that an explanation improves understanding, which then affects purchase, contains more than the single claim that the explanation changes purchase. Keep those questions distinct before selecting an analysis.
Imai, Keele and Yamamoto explain that causal mediation relies on additional identifying assumptions and that randomizing treatment does not automatically remove mediator-outcome confounding. Their work also emphasizes sensitivity analysis for unmeasured confounding. This is a reason to examine the mechanism claim carefully, not a ban on learning about how interventions work.
Check whether the intermediate measure represents the mechanism
Inspect the measure and its timing. A trust question asked after purchase may partly reflect the purchase experience rather than a cause of it. Recall measured only among people who finish a follow-up survey represents a selected subset. A click can be a convenient observable action without capturing the understanding the theory actually requires.
Review whether asking about the mediator could itself change later behavior. An intrusive survey about product risk might become an additional intervention. Use proportionate measurement and explain any such possibility in the design. Do not collect sensitive information simply because it might improve a model. Preserve the distinction between the underlying concept, its imperfect measurement and the customer population in which it was observed.
Ask what could produce both the mediator and the outcome
Work with the analyst to identify prior factors that could influence the intermediate measure and the outcome, such as product need or existing brand experience. Also examine factors changed by treatment that could affect both. These are not all solved by adding every available field to a regression; some adjustments can change the causal question or introduce bias.
Ask the analyst to state the assumptions needed for the chosen direct and indirect effects, the role of treatment-mediator interaction and the limits of the available design. A path diagram or a software output does not verify those assumptions. If they are implausible, consider whether a different experiment can isolate a component or whether the present evidence should remain a descriptive process account.
Report the mechanism at the strength the design supports
Separate a supported total treatment effect from a tentative explanation. If mediation estimates are reported, show their uncertainty and sensitivity to plausible unmeasured confounding. Be especially careful with a headline percentage mediated when the total effect is small, uncertain or composed of effects in different directions. Have the analyst explain whether that summary is interpretable at all.
Use the remaining uncertainty to plan a useful next test. A revised execution might isolate an informational component while preserving the service experience, or a better-timed measure might test a necessary step in the proposed pathway. Do not claim that deleting an associated intermediate behavior will preserve the campaign effect. The deliverable is a clear account of what changed, what mechanism is supported and what remains a hypothesis.
Make the reasoning concrete.
A fictional appliance seller randomizes an explanatory video and observes more purchases. It surveys trust only after buyers receive their orders and claims that improved trust explains most of the video effect.
Treat the post-purchase score as the cause
Opening: Fit a mediation model and publish its percentage as the explanation.
Hypothesis: The report appears to identify an active ingredient.
Risk: Purchase and delivery may have changed the score, and non-buyers are missing.
Review timing and identification
Opening: Separate the purchase effect from the mechanism hypothesis and redesign the measurement with an analyst.
Hypothesis: The next study can test a more meaningful pathway.
Risk: The current data may not identify the indirect effect.
The seller reports the supported assignment result separately and labels the trust explanation provisional. It does not treat the post-purchase survey as proof that trust caused the purchase.
Take the questions into your next brief.
Use these prompts in your own notes. Nothing is submitted from this page.
Path
What intermediate change is proposed between treatment and outcome?
Timing
Was it measured before the outcome could influence it?
Confounding
What could affect both mediator and outcome?
Assumptions
Which direct or indirect effect can the design identify?
Sensitivity
What plausible violation would change the mechanism conclusion?
Watch for the wrong conclusion.
- Assuming randomized treatment makes every later variable randomized.
- Measuring the proposed cause after the outcome.
- Publishing a percentage mediated without checking its interpretation.
Your next action
Draw the timed pathway behind one mechanism claim and label each arrow as tested, assumed or still unknown.
Sources & scope
Workflows, worksheets and fictional examples are original editorial analysis. Named research guidance supports identified method principles. This prepares a review with a qualified analyst and does not claim an observed business result or a named human reviewer.
- Imai, Keele and Yamamoto: Identification, Inference and Sensitivity Analysis for Causal Mediation Effects →Primary research-method guidance · Supports: Additional identification assumptions, mediator-outcome confounding despite treatment randomization, and sensitivity analysis. The commercial example and review process are original analysis.