Correlation describes how observations move together. A causal explanation also needs a credible account of what would have happened without the proposed cause. Check timing, shared influences, selection and measurement before using an association to justify an intervention or claim its effect.
State the observation without smuggling in an explanation
Write the relationship in terms of measured quantities, population and period. Orders and support contacts might both rise after a launch. That statement does not establish that orders caused confusion, that support created demand or that the launch caused either change. Check whether contacts count unique customers, conversations or messages, because a logging change can create an apparent relationship without changing customer experience.
Plot the observations at a useful interval and inspect the underlying records. Weekly totals may conceal a one-day outage or a promotion affecting only one market. Determine which event happened first, allowing for reporting delays. Timing can rule out some stories, but an earlier event is not automatically the cause of a later event. Keep the descriptive chart separate from the explanation proposed beside it.
Ask what else could produce the same pattern
Build a short list of alternatives that imply different observable evidence. More customers could produce more orders and more contacts even if contact rates stay constant. A difficult release could raise contacts among existing customers while a campaign independently raises sales. A change in the help widget could make previously unrecorded requests visible. Each explanation suggests a different denominator, comparison or source check.
Draw a simple chain of events before adjusting statistically for other variables. A variable can be a shared influence, an intermediate step or a consequence of the outcome. Treating all available fields as controls can therefore obscure the question. Record which factors were observed before the exposure, which were measured later and which important influences remain unavailable. Seek specialist design advice when the decision warrants causal estimation.
Choose an investigation that could change the decision
A useful next step distinguishes explanations rather than producing another correlation. Compare contact rates for customers using the changed feature with rates for other customers, while acknowledging differences between those groups. Review a sample of contact reasons and check release timing. If an experiment is feasible and appropriate, specify assignment, outcome and observation window before seeing results. An observational comparison needs an explicit account of its remaining limitations.
State the action proportionately. A temporary clarification in the help centre may be justified by repeated confusing instructions without claiming a quantified causal effect on sales. A costly product reversal needs stronger evidence. Name the person collecting the missing evidence and the decision deadline. Preserve the original hypothesis, alternatives and findings so the final report does not retroactively present the winning story as inevitable. A useful review also records which expected observation would weaken the preferred explanation and where that observation would be found.
A fictional launch with more orders and contacts
A fictional scenario to explain the method, not a customer result or a market benchmark.
A fictional equipment retailer records 200 orders and 40 support conversations in the week before a launch, then 400 orders and 80 conversations afterward. Both totals double. Dividing conversations by orders gives 0.20 in each week, or 20 conversations per 100 orders. That ratio is a workload description; conversations may come from earlier purchasers, so it is not automatically the proportion of new buyers needing help.
The analyst links conversations to order cohorts where identifiers permit. Of the 80 later conversations, 30 concern older orders, 20 concern the new product and 30 have no reliable order link. The team cannot allocate all contacts to the launch cohort. It publishes the two totals and stable crude ratio, then assigns review of the 20 product conversations and the 30 unlinked records. A clearer setup guide is a plausible response to documented confusion; these invented counts do not prove that the launch caused support demand.
TAKE THIS INTO YOUR NEXT CONVERSATION
A practical next-step checklist.
- Define the unit and time window for both variables.
- List at least two explanations with different predictions.
- Check timing and changes in recording.
- Match the strength of the action to the evidence.
Common questions.
Does a stable contact-to-order ratio rule out a product problem?
No. Problems in one cohort can be hidden by other contacts, changing customer mix or incomplete linking.
Can a chart establish which variable caused the other?
A chart can reveal timing and patterns, but causal attribution also needs a defensible design and assumptions.
Further reading & source context.
These primary references provide background. They are not evidence for the fictional example above and do not endorse Accuralytix.
Found an error or a missing qualification? Send an editorial correction.
Check my data readiness.
Use a practical tool to turn this idea into a clearer starting point.
Check my data readiness