DOCS / TRACE

Trace

Causal evaluation of interventions — designs, diagnostics, and approvals.

Trace is Marres' causal evaluation function. It helps teams investigate whether an observed change — in sales, complaints, retention, cost or any measured outcome — can reasonably be attributed to a specific intervention, rather than to timing, seasonality or coincidence.

Most analytics describe what happened. Trace asks a harder question: what changed because you acted? It compares observed outcomes with a carefully defined estimate of what might have happened without the intervention, and returns a structured, reviewable account of the evidence, the assumptions behind it and the limits of what can be claimed.

Why timing alone is not evidence

Sales rose after a campaign. Complaints fell after a service change. One outlet improved while another did not. None of these observations establish cause on their own. Other promotions, seasonal patterns, market shocks or differences between locations may offer competing explanations.

Trace makes those competing explanations part of the investigation instead of hiding them behind a headline number.

The evaluation workflow

Every Trace evaluation moves through six explicit stages:

1. Declare the intervention. Record what changed, when it changed, where it was applied and which outcome it was expected to influence.
2. Define the comparison. Choose the treated population, the comparison group or the pre-intervention period. Trace checks whether the available data can support the proposed design.
3. State the assumptions. Document the conditions required for the comparison to be meaningful, including possible spillovers, anticipatory behaviour and concurrent events.
4. Estimate the effect. Where the design is eligible, Trace applies an appropriate causal comparison. The initial engine supports interrupted time series and difference-in-differences evaluations.
5. Challenge the result. Diagnostics test whether the evidence behaves as the design expects. Warnings, uncertainty and failed checks remain attached to the estimate.
6. Review before use. No result becomes a causal claim automatically. Analyst approval is required before estimation, and final sign-off is required before the finding informs a client decision or model update.

What an evaluation returns

• An estimate of the intervention's effect
• An uncertainty interval around that estimate
• A record of the population, outcome and observation window examined
• Diagnostics relevant to the selected design
• Warnings when the data or comparison are fragile
• The assumptions on which the interpretation depends
• A clear boundary around what the result does not establish
• An auditable approval record

Built to say "we do not know"

Trace does not discover causality from any dataset placed in front of it. It does not turn correlation into proof, and it cannot rule out events that were never recorded.

If the intervention cannot be separated credibly from other explanations, the investigation should stop. "Not identifiable from the available evidence" is a valid result — and often a more valuable one than false certainty.

Closing the learning loop

Marres helps teams learn, allocate, simulate and respond. Trace completes the cycle by testing what happened after a decision became an intervention.

The evidence can then support a governed update to future assumptions and responses. Models may learn from the result, but they may not quietly rewrite themselves. The estimate, its limitations and the person who approved its use remain visible.

Interpreting Trace results responsibly

• An estimate is conditional on its design and assumptions; report both alongside the number.
• Diagnostics and warnings are part of the result, not an appendix — a passed estimate with failed diagnostics should not travel.
• Uncertainty intervals describe the precision of the comparison, not a guarantee about the future.
• A causal finding for one population, period or location does not automatically transfer to another.

Related product page: Trace — product overview

Documentation draft · Source baseline 16 August 2026