Challenge the counterfactual behind a paired change contrast

Define the decision and evidence boundary

A before–after change can reflect weather, background trends or measurement changes as well as an intervention. A comparison site helps estimate what might have happened without treatment. Pairing can reduce nuisance variation, but causal interpretation still depends on design, interference, measurement comparability and the relevant counterfactual.

Reconstruct the transparent method

Use corrected packet contrasts−4,−2,−2,−2 mg/L. Their mean is−2.5. If an unmeasured treatment-minus-comparison background trend would have been d even without intervention, the adjusted effect is observed contrast minus d. That parameter is a sensitivity assumption, not information learned from the same four contrasts.

Interpret the comparison

Leaving out one pair shows how much the summary depends on a particular independent unit. Omitting pair 1 gives mean−2; omitting any other pair gives−8/3≈−2.667. Report all planned sensitivities rather than choose the exclusion producing the strongest claim. Random assignment supports exchangeability in expectation, not perfect balance in every small sample.

Predeclare limitations and revision

Concentration improvement is not automatically load reduction: streamflow and sampling representativeness matter. Spillover from treated to comparison sites, differential changes and a mismatched time window can defeat the intended counterfactual. The virtual analysis does not authorize an environmental intervention or prove actual restoration success.

Worked original packet

With all pairs and a hypothetical background differential trend−1 mg/L, adjusted intervention effect is−2.5−(−1)=−1.5 mg/L. Ignoring that alternative would attribute too much of the decrease to treatment.

Investigation

Compare the full packet with each single-pair omission, then vary the background trend. Identify the assumption needed to call the contrast an intervention effect and propose an independent observation testing it.

Changed-case transfer

Omit pair 1 and assume treatment sites would have declined 2 mg/L more than comparisons without intervention. What adjusted effect remains?

Feedback

Retained mean is−2 mg/L; subtract background differential−2 to obtain zero. This sensitivity does not prove no effect, but shows that the stated unmeasured trend could explain the entire remaining contrast.

Sources and evidence boundary

Original prose, dataset and numerical exercises. Primary method/mechanism source, checked 30 October 2026. The versioned packet is synthetic, not observed fieldwork. No treatment implementation, site access, pollutant handling or real management authorization is requested. Preserve first reasoning and critique in the course workspace.