Before/After Analysis¶

Temporal comparison method — instantiates Contrastive Differentiation

Distinguishes a changed state from its prior state by holding the earlier condition as a baseline and reading the difference the intervening change actually made.

Before/After Analysis makes a difference legible along time: it fixes the prior condition as an explicit baseline and reads the changed condition against it, so the effect of some intervening event or intervention becomes visible. Its defining feature — and its defining hazard — is that the two things being contrasted are the same subject at two moments, not two separate items. That makes the baseline the load-bearing part: the whole analysis is only as honest as the claim that the "before" is a fair stand-in for what would have happened without the change. Where the other mechanisms in the family lay confusable items side by side in space, this one lays a single thing against its own past, and lives or dies on whether the difference can be attributed to the change rather than to everything else that also moved.

Example¶

A city rebuilds a dangerous four-way stop as a roundabout and wants to know whether it worked. Before/After Analysis fixes the baseline first: collisions at that intersection over the three years before construction — say ≈14 per year, with a note on the mix of severities. After a full year of operation the count is ≈5. The raw difference is stark, but the method's discipline is to keep the surrounding context attached to both readings, because a naïve before/after is exactly how false credit gets assigned. Traffic volume fell 8% city-wide that year; a nearby arterial closed and diverted through-traffic; winter was milder. Each of these is recorded next to the numbers so the reader can see what the "before" does and does not control for. The honest output is not "the roundabout cut crashes by ~64%" but "crashes fell from ≈14 to ≈5; roughly this much is plausibly the roundabout after accounting for the volume drop, and the rest is uncertain" — a distinction the bare two numbers would have overstated.

How it works¶

Its distinguishing steps are about anchoring time and defending attribution:

Fix the baseline explicitly. Capture the prior state as a stable, dated reference — ideally averaged over enough history to represent "normal," not a single lucky or unlucky moment.
Read the same dimension after. Measure the changed state on the identical dimension and units, so the two moments are actually comparable.
Attach the confounds. Record what else changed between the two moments — this is the step that separates a credible before/after from a misleading one.
Attribute cautiously. State how much of the difference is plausibly the intervention versus the co-movers, and carry that uncertainty forward.

Tuning parameters¶

Baseline window — how much prior history defines "before." A longer window smooths out flukes but can absorb a slow pre-existing trend; a short window is responsive but noisy.
Confound accounting — how hard you look for other things that moved. More rigor (a comparison site, a seasonal adjustment) buys credible attribution at the cost of effort.
Measurement horizon — how long after the change you read the "after." Too soon catches transient novelty effects; too late lets unrelated drift creep in.
Attribution stance — whether the difference is claimed as caused by the change or merely associated with it. The honest default is the weaker claim unless confounds are genuinely controlled.

When it helps, and when it misleads¶

Its strength is that it answers "did this change do anything?" in the subject's own terms, using history you already have — cheap, intuitive, and often the only comparison available when a true control group is impossible.

Its central failure is the post hoc trap: because the change and the outcome are separated only by time, it is dangerously easy to credit the intervention for a difference that trend, season, or a coincident event actually produced.^[1] A cherry-picked baseline (an unusually bad "before") inflates any improvement, and the method is a favourite for manufacturing a success story after the fact. The guarding discipline is to fix the baseline before the outcome is known, keep the confounds attached to the numbers, and downgrade the claim from "caused" to "associated" whenever the before cannot stand in for a real counterfactual.

How it implements the components¶

baseline_reference — its signature: the prior state is captured as the explicit, dated benchmark the change is read against.
contrast_dimension — the difference is made along time, on one consistent measured dimension held identical before and after.
context_preservation — the co-occurring changes and uncertainty are kept attached to both readings, guarding against misattribution.

It does not build a multi-item comparison_set in space or turn the result into a classification_or_choice_link rule — those belong to A/B Comparison and Decision Rubric with Distinguishing Criteria; it contrasts one subject across two moments, not several alternatives at once.

Instantiates: Contrastive Differentiation — the temporal, baseline-anchored case of making a difference explicit.
Sibling mechanisms: A/B Comparison · Contrast Table · Confusion Audit · Differential Diagnosis · Product or Option Comparison Matrix · Concept Disambiguation Examples · Near-Miss Case Pairing · Annotation and Callout Layer · Signal Highlighting · Visual Contrast Encoding · Decision Rubric with Distinguishing Criteria

Notes¶

The baseline is the whole game. A before/after with a fairly chosen, well-dated baseline and honestly attached confounds is a legitimate causal argument; the same arithmetic with a cherry-picked "before" and no confound accounting is the most common way people are misled by a true number.

References¶

[1] Post hoc ergo propter hoc — "after this, therefore because of this" — the fallacy of inferring causation from mere temporal succession. It is the standing hazard of any before/after comparison, which is why confounds are recorded beside the numbers rather than dropped. ↩