Causal Identification Probe¶
Identification strategy — instantiates Observational Equivalence Resolution
Separates rival causal stories for the same outcome by pairing the predictions each makes over naturally occurring variation, then reading which pattern the world actually shows.
The Causal Identification Probe resolves an observational tie between rival causal stories without touching the system. Its defining move is to find variation that already exists in the world — a policy that switched on in one region and not another, a threshold that sorts otherwise-similar cases, a shock that hit some units and missed others — and to ask what pattern each candidate cause would leave across that variation if it, and not its rival, were operating. The two stories are written down as a matched pair of predictions; the data are then read to see which prediction holds and which is contradicted. Where an interventionist mechanism would create the contrast by acting, this probe harvests a contrast nature has already produced, which is both its strength (nothing is disturbed) and its constraint (it can only separate stories that the available variation happens to distinguish).
Example¶
A city runs a subsidized job-training program and, a year later, the trainees' employment rate has risen. The shared observation — "employment went up among participants" — is compatible with several generators: the program worked; or the people who enrolled were already on their way up (selection); or the whole regional labor market improved (a trend that would have lifted them anyway); or the earlier low numbers were an unlucky dip now correcting (regression to the mean).
The probe declines to attribute the rise and instead looks for separating variation. Neighboring towns with similar labor markets but no program become a comparison group; the question becomes not "did participants improve?" but "did participants improve relative to otherwise-similar non-participants over the same window?" — a difference-in-differences contrast.[1] Each candidate now makes a distinct prediction: if the program caused the gain, participants should diverge from the comparison towns only after enrollment; if a regional trend drove it, both groups should rise together; if selection drove it, participants should already have been climbing before the program began. The pre-period is the discriminating observable — parallel trajectories before the program, divergence only after, is a pattern the trend and selection stories cannot easily reproduce. If the contrast clears the pre-registered robustness bar (placebo periods, alternative comparison groups), the causal story is credited; if the pre-trends were already diverging, the probe reports the effect as unidentified rather than manufacture a number.
How it works¶
The probe begins from already-enumerated rival causes and does not itself widen the field. It locates exploitable variation — a natural experiment, a discontinuity, a staggered rollout, an instrument — where the candidates would behave differently. It then writes the counterfactual pair: for each rival, the specific pattern the data should show if that cause were the true generator, stated before looking. It reads the discriminating contrast — the timing gap, the treated-versus-control divergence, the jump at a threshold — where the pairs come apart. Finally it checks the contrast against a threshold of robustness: placebo tests, alternative specifications, and falsification checks that a spurious pattern would fail. Only a contrast that survives is allowed to break the tie; a fragile one leaves the case unresolved.
Tuning parameters¶
- Source of variation — natural experiment, regression discontinuity, staggered adoption, or instrument. Each buys identification under a different assumption; the credibility of the whole probe rests on which assumption is defensible here.
- Comparison construction — how the counterfactual group or period is built. A tighter match strengthens the contrast but shrinks the usable sample.
- Pre-registration stance — whether the discriminating prediction is fixed before the data are examined. Pre-committing blocks the temptation to fit the story to the pattern after the fact.
- Robustness bar — how many placebo and falsification checks a contrast must survive before it is credited — the dial between crediting noise and dismissing a real effect.
- Assumption transparency — how explicitly the identifying assumption (parallel trends, exogeneity) is stated and stress-tested, since the conclusion is only as good as that assumption.
When it helps, and when it misleads¶
Its strength is that it can separate causal stories that look identical in a raw before-and-after, using variation that is already lying in the data and disturbing nothing — invaluable exactly where an experiment is impossible, unethical, or too slow. It makes the identifying assumption the object of debate, which is where the real disagreement usually is.
Its failure modes are all failures of the assumption behind the variation. A pseudo-discriminator is the sharpest: a contrast is computed, but it would not actually separate the live stories because the comparison group was not truly comparable — the pre-trends were already diverging, and the "effect" is confounding wearing an identification costume. The probe is also run backwards when the natural experiment is chosen after seeing which one yields the desired result, and it is easily overstated when a single specification is reported as if it were the only one. The guarding discipline is to fix the discriminating prediction before looking, to test the identifying assumption directly (placebo outcomes, pre-period parallelism), and to report the effect as unidentified when the assumption fails rather than dress a correlation as a cause.
How it implements the components¶
The Causal Identification Probe realizes the counterfactual-contrast face of the archetype — separating rival causes by what they predict over variation that already exists:
counterfactual_prediction_pair— for each rival cause, the explicit pattern the data should show under the available variation if that cause (and not the other) were operating.discriminating_observable— the specific contrast — a timing divergence, a treated-versus-control gap, a jump at a threshold — where those paired predictions come apart.evidence_threshold— the robustness bar (placebo checks, alternative specifications) the contrast must clear before one story is credited.
It does not create the variation by intervening on the system (discriminating_test_design, harm_constraint) — that is its nearest twin, Ablation or Perturbation Test, which manufactures the contrast rather than harvesting it; nor does it enumerate and thin an open field of candidates (candidate_explanation_set, observational_equivalence_class), which is Differential Diagnosis Protocol.
Related¶
- Instantiates: Observational Equivalence Resolution — this probe is the archetype's non-interventional resolver for rival causal stories.
- Sibling mechanisms: Differential Diagnosis Protocol · Ablation or Perturbation Test · Forensic Discriminator · Frame-of-Reference Shift · Side-Channel Measurement · Decision Tree with Hold State · Controlled Disambiguation Test · Ambiguity Register
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: The mechanism specifies rival predictions over naturally occurring variation and reads the discriminating contrast to infer which causal story fits, so its operative form is causal identification analysis.
Nearest alternative: Experiment, Test & Rehearsal — It exploits natural experiments and staggered variation but does not deliberately perturb the target to generate the variation.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Economics & Finance
Origin pattern: Convergent development
Present-day reach: Multi-domain
Rationale: Econometrics established natural experiments, difference-in-differences, and discontinuities for separating rival causal stories.
Related originating lineages:
- Political Science — Comparative policy research contributes naturally occurring institutional variation and rival-story discrimination.
- Statistics & Experimental Design — Statistical causal inference contributes identification assumptions, placebo tests, robustness checks, and uncertainty.
Review resolution: Economics and finance is the agreed primary lineage because econometrics developed natural experiments, instruments, discontinuities, and staggered rollouts as identification strategies. Statistics and political science independently shaped their validation and policy use, making the lineage convergent and multi-domain rather than specialized to economics.
Review outcome: Reconciled after independent review; high confidence.
References¶
[1] Angrist, J. D., & Pischke, J.-S. Mostly Harmless Econometrics: An Empiricist's Companion. Princeton University Press (2009). Identifies the comparison as a difference-in-differences contrast. registry ↩