Skip to content

Rival Explanation Elimination Table

Structured review — instantiates Structured Comparative Case Design

Lays every candidate explanation for an outcome side by side and rules each out by the evidence it would predict but the cases do not show.

A comparison can look decisive and still be wrong, because a difference between cases may be explained several ways at once. The Rival Explanation Elimination Table is the adjudication layer that keeps resemblance from being mistaken for cause: one row per candidate explanation, columns for the evidence that explanation predicts, the disconfirming evidence sought, and a verdict — survives, eliminated, or undecided. Its distinguishing move is that it is organized by explanation, not by case or by variable: the unit is the rival hypothesis, and a rival earns elimination only when a specific observation it requires turns out to be absent. Where the evidence matrix displays the data, this table argues over what the data rule out.

Example

A cluster of gastrointestinal illness turns up across several restaurants in a city over one weekend. Investigators list the rivals: contaminated lettuce from a shared supplier, a single infected food-handler, cross-contamination at a common commissary, and a local water-supply fault. For each, the table records what you would expect to see if it were the cause: if the lettuce, every case ate a salad and salad-skippers are spared; if a food-handler, cases cluster at one location on that person's shifts; if the water, illness tracks a service zone.

Then each prediction meets the interviews and the traceback. The food-handler rival is eliminated — cases span three restaurants that share no staff. The water rival fails — the affected restaurants sit on different mains. Lettuce survives: every case ate it, and the one restaurant that stayed clean happened to receive a different lot. The output is not a hunch but an audit trail of exclusions, with the surviving explanation and the reason each rival was retired written down.

How it works

  • Enumerate rivals wide enough to include the boring ones — chance, reverse causation, and a common prior cause that produces both the supposed cause and the effect.
  • State a diagnostic prediction per rival — something that rival implies and the others do not, so the observation can discriminate rather than merely fit.
  • Seek the disconfirming evidence first. A rival is retired when evidence it requires is missing, not when the favoured story feels strong.
  • Record a verdict with residual uncertainty, so "undecided" is a visible state rather than a silent pass.

Tuning parameters

  • Rival breadth — how many candidate explanations you admit, including nulls; wider catches the sneaky one but spreads effort thin.
  • Diagnosticity bar — how discriminating a prediction must be to count as a real test; demanding uniquely-predicted evidence is stronger but harder to source.
  • Elimination threshold — how much disconfirming evidence retires a rival — one clean miss versus a preponderance; strict thresholds keep rivals alive longer.
  • Residual-uncertainty tolerance — whether an undecided rival blocks the conclusion or is carried forward as a caveat.

When it helps, and when it misleads

Its strength is that it forces the analyst to argue against the favoured story and leaves the reasoning auditable — every exclusion carries a stated reason, so a later reader can re-open any verdict. Its failure modes are twofold: the table can only eliminate rivals it listed, so an unimagined explanation is invisible to it; and it tempts confirmation, where predictions are quietly built so the favoured story passes rather than so a wrong story would fail. The classic misuse is running it backwards — populating the table after the conclusion to retire inconvenient rivals on thin evidence. The discipline that guards against this is to specify each rival's disconfirming test before looking, in the spirit of strong inference,[1] and to keep undecided rivals visibly open rather than rounding them down.

How it implements the components

  • rival_explanation_register — the table is the register: it enumerates the competing explanations and tracks each one's status from candidate through eliminated or surviving, with the evidence attached.

It does not assemble the cross-case data those verdicts rest on (that's Cross-Case Evidence Matrix Tool), nor produce the temporal and within-case traces a rival's prediction is checked against (that's Comparative Historical Timeline and within-case process tracing); it consumes both and adjudicates.

Notes

The table adjudicates rivals but does not generate them — a gap in the rival list is a gap it cannot see. That is why it pairs naturally with mechanisms that surface anomalies from outside the current list, such as deviant-case follow-up: they widen the register the table then works through.

References

[1] Strong inference (John Platt) — framing several alternative hypotheses at once and designing observations expressly to exclude them, rather than piling up support for one favoured explanation. The elimination table is a bookkeeping form of the same discipline. ↩