Skip to content

Differential Diagnosis

Diagnostic method — instantiates Contrastive Differentiation

Enumerates the plausible explanations for a case in hand and eliminates among them by the features that would distinguish one from another.

Differential Diagnosis takes a single case that could have several explanations, lists the candidates, and then hunts for the discriminating features — the findings that would be present under one explanation and absent under its rivals — to narrow the field to the most likely cause. Its defining move is that the comparison set is a set of competing explanations for one present situation, not a menu of options to adopt, and the whole method is elimination by distinguishing evidence rather than accumulation of confirming evidence. It is inference to the best explanation, disciplined: you don't stop at the first plausible cause, you ask what else it could be and what observation would tell them apart.

Example

An on-call engineer sees request latency on a service jump from ≈40 ms to ≈900 ms. Rather than seizing the first suspect, she runs a differential. The candidates: a garbage-collection pause, network saturation on the node, lock contention in the database, or a slow downstream dependency — and, deliberately included, the "can't-miss" one that's cheap to check and expensive to overlook: a bad deploy that shipped minutes ago. Each candidate has a discriminating feature. GC pauses would show sawtooth heap graphs and correlate with collection cycles; network saturation would show in interface counters and hit all traffic on the node, not just this service; DB lock contention would surface as long-held locks and would spare requests that don't touch the database; a slow dependency would show the latency living in the outbound call span. She pulls the one observation that splits them fastest — a trace showing where the time is actually spent — and it lands almost entirely in the database call span with no dependency involvement, pointing at lock contention. The response (kill the long-running migration holding the lock) follows from the distinction, and she keeps the differential open in case a second cause is riding along.

How it works

Its distinguishing steps are enumeration then elimination:

  • Cast the candidate set wide. List the plausible explanations, deliberately including the low-probability, high-cost ones that must not be missed — breadth first, before fixation sets in.
  • Name each candidate's discriminator. For every explanation, state the feature that would be present under it and absent under the others; a candidate with no discriminator can't be tested.
  • Gather the finding that splits fastest. Choose the observation that eliminates the most candidates for the least effort, rather than the one that merely confirms a favourite.
  • Converge, but keep the differential open. Land on the working explanation that best fits, while holding that more than one cause may be present.

Tuning parameters

  • Candidate breadth — how wide the initial list is. Wider lists resist tunnel vision but cost time to work through; narrow lists are fast but risk missing the true cause entirely.
  • Test ordering — whether to run the cheapest discriminating test first or to rule out the most dangerous candidate first. Efficiency versus safety, and they often point at different first moves.
  • Stopping rule — how confident you must be before acting. Stopping early treats faster but risks acting on the wrong cause; demanding certainty delays a response that may be time-critical.
  • Multiplicity stance — whether you accept the single best explanation or actively look for a second concurrent cause once one is found.

When it helps, and when it misleads

Its strength is that it structurally resists the single most common reasoning error under pressure — grabbing the first explanation that fits — by forcing the question "what else could this be, and what would tell them apart?"

Its failure modes are the classic diagnostic traps. Premature closure (anchoring on the first plausible cause and stopping) and its mirror, confirmation bias, run the method backwards: start from the favoured diagnosis and collect only features that fit it.[1] The method also tends to hunt for one cause and miss the case where two are present at once, and it is only as good as the candidate list — a cause never listed is never found. The guard is to generate the differential before testing, to seek disconfirming evidence rather than confirming, and to keep the differential open until the discriminating findings — not a hunch — have actually closed it.

How it implements the components

  • comparison_set — the differential itself: the set of competing explanations for the present case, must-not-miss candidates included.
  • distinguishing_feature — for each candidate, the discriminating finding that would separate it from its rivals; the engine of elimination.
  • classification_or_choice_link — the method converges on a working explanation that directly drives the response.
  • exception_note — it flags that more than one cause can be present at once, where a clean either/or diagnosis would mislead.

It does not teach a reusable category boundary (that is Concept Disambiguation Examples), lay a neutral reference grid (that is Contrast Table), or standardize a recurring call across many raters (that is Decision Rubric with Distinguishing Criteria).

Notes

Differential Diagnosis differs from A/B Comparison in what the alternatives are: A/B compares two options you might choose to adopt, whereas the differential compares rival explanations for something that is already happening and that you cannot pick — you can only infer which one is true. The consequence is that the differential's job is elimination by evidence, not preference among options.

References

[1] Premature closure — accepting a diagnosis before it is fully verified — and anchoring are among the best-documented sources of diagnostic error; both amount to running a differential backwards from a favoured answer. Naming candidates before testing, and seeking disconfirming findings, are the standard correctives. ↩