Reverse conditional probability ABC data errors occur when P(response | antecedent) is treated as P(antecedent | response). The first divides response-following exposures by all antecedent exposures. The second divides responses preceded by the antecedent by all responses. They share a joint cell but use different denominators. Label direction, raw cells, timing, and coverage for both.

Translate the conditioning bar into words

In P(A | B), the event after the bar, B, defines the denominator universe. P(response | demand) asks what proportion of valid demand exposures were followed by the response. P(demand | response) asks what proportion of valid responses were preceded by demand. The same linked sequence may contribute to both numerators, but each question begins from a different list.

Write the conditioning event, outcome, anchor, timing window, and eligibility rule before inserting a count. This habit prevents a percentage from losing its direction when it moves from a worksheet to a graph or report. Include numerator and denominator labels in the data field, rather than storing only a decimal.

Build two worklists from the raw events

For response given demand, start with every valid demand exposure. Mark whether the response occurs within the defined forward window. For demand given response, start with every valid response. Look backward and mark whether demand meets the corresponding rule. Keep identifiers that allow matched sequences to be reconciled without exposing more personal information than the review requires.

The joint count is reusable only if both calculations use the same response, demand, and timing definitions. A demand recorded as ongoing for the backward analysis may not satisfy an onset-only forward rule. When rules differ, label the two joint cells separately. Do not force them to match for a tidy report.

Calculate Benji's two directions

Benji has 20 valid demand exposures. The target response occurs within the forward window after 8 of them. Twelve demands are not followed by the response.

P(response | demand) = 8 / 20 = 0.40, or 40.0%.

Benji also has 14 valid response events with complete look-back windows. Eight are preceded by demand and six are not.

P(demand | response) = 8 / 14 = 0.5714, or 57.1%.

The shared sequence count is 8. The probabilities differ because 20 demand exposures and 14 response events are different denominators. The 17.1 percentage-point numerical gap does not represent a contingency effect; it compares reciprocal conditional questions.

Reconcile exposure and coverage separately

Each worklist needs its own planned and valid coverage. Demand exposures may be missed during transitions, while response anchors may be truncated at observation start. Report observed, valid, invalid, ambiguous, and truncated counts for each direction. Do not let a response with incomplete history enter P(demand | response), and do not treat a demand with a cut-off forward window as response absent.

Review where observations occurred. Twenty demand exposures from one staff routine and 14 responses from several routines can yield correct arithmetic yet answer a narrow question. Representative coverage depends on the predeclared sampling frame and the pattern of missingness across contexts, people, times, access conditions, and relevant health states.

Handle zero denominators and sparse sequences

If no valid demand exposures occur, P(response | demand) is undefined. If no valid responses occur, P(demand | response) is undefined. Neither should be shown as zero. When the denominator is positive and the linked numerator is zero, the conditional probability is zero under the chosen rule. Uncertain or unmatched events remain unresolved and should be shown in coverage.

A high reciprocal conditional can arise from a small or common-event denominator. For example, 2 of 2 responses preceded by demand gives P(demand | response) = 1.0, even if only 2 of 30 demands were followed by response. Show both counts and the underlying event prevalence before describing a pattern as strong.

Keep compound and repeated events visible

A response may follow several antecedents, and one demand episode may include repeated prompts. Predeclare whether the unit is an onset, episode, interval, or opportunity. Preserve component codes when noise, pain indicators, denied access, or communication breakdown co-occur with demand. Post hoc selection of one convenient antecedent can distort both directions.

Overlapping forward windows can link one response to multiple demands. A backward rule may choose the most recent demand, any demand, or every demand within the window. These are different analyses. Retain timestamps and document the linkage rule so the numerator can be reproduced.

Interpret without converting association into cause

Benji's 40.0% result describes response occurrence across sampled demands. The 57.1% result describes demand presence across sampled responses. Neither result shows that demand caused the response or identifies a behavioral function. Demand prevalence, shared contexts, scheduling, response duration, observer behavior, access barriers, or another event can shape the sequence.

Qualified clinical interpretation draws on Benji's account, health and safety context, other assessment components, and implementation evidence. Descriptive probabilities can suggest what to observe next. They do not authorize provoking unsafe events, restricting rights, withholding communication or care, or choosing treatment from a percentage alone.

Report direction in every display

Use two panels or clearly labeled rows: "response after demand, 8/20" and "demand before response, 8/14." Include the timing window and coverage beneath each. Avoid labels such as "demand-response probability" because they hide which event conditions the result. When data are missing, use a gap or a separate state rather than plotting an observed zero.

Appropriate wording is: "The response followed 8 of 20 valid demand exposures (40.0%). Demand preceded 8 of 14 valid responses (57.1%). The calculations share eight matched sequences but use distinct denominator universes and answer different descriptive questions. Neither conditional establishes function." State exclusion reasons and context balance.

Review the result accessibly with Benji

Present the two worklists and graphs in a form Benji can inspect using the person's established communication method. Keep AAC, interpreters, mobility, food, water, bathroom use, health care, rest, relationships, and emergency support available. The ASHA AAC practice resource supports continuous system access. Record Benji's corrections, assent, dissent, discomfort, and priorities with a clear source label.

Privacy applies to the raw sequence record as well as the report. Separate direct observation from partner report, imported timestamps, and clinician inference. Limit access, preserve the authorized audit trail, and act on urgent medical or safety needs without waiting for an observation window to finish.

Place the sources around the right claims

The BACB ethics materials, CASP practice-guideline summary, and BCBA Test Content Outline give professional context for assessment and measurement. Predictor research and contingency-space methods demonstrate why background and conditional cells matter. Lag-sequential analysis supplies timing context. The difference between reciprocal denominators is especially visible in precursor sequence research, while comparative descriptive research supports a cautious noncausal interpretation.

Clinician review checklist

When checking reverse conditional probability ABC data, confirm both questions in words, conditioning events, response and demand definitions, forward and backward anchors, timing rules, joint-cell compatibility, demand-exposure worklist, response-anchor worklist, valid and truncated counts, 8/20 and 8/14 arithmetic, rounding, graph labels, source labels, access review, safety actions, qualified interpretation owner, next step, and review date.

Practical limits

Reciprocal conditionals depend on event prevalence, selected windows, observation coverage, and coding quality. They cannot establish causality, functional relation, clinical importance, treatment integrity, or future behavior. A shared numerator does not make denominator universes comparable. Clock error, overlapping windows, repeated events, and missing contexts can change both. Retain raw worklists and create a separately labeled series when definitions or linkage rules change.

Related resources

Sources