An ABA event rate confidence interval calculator can pair a count-per-time estimate with an exact Poisson interval when the event stream and exposure clock reasonably fit that model. This worksheet keeps the event definition, valid exposure, excluded time, conversion unit, and Garwood limits visible. It also compares declared exposure views without pretending they are interchangeable. The output is not a forecast, mastery rule, safety boundary, treatment-effect estimate, or causal conclusion.

Clinicians & ABA Professionals / Data, Outcomes and Clinical Decision-Making.

Rate begins with a clock rule

A frequency of six events is incomplete until the record states the exposure during which those events could be detected. Six events in three observed hours equals two per hour. Six during two hours of a specific activity equals three per activity-hour. Both arithmetic statements can be true, but they describe different denominators and may answer different clinical questions.

This calculator asks: under a homogeneous Poisson model, what exact two-sided interval follows from a nonnegative integer count k and a positive fixed exposure T? Its sensitivity view shows how a declared exposure or count-exposure pair changes the point rate and model-based limits. It does not choose the clinically correct clock.

The NIST Poisson distribution reference defines the model for counts within a fixed interval. Its probability model uses one average event count for that interval. Translating that model into ABA requires a defensible event boundary, opportunity to observe, exposure unit, and rate-stability judgment.

Exact Poisson limits and unit conversion

Let k be the observed event count, T the valid exposure in a chosen base unit, and alpha the complement of the selected two-sided confidence level. The point rate is:

rate = k / T

For k > 0, the Garwood lower limit for the Poisson mean count is:

lower-count = 0.5 x chi-square quantile(alpha / 2, 2k degrees of freedom)

The upper limit is:

upper-count = 0.5 x chi-square quantile(1 - alpha / 2, 2(k + 1) degrees of freedom)

When k = 0, set the lower count limit to 0; the same upper formula applies. Divide both count limits by T to obtain rate limits in the base unit. If T is recorded in minutes and the display is events per hour, multiply the point estimate and both limits by 60. Apply exactly the same conversion to all three values.

The CDC-linked Fay and Feuer methods paper explains the relationship between Poisson limits, gamma quantiles, and chi-square quantiles. Current NCHS presentation standards for rates and counts use gamma intervals when a vital-statistics denominator can be treated as fixed. Those population-rate applications provide a calculation reference, not evidence that a client's event stream follows a homogeneous Poisson process.

Conditions that should leave the result blank

Do not calculate when k is negative or fractional, T is zero or negative, or the event-count window does not match the exposure window. Pause when:

  • the event definition or offset rule changed midstream;
  • observers could not detect the event during part of the recorded exposure;
  • excluded time was removed from the denominator but its events remained in the numerator;
  • state duration was converted into a frequency without a valid event boundary;
  • the exposure mixes contexts with visibly different event processes and no declared reason;
  • observations are bounded by discrete opportunities better represented as binary trials;
  • counts are strongly clustered, serially dependent, or more variable than the Poisson model allows;
  • a comparison uses a different client, behavior, support, phase, observer, or unit without saying so; or
  • the decision concerns immediate safety, medical risk, crisis response, experimental control, or a payer rule outside this arithmetic tool.

The BCBA Test Content Outline includes event recording, rate, measurement validity and reliability, graphing, single-case design, client-informed goals, and data-based decisions. The BACB ethics-code page supplies the current professional code. Neither makes an exact Poisson interval a default clinical decision threshold.

Copyable event-rate record

Freeze the measurement and exposure rules before completing this ABA event rate confidence interval calculator:

FieldPrespecified recordRecord ID, program version, and calculator versionClient-selected or client-informed purposeObservable event onset and offsetCount unit and count-matching ruleExposure clock start, pause, resume, and stop rulesSetting, activity, partner, support, and scheduleObservation window and excluded-time reasonsBase exposure unit and display unitConfidence level and quantile sourceSensitivity view purpose and status as observed or hypotheticalObserver, reviewer, and calculation date

Retain the arithmetic for each declared view:

CalculationPrimary viewSensitivity viewEvent count kValid exposure TBase exposure unitPoint rate k / TConfidence level and alphaLower chi-square degrees of freedomLower chi-square quantileUpper chi-square degrees of freedomUpper chi-square quantileLower count limitUpper count limitLower rate limitUpper rate limitDisplay conversion factorDisplayed point and intervalModel-fit or comparability concerns

If a software package labels the interval “exact,” record its function, version, tail convention, confidence level, and quantile output. The discrete Poisson distribution can produce conservative coverage, so “exact” identifies the inversion method rather than a promise that every real process matches the model perfectly.

Fictional example: whole-session and activity exposure

The following numbers are invented. A team records six occurrences across three valid observed hours. For a two-sided 95 percent interval, alpha = 0.05.

rate = 6 / 3 = 2 events per hour

The required quantiles are:

chi-square quantile(0.025, 12 df) = 4.4037885070

chi-square quantile(0.975, 14 df) = 26.1189480450

Therefore:

lower-count = 0.5 x 4.4037885070 = 2.2018942535

upper-count = 0.5 x 26.1189480450 = 13.0594740225

Dividing by three hours gives 0.7339647512 to 4.3531580075 events per hour. A declared two-decimal display would read 2.00 events/hour; 95% exact Poisson interval 0.73 to 4.35.

The team also asks about the rate during two hours of a specified activity. All six fictional events occurred within that activity, and the count and exposure rules were frozen together. The conditional activity rate is 3.00 per hour, with interval 1.1009471267 to 6.5297370113 per hour. This is not a correction of the whole-session estimate. It is a different rate with a different exposure definition.

Reporting both views prevents a convenient denominator from quietly replacing the planned one. The comparison still says nothing about why the events occurred or whether one denominator is more socially meaningful.

More exposure can narrow the model interval without fixing the design

Consider a separate planning illustration, also fictional. Seven events in three hours and 14 events in six hours both produce 2.3333333333 events per hour. The first exact 95 percent interval is approximately 0.9381 to 4.8076; the second is approximately 1.2757 to 3.9149.

The larger count-exposure record has a narrower interval under the model. It does not prove that doubling session time will reproduce the same count, that the next six hours will be homogeneous, or that longer observation repairs definition drift, missing exposure, observer disagreement, or phase confounding. A hypothetical planning row must remain labeled hypothetical until actual data exist.

Zero observed events also do not establish a zero underlying rate. With k = 0 over three hours, the two-sided 95 percent Garwood interval is 0 to about 1.2296 events per hour. The upper limit follows from the 97.5th percentile for a mean with two degrees of freedom. For safety-critical behavior, this arithmetic must never delay an applicable safety or medical response.

Homogeneity and independence are substantive assumptions

A homogeneous Poisson process assumes a stable event rate over the exposure and an event structure compatible with the model. Many ABA observations have bursts, refractory periods, changing motivation, scheduled access, prompt sequences, unequal activity lengths, or strong session-to-session variation. When the variance exceeds the Poisson mean, an unadjusted interval may be too narrow.

A broad review of single-case experimental designs describes autocorrelation and the inferential problems that arise when sequential observations are treated as independent. A Poisson interval around a pooled event rate does not evaluate level, trend, variability, immediacy, overlap, replication, or treatment integrity. Keep raw event times and session-order graphs available.

This interval estimates a model parameter under repeated sampling. It is not the probability range for the next session, a Bayesian credible interval, a prediction interval for future counts, or a treatment-effect interval. Comparing two overlapping or nonoverlapping Poisson intervals is also not a valid stand-alone significance test.

Retain the event trail, not just the hourly number

Store the event record, valid-exposure log, exclusions, observation notes, unit conversion, quantile source, software version, full-precision bounds, display rule, and reviewer. Create a new version when the event or clock definition changes. A result should be reproducible from retained inputs without reconstructing the denominator from memory.

The Standards for Educational and Psychological Testing frame validity, reliability, fairness, and intended-use evidence. They do not certify this calculator or its use with a particular behavior, client, or decision.

Use coded identifiers and approved systems for protected information. The HHS Privacy Rule summary and Security Rule summary describe federal protections at a high level. Local consent, privacy, security, access, retention, accessibility, payer, regulator, legal, and organizational requirements remain separate review obligations.

Related resources

Sources