Frequency, Rate, Duration, Latency and IRT: Choosing the Continuous Measure That Answers Your Question

Continuous measurement means capturing every occurrence of a behavior rather than sampling. But which property of the behavior you capture is a separate decision, and it is the one that determines whether your data can answer the question you actually have.

Five measures cover nearly everything in practice.

Frequency (count)

What it is: the number of times the behavior occurred.

The question it answers: how often?

Frequency is the default and it is usually right for discrete behaviors with a clear beginning and end — requests, aggression episodes, hand raises, correct responses.

The trap: a raw count is only interpretable when the observation period is fixed. Twelve instances means nothing without knowing whether the session was twenty minutes or four hours. If your session lengths vary at all — and in ABA they always do, between cancellations, short days and partial sessions — a count graph will show variability that is entirely an artifact of session length.

Rate

What it is: count divided by observation time. Responses per minute, or per hour.

The question it answers: how often, adjusted for opportunity?

Rate is what frequency should usually have been. It survives variable session lengths, which makes it the more defensible default for almost every count-based target in a clinical setting.

The trap: rate assumes the behavior could have occurred throughout the observation. For a behavior that can only occur when an opportunity is presented — say, responding to an instruction — rate across a whole session is misleading. Use percentage of opportunities instead.

Duration

What it is: how long the behavior lasted. Either total duration across the session or duration per occurrence.

The question it answers: how long?

Duration is correct whenever the clinical concern is time rather than instances. Tantrums, on-task engagement, time out of area, self-injury episodes, sustained attention.

The trap: total duration and duration per occurrence tell different stories and are routinely confused. A learner whose tantrum total drops from sixty to thirty minutes might be having the same number of episodes at half the length — real progress — or half as many episodes at the same length, which is also progress but a different kind. Graph both when the distinction matters, and it usually does.

Latency

What it is: the time between the antecedent — usually an instruction — and the start of the behavior.

The question it answers: how long before they start?

Latency is badly underused. It is the correct measure whenever the problem is not whether a learner responds but how long they take to. A learner who eventually complies with every instruction but takes four minutes each time has a serious clinical problem that frequency data records as 100% compliance.

It is also the right measure for the reverse case — reducing the latency to request help, or to initiate a transition.

The trap: latency requires a clearly defined antecedent and a clearly defined behavior onset. If the operational definition does not say precisely when the clock starts and stops, latency data will have terrible agreement.

Interresponse time (IRT)

What it is: the time between the end of one occurrence and the start of the next.

The question it answers: how spaced out is the responding?

IRT is the specialist measure of the group. Its main uses are in schedule thinning and in differential reinforcement — DRL programs are defined by IRT, since the whole intervention is about reinforcing responses that occur after a minimum gap.

It is also diagnostically useful for behavior that comes in bursts: a stable rate can hide the difference between evenly spaced responding and long quiet periods punctuated by clusters.

Choosing, as a decision tree

1. Is the clinical concern about how much time the behavior takes up? Use duration. Graph total and per-occurrence separately.
2. Is it about how long before the behavior starts? Use latency.
3. Is it about the spacing between responses, or are you running a DRL? Use IRT.
4. Is the behavior discrete, countable, and able to occur any time? Use rate, not raw frequency, unless your observation period is genuinely fixed.
5. Does the behavior only occur when an opportunity is presented? Use percentage of opportunities.

Most programs need two of these, not one. A tantrum program that graphs only frequency is missing the thing families care most about, which is how long they last.

Dimensional quantities you are not measuring, and probably should not

Two more properties get discussed and rarely measured well in applied settings:

• Magnitude or intensity — force, volume, severity. Genuinely important for behaviors like self-injury, and genuinely hard to measure reliably without instrumentation. Where it is attempted, it is usually via a rating scale, which means agreement should be checked carefully and claims kept modest.
• Topography — the form of the response. Descriptive rather than quantitative, useful for defining response classes, not something you graph.

Being honest about what you cannot measure well is part of a defensible measurement plan.

Making multiple measures survivable

The reason most programs measure one thing when they need two is workload. Asking an RBT to run a frequency count and two duration timers concurrently while delivering instruction is asking for fabricated data.

This is the practical constraint that should shape the measurement plan: the number of concurrent measures a person can run at fidelity is small — realistically one duration timer plus one count, or a couple of counts. Beyond that, either the measurement degrades or the teaching does.

Where the data collection system earns its place is in collapsing that workload — a duration timer that starts and stops with one tap on the same screen as the trial data, counts that increment without leaving the program, and the resulting numbers flowing into the graph and the session note without re-entry. That does not make three concurrent measures free, but it moves the ceiling from one to two or three, which is usually the difference between measuring what matters and measuring what was easy.

The summary

Frequency answers how often, rate answers how often per unit time and should usually replace raw frequency, duration answers how long, latency answers how long until it starts, and IRT answers how spaced out responding is. Choose from the clinical question rather than from habit, expect to need two measures on most meaningful targets, and keep the concurrent load low enough that the data is real.