Partial Interval vs Whole Interval Recording: Why Your Choice Changes the Answer

Partial and whole interval recording are the two most commonly confused measurement systems in applied behavior analysis, and the confusion is expensive. They use the same data sheet, the same interval timer and the same scoring marks. The only difference is one word in the rule — and that word inverts the direction of the error in your data.

The rule for each

Partial interval recording: divide the observation into equal intervals. Score the interval if the behavior occurred at any point during it, even once, even briefly.

Whole interval recording: same intervals. Score the interval only if the behavior occurred for the entire duration of it, with no break.

That is the whole distinction. Everything that follows comes from it.

Why partial interval overestimates

Consider a learner who engages in two seconds of hand flapping at the start of each thirty-second interval, and nothing else. Actual behavior: roughly 7% of the observation. Partial interval score: 100% of intervals.

You have recorded a behavior occupying 7% of the session as though it occupied all of it. That is a fourteenfold overestimate, and it is not a mistake in execution — it is the method working exactly as designed.

The bias grows with interval length. Sixty-second partial intervals overestimate far more aggressively than ten-second ones, because a longer interval gives a brief behavior more opportunity to score the whole block.

Why whole interval underestimates

Now flip it. A learner is on task for twenty-eight of every thirty seconds, pausing twice for a second each. Actual on-task behavior: about 93%. Whole interval score: 0% of intervals, because the behavior did not persist through any complete interval.

You have recorded a learner who is on task nearly all the time as being on task never.

The clinical consequence

These are not academic biases. They change decisions.

Use partial interval on a behavior-reduction program and your baseline will look inflated. Any subsequent decrease may partly reflect regression toward what was always true. Worse, once the behavior becomes brief but still frequent, partial interval will keep scoring near 100% and the graph will show no progress on a program that is working well. Families see a flat line. Funders see a failed intervention.

Use whole interval on a skill-acquisition program and the reverse happens. A learner making genuine gains in sustained attention scores zero until the behavior becomes almost perfectly continuous, at which point the data jumps from nothing to something. That is a graph nobody can make a timely decision from.

Choosing correctly

The convention that actually holds up:

• Whole interval for behaviors you want to increase and want sustained — on-task behavior, engagement, in-seat, sustained attention, cooperative play. The conservative bias works in your favor here: it will not credit you with success until the behavior is genuinely continuous.
• Partial interval for behaviors you want to decrease and want to catch every occurrence of — aggression, elopement, self-injury. The liberal bias also works in your favor: you would rather over-detect a dangerous behavior than miss instances of it.
• Momentary time sampling when you want the least biased estimate of how much time a behavior occupies, and you do not need to catch every episode.

Notice the logic: in each case you choose the method whose bias errs on the safe side for that clinical purpose. That is the actual decision rule, and it is more useful than memorizing which method overestimates.

Setting the interval length

Shorter intervals mean less bias in both systems. The trade is that shorter intervals are harder to run — a ten-second interval means six scoring decisions a minute while you are also delivering instruction.

Practical guidance:

• Ten to fifteen seconds for high-rate or brief behaviors where bias would otherwise be severe.
• Thirty seconds as a workable default for most classroom and clinic observations.
• Sixty seconds only when the behavior is long in duration, so the bias stays modest.

Write the interval length into the program. An interval length that drifts from thirty seconds to a minute because a new RBT set the timer differently is a silent change in what your data means.

Interobserver agreement on interval data

Interval data uses interval-by-interval agreement, sometimes called point-by-point: compare the two observers' scores interval by interval, count the intervals they agreed on, divide by the total number of intervals, multiply by 100.

Be aware this inflates when behavior is very rare or very frequent, because two observers can agree on a long run of empty intervals without demonstrating any real reliability. For behaviors at the extremes, calculate scored-interval agreement (only intervals where at least one observer scored occurrence) or unscored-interval agreement. Reporting the appropriate one is the difference between an IOA figure that means something and one that flatters you.

Making it run in a real session

Interval recording fails in practice for a mundane reason: the timer and the data sheet are in different places, so the RBT is watching a phone, a paper sheet, and a learner at once. Fidelity degrades within a week and the drift is invisible until someone runs IOA.

The fix is not more training. It is a data collection setup where the interval timer prompts on the same screen the score is entered on, the interval length is set by the program rather than by the person, and the scored data flows into the graph without re-entry. When the prescribed method is also the path of least resistance, the method that gets run is the method that was written.

The summary

Same sheet, same timer, opposite errors. Partial interval inflates, whole interval deflates, and both get worse as intervals get longer. Match the bias to the clinical purpose, fix the interval length in writing, and calculate the right flavor of agreement for how often the behavior actually occurs.