Every clinical decision in ABA rests on data, and the method you choose to collect that data decides whether it answers the question or quietly misleads you. Frequency, duration, interval sampling, task analysis: each measures a different dimension of behavior, and a method that fits one target can distort another. A high-rate behavior recorded with the wrong interval method can look like it is improving when it is not. A teaching program tracked with the wrong measure can show mastery that has not happened. This guide walks through the 12 data collection methods that cover almost everything an ABA practice records, what each one actually measures, when to reach for it, and the error it tends to introduce. It is one discipline within a larger picture; the ABA practice operations guide covers how clean data fits alongside documentation, compliance, and the other systems that keep a practice running.
Reviewed by the VG Soft Co Clinical team. Last updated June 2026.
TL;DR
- The 12 methods sort into four families. Continuous measurement (frequency, rate, duration, latency, inter-response time), discontinuous measurement (partial-interval, whole-interval, momentary time sampling), product and outcome measures (permanent product, trials-to-criterion, task analysis), and descriptive recording (ABC).
- Continuous is precise, discontinuous is feasible. Continuous methods count or time every instance. Discontinuous methods sample, which is easier across a long session but introduces predictable error.
- Partial-interval overestimates, whole-interval underestimates, momentary time sampling is closest to true. This is one of the most replicated findings in behavioral measurement, and it changes how you read the data.
- Match the method to the dimension that matters. How often, how long, how fast, how efficiently learned: each question points to a different method. The wrong method can defend the wrong decision.
- Data quality lives in consistency, not method choice. A well-chosen method collected differently by every RBT is worse than a simple method collected the same way by everyone.
How to think about ABA data collection methods
Before reaching for any specific method, two questions settle most of the decision. Which dimension of the behavior matters clinically, and how completely can you record it given everything else happening in the session?
The first question points to a measurement. Behavior can be counted, timed, measured by how long it takes to start, or judged by what it produces. Picking the dimension that matters is a clinical judgment, not a clerical one. The second question is practical. An RBT running a discrete-trial program cannot also watch a separate behavior every second, so sometimes the honest choice is a method that samples rather than captures everything.
The field organizes these methods into continuous and discontinuous measurement, a split the BACB's RBT requirements make explicit: task A.1 covers continuous measurement, task A.2 covers discontinuous. Continuous methods record every occurrence. Discontinuous methods estimate from samples. A third group measures the product or efficiency of behavior rather than the behavior in real time, and ABC recording stands apart as a descriptive tool for understanding why a behavior happens at all. Choosing well across these families is also a medical-necessity issue, since the progress data that defends an authorization renewal has to actually reflect the behavior, a point the Council of Autism Service Providers guidelines treat as central to documenting outcomes.
The 12 methods at a glance
| # | Method | Family | What it measures | Best for | Watch out for |
|---|---|---|---|---|---|
| 1 | Frequency (event) | Continuous | Count of occurrences | Discrete, countable behaviors | Needs equal observation time to compare |
| 2 | Rate | Continuous | Count per unit time | Sessions of unequal length | Can hide within-session patterns |
| 3 | Duration | Continuous | How long behavior lasts | Tantrums, on-task, stereotypy | Requires continuous timing |
| 4 | Latency | Continuous | Time from cue to response onset | Responsiveness to instructions | Needs a clear antecedent onset |
| 5 | Inter-response time | Continuous | Time between successive responses | Pacing, DRL programs | Niche, demands precise timing |
| 6 | Partial-interval | Discontinuous | Did it occur at all in the interval | Behaviors targeted for reduction | Overestimates occurrence |
| 7 | Whole-interval | Discontinuous | Did it occur the entire interval | Behaviors targeted for increase | Underestimates occurrence |
| 8 | Momentary time sampling | Discontinuous | Occurring at the interval's end | Long sessions, group settings | Misses behavior between samples |
| 9 | ABC (narrative) | Descriptive | Antecedent, behavior, consequence | Hypothesizing function | Correlational, not a functional analysis |
| 10 | Permanent product | Product | Tangible result of behavior | Work output, completed tasks | Must confirm who produced it |
| 11 | Trials-to-criterion | Outcome | Trials needed to reach mastery | Comparing teaching efficiency | Depends on how strict the criterion is |
| 12 | Task analysis | Outcome | Performance on each chained step | Multi-step routines | Only as good as the step breakdown |
Continuous measurement: recording every instance
Continuous methods capture the behavior in full. They are the most precise tools available, and they are the right default whenever the behavior is discrete enough and infrequent enough that an observer can catch every instance.
1. Frequency (event) recording
Frequency is the simplest measure in ABA: a tally of how many times a behavior occurs. It fits behaviors with a clear beginning and end that happen at a manageable pace, like raising a hand, requesting an item, or hitting. An RBT might record nine instances of spontaneous manding across a session. The catch is comparability. A raw count only means something against a fixed observation window, so if your session lengths vary, frequency alone will mislead you, which is exactly the gap rate fills.
2. Rate
Rate takes the frequency count and divides it by time, giving a per-minute or per-hour figure. Twelve instances across a four-hour session is a rate of three per hour, and that number is comparable to a different session of a different length. Rate is the honest way to report countable behaviors in a practice where sessions run anywhere from one to four hours. Its one blind spot is that an average can smooth over a behavior that clusters at a particular time, so a rate that looks stable can hide a pattern worth seeing.
3. Duration recording
Duration measures how long a behavior lasts, from onset to offset. Use it when length is the clinical story: a tantrum, a stretch of on-task engagement, an episode of self-stimulatory behavior. A 6-minute tantrum and a 30-second one are both one instance by frequency, but duration is what separates them. The cost is attention. Duration needs continuous observation with a timer, which is why long or near-continuous behaviors often get estimated with momentary time sampling instead of timed directly.
4. Latency recording
Latency is the time between a cue and the start of the behavior. It answers how quickly a client responds, which makes it the measure of choice for instruction-following and responsiveness. If you say "time to clean up" and the client begins 45 seconds later, that 45 seconds is the latency. It requires a clearly marked antecedent, because without a clean onset you cannot time from anywhere. Latency captures only the start of the response, not what happens after, so it usually rides alongside another measure rather than standing alone.
5. Inter-response time (IRT)
Inter-response time is the gap between the end of one response and the start of the next. It describes pacing rather than amount, and it is inversely related to rate: as IRT shrinks, rate climbs. IRT earns its keep in programs built around spacing, like differential reinforcement of low rates, where the goal is to stretch the time between responses. It is the most specialized method on this list and demands precise timing, so most practices reach for it only when the pacing of a behavior is itself the target.
Discontinuous measurement: sampling the behavior
Discontinuous methods divide the observation into intervals and record whether the behavior happened, rather than counting every instance. They trade precision for feasibility, and that trade is often worth making, but only if you know which direction each method bends the data. A 2019 analysis in Behavior Analysis in Practice found these methods are widely used in everyday clinical work, which makes understanding their built-in error a practical necessity, not an academic one.
6. Partial-interval recording
Partial-interval recording scores an interval as an occurrence if the behavior happened at any point within it. Because one brief instance flags the whole block, it tends to inflate the apparent prevalence of a behavior. That upward bias is useful in one specific way: for behaviors you are trying to reduce, partial-interval recording is conservative, since it will not let you understate the problem. Use it for reduction targets, and read the resulting percentage as an upper bound rather than a true value.
7. Whole-interval recording
Whole-interval recording is the mirror image. It scores an interval only if the behavior occurred for the entire interval, which means any pause breaks the count and the method underestimates. That downward bias suits behaviors you want to increase, like sustained on-task engagement, because it sets a demanding bar and will not let you overstate progress. The same property that makes it conservative for skill-building makes it a poor fit for anything brief or intermittent.
8. Momentary time sampling (MTS)
Momentary time sampling records whether the behavior is happening at the exact moment an interval ends, ignoring everything in between. That single-snapshot approach is why it stays closer to the truth than the interval methods. Research comparing it to continuous duration recording, including the foundational Powell, Martindale, and Kulp study and later accuracy analyses, consistently finds momentary time sampling produces far less biased estimates than partial- or whole-interval recording. Accuracy improves as intervals shorten, and it is the most practical choice when an RBT has to monitor a behavior while running another program or watching a group. Its limit is structural: anything happening only between the sample points is invisible to it.
A worked example: three interval methods, one behavior, three different answers
The reason method choice is not academic shows up the moment you run the interval methods side by side on the same behavior. Suppose you observe vocal stereotypy for 10 minutes, divided into twenty 30-second intervals, and the behavior actually occurs for 4 of those 10 minutes, a true value of 40% of the observation.
| Method | What gets scored | Intervals scored | Reported result | vs. true 40% |
|---|---|---|---|---|
| Partial-interval | Any occurrence in the interval | 13 of 20 | 65% | Overestimates |
| Whole-interval | Occurrence for the full interval | 4 of 20 | 20% | Underestimates |
| Momentary time sampling | Occurring at the interval's end | 8 of 20 | 40% | Matches |
Same behavior, same 10 minutes, three different answers ranging from 20% to 65%. The numbers here are illustrative, but the direction of the error is not. It is one of the most replicated patterns in behavioral measurement, documented in a JABA comparison of momentary time sampling and partial-interval recording. If a treatment decision hinges on whether stereotypy is at 20% or 65%, the method you picked made that decision for you. This is why a BCBA should specify the measurement method in the program, not leave it to whoever is collecting that day.
Product and outcome measures
Some behaviors are best measured by what they leave behind or how efficiently they are learned, rather than watched in real time. These three methods answer different questions than the dimensional measures above.
9. ABC (narrative) recording
ABC recording sits in the descriptive family. You document the Antecedent that came before a behavior, the Behavior, and the Consequence that followed, building a record that surfaces patterns: demands tend to precede screaming, attention tends to follow it. That pattern becomes a hypothesis about function that informs a behavior intervention plan. The essential caveat is that ABC data is correlational and shaped by what the observer notices. It is not a functional analysis, which experimentally tests function, and treating an ABC hunch as proven function is one of the more consequential errors in assessment. The BACB Ethics Code frames this kind of assessment rigor as a professional obligation, not an option.
10. Permanent product recording
Permanent product measures the tangible result a behavior produces instead of the behavior itself. Worksheets completed, items put away, words written: if the behavior leaves a durable trace, you can measure the trace after the fact, which frees the observer from watching in real time. The discipline it requires is confirming that the client actually produced the product, and produced it the way intended. A stack of finished worksheets says nothing useful if someone else completed half of them, so permanent product works best when the link between the client's behavior and the result is unambiguous.
11. Trials-to-criterion
Trials-to-criterion counts how many learning trials a client needs to reach a predefined mastery standard, such as 80% correct across three consecutive sessions. It is a measure of learning efficiency, which makes it valuable for comparing teaching procedures or tracking how quickly a client acquires new skills over time. Because it is a derived measure, it is only as meaningful as the criterion behind it. A loose criterion produces flattering trial counts and a strict one produces sobering ones, so the number means nothing without the standard it was measured against. Research on data collection during discrete-trial teaching shows even the sampling approach within a teaching session can change how fast a skill appears to be acquired.
12. Task-analysis (chaining) data
Task-analysis data tracks performance on each step of a multi-step skill. You break a routine like handwashing or tooth-brushing into its component steps, then score each one, typically as independent, prompted, or incorrect. The payoff is precision about where teaching needs to focus: a client may sail through eight steps and stall on the ninth, and a single pass-or-fail score for the whole routine would hide that. The data depends entirely on the quality of the task analysis. Too few steps and you cannot see where the chain breaks, too many and collection becomes unworkable.
Matching the method to the behavior
No single method is correct in the abstract. The right one is whichever captures the dimension that matters with acceptable accuracy and realistic effort. A practical sequence: name the behavior precisely enough that two people would record it identically, decide which dimension carries the clinical question, then pick the most accurate method you can actually sustain across the team. Guidance for practitioners on selecting discontinuous methods makes the same point from the measurement side: the method has to fit both the behavior and the conditions you collect it under.
That last clause carries more weight than method choice itself. A perfectly chosen method collected three different ways by three different RBTs produces worse data than a simple method everyone runs the same way. Inter-observer agreement is the real currency of data quality, and it depends on clear behavioral definitions and consistent training far more than on method sophistication. This is partly why supervision matters operationally, not only clinically; the BACB supervision and training standards exist in large part to keep delivered services, and the data behind them, consistent across a team.
Where data collection fits in practice operations
Good measurement does not end at the data sheet. The point of collecting clean data is to make better clinical decisions and to defend the medical necessity of continued services, and both of those depend on the data being usable, not only collected. Demand for ABA keeps climbing, with the CDC now identifying autism in 1 in 31 children, and payers are moving toward outcomes-based authorization that rewards practices able to show clean, consistent progress. Data that sits unread in a binder defends nothing.
This is where the collection method and the system that holds it stop being separate concerns. Paper-based collection adds transcription lag and error between the session and the decision, and it makes the inter-observer consistency described above harder to maintain. Digital, real-time collection closes that gap. VGPM's data collection supports the full range of methods covered here across eight goal types, from trial-by-trial percent-correct to duration and interval sampling and task analysis, and it counts consecutive sessions against mastery criteria automatically so a goal does not sit unadvanced after the client has already met it. Phase-change markers on the progress graph separate data collected under different intervention conditions, and role-aware reporting turns that data into the progress view an authorization renewal needs to show. For a vendor-neutral look at the options, the best ABA data collection software comparison covers the landscape.
The methods are the foundation, and they are the same regardless of which platform a practice uses. Get the behavioral definition tight, choose the method that fits the dimension and the conditions, train the team to collect it the same way, and the data will earn its place in clinical and billing decisions alike. The same discipline that produces trustworthy progress data also produces session notes that survive review, which is why measurement and documentation tend to improve or decay together. We cover that connection in the eight session note mistakes that trigger denials, and the wider operational system that both feed into is mapped in the ABA practice operations guide.
