DRAGENwiki · Article

Continuous Measurement: Recording Every Time a Behavior Happens

Sources checked. Clinical review pending.

Continuous measurement means recording every occurrence of a behavior during an observation, and, for behaviors that last, when each one starts and stops.

Continuous measurement in 60 seconds

  • Nothing skipped. Continuous systems are "those that capture every possible behavioral occurrence." [7]
  • One record, many numbers. From one timed record you can get count, rate, duration, latency and interresponse time. Technicians are expected to "Implement continuous measurement procedures (e.g., frequency, duration, latency, interresponse time)." [2]
  • Preferred when possible. "When it is possible to capture information on each event, continuous measurement procedures are strongly preferred." [6]
  • It takes attention. Very fast behavior, several behaviors at once, or an observer with other jobs may lower its accuracy. [7]
  • Check it two ways. Agreement asks whether two observers match each other. Accuracy asks whether they match a known record. [4]

Jump to: What makes it continuous · What one record gives you · Tools · When it fits · Agreement · Calibration · Measuring yes and no · Example · Myths · Words to know · Need a PDF? · More from DRAGEN · Sources

What makes it continuous

Continuous measurement is a promise about coverage. During the observation, every occurrence of the behavior gets recorded. None are skipped because the observer was between intervals or glancing at a timer.

Mudford, Taylor and Martin (2009) used a precise test when they reviewed ten years of research articles. They counted data as continuous when the records "contained second-by-second records of occurrences of discrete behaviors or the onsets and offsets of behaviors with duration." [3] They counted methods as discontinuous when they "recorded behaviors in time samples or intervals of more than 1 s." [3]

That test sorts behavior into two kinds.

  • Events. Brief behaviors you can count one by one: a raised hand, a spoken request, a thrown block. You record each one as it happens.
  • Behaviors that last. Crying, working on a task, sitting with a group. You record when each one starts (the onset) and when it stops (the offset).

The BCBA exam outline asks behavior analysts to "Distinguish between continuous and discontinuous measurement procedures." [1] The DRAGENwiki page on discontinuous measurement covers the sampling methods.

Continuous does not mean all day. It means every occurrence within an observation window you choose, such as circle time or the first 20 minutes of homework. Write down when the window started and when it ended. [8]

What one record gives you

A continuous record with times on it holds a lot. From one record you can find:

  • Count: how many times the behavior happened.
  • Rate: count divided by observation time. Merbitz, Merbitz and Pennypacker (2016) write that "A count of some behavior or class of behavior recorded over a known period of time provides the two components of rate." [8]
  • Duration: how long each occurrence lasted, and the total time across the session.
  • Latency: the time from a cue, such as "time to clean up," to the start of the behavior.
  • Interresponse time (IRT): the time between one occurrence and the next.

The BCBA outline lists "Measure occurrence" and "Measure temporal dimensions of behavior (e.g., duration, latency, interresponse time)" as separate tasks. [1] RBTs are expected to "Calculate and summarize data in different ways (e.g., rate, mean duration, percentage)." [2]

Sampling methods cannot give you all of this. Mudford et al. write that "Discontinuous methods do not allow the basic dimensions of behaviors to be quantified accurately in standard scientific units." [3] In their review, 95% of the articles with continuous data reported frequency measures, usually responses per minute, and 36% reported duration. [3]

The times matter as much as the totals. A record that shows six short episodes in the first ten minutes tells a different story from one long episode at the end of the day, even when the total minutes match.

Tools: paper, counters, timers and devices

Continuous recording can be simple.

  • Paper and a clock. A tally mark for each event, with the start and end time of the observation written at the top.
  • A counter. A handheld clicker for fast events, so your eyes stay on the person.
  • A stopwatch. For duration and latency: start it at the onset or the cue, stop it at the offset.
  • A phone, tablet or laptop. Data apps can stamp each tap with a time, which makes rate, duration and IRT easy to work out later.

Handheld electronic devices for entering and storing data have made continuous recording more practical and affordable for research and clinical work. [3] Of 168 research articles that recorded free-operant behavior (behavior a person can do at any time, not only on a teacher's cue), 93 (55%) used continuous recording. [3] A later paper from the same group notes that "Observers collected data on portable laptop or handheld computers in the majority (95%) of those applications." [4]

In one calibration study, experienced observers kept their hands over the keyboard so they could type codes without looking away from the video they were scoring. [5] The authors add that more calibration study is needed to learn whether changing the way an observer records, such as with a touch-screen device, affects accuracy. [5]

Whatever the tool, the RBT outline expects technicians to "Enter data and update graphs." [2] A record that never reaches a graph cannot guide a decision.

When it fits, and when it is hard

When it fits. LeBlanc, Raetz, Sellers and Carr (2016) put it plainly: "When it is possible to capture information on each event, continuous measurement procedures are strongly preferred." [6] Fiske and Delmolino (2012) agree that continuous methods "offer a more accurate (i.e., higher quality) measurement of the target behavior and are preferred relative to discontinuous methods under most circumstances." [7]

Counting events works best when each event has clear edges. LeBlanc et al. describe event recording as "only appropriate for behaviors that have clear beginnings and endings." [6] When the edges are blurry, write onset and offset rules into the definition first. The DRAGENwiki page on operational definitions shows how.

When it is hard. Fiske and Delmolino name three conditions: counting behavior that happens at very high rates, such as words in a conversation; recording several responses at the same time; and collecting data while doing other work, such as a teacher collecting data during the school day. [7] They write that these "may all adversely affect the accuracy of continuous measurement." [7]

Some ways to keep it workable:

  • Shrink the window, not the coverage. Record every occurrence during a set 15 minutes instead of trying to cover the whole day.
  • One behavior at a time when you can.
  • Pick an observer who can watch. A second adult can record, or the team can score a video later, if the family or school agrees to the recording.
  • Choose sampling on purpose when none of that is possible. Fiske and Delmolino note that practitioners "may choose discontinuous systems (which tend to be easier to use) under conditions in which there are multiple demands upon data collectors." [7]

Checking agreement between two observers

A continuous record can still be wrong. Two ideas help, and they are easy to mix up.

Mudford, Martin, Hui and Taylor (2009) separate them: "Interobserver agreement concerns the extent to which observers' records agree with one another. By contrast, observer accuracy concerns agreement between observers' records and criterion records." [4] A criterion record is one whose answer is already known or has been checked with great care.

Most research checks agreement only. In the 2009 review, every article with continuous data reported interobserver agreement, and "none reported observer accuracy measures." [3]

Three ways of comparing two observers' continuous records each appeared in more than 10 articles in that review. [3]

  • Exact agreement. The session is cut into 10-second intervals. An interval counts as an agreement only when both observers recorded the same number of events in it, including both recording none. [3]
  • Block-by-block agreement. The session is cut into the same 10-second intervals. "In the block-by-block method, the smaller of the two observers' totals in a 10-s interval is divided by the larger." [3] The interval scores are then averaged, with intervals where both observers recorded none scored as full agreement. [3] It is also known as mean count-per-interval agreement. [4]
  • Time-window analysis. "One-second intervals are imposed on two observers' data streams, and second-by-second comparisons are made between them." [3] The method's designers found it too strict for brief events and recommended a small tolerance, so an event the other observer marked a few seconds away still counts as agreeing. In the articles reviewed, that tolerance ranged from 1 to 5 seconds. [3]

Each formula has blind spots. Mudford, Martin, Hui and Taylor tested all three against known records and found that "Block-by-block and exact agreement algorithms were susceptible to inflated agreement and accuracy estimates at lower rates and durations." [4] At high rates and durations the problem flipped: exact agreement was "excessively conservative" and time-window analysis "too liberal." [4]

Cutting a record into fixed intervals causes its own trouble. Two observers' marks "could be up to 9 s apart and count as an agreement or, at the other extreme, could be 1 s apart and count as two disagreements." [4]

The practical lesson: name the formula you used, and look at how often the behavior happened before trusting a high percentage. Mudford, Taylor and Martin found that agreement methods in published articles "were not always fully described or consistently named." [3] The DRAGENwiki pages on interobserver agreement go further.

Checking accuracy against a known answer

Most sciences check a measuring tool against a known value, the way a scale is checked with a known weight. Mudford, Zeleny, Fisher, Klum and Owen (2011) point out that "The quality of measurement systems used in almost all natural sciences other than behavior analysis is usually evaluated through calibration study rather than relying on interobserver agreement." [5]

They tried it. They scripted 10 videos, each 10 minutes long, with known rates from 0 to 8 responses per minute. Five new observers and five experienced observers recorded them on laptops, and each record was compared with the known values. [5]

The results were encouraging for that setup: "all of the observers recorded rates that were accurate to within ±0.4 responses per minute and 5 were accurate to within ±0.1 responses per minute." [5]

They separate two kinds of error.

  • Accuracy is "the absence of systematic error," meaning the observer does not lean steadily too high or too low. [5]
  • Precision is "the absence (or minimization) of random error in measurement." [5]

Three cautions from the same paper:

  • It may not transfer. "We must emphasize that the results of this calibration study cannot be generalized or extrapolated beyond the particular measurement systems assessed." [5] Their task used three active recording keys, while the protocols behind their videos typically used 7 to 16, so the authors see the study task as less demanding than the experienced observers' usual work. [5]
  • Train first. When an observer's numbers run steadily high or low, they recommend "training observers in the first instance," not adjusting the numbers afterward. [5]
  • Keep agreement too. An observer could match the known total while actually recording some other behavior at different times, and whole-session totals cannot reveal that. So they recommend that calibration studies keep agreement measures as well. [5]

For a team, the simplest version is a practice video that one skilled observer has scored with great care. The authors suggest that "The most economical method for obtaining reference values would be to employ a single expert observer to measure reference values for calibration." [5]

Measuring yes and no

A continuous record is a good place to keep a learner's voice. Each "no," each request for a break and each push-away can go on the record with its time, next to what the adult did in response.

Breaux and Smith (2023) describe practice in which "both assent and assent withdrawal are individually defined for the learner, continuously evaluated, and responded to." [9] A timed record makes that possible. It shows when a "no" tends to come, and whether it was honored.

DRAGEN's rule. We record a learner's "no" as information about the task, the setting or the relationship. It is never "noncompliance" and never a target to reduce. Breaux and Smith write that "any form of the learner's assent withdrawal behaviour should be honored." [9] When a behavior can cause harm, it gets its own definition, its own measure and its own safety plan, kept separate from the learner's ways of saying no.

A made-up example: Student B

Student B is an invented 6-year-old who cries at school drop-off. The teacher says Student B "cries all morning." The team, which includes the parent, decides to record continuously during the first 30 minutes after arrival.

The teacher's aide records each start and stop on a tablet. Requests go on the record too. Each time Student B asks for a hug or for the family photo kept in the cubby, the aide marks it, and the request is honored.

On Monday the record shows 5 crying episodes, 9 minutes in all, every one in the first 12 minutes. The gaps between episodes grow longer each time. After minute 12 there is no more crying, and Student B joins the morning song.

"All morning" was really the first 12 minutes. That points the team to the start of the day: a picture schedule of the morning, the same short goodbye routine each day, and the family photo offered right away instead of waiting for a request.

On Thursday a second adult records the same window. The two totals are close, but the times are far apart. The two adults had read "starts crying" differently: one marked the first sound, the other the first tears. The team rewrites the definition: crying starts at the first tears or crying sound, whichever comes first, and stops after 10 quiet seconds. At the next check, the two records line up.

Two weeks later the record shows 2 episodes, 3 minutes in all, both in the first 5 minutes. Requests for the photo are up, and each one is still honored. The numbers do not explain everything about Student B's mornings. They showed the team where to look.

Common myths

  • "Continuous means all day." It means every occurrence within the window you pick. Write down when the window starts and ends. [8]
  • "Continuous data are always accurate." Speed, several behaviors at once and an observer's other jobs may lower accuracy. [7]
  • "If the total is right, the record is right." A total can match the known value even when the observer was recording some other behavior at different times. [5]
  • "Any agreement formula will do." Each common formula can run high or low depending on how often the behavior happens. [4]
  • "You need a computer." Paper, a pencil and a clock can record every occurrence. Whether laptops beat pencil and paper is still open; Mudford et al. call that comparison one "that should be the focus of future investigation." [5]

Words to know

  • Continuous measurement: recording every occurrence during an observation window.
  • Discontinuous measurement: recording a sample, such as intervals or moments.
  • Event: a brief behavior counted one by one.
  • Onset and offset: the moment a behavior starts and the moment it stops.
  • Observation window: the stretch of time you record, with a written start and end.
  • Interresponse time (IRT): the time between one occurrence and the next.
  • Interobserver agreement (IOA): how closely two observers' records match each other.
  • Accuracy: how closely a record matches what truly happened, with no steady lean high or low.
  • Precision: how little a record's errors scatter at random.
  • Calibration: testing a measuring system against known values.
  • Tolerance: the few seconds by which two records may differ and still count as agreeing.

Need a PDF?

Create it here, for free: the DRAGEN app. Use it at your next session or team meeting.

Or print one we already made, also free:

More from DRAGEN

Related DRAGENwiki pages: measuring behavior · discontinuous measurement · operational definition · frequency · duration · latency · interobserver agreement · accuracy

What remains unknown

  • Fiske and Delmolino note that the effect of fast behavior, several behaviors at once and an observer's other jobs on continuous recording "has not been definitively examined in research." [7]
  • The calibration results apply to that study's videos, laptops, observers and rates up to 8 per minute. [5] The authors call the benefits of calibration for applied researchers "hypothetical." [5]
  • No agreement formula in the comparison study worked well across every rate and duration, and the authors suggest that other, less common methods should be studied too. [4]
  • Whether laptops or pencil and paper give more accurate records has not been settled. [5]
  • The 2009 review covers research articles from 1995 to 2005. [3] How often everyday clinics and schools record continuously today is not known from it.
  • This page is general information, not medical or legal advice.
  • This page was last checked on November 12, 2026.

Sources

  1. Behavior Analyst Certification Board. BCBA Test Content Outline (6th ed.). https://www.bacb.com/wp-content/bcba-outline-6thEd/
  2. Behavior Analyst Certification Board. RBT Test Content Outline (3rd ed.), effective January 1, 2026. https://www.bacb.com/rbt-test-content-outline-3rd-ed
  3. Mudford, O. C., Taylor, S. A., & Martin, N. T. (2009). Continuous recording and interobserver agreement algorithms reported in the Journal of Applied Behavior Analysis (1995–2005). Journal of Applied Behavior Analysis, 42(1), 165–169. https://doi.org/10.1901/jaba.2009.42-165 (free full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC2649837/)
  4. Mudford, O. C., Martin, N. T., Hui, J. K. Y., & Taylor, S. A. (2009). Assessing observer accuracy in continuous recording of rate and duration: Three algorithms compared. Journal of Applied Behavior Analysis, 42(3), 527–539. https://doi.org/10.1901/jaba.2009.42-527 (free full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC2741073/)
  5. Mudford, O. C., Zeleny, J. R., Fisher, W. W., Klum, M. E., & Owen, T. M. (2011). Calibration of observational measurement of rate of responding. Journal of Applied Behavior Analysis, 44(3), 571–586. https://doi.org/10.1901/jaba.2011.44-571 (free full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC3177337/)
  6. LeBlanc, L. A., Raetz, P. B., Sellers, T. P., & Carr, J. E. (2016). A proposed model for selecting measurement procedures for the assessment and treatment of problem behavior. Behavior Analysis in Practice, 9(1), 77–83. https://doi.org/10.1007/s40617-015-0063-2 (free full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC4788644/)
  7. Fiske, K., & Delmolino, L. (2012). Use of discontinuous methods of data collection in behavioral intervention: Guidelines for practitioners. Behavior Analysis in Practice, 5(2), 77–81. https://doi.org/10.1007/BF03391826 (free full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC3592492/)
  8. Merbitz, C. T., Merbitz, N. H., & Pennypacker, H. S. (2016). On terms: Frequency and rate in applied behavior analysis. The Behavior Analyst, 39(2), 333–338. https://doi.org/10.1007/s40614-015-0048-z (free full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC6701258/)
  9. Breaux, C. A., & Smith, K. (2023). Assent in applied behaviour analysis and positive behaviour support: Ethical considerations and practical recommendations. International Journal of Developmental Disabilities, 69(1), 111–121. https://doi.org/10.1080/20473869.2022.2144969 (free full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC9897747/)

Pages that link here