Data Collection and Graphing

Behavior-analytic decisions depend on observable evidence. A technician’s job is not to decide whether a person is doing “better” based on impression; it is to record defined events in a way that allows the supervisor to evaluate change. Good measurement is objective, consistent, feasible in the service setting, and matched to the dimension of behavior that matters. Before collecting data, review the operational definition. It should identify what counts, what does not count, and where an episode begins and ends. When a definition is unclear, request clarification rather than creating a personal interpretation. Small differences in how staff apply a definition can produce large differences in the graph.

Continuous systems attempt to capture every relevant occurrence during the observation period. Count is the total number of responses. Rate adjusts count for observation time, such as responses per hour. Duration measures how long behavior lasts. Latency measures the delay from a specified antecedent to response onset. Interresponse time measures the time between successive responses. Other useful measures include percentage of opportunities and trials to criterion. Percentage is calculated by dividing successful opportunities by total opportunities and multiplying by 100. Trials to criterion records how many teaching opportunities are needed before a predefined standard is met. Record units clearly; a number without its unit can be misleading.

Some behavior leaves a measurable result after the response is complete. Examples include assembled items, completed forms, dishes placed in storage, or correctly sorted materials. Permanent-product measurement can be efficient because the observer may not need to watch the behavior occur. The product must be caused by the target behavior, distinguishable from products made by others, and available for reliable counting. A completed worksheet may not be a valid product of independent work if another person filled it in. When the plan requires direct observation of quality, prompts, or safety, a permanent product alone may be insufficient.

Discontinuous systems estimate behavior from samples. Partial-interval recording scores an interval if behavior occurred at any time during it; it can overestimate the proportion of time occupied by behavior. Whole-interval recording scores an interval only if behavior occurred throughout it; it can underestimate occurrence. Momentary time sampling records whether behavior is occurring at a specific instant, often at the end of the interval. These systems are useful when continuous recording is impractical, but they do not produce the same information as exact count or duration. Interval length, observation schedule, and the behavior’s pattern affect the estimate. Follow the specified timing and avoid watching only when behavior seems likely.

Accurate data require correct definitions, reliable timing tools, complete records, and consistent implementation. Observer drift occurs when a person gradually changes how they apply a definition. Periodic retraining and comparison with examples can reduce drift. Interobserver agreement compares independent records from the same observation. High agreement supports confidence that observers applied the system similarly, but it does not prove that the definition itself is clinically useful. Treatment integrity or procedural fidelity is a separate question: Was the intervention delivered as written? A session can have accurate measurement and poor implementation, or the reverse.

On a typical line graph, the horizontal axis represents sessions, dates, or time, and the vertical axis represents the measured value. Data points are plotted in sequence and connected within the same condition when appropriate. Condition-change lines and labels show when a meaningful change in baseline, teaching, or intervention occurred. Visual review often considers level, trend, and variability. Level is the general magnitude of the data. Trend is the overall direction. Variability is the amount of fluctuation. Technicians should enter data accurately and promptly, not interpret beyond their assigned role or alter data to make a pattern look cleaner.

Rate equals count divided by observation time. Convert time into the same unit requested by the program. For example, 15 responses in 30 minutes equals 0.5 responses per minute or 30 responses per hour. Percentage equals correct responses divided by total opportunities multiplied by 100. For total-count agreement, divide the smaller count by the larger count and multiply by 100. Other agreement formulas exist, so use the formula specified by the supervisor or system. Keep raw values available; a calculated percentage should be traceable to its numerator and denominator.

Never estimate data from memory when direct recording was required. If data are missing, mark them according to policy rather than entering zero; zero means the behavior was observed and did not occur. If an entry is wrong, use the approved correction process. Electronic systems may preserve an audit trail, while paper systems may require a single-line correction, date, and initials. Report equipment failures, unclear definitions, interrupted observations, or deviations that may affect interpretation. Honest documentation of a limitation is more useful than a precise-looking value that was not actually measured.