Data Collection and Measurement
Measurement starts with an operational definition that is observable, measurable, and complete enough that two trained observers would agree. Then choose a dimension that matches the question you are asking.
Continuous measures include frequency (a count), rate (count divided by observation time), duration (total or per occurrence), latency (time from the SD to the start of the response), and inter-response time. Use rate whenever observation lengths differ across sessions; comparing raw counts from a 10-minute and a 30-minute observation is a classic trap.
Discontinuous or time-sampling measures are for behavior that is too frequent, too brief, or too hard to watch continuously. Whole-interval recording tends to underestimate, partial-interval recording tends to overestimate, and momentary time sampling is a spot-check that is least biased for duration but a poor index of frequency. Never convert interval data into a frequency count.
Also know accuracy versus reliability, observer drift, and reactivity, and the interobserver agreement methods: total count, mean count-per-interval, exact count, trial-by-trial, total duration, and mean duration-per-occurrence. Agreement of 80 to 90 percent is generally acceptable, with higher standards for high-stakes decisions.
On graphs, the x-axis is time or sessions, the y-axis is the behavior, and condition change lines plus phase labels separate conditions. Read level, trend, and variability before claiming an effect.
Worked example: in a 10-second whole-interval recording, the behavior occurs during seconds 2 through 7. Because the behavior did not occur for the entire interval, the interval is not scored, which illustrates why whole-interval data underestimate behavior. Partial-interval recording would have scored it.
Rule of thumb: count it if it has clear onset and offset and is countable; time it if duration matters; sample it when continuous observation is impractical.