What Actually Happens When You Collect Data in a Functional Analysis
Most people treat functional analysis data collection like it is some kind of sacred ritual. You set up conditions, you press record, you stare at ABC charts until your eyes glaze over, and somehow insight is supposed to materialize. It does not work that way unless you are paying attention to the right things at the right time. The phrase sounds pretentious when you read it in a textbook, but it is actually describing a very mechanical process. You observe a behavior under controlled conditions. You note what happens before and after. You look for patterns across trials. Then you decide what to do about it. The gap between observation and action is where most people fail, not the observation part itself. I spent years watching clinicians collect beautiful data that went nowhere. They would record frequency, duration, latency with perfect integrity, produce graphs that looked gorgeous, and then have no idea what intervention to recommend. The data was accurate and completely useless. The problem was they were collecting everything instead of collecting what mattered for the decision at hand.
Here is what I learned the hard way. Before you run a single trial, write down the exact decision you need to make after the data comes in. Not a vague goal like "understand the behavior." A specific decision: do we test attention function, tangible access, automatic reinforcement, or escape demand? Your data collection plan should be designed backward from that single question. Everything else is noise.
Condition Design and What to Actually Record
In a standard functional analysis you run conditions like alone, attention, escape, and play. That is the Iwata framework most people reference. Under each condition you manipulate antecedents and consequences to see if the target behavior changes rate. The data you collect is almost always instantaneous or partial interval recording of the behavior during the condition periods. Response latency can be useful in escape conditions. Frequency works fine for high-rate behaviors but falls apart when the behavior happens three times in an hour. The counter-intuitive part that beginners miss is that condition length matters more than most people realize. Twenty second sessions with thirty second intertrial intervals are standard in many labs, but in natural classroom or home settings that is often impossible to implement with fidelity. I ran a case where the classroom teacher could not manage twenty second trials because of transition demands and student schedules. We extended sessions to two minutes and adjusted the intertrial interval to fifteen seconds. The data patterns came out essentially the same. The function was still clear. Longer sessions actually improved data quality here because the behavior had room to emerge naturally rather than being compressed into artificial windows. Another thing nobody warns you about: the play condition is not just a control. It is your baseline for comparison. If the play condition data looks weird or inflated, every other condition is going to look weird too. I once had a client where the play condition showed elevated problem behavior because the reinforcer presentation was accidentally contingent on compliance rather than being freely available. Every condition looked like it produced behavior relative to that inflated baseline. We caught it only because I manually checked the play condition session logs against the coding sheet and noticed the contingency drift. That cost us three extra sessions to redo properly.
Get the Full Details
The Workaround I Use Now
When I am doing functional analysis data collection in the field, I use a dual-coding system. One person runs the condition and delivers contingencies. A second person records data from a separate room or across the room using a simple timing app. This eliminates the temptation to eyeball data while managing the environment. The tradeoff is you need two trained people, which is not always available. When I do not have a second person, I use brief momentary time samples rather than continuous recording. I set a timer for ten second pulses and scan for the behavior at each pulse. It is less precise than continuous recording, maybe five to ten percent error in my experience, but it is dramatically more feasible in busy settings. The pattern of results across conditions still comes out reliable enough for decision making. You lose the ability to detect very low rate behaviors reliably, but that is a known limitation you accept for the sake of implementation feasibility. I also stopped relying on visual inspection of graphed data alone. I started calculating simple rate ratios between each condition and the play condition. If the attention condition rate is more than twice the play condition rate and the escape condition rate is three times the play condition rate, you have a clear multiple function even if the visual overlap looks messy. Ratios cut through a lot of the ambiguity that makes clinicians second guess themselves.
When This Approach Completely Fails
Functional analysis data collection breaks down in a few specific scenarios and you need to know them before you walk into a room. First, behaviors with extremely low response rates, like once per day or less, are nearly impossible to analyze functionally in a single session. You would need days or weeks of baseline data just to establish a stable rate before you could even start manipulating conditions. Second, dangerous behaviors that require immediate protective equipment or safety protocols can make standard session timing impossible. Third, medical or neurological conditions that produce stereotypic or self-injurious behavior indistinguishable from automatically reinforced behavior on surface observation. In those cases the functional analysis might show behavior in the alone condition but you cannot tell if it is truly automatic reinforcement or a seizure variant without medical collaboration and possibly EEG monitoring. If any of those apply, you do not push through with standard FA procedures. You adapt or you refer. There is no honor in forcing a method that is not suited to the case.
Practical Steps to Run a Clean Session
Before the session starts, define the target behavior operationally so that anyone watching the same video would code it the same way. Not "aggression" but "any instance where the client made forceful physical contact with another person's body using hand, foot, or object." Write that definition on a card and keep it at the recording station. Set up your timing device and testing sheet before the client enters the room. I use a simple spreadsheet with columns for condition, trial number, onset time, offset time, and a notes field for any anomalies. If something goes wrong during a trial, like a staff member accidentally provides attention during an escape condition, you note it immediately and consider discarding that trial rather than trying to correct it retroactively. Run conditions in the standard sequence: alone, attention, escape, play, then repeat if needed. Do not skip the alone condition even if you think you already know the function. The alone condition is the test for automatic reinforcement and skipping it means you will never know whether your attention and escape findings are contaminated by an undetected automatic component.
After each condition block, calculate the rate per minute immediately while the data is fresh. If a condition shows zero behavior, that is meaningful data, not a missing data point. Record it explicitly. Zero behavior in attention while the same behavior is high in escape is a cleaner result than a clinician who only recorded the trials where behavior actually occurred.
The Decision Point
Once you have data across all conditions, the actionable insight comes from comparing rates. Dominant function is the condition with the highest rate. Multiple functions exist when two or more conditions show elevated rates above the play baseline. No clear function means either the behavior is maintained by something your conditions did not capture or the behavior rate is too low and variable for the analysis to detect a pattern. In that last case, you extend the number of sessions, you consider alternative assessments like descriptive assessment or scatterplot analysis, or you accept that you may need more information before intervention decisions are safe to make. The whole process from setup through data interpretation usually takes about two to three hours for a standard four-condition FA with three repetitions per condition. With practice and good team coordination you can compress that to around ninety minutes. The data integrity drops sharply if you rush below that threshold, so do not compress further unless you have a very controlled environment and well-trained staff. Write the intervention recommendation based on the identified function. Not based on your hunch. Not based on what worked last time. Based on the data you collected under the conditions you designed. That is the entire point of turning observations into actionable insight, even if the sentence sounds like it was written by someone who billable hours were running low.
