Translated from the Portuguese original. Read the original (Português)
In Behavior Analysis, the decision to keep, adjust, or abandon an intervention usually comes from a graph, not from a statistical test. Visual analysis is the disciplined procedure of looking at that graph and answering two questions: did the behavior change? and can the change be attributed to what was done? To answer them, six properties of the data are examined — level, trend, and variability within each phase; immediacy of effect, overlap, and consistency between phases. A causal relation is only considered documented when the same pattern repeats at three different points in the course of treatment. It is a transparent and conservative method — and precisely because it depends on judgment, it requires training and well-built graphs.
Visual analysis (or visual inspection) is the primary method of data interpretation in single-case designs: the systematic examination of behavioral data represented graphically — generally a line graph with time on the abscissa and the behavior measure on the ordinate — to judge whether a behavioral change occurred and whether that change can be attributed to the independent variable, characterizing a functional relation. The judgment rests on six characteristics of the data: level, trend, variability, immediacy of effect, overlap between phases, and consistency of patterns across similar phases. Under the What Works Clearinghouse (WWC) criteria, these six characteristics are used to check whether the set of phases documents at least three demonstrations of the effect at three different points in time.
Why a graph, not an average
In group research, the result is usually a number that summarizes many people: the average of the group that received the treatment compared to the average of the group that did not. Behavior Analysis works with a different question — is this person improving with this procedure? — and so it tracks the same individual repeatedly over time, measuring the behavior session after session.
With data like this, the graph is not an illustration of the result: it is the result. It shows the change happening, at the pace it happened, with the stumbles along the way. And unlike a p-value, it allows a decision while the intervention is still underway: if three weeks of data show no movement, no one waits for the end of the study to change the plan.
Before reading the graph, it helps to recognize its parts. A single-case line graph, in the format described by Cooper, Heron, and Heward, has:
- Abscissa (horizontal axis, x) — the passage of time: sessions, days, weeks.
- Ordinate (vertical axis, y) — the behavior measure, in the chosen dimension: count, rate, duration, percentage.
- Condition-change lines — vertical marks that separate the phases (baseline, intervention, withdrawal), each with its own label.
- Data points and data path — each measure and the line connecting consecutive measures. By convention, the path does not cross a condition-change line: connecting the points across it would suggest a continuity that was not observed.
- Figure caption — the sentence that states what was measured, in whom, and under which conditions.
Two practical construction recommendations make a huge difference in reading the graph: do not crowd more than about three data paths onto the same graph, and keep a balanced ratio between the height of the ordinate and the length of the abscissa (the usual convention is around two-thirds). A squeezed or stretched graph misleads the reader — including whoever drew it.
The six properties that support the judgment
Reading the graph is not a general impression of the drawing: it is a sequence of comparisons. First you look within each phase, then between neighboring phases, and finally you put it all together. Explore each step:
Level, trend, and variability
Before comparing anything, each phase must be described on its own. Three questions are enough:
- Level — where are the data on the vertical axis? It is the typical value of the phase, usually read as the mean or the median of the points. It answers "how much of the behavior is happening."
- Trend — where are the data heading? Rising, falling, or with no clear direction, and at what slope. A phase can have a high level and still be falling steadily.
- Variability — how much do the points spread around that level or trend? Data clustered tightly together allow reliable predictions; data that jump from one extreme to the other make any prediction fragile.
This is where it is decided whether the baseline is ready. Until it shows a reasonably predictable pattern, there is nothing to compare against — and changing phases too early usually costs the whole conclusion.
Immediacy, overlap, and consistency
Once the phases are described, the adjacent ones are compared. Besides changes in level and trend, three properties come into play:
- Immediacy of effect — how quickly the change appears after the phase switch. In practice, the last points of one phase are compared with the first points of the next. Immediate changes strengthen the attribution to the intervention; changes that only show up many sessions later leave room for other explanations. (There are legitimate exceptions: procedures that by nature produce a gradual effect, such as shaping.)
- Overlap — how much the values of one phase invade the range of values of the previous phase. The less overlap, the sharper the separation between the two conditions. High overlap means that a good part of the intervention sessions is indistinguishable from the baseline sessions.
- Consistency of patterns across similar phases — do all the baseline phases resemble one another? And all the intervention phases? When they do, the pattern is reliable. When the second baseline does not reproduce the first, something besides the intervention is at work.
The verdict: three demonstrations, three points in time
The six properties are not a checklist: they feed into a single conclusion. Under the WWC criteria, a demonstration of effect is considered to have occurred when the pattern of data in one phase differs from the pattern of the previous phase by more than would be expected from the simple continuation of that pattern. And a causal relation is considered documented when the set of phases gathers at least three demonstrations of the effect at three distinct points in time — something an AB design, with a single phase change, cannot offer.
In practice, the reading follows this order: (1) check whether the baseline establishes a predictable pattern; (2) describe what happens within each phase; (3) compare adjacent phases; (4) integrate all phases to count how many demonstrations exist.
There is a third possible verdict, and it is an honest one: "it's not possible to say yet." Highly variable data, phases that are too short, or a baseline that was already changing on its own can legitimately lead to this conclusion — which is not a failure, but an invitation to collect more data.
What the methodological standards sayThe technical documentation from the What Works Clearinghouse for single-case designs, coordinated by Kratochwill and colleagues, defines exactly the six characteristics examined in visual analysis — level, trend, variability, overlap, immediacy of effect, and consistency of patterns across similar phases — and establishes that a study needs to gather at least three attempts to demonstrate the effect at three different points in time. As for phase length, the most demanding criterion reserves the top rating for studies with five or more points per phase. None of this replaces clinical judgment, but it provides a common floor: it is the yardstick by which a published result is judged to support what it claims.
Where the reading slips
Visual analysis is conservative by design — small and inconsistent effects tend not to convince anyone. But it is only reliable if the graph is honest and the reader knows the pitfalls. Flip the cards:
The Achilles' heel: it's a human judgmentBecause the decision depends on who is looking, it is fair to ask whether two competent people would reach the same verdict. In 1979, DeProspero and Cohen sent 36 ABAB design figures to reviewers of behavioral journals and found a mean agreement of 0.61 among the judges — an uncomfortable result that became the classic critique of the method. Decades later, Kahng and colleagues (2010) repeated the procedure with members of the editorial board of the Journal of Applied Behavior Analysis and found high agreement. Read together, the two studies say less about the method and more about who applies it: visual analysis done by well-trained people is consistent; done by eyeballing it, it is not. Training, well-built graphs, and explicit criteria are not a luxury — they are the condition for the method to work.
Structured aids — and the limits of what a graph answers
To reduce the margin of subjectivity, structured criteria have been proposed that add explicit rules to graph inspection. The most used in Behavior Analysis are the dual-criteria (DC) and conservative dual-criteria (CDC) methods, described by Fisher, Kelley, and Lomas in 2003 as refinements of the older split-middle method. The logic is straightforward: a mean line and a trend line are drawn for the baseline and both are extended over the intervention phase; the number of intervention points that fall on the expected side of both lines is counted; and that number is compared to a critical value derived from the binomial distribution. In the CDC, the two lines are further shifted by 0.25 standard deviations in the direction of the expected effect, making the criterion harder to satisfy. In the original study, both methods controlled the error of "seeing an effect where there is none" better than the split-middle method did.
These aids do not replace visual analysis: they discipline it, offering a second pair of eyes with a fixed rule when the graph is ambiguous. The same goes for overlap indices and statistical models applied to single-case data — they are complements, useful above all when results need to be pooled with those of other studies in a systematic review.
Finally, it is worth separating two questions that the graph tends to blur together. Visual analysis answers whether the change is reliable and whether it can be attributed to the intervention. It does not answer whether the change matters to the person's life — whether the new level of behavior makes a real difference at school, at home, in relationships, and whether the people involved consider the result valuable. That second question has its own name: social validity. An impeccable graph with an irrelevant goal is still an impeccable graph — and a pointless intervention.
Test your understanding
Quick quiz
Three questions to lock in the concept. You see the explanation right away.
Further reading
- Cooper, J. O.; Heron, T. E.; Heward, W. L. Applied Behavior Analysis. 3rd ed., 2020 — graph construction and visual interpretation of behavioral data.
- Kratochwill, T. R. et al. Single-Case Designs Technical Documentation. What Works Clearinghouse / Institute of Education Sciences, 2010 — criteria for visual analysis and demonstration of effect.
- Fisher, W. W.; Kelley, M. E.; Lomas, J. E. "Visual aids and structured criteria for improving visual inspection and interpretation of single-case designs". Journal of Applied Behavior Analysis, 36(3), 387–406, 2003 — dual-criteria methods (DC and CDC).
- DeProspero, A.; Cohen, S. "Inconsistent visual analyses of intrasubject data". Journal of Applied Behavior Analysis, 12(4), 573–579, 1979 — the classic critique of agreement among judges.
- Kahng, S. et al. "Consistent visual analyses of intrasubject data". Journal of Applied Behavior Analysis, v. 43, 2010 — a replication with trained raters, with much more favorable results.
This entry is educational and does not replace assessment, diagnosis or treatment by a qualified health professional. Every person is unique — for a specific case, seek individual guidance.