A graph can look like a success story and still prove very little. Your job is to read single-case data like a careful skeptic: did behavior change, and did it change because of the intervention?
Why it matters
Single-case designs rely mainly on visual analysis, not statistical tests. Whether the graph comes from your caseload, a supervisee, or a published study, you decide whether to keep, change, or drop an intervention. A misread graph can keep a client in a treatment that isn't working or end one that is.
This task sits in Experimental Design (domain D, 13 of 175 scored questions). Practice items describe a graph and ask what you can conclude, what is wrong with the evidence, or what to do before changing phases.
Key ideas
- Look within each phase first. Before comparing phases, describe each one on its own terms: its level, its trend, and its variability.
- Then compare adjacent phases. Ask how level, trend, and variability shifted from one phase to the next, how quickly the shift appeared (immediacy), and how much the data paths overlap.
- Stable baselines make prediction possible. A baseline should be steady enough to predict what would happen without intervention. If baseline data are highly variable or trending in the direction you hope treatment will produce, a later "improvement" may simply continue what was already happening.
- Experimental control needs more than one change. A single shift from baseline to treatment shows a change, not a functional relation. Confidence grows when the effect is verified (for example, behavior returns toward baseline when treatment is withdrawn, or untreated tiers stay steady while one tier changes) and replicated (the effect appears again in another phase, tier, or condition).
- Consistency across similar phases counts. If two treatment phases look alike and two baseline phases look alike, the case for the intervention is stronger. Very different results in phases that should match weaken it.
- Check the quality of the data, not just the shape. Interobserver agreement, procedural fidelity, and whether only one variable changed at each phase change all affect what the graph can tell you.
- Experimental control and social significance are separate questions. A clear, well-controlled effect can still be too small to matter in the client's daily life, and a meaningful change can occur without a design that shows why it happened.
Definitions
- Visual analysis: systematically inspecting graphed data to decide whether behavior changed meaningfully and whether the change can be attributed to the independent variable.
- Level: the value on the vertical axis around which the data in a phase cluster, often summarized with a mean or median line.
- Trend: the overall direction of a data path (increasing, decreasing, or flat) and how steep it is.
- Variability: how much the data points bounce around from one measurement to the next within a phase.
- Immediacy of effect: how quickly behavior changes after a phase change, judged by comparing the last few points of one phase with the first few of the next.
- Overlap: the extent to which data points in one phase fall within the range of the adjacent phase. Less overlap generally means a clearer effect.
- Experimental control: a demonstration that changes in behavior reliably follow manipulation of the independent variable, shown through prediction, verification, and replication.
Examples from a session
- A baseline that is already improving. Lucía, a BCBA, plans a token system for Andre, age 9, who calls out in class. His baseline call-outs drop each day: 14, 11, 9, 7. If call-outs keep falling after treatment starts, she won't know whether the tokens did anything. She extends baseline until the data level off.
- A clear, immediate change. In a clinic, Wen, age 5, independently requests items in 0 to 2 opportunities per session across a stable baseline. Within two sessions of a new prompting procedure, she requests in 8 to 10 opportunities, and no treatment point falls in the baseline range. The change is large, immediate, and has no overlap.
- Lots of overlap. At home, Mr. Haddad records his son Sami's on-task homework time. Baseline ranges from 10 to 25 minutes per evening. During a new visual schedule, it ranges from 12 to 27 minutes. Almost every treatment point falls within the baseline range, so the graph doesn't support a convincing effect.
- A confound at the phase change. Danielle, age 34, receives coaching at a vocational site. Her task completion rises right when the job coach begins self-monitoring with her, but the same week she moves to a quieter workstation. Two variables changed at once, so the graph can't show which one mattered.
Example
When you describe a graph out loud, use a fixed order: level, trend, variability within each phase, then immediacy and overlap between phases, then whether the effect was verified and replicated.
Common mistakes / exam traps
- Changing phases during high variability. If you can't predict where the data are headed, you can't tell what the next phase did.
- Forgetting data quality. Low agreement between observers or poor treatment fidelity limits any conclusion, however clean the graph looks.
Exam trap
Change vs. functional relation. One AB comparison shows that behavior changed, not why; history and maturation remain open. Distrust options that say one phase change "proved" an effect. The discriminating feature is whether the effect was verified and replicated.
Exam trap
Treatment effect vs. continuing baseline trend. If baseline was already moving in the hoped-for direction, a treatment phase that keeps the same slope shows no clear effect. The discriminating feature is a change in level or slope at the phase change, not improvement alone; the fix is more baseline until the data stabilize.
Exam trap
Experimental control vs. social significance. A clean ABAB with verification and replication answers "did the intervention cause this?" It does not answer "is the change big enough to matter in the client's daily life?"
Quick check
A BCBA reviews a graph in which a client's baseline data are highly variable, ranging from near zero to the maximum possible score. What is the best interpretation before introducing treatment?
Related topics
- Interpreting graphed data (C.11)
- Threats to internal validity, such as history and maturation (D.3)
- What makes a design single-case: repeated measures, prediction, verification, replication (D.4)
- Reversal, multiple-baseline, multielement, and changing-criterion designs (D.7)
- Putting single-case designs to work (D.9)