/ Single-Case Design Graphs

Overview › Visual analysis › Lesson 3

Lesson 3 of 7About 45 minutes

Design logic

Every design asks the same question: does behavior change when, and only when, the conditions change? Each design answers it differently. For each one, this lesson explains where its demonstrations come from, shows graphs that do and don't show experimental control, covers the exceptions and judgment calls analysts make, and ends with a knowledge check.

You will be able to

  • Explain where each design gets its prediction, verification, and replication.
  • Tell experimental control from its look-alikes in each design, including the accepted exceptions: a baseline trending away from the goal, a partial reversal, a separation that emerges after a few alternations, and behavior that meets a criterion without matching it exactly.
  • In a functional analysis, compare every test condition with the control condition, and explain why responding in the demand condition can be capped.
  • Answer the knowledge check for each design.

In this lesson: Before any design: trend and when to start · Part A: AB · Part B: Reversal · Part C: Multiple baseline · Part D: Alternating treatments · Part E: Functional analysis · Part F: Changing criterion

Before any design: trend and when to start

Every design begins with a prediction. Baseline data predict what behavior would do if nothing changed, and each later phase is judged against that prediction (Lesson 2). So the first judgment in any design is whether baseline predicts well enough to start the intervention. Usually that means stable data, at least 3 to 5 points, with no trend. One kind of trend is an exception.

The same rule applies inside every design below: when a phase or tier is trending, ask which way. Only a trend toward the change you expect can masquerade as an effect.

Part A: AB design

The logic. Baseline predicts; the intervention phase tests that prediction once. A change when the intervention starts is affirmation of the consequent, but nothing on the graph verifies it or replicates it. Anything else that changed at the same time, such as a new medication, a new teacher, or the end of the school year, could explain it. An AB graph can show change. It can't show experimental control. For standards and practice, see the AB design page.

Change, not control AB graph of percentage of intervals on task. Baseline data run from 22 to 30 percent over 6 sessions; intervention data run from 58 to 86 percent over 8 sessions.
One clear change. The level rises immediately, with no overlap. It is a promising result, but it happened once, so nothing rules out another event between sessions 6 and 7.
Change, not control AB graph of aggressive behaviors per session. Baseline rises from 6 to 13 over 5 sessions. In intervention, aggression drops to 7 in the first session and falls to 2 by the end.
Exception: a rising baseline. Aggression was getting worse (6 to 13), the opposite of what the intervention should do, so starting without waiting was right. The drop breaks the baseline's prediction sharply. It is still one demonstration.
No effect shown AB graph of percentage of intervals on task. Baseline rises steadily from 20 to 46 percent over 6 sessions, with a solid trend line fitted to it. The trend line continues as a dashed line through the intervention phase, where the data rise from 52 to 82 percent and stay within 1.5 points of that line.
A trend toward the goal. On-task behavior was already rising about 5 points per session. Extended into intervention (dashed), the baseline trend predicts every intervention point within 1.5 points, so the higher level can't be credited to the intervention.

Where AB graphs still help. In practice, an AB graph is a reasonable way to track one client's progress and decide whether to keep an intervention. Describe it as showing change. To show that the intervention caused the change, withdraw and reintroduce it (Part B), or stagger its start across three or more tiers (Part C).

Knowledge check: AB design

1. An AB graph shows self-injury dropping immediately and staying low, with no overlap between phases. What can you conclude?

  • The intervention caused the drop.
  • Behavior changed when the intervention began, but the graph can't show the intervention caused it.
  • Nothing, because a change that large is unlikely to be real.

2. Baseline aggression is rising: 6, 8, 9, 11, and 13 per session. The intervention is meant to reduce it. Should you wait for baseline to stabilize before starting?

  • Yes. A trending baseline can never predict.
  • No. The trend runs opposite to the change the intervention should produce, so you can start.
  • Yes. Wait until aggression peaks, then start.

3. Which change would let this AB study show experimental control?

  • Collect more intervention sessions.
  • Add a trend line and the dual-criteria lines.
  • Withdraw and reintroduce the intervention, or stagger its start across three or more tiers.

Part B: Reversal (ABAB) design

The logic. Each phase has a job. The first baseline predicts. The first intervention phase departs from the prediction (affirmation). Withdrawing the intervention tests the prediction again: if behavior goes back toward baseline, that verifies baseline would have continued without the intervention. Reintroducing it replicates the effect. Introduce, withdraw, and reintroduce give three demonstrations at three different times. For standards and practice, see the reversal design page.

What control can look like

Experimental control Reversal graph of percentage of intervals on task. Data are 25 to 35 percent in the first baseline, 60 to 85 percent in the first intervention phase, 30 to 45 percent in the second baseline, and 80 to 92 percent in the second intervention phase.
The textbook case. The level changes immediately at every phase change, adjacent phases don't overlap, and the two baselines and the two intervention phases look alike.
Experimental control Reversal graph of aggressive behaviors per session. The first baseline rises from 6 to 13. Aggression drops to 2 to 5 in the first intervention phase, rises again to 8 to 15 when the intervention is withdrawn, and drops to 1 to 4 when it is reintroduced.
Exception: aggression rising in baseline. Baseline was worsening, away from the goal, so starting the intervention without waiting was justified. Aggression dropped (mean 3), climbed back when the intervention was withdrawn (mean 11.2), and dropped again when it returned (mean 2.4). The rising baseline makes the case stronger, not weaker.
Control, with a partial reversal Reversal graph of percentage of intervals on task. The first baseline is 25 to 35 percent, the first intervention phase 70 to 85 percent, the second baseline 50 to 62 percent, and the second intervention phase 82 to 88 percent.
Judgment call: a partial reversal. When the intervention was withdrawn, on-task behavior fell from a mean of 79% to 55%, about halfway back to the first baseline's 30%, and it rose again (mean 85%) when the intervention returned. Each phase change moved behavior clearly in the predicted direction, which most analysts accept as verification. A full return is more convincing.

What it doesn't look like

No reversal Reversal graph of percentage of intervals on task. Data rise from 25 to 35 percent in baseline to 60 to 85 percent in the first intervention phase, then stay at 80 to 86 percent after the intervention is withdrawn and at 82 to 88 percent when it is reintroduced.
Behavior didn't return. In the second baseline, behavior stayed at intervention levels. There is no verification, so the first change could be history or maturation.
Not replicated Reversal graph of percentage of intervals on task. Data are 25 to 35 percent in the first baseline, rise to 60 to 85 percent in the first intervention phase, fall to 30 to 45 percent in the second baseline, and stay at 29 to 34 percent when the intervention is reintroduced.
Behavior didn't change again. The effect appeared and reversed, but didn't return with the intervention. That is two demonstrations, not three.
Baseline trend Reversal graph where percentage of intervals on task rises steadily from 20 to 84 percent across all four phases, about 3 to 4 points per session, with no jump at any phase change.
A trend toward the goal. Behavior rose steadily from the first session, and no phase change altered the data path. Compare the aggression example above, where the trend ran the other way.
Too variable Reversal graph where percentage of intervals on task swings between about 20 and 80 percent in every phase, so the four phases overlap almost completely.
No difference in level. The phases overlap almost completely. Variability alone doesn't rule out control, but here no phase is distinguishable from the next.

Other judgment calls. A behavior that won't reverse, such as a skill once learned, can't show control in this design, so use a multiple baseline instead. Withdrawing an effective treatment for a dangerous behavior raises ethical questions: keep the withdrawal short, or choose a design that doesn't need one. And when a few sessions pass before behavior shifts at a withdrawal, look at whether the change follows each phase change consistently. A delay that repeats at every change weakens the case less than one that appears only once.

Knowledge check: reversal design

1. Which phase provides verification?

  • The first baseline
  • The first intervention phase
  • The second baseline, after the intervention is withdrawn
  • The second intervention phase

2. Baseline aggression was rising. It dropped when the intervention began, rose when it was withdrawn, and dropped again when it returned. Does the rising first baseline undermine control?

  • Yes. Any baseline trend rules out control.
  • No. The trend ran away from the goal, so it strengthens the case.
  • It depends on whether the second baseline also rose.

3. In the second baseline, behavior moves most of the way back toward the first baseline's level but doesn't reach it. Does that rule out control?

  • Yes. Behavior must return all the way.
  • No. A clear change back toward baseline can still verify, though the evidence is weaker.
  • No. Reversals are never judged on how far behavior returns.

4. Why is a reversal design a poor choice for teaching a child to read sight words?

  • A learned skill doesn't reverse when teaching stops.
  • Reading data are too low to graph.
  • It needs more than four phases for skills.

Part C: Multiple baseline design

The logic. Three or more tiers, such as participants, behaviors, or settings, start baseline together, and the intervention starts in each at a different time. Read the graph two ways. Horizontally, each tier is its own AB comparison. Vertically, at each start, look down at the tiers still in baseline: they are the verification. If they stay where they were while the treated tier changes, time alone isn't producing the change. Each staggered start replicates the effect. For standards, multiple probe rules, and practice, see the multiple baseline design page.

What control can look like

Experimental control Three-tier multiple baseline graph of percentage correct. Baselines last 5, 8, and 11 sessions, with data between 7 and 23 percent. Each participant's data rise to 40 percent in the first intervention session and reach 94 to 99 percent, only after that participant's intervention begins.
Each tier changes only when it is treated. At every start, the tiers still in baseline stay flat, and each changes right after its own start: three demonstrations at three times.
Experimental control Three-tier multiple baseline graph of aggressive behaviors per session in the classroom, lunchroom, and playground. Classroom baseline is 9 to 12, lunchroom baseline rises from 5 to 12, and playground baseline is 7 to 10. Each setting drops to 0 to 6 only after its own intervention begins, at sessions 6, 9, and 12.
Exception: a tier worsening while it waits. Lunchroom aggression rose from 5 to 12 during its baseline, away from the goal. That trend can't explain the drop at session 9, so the tier still counts. A tier improving before its start would not.
Experimental control Multiple probe graph with three skill sets. All three are probed in sessions 1 to 3, the sets still in baseline are probed again when another set starts teaching, and each set gets three consecutive probes before its own teaching starts. Each set improves only after its teaching begins.
Multiple probe: the same logic with fewer baseline sessions. Probes at the start, at each other tier's start, and just before a tier's own start show that the waiting tiers didn't change.

What it doesn't look like

Untreated tiers changed Three-tier multiple baseline graph of percentage correct. All baselines are between 8 and 23 percent for the first 5 sessions. At session 6, when Participant 1 starts intervention, Participants 2 and 3 rise to about 50 percent while still in baseline and keep rising to 74 and 84 percent before their own interventions begin.
The vertical check fails. When Participant 1 started, Participants 2 and 3 improved while still in baseline. Something else changed for everyone, or the tiers weren't independent.
Only two demonstrations Three-tier multiple baseline graph of percentage correct. Participants 1 and 2 rise from below 25 percent to above 90 percent after their interventions begin. Participant 3 stays between 7 and 21 percent before and after intervention begins at session 12.
One tier didn't respond. Two demonstrations and a non-effect fall short of the three needed.
Baseline trend Three-tier multiple baseline graph of percentage correct. Every participant improves steadily from session 1, about 4 to 5 points per session, at the same rate before and after their own intervention begins. Participant 1 rises from 13 to 31 percent in baseline and from 35 to 90 percent in intervention, Participant 2 from 18 to 46 percent and from 51 to 88 percent, and Participant 3 from 7 to 58 percent and from 62 to 91 percent.
Every tier improving toward the goal. All three rose at the same rate before and after their starts, so the changes don't line up with the staggered starts.

Other judgment calls. If a tier starts improving before its own intervention, extend its baseline until it levels off rather than starting on schedule. Tiers should be similar enough to respond to the same intervention but independent enough that treating one doesn't change another; behaviors of the same person are the most likely to covary. And concurrent tiers, measured over the same calendar time, rule out outside events far better than participants enrolled at different times.

Knowledge check: multiple baseline design

1. When Participant 1 starts intervention, what should Participants 2 and 3 do?

  • Improve a little, showing the intervention is spreading.
  • Stay at their own baseline levels until their own interventions start.
  • Get worse, to show a contrast.

2. While it waits in baseline, the lunchroom tier's aggression rises from 5 to 12 per session. It drops when the lunchroom intervention starts. Can this tier count as a demonstration?

  • Yes. The trend ran away from the goal, so it can't explain the drop.
  • No. Every tier needs a stable baseline.
  • Only if the other tiers were also rising.

3. A tier begins improving steadily two sessions before its scheduled start. What should you do?

  • Start on schedule; the plan is set.
  • Hold its intervention and extend its baseline until the data level off.
  • Drop the tier from the study.

4. Why does a multiple baseline need at least three tiers?

  • Two tiers are enough if both change.
  • Each tier gives one demonstration, and control needs at least three at different times.
  • A graph with fewer tiers can't share an axis.

Part D: Alternating treatments design

The logic. Two or more conditions alternate rapidly, often session by session, in a counterbalanced order with a distinct cue for each. Each condition's data path predicts its next point, and every switch tests that prediction again, so each alternation is a small replication. Control is consistent separation: one path stays above or below the other through many alternations. The conditions are compared with each other, and with baseline if there is one. For standards and the adapted design for teaching skills, see the alternating treatments page.

Experimental control Alternating treatments graph of responses per minute. Baseline is 1.9 to 2.5 over 4 sessions. In the alternating phase, Treatment A falls from 1.2 to 0.2 while Treatment B stays between 1.6 and 2.0, with no overlap.
Clear, consistent separation. Every Treatment A point is below every Treatment B point, from the first alternation on.
Control, once it holds Alternating treatments graph of responses per minute. Baseline is 1.9 to 2.5 over 4 sessions. Treatment A and Treatment B are close for their first two sessions each (1.7 to 2.0), then Treatment A falls to 1.2, 0.8, 0.6, and 0.4 while Treatment B stays between 1.8 and 2.1.
Judgment call: separation that emerges. The treatments overlap for their first two sessions each, then Treatment A falls away and stays below for the remaining four. Early overlap is common while a participant learns which condition is in effect. Judge control from the alternations after the paths separate, and make sure enough of them remain to replicate the difference.
No difference Alternating treatments graph of responses per minute. Baseline is 1.9 to 2.5 over 4 sessions. In the alternating phase, Treatment A (1.2 to 1.9) and Treatment B (1.3 to 1.9) overlap and cross repeatedly.
Overlapping paths. The treatments cross repeatedly, so neither is shown to be more effective. Both are below baseline, but that drop happened once, when the alternating phase began: like an AB comparison, it can't be credited to either treatment.

Other judgment calls. If one condition's effect carries over into the next session, the paths converge and can hide a real difference. Counterbalance the order, so that order can't explain the difference, and give each condition its own cue. A skill can't be unlearned between sessions, so compare ways of teaching skills with the adapted design, which gives each procedure its own equated set of targets.

Knowledge check: alternating treatments design

1. What shows experimental control in an alternating treatments design?

  • Both treatments are lower than baseline.
  • One condition's path stays clearly above or below the other's across many alternations.
  • One condition has a higher mean.

2. The two paths overlap for the first two alternations, then separate and stay apart for the next four. What should you conclude?

  • No control, because the paths overlapped.
  • Control is possible, judged from the alternations after the paths separate.
  • Restart the comparison with a different order.

3. Why counterbalance the order of the conditions?

  • So the order of sessions can't explain the difference between conditions.
  • So each condition gets more sessions.
  • So the data are less variable.

Part E: Functional analysis (multielement design)

The logic. A multielement functional analysis rapidly alternates test conditions, each of which arranges one possible reinforcer after problem behavior, with a control condition (Iwata et al., 1982/1994).

Each test condition is compared with Play, in every cycle, and not with the other test conditions. A test condition that is consistently above Play shows that its reinforcer maintains the behavior; the analysis is then differentiated. Every condition consistently above Play counts, not only the highest one. Structured criteria for these comparisons draw their lines from the control condition for the same reason (Hagopian et al., 1997). For more graphs and practice, see the functional analysis section of the design guide.

Differentiated

Escape function Functional analysis graph of responses per minute with five cycles of five conditions. Demand is 2.5 to 3.6 in every cycle. Play, Attention, Tangible, and Alone all stay between 0 and 0.3 and overlap.
One condition stands out. Demand is above Play in all five cycles; Attention, Tangible, and Alone stay with Play.
Attention and escape Functional analysis graph of responses per minute with five cycles of five conditions. Attention is 2.6 to 3.4 and Demand is 1.2 to 1.6 in every cycle. Play stays between 0 and 0.2, and Tangible and Alone stay with it.
Two functions, not one. Attention (2.6 to 3.4) and Demand (1.2 to 1.6) are both above Play in every cycle. Demand is lower than Attention, but that doesn't make escape a lesser function, or no function: each is compared with Play. The behavior is maintained by both attention and escape.
Automatic reinforcement Functional analysis graph of responses per minute with five cycles of five conditions. Alone is 1.8 to 2.6 in every cycle. Play, Attention, Demand, and Tangible all stay between 0 and 0.3 and overlap.
Only Alone is elevated. Responding is highest when no one provides any consequence, and the social test conditions stay with Play. That points to automatic reinforcement.

Undifferentiated

Undifferentiated Functional analysis graph of responses per minute with five cycles of five conditions. Play, Attention, Demand, Tangible, and Alone all fall between 1.7 and 2.9 and overlap throughout.
High everywhere, including Play. No test condition is above the control, so no social function is shown. High responding in every condition can suggest automatic reinforcement; a series of extended alone sessions can test it.
Undifferentiated Functional analysis graph of responses per minute with five cycles of five conditions. Play, Attention, Demand, Tangible, and Alone all stay between 0 and 0.4 and overlap throughout.
Too little responding to compare. The behavior rarely occurred in any condition. Longer sessions, or test conditions closer to the situations where it happens, may be needed.

Why the demand condition can be capped

Rates in different test conditions aren't directly comparable, because each condition limits how often the behavior can occur. The demand condition limits it most. Each response ends the current task and starts a break with no demands, so responding can happen only as often as demands are presented. Suppose presenting a demand takes about 5 s and each response earns a 30-s break. Each demand, response, and break then takes at least 35 s, and a 10-min (600-s) session holds about 17 of them, however strong the escape function. Responding can't go much above 1.7 per minute. In the Attention condition, attention is brief, so the behavior can recur almost at once and its rate has far more room.

So a demand rate lower than the attention rate, as in the two-function example above, doesn't mean escape is the weaker function. What matters is whether Demand is consistently above Play. Some analysts also report the percentage of demands followed by problem behavior, which isn't limited by session length.

Other judgment calls. Look for consistency across cycles, not one high session: a single high Attention point among four at Play's level doesn't show a function. If Play itself is high, nothing can stand out against it, so check whether the control condition really removed the test reinforcers. And some behaviors can't occur in the Alone condition: aggression needs another person. Zero aggression in Alone says nothing about automatic reinforcement, which is why many analyses of aggression use an Ignore condition, with a person present who provides no consequences, instead.

Knowledge check: functional analysis

1. Attention is 2.6 to 3.4 responses per minute, Demand 1.2 to 1.6, and Play 0 to 0.2, each in every cycle. Which functions are shown?

  • Attention only, because it is the highest.
  • Attention and escape, because both are consistently above Play.
  • None, because the two test conditions differ.

2. Why is each test condition compared with Play rather than with the other test conditions?

  • Play is the control: it differs from each test condition only in that condition's reinforcer.
  • Play always has the lowest rate.
  • Play comes first in every cycle.

3. Why can responding in the Demand condition look lower than in Attention, even when escape is a strong reinforcer?

  • Escape is usually the weaker function.
  • Each response starts a demand-free break, so the number of demands presented caps how often it can occur.
  • Demand sessions are shorter.

4. Responding is high in every condition, including Play. What does the analysis show?

  • An attention function, because Attention is high.
  • It is undifferentiated; automatic reinforcement is possible and needs follow-up.
  • The data are invalid and should be discarded.

5. In a functional analysis of aggression, responding is zero in every Alone session. What does that tell you?

  • Aggression isn't maintained by automatic reinforcement.
  • Little: aggression can't occur with no one there, so an Ignore condition is a better test.
  • The whole analysis is undifferentiated.

Part F: Changing criterion design

The logic. After baseline, the intervention sets a performance criterion and changes it in steps. Each new criterion predicts a new level of behavior. When behavior moves to the new criterion promptly and stays near it until the next change, that step is a demonstration, and at least three criterion changes are needed. Subphases of different lengths, steps of different sizes, and a temporary step back down show that behavior follows the criterion, not the passage of time (Hartmann & Hall, 1976). For standards and practice, see the changing criterion page.

Experimental control Changing criterion graph of minutes of exercise. Baseline is 7 to 10 minutes. In each subphase, data stay within 1 minute of the criterion of 12, 20, 16, 26, and 30 minutes, including the temporary return to 16 minutes. The subphases last 3, 5, 3, 4, and 3 sessions.
Behavior tracks every step. Exercise moves to each criterion as soon as it changes, including the step back down to 16 minutes.
Control, above the criterion Changing criterion graph of minutes of exercise. Baseline is 7 to 10 minutes. In each subphase, data stay 2 to 3 minutes above the criterion of 12, 20, 16, 26, and 30 minutes, including the temporary return to 16 minutes.
Judgment call: meeting, not matching. Exercise runs 2 to 3 minutes above every criterion, but shifts with each change and levels off until the next one, including the drop to 16 minutes. Behavior doesn't have to hit the criterion exactly; it has to follow it.
Didn't follow the steps Changing criterion graph of minutes of exercise. Baseline is 7 to 10 minutes. In the first subphase, with a 12-minute criterion, behavior jumps to 28 to 31 minutes, and it stays between 28 and 32 minutes through every later criterion, including 16 minutes.
An immediate jump to the final level. Behavior improved, but not in step with the criterion, and it didn't drop when the criterion did.
Steady drift Changing criterion graph where minutes of exercise rise steadily, about 1.5 minutes per session, from 11 to 37 minutes across subphases with criteria of 12, 20, 16, 26, and 30 minutes. Behavior keeps rising to 23 to 26 minutes when the criterion drops to 16, and ends at 34 to 37 minutes, above the final 30-minute criterion.
Rising regardless of the criterion. Behavior climbs within every subphase instead of leveling off near each criterion, and keeps rising when the criterion drops.

Other judgment calls. If behavior moves to the next level before the criterion changes, the criterion isn't what moved it. The step back down is optional, but it is the strongest test: steady improvement over time can't explain a drop. And when the criterion is set as a range rather than a single value, judge whether behavior stays inside each new range.

Knowledge check: changing criterion design

1. What counts as one demonstration in a changing criterion design?

  • The change from baseline to the final criterion.
  • Each criterion change that behavior follows promptly and then levels off near.
  • Each session in which behavior meets the criterion.

2. Exercise runs 2 to 3 minutes above every criterion, but shifts with each change, including a step back down. Does the graph show control?

  • Yes. Behavior follows each criterion even though it doesn't match it exactly.
  • No. Behavior must match each criterion exactly.
  • No. The step back down makes the result unclear.

3. Why should subphases differ in length and steps differ in size?

  • To keep the participant from getting bored.
  • So steady improvement over time can't line up with the criterion changes by chance.
  • So each subphase has enough data points.

Next, structured visual aids and the dual-criteria method in Lesson 4. For mixed graphs from every design, use the practice at the end of Lesson 2.