Skip to content
Back to search results

Percentage of Non-Overlapping Data

Introduced by
Thomas E. Scruggs, Margo A. Mastropieri & Glendon Casto
Year
1987
Type
Nonoverlap effect-size index for single-case (single-subject) research
Family
Single-case effect-size / nonoverlap statistics
Font size

What does it do?

The percentage of non-overlapping data (PND) is a simple nonoverlap effect-size index for single-case (single-subject) research, introduced by Scruggs, Mastropieri, and Casto in 1987 to enable quantitative synthesis of studies that report data graphically rather than as group statistics.

PND quantifies how strongly an intervention shifted a behavior by computing the percentage of intervention-phase data points that exceed the single most extreme baseline data point in the direction of desired change. Because it requires only the graphed data and a ruler, PND became one of the most widely used effect sizes in meta-analyses of single-case research in special education, rehabilitation, and behavior analysis. Its very simplicity, however, brings well-documented weaknesses — a ceiling at 100 percent, extreme sensitivity to a single outlying baseline point, and blindness to trend — which motivated later nonoverlap-plus-trend indices such as Parker and colleagues' 2011 Tau-U. PND is best understood as a fast, interpretable, but coarse first look at single-case effect magnitude.

The intuition behind the method

Picture a single-case graph with a baseline phase on the left and an intervention phase on the right, and suppose you want the behavior to increase. The intuition behind PND is to ask a blunt question: how often did the intervention push the behavior above its previous best? You find the highest baseline point and draw a horizontal line through it; then you count how many intervention points rise above that line and express that as a percentage of all intervention points. If almost every intervention point clears the line, the intervention seems to have produced a level of behavior the baseline essentially never reached, which feels like a strong effect. If only half do, the phases overlap heavily and the effect looks weak. The appeal is that you can compute it by eye from any published graph, which is exactly why it spread; the catch is that hanging everything on the one most extreme baseline point makes the index fragile.

History

Quantitative synthesis of single-case research lagged behind group meta-analysis because single-case studies report time-series graphs rather than means and standard deviations. Scruggs, Mastropieri, and Casto addressed this in 1987 with the percentage of non-overlapping data, a nonoverlap index that could be read directly from published graphs, validating it as a tool for the quantitative synthesis of single-subject research. PND quickly became the dominant single-case effect size, but methodologists documented its sensitivity to outliers, its ceiling, and its neglect of trend. This prompted a generation of alternative nonoverlap indices, culminating in Parker and colleagues' 2011 Tau-U, which combined nonoverlap with a baseline-trend correction. PND thus remains historically central as the first practical single-case effect size while ceding ground to more robust successors.

When to use it

Use PND when you need a quick, transparent, and widely understood index of how much an intervention shifted a behavior in a single-case study, especially when the only available information is the published graph. It is most useful for rapid appraisal and for synthesizing many single-case studies in a meta-analysis where a common, easily computed metric is needed and where the direction of desired change is clear. PND is appropriate when baselines are relatively stable and free of extreme outliers, since the index depends on a single extreme baseline point. It is a poor choice when baselines show strong trend, when a single aberrant baseline observation would dominate the benchmark, or when effects are large enough to bump against the ceiling; in those situations a trend-aware nonoverlap index such as Tau-U or a regression-based effect size is preferable, and PND should at most be reported as a coarse complement.

Limitations and considerations

  • Hinges entirely on the single most extreme baseline point, so one outlier can drastically distort the index.
  • Has a hard ceiling at 100 percent that compresses and cannot distinguish among strong effects.
  • Ignores trend in both baseline and intervention phases, misjudging effects when slopes are present.
  • Lacks a known sampling distribution, so it does not support confidence intervals or formal significance testing.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Extremely simple and transparent — computable by eye from any published single-case graph without raw data.
  • Widely used and understood, making results comparable across decades of single-case meta-analyses.
  • Provides a single interpretable percentage with conventional effectiveness bands for quick appraisal.
  • Requires no distributional assumptions and works directly on the graphed time-series points.

Applications

  • Summarizing the magnitude of an intervention effect in an individual single-case study.
  • Meta-analyzing and comparing effectiveness across many single-subject studies in special education and rehabilitation.
  • Screening published graphs quickly when raw data are unavailable.
  • Communicating single-case results to practitioners through an interpretable effectiveness percentage and bands.
  • Serving as a baseline comparison against which improved nonoverlap indices such as Tau-U are benchmarked.

How it works

How it works

1

The procedure begins by identifying the single most extreme baseline data point in the direction of the intended change. If the goal is to increase the behavior, this is the highest baseline value; if the goal is to decrease it, the lowest baseline value. This extreme point becomes the benchmark line against which the intervention phase is judged, encoding the idea that a real effect should move the behavior beyond anything seen at baseline. Because the entire index pivots on this one observation, the analyst must determine it carefully and recognize immediately that an unusually high (or low) baseline point will set a demanding benchmark that depresses the resulting PND.

Formula
y^{*} = \max_{t\in A} y_t \ \text{(for an increase)}\quad\text{or}\quad y^{*} = \min_{t\in A} y_t \ \text{(for a decrease)}
2

Next the analyst counts how many intervention-phase points fall beyond that extreme baseline benchmark — strictly above it for a targeted increase, or strictly below it for a targeted decrease. These are the non-overlapping points, the intervention observations that the baseline never matched. Points that merely equal or fail to exceed the benchmark are not counted, since they overlap with the baseline range and therefore do not demonstrate a level of performance unique to the intervention. This count is the raw evidence of separation between phases that PND will turn into a percentage.

Formula
m = \#\{\,t\in B : y_t > y^{*}\,\}\ \text{(or } y_t < y^{*} \text{ for a targeted decrease)}
3

PND is then computed as the number of non-overlapping intervention points divided by the total number of intervention points, multiplied by one hundred. Formally PND equals the count of intervention points exceeding the most extreme baseline point over the number of intervention observations, times 100, yielding a value from 0 to 100 percent. The result is interpreted as the proportion of the intervention phase that surpassed the best the baseline ever achieved, a single number that summarizes phase separation and can be compared and averaged across studies in a meta-analysis.

Formula
\mathrm{PND} = \frac{\#\{\,y_t^{B} > \max y^{A}\,\}}{n_B}\times 100 = \frac{m}{n_B}\times 100
4

Scruggs and colleagues proposed interpretive bands to give the percentage practical meaning: PND above 90 percent is taken to indicate a very effective intervention, 70 to 90 percent an effective one, 50 to 70 percent questionable effectiveness, and below 50 percent an ineffective intervention. These conventional benchmarks let researchers and reviewers translate a raw nonoverlap percentage into a qualitative judgment and aggregate effectiveness across a body of single-case studies. The bands are heuristic rather than derived from a sampling distribution, so they support description and synthesis but should not be treated as formal significance thresholds.

Formula
\mathrm{PND}>90\ \text{very effective},\ 70\text{–}90\ \text{effective},\ 50\text{–}70\ \text{questionable},\ <50\ \text{ineffective}
5

Finally, responsible use of PND acknowledges its limitations and the indices designed to address them. Because PND is anchored to a single extreme baseline point, one outlier can collapse the index toward zero; because it is a percentage, it has a hard ceiling at 100 that compresses differences among strong effects; and because it counts points without modeling slope, it ignores baseline and intervention trend entirely. Parker and colleagues' Tau-U was developed to overcome these issues by combining nonoverlap between phases with a correction for baseline trend, producing a more robust, less ceiling-bound effect size. Reporting PND alongside or in favor of such successors gives a more defensible picture of single-case effect magnitude.

Formula
\text{ceiling at }100,\ \text{outlier-driven via }\max y^{A},\ \text{trend-blind} \Rightarrow \text{Tau-U} = \frac{S_{\text{nonoverlap}} - S_{A\text{trend}}}{n_A n_B}

Common pitfalls

  • Computing PND when the baseline contains an extreme outlier, which sets an artificially harsh benchmark and understates the effect.
  • Applying PND to data with a strong baseline trend, where a trend-blind index gives a misleading picture of intervention impact.
  • Treating the 90/70/50 effectiveness bands as statistical significance thresholds rather than descriptive heuristics.
  • Reporting PND alone for large effects near the ceiling instead of pairing it with a trend-aware index like Tau-U.

Frequently asked questions

How is the percentage of non-overlapping data calculated?

First identify the single most extreme baseline point in the direction of desired change — the highest baseline value if you want the behavior to increase, the lowest if you want it to decrease. Then count how many intervention-phase points fall beyond that benchmark, divide by the total number of intervention points, and multiply by 100. The result is the percentage of the intervention phase that surpassed the best the baseline ever achieved, interpreted with bands where above 90 percent is very effective and below 50 percent is ineffective.

Why is PND criticized despite being so popular?

PND rests on a single extreme baseline point, so one outlying baseline observation can drag the index down regardless of how strong the intervention was. It also has a ceiling at 100 percent that cannot distinguish among very large effects, and it ignores trend in both phases, so an improving baseline can masquerade as an intervention effect or mask one. These weaknesses, documented over decades, are why researchers increasingly prefer trend-aware nonoverlap indices such as Tau-U, which Parker and colleagues introduced specifically to address them.

What replaced or improved on PND?

Several nonoverlap indices were developed to address PND's shortcomings, and Tau-U from Parker and colleagues in 2011 is the most prominent. Tau-U combines the nonoverlap between baseline and intervention phases with a correction for baseline trend, which makes it more robust to outliers, less affected by the ceiling, and able to account for slope that PND ignores. Other alternatives include the improvement rate difference and percentage of data exceeding the median. PND is still reported for continuity and simplicity, but trend-aware indices are now generally preferred.

Related methods

Sources

  1. Scruggs, T. E., Mastropieri, M. A., & Casto, G. (1987). The quantitative synthesis of single-subject research: Methodology and validation. Remedial and Special Education, 8(2), 24-33.

    Open DOI reference
    Copied.
  2. Parker, R. I., Vannest, K. J., Davis, J. L., & Sauber, S. B. (2011). Combining nonoverlap and trend for single-case research: Tau-U. Behavior Therapy, 42(2), 284-299.

    Open DOI reference
    Copied.