P100 Scale: Normalized Percentile Ranking

• P100 Scale is a percentile-based method that linearly maps unique data values to a fixed 0–100 range, ensuring consistent endpoints.
• Variants like P100′ and P100″ refine the original approach by incorporating frequency weighting and interpolation for improved tie handling.
• The scale is widely applied in bibliometrics and effect-size analysis, enabling standardized comparisons in regression and machine learning.

The P100 scale is a percentile-based rating methodology originating in bibliometrics and extending to generalized effect-size analysis via normalized percentage scaling. Distinct from conventional percentiles or parametric standardization, P100 methods seek to map observed data onto a fixed [0,100] or [0,1] interval, ensuring interval comparability and interpretability across diverse measurement regimes. Various P100 variants are recognized: the original rank-by-unique-value (P100), the frequency-weighted refinement (P100′), and the interpolated percentile-bound version (P100″). Closely related is the percent of maximum possible (POMP) transformation and the percentage coefficient ($b_p$), which operationalize effect-size regression on percentage-normalized variables. P100 concepts interface with research on scientific impact, effect size scaling in the behavioral and social sciences, and modern machine learning normalization pipelines (Bornmann et al., 2013, Bornmann et al., 2013, Schreiber, 2014, Zhao et al., 18 Jul 2025).

1. Mathematical Definitions and Core Algorithms

The canonical P100 scale is defined as follows: given $N$ entities (papers, observations, etc.) with nonnegative scores (e.g., citations), extract the sequence of unique observed values $c_0 < c_1 < \dots < c_{i_{\max}}$. The P100 score for a value $c_i$ is

$\mathrm{P100}(c_i) = 100 \cdot \frac{i}{i_{\max}},$

where $i$ is the index of $c_i$ in the sorted unique list and $i_{\max}$ is the total number of unique ranks minus one. Unlike percentiles, which are based on empirical cumulative distributions, P100 distributes scale values linearly over the rank of unique scores, ensuring the lowest and highest observations always map to 0 and 100, respectively (Bornmann et al., 2013, Bornmann et al., 2013).

The P100′ variant refines this by incorporating the frequency of each observed value. For each $c_i$ with frequency $n_i$, a cumulative count $j_i = \sum_{i'=0}^{i-1} n_{i'}$ is assigned, yielding

$\mathrm{P100'}(c_i) = 100 \cdot \frac{j_i}{N-1}.$

This assignment ensures stability against tie-induced artifacts and aligns more closely with empirical percentiles (Schreiber, 2014, Bornmann et al., 2013).

Additionally, P100″ interpolates between lower and upper percentile bounds ("inverted-InCites" and Rousseau's percentiles) to create an indicator always confined to the standard percentile "uncertainty interval" (Schreiber, 2014):

$$\mathrm{P100''}(j_i) = 100 \cdot \frac{j_i}{N} + 100 \cdot \frac{n_i}{N} \cdot \frac{i}{i_{\max}}.$$

This approach avoids paradoxes that affect previous renditions when the composition of the leading citation group changes.

For general scale transformation, P100 or POMP (percent of maximum possible) is defined as:

$$\mathrm{P100}(x) = \frac{x-x_{\min}}{x_{\max} - x_{\min}} \times 100,$$

mapping any bounded variable to $[0,100]$ (Zhao et al., 18 Jul 2025). The corresponding unit-interval form ("percentage score," $ps$) is:

$$S_{p} = \frac{S_{w}-C_{n}}{C_{x}-C_{n}}, \quad S_{p} \in [0,1].$$

2. Historical Origin and Theoretical Motivation

The intellectual provenance of the P100 scale embeds it within a lineage of decimalization and normalization practices. Roman fiscal ledgers, Simon Stevin's La Thiende (1585), the European metrication movement, and standardized grading experiments at Yale and Harvard in the 1800s all contributed to the ascendancy of "[0,100]" percentage scaling for commensurate reporting and assessment (Zhao et al., 18 Jul 2025).

In applied research, Cohen et al. (1999) formalized the POMP index as a solution to the heterogeneity-of-units problem: any closed interval scale can be mapped to [0,100], enabling regression coefficients to be interpreted as "percent-point changes" in outcome for an increment of predictor (Zhao et al., 18 Jul 2025). The P100 rank metric emerged in bibliometrics to supply an anchor-fixed, tie-robust scale for citation comparisons (Bornmann et al., 2013, Bornmann et al., 2013).

3. Properties, Advantages, and Limitations

P100 scales guarantee several desirable properties for rating and impact analysis:

• Fixed endpoints: The minimum observed value always scores 0; the maximum always 100, regardless of dataset skew or distribution (Bornmann et al., 2013, Bornmann et al., 2013).
• Monotonicity: Higher scores receive equal or greater P100 values (strict or weak monotonicity, depending on ties).
• Tie handling: All observations sharing a value get identical P100 or P100′ scores, with no randomized or fractional assignment (Bornmann et al., 2013).
• Comparability: Scores are strictly comparable across datasets, fields, or time windows as long as the reference set definition is consistent (Bornmann et al., 2013).

However, limitations and counterintuitive behaviors have been exhaustively cataloged:

• Instability to rare value perturbations: The creation or deletion of a unique value, particularly in the high tail, redistributes all P100 scores except the extremes (Schreiber, 2014).
• Insensitivity to frequency profiles: If the set of unique values is identical across two distributions, P100 assignments are identical, regardless of underlying citation or frequency patterns (Schreiber, 2014).
• Longitudinal and cross-field anomalies: Identical citation counts can possess sharply different P100 values in different years or fields, owing to shifts in the unique value ladder (Schreiber, 2014, Bornmann et al., 2013).
• Gaming potential: Authors can select publication fields or journals to exploit sparse unique-value ladders for artificially high P100 ranks (Schreiber, 2014).
• Interpretation ambiguity: P100 values lack the direct cumulative meaning of percentiles; a P100 of 60 does not indicate 60% of observations are lower, but only encodes unique-value position (Bornmann et al., 2013).

P100′ and P100″ mitigate some of these artifacts by weighting according to cumulative frequency or interpolating within robust percentile intervals (Schreiber, 2014).

4. Comparison with Alternative Percentile-Based Approaches

P100 is distinguished from main percentile ranking schemes by its construction on unique value ranks, not cumulative distributions. The following table synthesizes key features:

Method	Tie Handling	Endpoints Fixed	Interpretation
P100	Same index, equal P100	Yes (0, 100)	Unique-rank position
P100′	Frequency-weighted	Yes (0, 100)	Cumulative underrank (≈percentile)
Hazen/SCImago	Averaged/secondary-key	Variable	Fraction ≤ focal value
InCites	Maximum rank in tie	Often No	"Optimistic" multi-category
CWTS	Fractional/aggregate	Class-specific	Probability in top x%

Empirical studies show P100 scale's stability and predictive power in bibliometric impact tracking is typically weaker in early windows than SCImago or InCites methods, due to coarse grading at the low end and sensitivity to rare-value modifications (Bornmann et al., 2013). P100 scales nevertheless provide unique advantages when strict anchoring and tie manageability are paramount.

5. Application to Effect Size Analysis and General Percentage Scaling

The generalized P100 transformation appears as POMP scaling in effect-size estimation, and as min–max normalization in data mining and machine learning. The canonical regression framework is:

$$\mathrm{P100}(Y) = a + b\,\mathrm{P100}(X) + \varepsilon,$$

with slope $b$ interpreted as the percent-point gain in $Y$ per percent-point increase in $X$ across a full 0–100 range (Zhao et al., 18 Jul 2025). In the 0–1 interval form, this is reported as the percentage coefficient $b_p$. Such normalization ensures that all features or outcomes are measured equitably, making coefficients directly comparable and stabilizing machine-learning optimization (Zhao et al., 18 Jul 2025).

6. Practical Implementation, Illustrative Examples, and Computation

Stepwise procedure for P100 assignment:

1. Collect the reference set and enumerate all unique values.
2. Index unique values ascendingly, assign equal rank index to all observations with a given value.
3. Compute P100 as $100 \cdot (i/i_{\max})$ for each (Bornmann et al., 2013).
4. For frequency-refined methods (P100′): compute cumulative frequency $j_i$, then $100 \cdot (j_i/(N-1))$, ensuring mean P100′ is 50.

Example (Bornmann et al., 2013):

For nine papers with citations: 0, 1, 2, 2, 4, 4, 5, 6, 8:

• Unique values: {0,1,2,4,5,6,8}, $i_{\max}=6$
• P100 values: 0, 16.7, 33.3, 50.0, 66.7, 83.3, 100
• P100′ assigns values per cumulative count, e.g., both papers with 2 citations receive 25.0.

In effect-size regression (Zhao et al., 18 Jul 2025):

Given variables on heterogeneous raw scales, they are mapped to $[0,100]$ via min–max normalization, enabling interpretable, unit-consistent regression:

$Y_\text{P100} = a + b_\text{POMP} X_\text{P100}$

with $b_\text{POMP}$ interpreted as the expected increase in $Y$’s percent-of-maximum for a one percent-of-maximum increase in $X$.

---

References:

• (Bornmann et al., 2013, Bornmann et al., 2013, Schreiber, 2014, Schreiber, 2014, Zhao et al., 18 Jul 2025)