Matthew Effect in Science and Research

• The Matthew effect is a cumulative-advantage phenomenon where initial advantages amplify into greater success, applicable in fields like science, education, and networks.
• The effect is broader than preferential attachment, encompassing concepts like success-breeds-success and self-organization, and can be modeled through mechanisms like proportional growth.
• The impact in science includes citation networks, journal prestige, and collaboration, with evidence showing higher-impact journals and prominent scientists garnering more recognition for similar work.

The Matthew effect is a cumulative-advantage process in which an existing advantage increases the probability of receiving further advantage: “the rich get richer and the potent even more powerful.” The term was introduced by Robert K. Merton in relation to unequal recognition in science, where established scientists may receive greater credit for contributions comparable to those of less-established researchers. Related concepts include cumulative advantage, success-breeds-success, preferential attachment, proportional growth, and self-organization. The Matthew effect is broader than preferential attachment: preferential attachment is a particular measurable mechanism, whereas the Matthew effect denotes the more general tendency for status, visibility, resources, or success to reproduce themselves across time and interaction contexts (Perc, 2014).

1. Historical origins and conceptual foundations

Merton’s formulation draws on the biblical expression “to all those who have, more will be given.” In science, the mechanism describes the disproportionate attribution of recognition to researchers who are already prominent. A comparable discovery may receive different levels of credit depending on the prior reputation of its discoverer. De Solla Price related this logic to cumulative advantage in citation networks, while later network research connected it to preferential attachment.

The Matthew effect should be distinguished from a static inequality or a power-law distribution. A highly unequal distribution may result from many processes, and a power law alone does not establish preferential attachment or cumulative advantage. The defining feature is dynamic positive feedback: prior success affects subsequent opportunities, visibility, allocation, or growth. A useful abstract representation is:

$$\text{initial advantage} \longrightarrow \text{greater subsequent success} \longrightarrow \text{reinforced advantage}.$$

Related mechanisms differ in specificity. Cumulative advantage describes the accumulation of an initial differential over time. Success-breeds-success emphasizes the recurrence of successful outcomes. Preferential attachment specifies that new links or resources are allocated with probability related to existing connectivity or status. Proportional growth, including Gibrat-type processes, relates growth rates to current size. Self-organization describes the emergence of macroscopic inequality from repeated local interactions without centralized control.

The effect may arise from quality, efficacy, visibility, copying, homophily, institutional resources, or chance. In scientific citation, for example, an already prominent paper may be cited because it is genuinely relevant, because it is easier to discover, because its journal confers legitimacy, or because authors copy references from existing literature. Consequently, observed cumulative advantage does not by itself identify whether success reflects merit, reputation, network position, or stochastic reinforcement.

2. Measurement and dynamical mechanisms

For empirical network data, preferential attachment is commonly represented by an attachment kernel:

$\Delta x \sim A x^\gamma,$

where $x$ is an entity’s current amount of a resource, $\Delta x$ its increase over a short interval, $A$ a growth-rate coefficient, and $\gamma$ the attachment-kernel exponent. Linear attachment corresponds to $\gamma=1$, sublinear attachment to $0<\gamma<1$, and superlinear attachment to $\gamma>1$.

Linear attachment is associated with power-law degree distributions. Sublinear attachment produces more homogeneous networks and stretched-exponential cutoffs, while superlinear attachment can generate “superhubs” and potential monopolization. In citation systems, superlinear accumulation can produce citation “immortality.” Because attachment is stochastic, deterministic growth is often supplemented by a noise term:

${\rm d}x \sim \lambda\,{\rm d}t+\sigma\,{\rm d}W,$

where $\Delta x \sim A x^\gamma,$0 is the ensemble-average growth rate, $\Delta x \sim A x^\gamma,$1 is the standard deviation of the random component, and $\Delta x \sim A x^\gamma,$2 is a Wiener-process increment.

Two empirical estimators are prominent. Cumulation integrates growth over entities with current value at most $\Delta x \sim A x^\gamma,$3:

$\Delta x \sim A x^\gamma,$4

with $\Delta x \sim A x^\gamma,$5. Conditional averaging instead bins entities by current $\Delta x \sim A x^\gamma,$6 and estimates

$\Delta x \sim A x^\gamma,$7

then fits $\Delta x \sim A x^\gamma,$8. Time reversal, random permutation, event-level analysis, self-consistent estimators, and Markov-chain Monte Carlo methods are used to assess temporal ordering, missing timelines, and stochastic effects (Perc, 2014).

The distributional signature of cumulative advantage is often a heavy tail. A power-law density can be written as

$\Delta x \sim A x^\gamma,$9

with complementary cumulative distribution $x$0. Maximum-likelihood estimation, Kolmogorov–Smirnov tests, logarithmic visualization, and examination of deviations across at least two or three decades are recommended. Exponential alternatives and finite-size cutoffs must be considered because a visually linear log-log plot is insufficient evidence.

A model of cooperation–competition systems provides a distinct dynamical formulation. It defines normalized total cooperation sharing $x$1 as the fraction of a cooperation-generated resource attributed to element $x$2, with $x$3. Across 12 systems—including journal impact factors, authors, Olympic athletes, firms, counties, restaurants, and world languages—the empirical distributions are fitted by a shifted power law:

$x$4

Longer evolution durations are associated with lower $x$5 and greater heterogeneity. An analytic rank-amplification model,

$x$6

treats existing success as increasing subsequent gain. The model predicts an inverse-time relation, $x$7, and reproduces the transition from exponential-like to power-law-like distributions as evolution proceeds. The fitted effective amplification timescale is approximately $x$8 years, although the model is phenomenological and the dataset contains only 12 heterogeneous systems (Xu et al., 2011).

3. The Matthew effect in science and research evaluation

Scientific communication provides one of the clearest examples of cumulative advantage. Collaboration networks show that the probability of acquiring a new collaborator generally increases with the number of existing collaborators, although the relationship is often sublinear because physical, cognitive, and organizational limits constrain collaboration. Reported collaboration exponents include values near $x$9 in biomedical networks and values around $\Delta x$0 in mathematics, neuroscience, Slovenia’s collaboration network, and genetic-programming coauthorship. Temporal ordering and finite-resolution effects alter estimates (Perc, 2014).

Citation networks exhibit linear, slightly sublinear, and superlinear regimes. Estimates include approximately $\Delta x$1 for Physical Review Letters, approximately $\Delta x$2 for several journals, and $\Delta x$3 in a later analysis of 40,195 physics papers. Citation dynamics can be modeled as a self-exciting point process: recent citations increase the probability of additional citations, producing citation bursts and potentially divergent citation lifetimes.

Journal prestige independently contributes to citation reception. Larivière and Gingras identified 4,532 pairs of duplicate papers—9,064 papers in total—using identical titles, the same first author, and the same number of cited references. The higher-impact-factor versions averaged 11.90 citations, compared with 6.33 for the lower-impact versions; the medians were 2.74 and 1.60, respectively. The high-impact versions were uncited in 30.5% of cases, compared with 41.1% for their counterparts. All reported differences were statistically significant at $\Delta x$4, and the result remained after field normalization. This natural-experiment design indicates that journal location, visibility, prestige, and discoverability contribute citation advantage beyond the intrinsic content of the presumed duplicates (0908.3177).

The effect also operates through evaluation indicators. The $\Delta x$5 index assigns a scientist credit for papers in the scientist’s $\Delta x$6-core when that scientist is the “alpha author,” operationalized in the simulations as the coauthor with the highest $\Delta x$7-index. Agent-based simulations with 200 scientists, 20 periods, and repeated runs show that initially high-$\Delta x$8 scientists acquire widening $\Delta x$9 advantages. Citation boosts, productivity correlations, and strategic collaboration intensify the divergence. The simulated mechanism is structural rather than empirical: the index allocates evaluative benefit to the already most reputable collaborator and may therefore reinforce pre-existing status inequalities (Bornmann et al., 2019).

Large teams can amplify the contribution–recognition gap. A dataset of 1,600,627 LaTeX preprints authored by 2,012,092 scientists inferred writing contributions from author-specific macros. The contribution measure was validated against 469 self-reported records, with precision 0.87 and recall 0.71. A separate credit-allocation algorithm used co-citation structure to infer socially perceived author credit. In two-author teams, the first author contributed approximately 0.2% more than the last author but received approximately 8% less credit. In six-author teams, the first author contributed approximately 42% more but received approximately 13% less credit. Career age and last-author position were more strongly associated with recognition than primary contribution, and the career-age advantage increased with team size (Yang et al., 5 Feb 2025).

Research globalization demonstrates that cumulative advantage and convergence can coexist. Analysis of 56,084,649 scholarly articles and 1,710,566,975 citations across 218 countries and regions from 1980 through 2019 shows catch-up by upper-middle-income countries relative to high-income countries, alongside widening gaps between upper-middle-income, lower-middle-income, and low-income regions. International collaboration can connect weaker systems to established networks, providing expertise, infrastructure, visibility, training, and access to high-recognition communities. Thus, preferential attachment can be exclusionary when access remains concentrated but convergence-producing when substantial collaboration enables entry into dominant networks (Mishra et al., 2021).

4. Cross-domain manifestations

The Matthew effect is not restricted to bibliometrics. In public-goods games, social power can be accumulated locally or transferred across groups. Under the intra-group mechanism, an agent’s power in group $A$0 equals its payoff from that group; under the inter-group mechanism, power is aggregated across all groups. A softmax allocation rule converts power into shares of the public good. Inter-group accumulation promotes cooperation at low productivity but can preserve successful defectors at intermediate and high productivity. Intra-group accumulation is more favorable to cooperation at high productivity, particularly for moderate Matthew strength. Mixtures of the two mechanisms often harm cooperation, showing that the scope of accumulated advantage matters (Wang, 2023).

In inhibitory neuronal networks, the Matthew effect appears as an adaptive feedback between synchronization and synaptic plasticity. In a scale-free network of inhibitory Hindmarsh–Rose neurons, well-synchronized populations undergo inhibitory long-term depression, weakening inhibition and becoming still better synchronized. Poorly synchronized populations undergo inhibitory long-term potentiation, strengthening inhibition and becoming still less synchronized. The mechanism is therefore reversed relative to excitatory plasticity: inhibitory LTD improves synchronization, whereas inhibitory LTP suppresses it. At noise intensities below approximately $A$1, inhibitory LTD dominates; above that threshold and below the nonplastic synchronization threshold $A$2, inhibitory LTP dominates (Kim et al., 2018).

In computation contests, finite search without replacement causes powerful players to win more than proportionally to computational power. For two players with powers 2 and 1 and one solution, the winning probabilities are $A$3 and $A$4, rather than $A$5 and $A$6. More generally, if $A$7, then

$A$8

with equality only for equal powers. Reinvestment of rewards therefore produces compounding concentration. Increasing the number of valid solutions makes winning probabilities approach computational-power proportions, while reducing the number of solutions aggravates the effect. Pooling further encourages concentration and admits equilibria in which a pool or individual controls at least half of total computational power (Zeng et al., 2019).

Educational evidence is mixed. Analysis of approximately 321,544 Madrid students across Grades 3, 6, and 10 finds a persistent positive parental-environment effect in mathematics, literature, and English. However, age interactions differ by subject: parental influence decreases in mathematics, follows a bell-shaped pattern in literature, and increases steadily in English. The English result is the clearest educational Matthew effect, consistent with the greater dependence of foreign-language learning on exposure, family support, cultural attitudes, and opportunities for communication (Aymeric et al., 21 Oct 2025).

5. Contemporary machine-learning and algorithmic systems

Machine-learning systems can create Matthew effects through training dynamics, representation imbalance, optimization, and feedback.

In differentiable neural architecture search, DARTS trains candidate operations in a weight-sharing supernet. Operations with fewer parameters often learn useful functions earlier. Their architecture weights increase, while slower operations receive smaller mixture coefficients and progressively smaller weight and architecture gradients. This creates a positive-feedback loop:

$A$9

DropNAS addresses this effect through grouped operation dropout, $\gamma$0-adjust correction, and partial decay. With a default drop-path rate of $\gamma$1, it reports 2.26% CIFAR-10 test error and 16.39% CIFAR-100 test error, compared with 3.00% and 17.76% for DARTS. On one edge, the correlation between architecture parameters and stand-alone operation performance is reported as 0.902 for DropNAS and approximately 0.2 for DARTS (Hong et al., 2022).

Collaborative filtering converts popularity into similarity connectivity. Under Zipf-distributed item popularity, popular items have larger overlap probabilities because shared-item probabilities scale with $\gamma$2. User-based collaborative filtering therefore gives popular items disproportionate influence over user similarity. Item-based filtering can preserve this bias under L2 cosine similarity, where the simplified independent-click model yields

$\gamma$3

Popular users and items have larger similarity neighborhoods, whereas rare entities suffer from sparse evidence. The resulting mechanism is structural amplification within similarity computation, even without modeling a complete repeated-exposure feedback loop (Wang et al., 2019).

Popularity bias also affects offline reinforcement learning for recommendation. Conservative methods penalize poorly supported state–action pairs to control value overestimation. Because long-tail items are rarely present in logged data, they receive larger uncertainty penalties, leading to:

$\gamma$4

Debiased model-based Offline RL (DORL) adds a state-entropy correction to ordinary uncertainty penalties. On KuaiRec and KuaiRand, DORL reports higher trajectory-level reward, longer interaction trajectories, and lower majority-category domination than MOPO, including KuaiRec trajectory reward increasing from 11.427 to 20.494 and majority-category domination decreasing from 0.479 to 0.379 (Gao et al., 2023).

Recommender-system bias can also be addressed through Zipf Matrix Factorization, which adds an estimated popularity-concentration term to matrix-factorization training. In a MovieLens experiment with 610 users and 9,742 movies, the reported MAE was 0.823113 for Zipf MF and 0.843682 for vanilla MF. However, the Zipf statistic, its sign convention, and its interpretation are internally ambiguous, and the method was evaluated on one dataset against one principal baseline (Wang, 2021).

Federated learning can generate inter-client Matthew effects when data-rich clients dominate the aggregate objective. Anti-Matthew FL formulates equal client accuracy and equal client decision bias as a constrained multi-objective optimization problem. Its three-stage multi-gradient procedure first enforces individual bias limits, then cross-client equality, and finally Pareto-oriented refinement. On synthetic, Adult, eICU, and ACSPublicCoverage datasets, it reduces cross-client dispersion in accuracy and bias while sacrificing relatively small amounts of average accuracy (Gao et al., 2023).

Differential privacy has no universal direction with respect to cumulative inequality. In synthetic-data generation, PrivBayes generally produces a Robin Hood effect by reducing majority–minority size gaps, PATE-GAN generally produces a Matthew effect by increasing them, and DP-WGAN has mixed behavior. Downstream minority accuracy can deteriorate under either size pattern. Thus, privacy guarantees individual influence but do not guarantee preservation of subgroup proportions, utility, or fairness (Ganev et al., 2021).

Recent work identifies analogous effects in AI programming assistants and LVLM self-improvement. Across 3,011 algorithmic problems, popular programming languages—Python, C++, Java, and JavaScript—outperform less-popular languages by 34.2 to 49.6 percentage points in reported Pass@1 across five models. Less-popular languages show more compile errors, suggesting weaker syntactic and ecosystem support. The framework experiment reports fewer correction rounds for mainstream stacks than for niche or emerging stacks, although it does not establish a causal training-data mechanism (Gu et al., 27 Sep 2025).

In LVLM self-improvement, successful trajectories are disproportionately generated for easy queries and retained for subsequent supervised fine-tuning. Head queries consequently receive more training, while tail queries receive fewer examples and shorter reasoning trajectories. Threshold clipping, repeat-based padding, adaptive-weighted resampling, and guided resampling counteract this feedback. Across Qwen2-VL-7B-Instruct and InternVL2.5-4B, the methods improve visual-reasoning performance over vanilla self-improvement, with a reported average improvement of 3.86 points (Guo et al., 30 Oct 2025).

6. Critiques, limitations, and mitigation

A Matthew-effect claim requires more than observing that successful entities are unequal. The principal methodological challenge is separating intrinsic quality from accumulated recognition. Duplicate-paper comparisons improve identification but rely on imperfect metadata rules and ethically problematic duplicate publication. Observational network estimates are sensitive to temporal resolution, missing timelines, finite-size constraints, field definitions, and unobserved heterogeneity. Journal impact factor is a measure of journal-level citation impact, not intrinsic paper quality, and citation counts remain heavily skewed.

Causal interpretation is particularly difficult. Journal placement, author status, collaboration, institutional resources, and quality are not randomly assigned. A power-law distribution is consistent with preferential attachment but does not uniquely identify it. Similarly, high citation counts among ChatGPT-selected references are consistent with a Matthew effect but do not establish that generated references subsequently receive additional citations. The environmental-science study reports a median cited-paper citation count of 1184.5, a median publication year of 2010, and a concentration in prestigious journals including Nature, but it lacks a control group, causal design, and complete reference-level data (Petiska, 2023).

Mitigation strategies depend on the domain. Bibliometric evaluation can use field normalization, age-sensitive indicators, contributorship information, and citation prediction based on manuscript characteristics rather than author and journal prestige. A biomedical prediction study shows that citation counts can be predicted from manuscript-level variables while excluding author and journal information, with test MAE values of 1.005 for weighted-citation linear models and 25.676 for the complete negative-binomial citation model. This approach mitigates one prestige channel but does not measure intrinsic scientific quality or eliminate bias encoded in language, open access, references, topic choice, or manuscript resources (Balzer et al., 2024).

Algorithmic systems require explicit monitoring of concentration, exposure, coverage, subgroup utility, and tail performance. Accuracy metrics alone may improve while popularity concentration worsens. Recommenders can use popularity-aware normalization, long-tail coverage constraints, diversity objectives, side information, entropy corrections, and distributional regularization. Offline RL should evaluate trajectory-level reward and retention rather than only immediate clicks. Federated learning can constrain client-level accuracy and decision-bias dispersion. Synthetic-data systems should audit rare-class representation, intersectional coverage, subgroup-specific precision and recall, mutual information, and variability across privacy budgets.

The Matthew effect is therefore neither synonymous with unfairness nor reducible to merit. Some cumulative advantage reflects genuine quality, competence, efficacy, or useful network centrality. Other components arise from visibility, institutional prestige, historical inequality, copying, algorithmic ranking, finite search, and self-reinforcing exposure. Its empirical characterization requires identifying the feedback mechanism, testing alternative explanations, and distinguishing persistent advantage from widening advantage. Across science, education, networks, markets, biological systems, and machine learning, the central analytical question is the same: whether an initial difference merely persists, or whether the surrounding system converts it into progressively greater access to recognition, resources, performance, and future opportunity.