More Practice Hours Linked to Bigger Math Gains in At-Risk Young Children

The study, published in the journal AI, Brain and Child, reanalyzed 171 effect sizes from 24 intensive mathematics intervention studies and found that intervention hours were strongly associated with learning gains, but only under specific conditions: when the samples contained only children at risk for mathematics learning disability and when the outcome measures tested the exact skills that had been taught. In that key stratum, the correlation between intervention hours and effect size was 0.52, a moderately strong association that held even when the analysis accounted for multiple effect sizes coming from the same study.

The analysis was conducted on a published meta-analytic database compiled by Miller and colleagues for intensive mathematics interventions in grades K through 3. Rather than extracting new data from primary studies, the researchers stratified the existing effect sizes along two dimensions. The first was sample composition: whether a study enrolled only at-risk children, usually identified by screening cutoffs at or below the 25th percentile on a mathematics measure, or also included not-at-risk peers. The second was outcome alignment: whether the test assessed what was actually taught, such as curriculum-based computation fluency probes, or broader achievement, such as standardized composites extending beyond the taught skill set. Nine effect sizes with unclear alignment coding were retained in the tables but excluded from alignment-specific interpretation.

The headline result emerged where both conditions converged. Across 53 effect sizes from at-risk samples assessed on skill-aligned outcomes, more intervention hours were associated with significantly larger gains, with the association visible across the studied dose range of roughly 30 to 40 hours without clear flattening. By contrast, in mixed or inclusive study contexts the overall correlation was essentially zero, at 0.05, and broad achievement measures failed to capture any stable positive dose-response pattern. The difference between the at-risk and mixed strata on skill-aligned outcomes was statistically significant, and a cluster-robust regression confirmed that outcome alignment significantly moderated the relationship. A large negative correlation in the mixed, broad-achievement cell was based on only eight effects from three studies and was not interpreted substantively.

Several sensitivity checks supported the central finding. When effect sizes were averaged within each study before correlating, eliminating within-study dependence entirely, the key-cell correlation strengthened to 0.76 across 13 studies. Bootstrap resampling at the study level, with 5,000 iterations, produced a 95 percent confidence interval of 0.21 to 0.78, excluding zero, and cluster-robust regression yielded a consistent slope estimate. Removing one extreme outlier reduced the key correlation to 0.40 but did not eliminate it, and a leave-one-study-out procedure kept the at-risk versus mixed contrast significant in 22 of 24 iterations. Restricting the analysis to the 162 effects with clearly coded alignment, matching the sample size of the original meta-regression, reproduced the key estimate exactly.

The shape of the dose-response curve also drew scrutiny. A curvilinear model fitted to the key stratum showed a linear relationship with no evidence of diminishing returns within the well-sampled range of up to 30 hours. An apparent acceleration above 30 hours, suggested by a significant squared-hours coefficient, was driven by just three effects from three studies and disappeared when those high-dose effects were excluded. The authors therefore concluded only that the pattern is consistent with the absence of diminishing returns, not that gains accelerate at higher doses. Domain-level moderation added further nuance: computation interventions showed a stronger association than word-problem interventions within the key cell, 0.62 versus 0.30, a difference the authors interpret descriptively, since computation practice maps most directly onto the skills that computation tests sample.

The theoretical backdrop comes from computational modeling. In a 2025 study, Strock and colleagues used personalized deep neural networks, or digital twins, calibrated to individual children, and found that networks calibrated to mathematics learning disability learned through mechanisms similar to typical networks but needed substantially more training to reach comparable performance, a difference of pace rather than kind. Imaging studies point the same direction: children with mathematics learning disability who receive intensive, aligned practice show brain activation and connectivity shifting toward typical patterns. The new correlational finding is qualitatively consistent with the slower-pace interpretation, but the authors are explicit that their analysis cannot determine whether children with mathematics learning disability need more practice to consolidate skills or face a ceiling on how far instruction can take them.

The practical stakes lie in how schools interpret progress-monitoring data. Curriculum-based measurement, a staple of multi-tiered systems of support, generates slopes that inform decisions about whether to intensify support or initiate comprehensive evaluation. A flat slope can mean several things: limited opportunity to learn, weak alignment between instruction and the probe, or genuine nonresponse despite adequate instruction. The new analysis cannot distinguish these mechanisms for any individual child, but it shows that a flat slope after limited or loosely aligned practice carries a different meaning than a flat slope after sustained, well-aligned intervention. The authors suggest that when a child’s slope is flat, schools should examine delivered minutes, session attendance, taught content, and probe alignment before treating the slope as nonresponse.

Existing data systems may already support this shift. Research by Maki and colleagues showed that growth-rate data are routinely collected in schools but rarely used in identification decisions, with demographic and achievement variables carrying far more weight. A 2026 school-based study by Newson and colleagues reported that intervention dose was likely below what the analysis could have detected, and that this under-dosing plausibly contributed to modest gains. The authors propose a modest addition to documentation: a brief implementation record at each progress-monitoring review covering minutes delivered, sessions attended, content taught, probe alignment, and interruptions, interpreted alongside the slope. They also flag an equity concern, since schools with fewer aligned materials and higher staff turnover may struggle to deliver and document consistent dose, risking that under-dosed children are flagged as nonresponders.

The authors extend these ideas toward closed-loop precision intervention, a research direction rather than a tested tool. The concept would pair computational estimates of a child’s learning pace with curriculum-based measurement data, re-fitting a child-level learning curve as each new probe arrives and flagging dose adequacy, content alignment, or fidelity when predicted and observed trajectories diverge. This would extend existing Data-Based Individualization practice, in which student response data guide intervention adjustments, by fitting and refitting learning curves online rather than reading slopes off fixed time windows. Realizing it requires prospective datasets pairing frequent, skill-aligned progress data with detailed records of intervention content, dosage, and fidelity, which retrospective meta-analyses cannot provide.

The limitations deserve equal weight. The findings are correlational and descriptive: studies were not randomized to dose, so unmeasured study-level differences could explain the association. Heterogeneity in the key stratum was high, with an I-squared of 86.7 percent, indicating the pooled correlation summarizes widely varying study contexts rather than a tight estimate. The effect size metric, a change-from-baseline measure dependent on control-group trajectories, cannot distinguish learning-rate from capacity accounts. Coding of sample composition was done by a single rater without inter-rater reliability data, and the mixed stratum aggregates studies with varying proportions of at-risk participants. Total hours are an incomplete dose measure, ignoring opportunities to respond, feedback rate, attendance, group size, spacing, and fidelity. The database covers K through 3 only, and its study pool reflects the original review’s inclusion criteria. The authors present the work as preliminary descriptive patterns that show dose and outcome alignment matter when progress-monitoring data are interpreted, not as evidence that settles competing accounts of mathematics learning disability.

Subject of Research: Dose-response patterns in mathematics interventions for at-risk kindergarten through third grade learners and their implications for progress monitoring

Article Title: Dose-response patterns in mathematics intervention for at-risk K–3 learners: implications for progress monitoring

Article References: Hite, J. (2026). Dose-response patterns in mathematics intervention for at-risk K–3 learners: implications for progress monitoring. AI, Brain and Child, 2(1), Article 11. https://doi.org/10.1007/s44436-026-00036-4

Image Credits: AI Generated

DOI: 10.1007/s44436-026-00036-4

Keywords: mathematics intervention, mathematics learning disability, dose-response, progress monitoring, curriculum-based measurement, MTSS, K-3 learners, meta-analysis reanalysis, skill-aligned outcomes, instructional dosage, digital twins, response to intervention