Psychometric properties of digital progress monitoring in special education: a COSMIN-guided systematic review

Introduction

Progress monitoring and curriculum-based measurement

Curriculum-based measurement (CBM) was developed as a method for indexing academic competence and progress through repeated, standardized assessment of locally relevant skills (Deno, 1985; Deno and Fuchs, 1987). Two design features distinguish CBM from other forms of assessment: it is built for repeated administration over time, and its data are intended to inform instructional decisions (Fuchs and Fuchs, 2004). These features place CBM within the broader practice of formative assessment (Black and Wiliam, 2009). Across research traditions, this practice appears under partially overlapping labels, including progress monitoring, learning progress monitoring or learning progress assessment (in the German literature: Lernverlaufsdiagnostik; Förster and Souvignier, 2014), curriculum-based measurement, and formative or continuous assessment. These labels denote the same family of repeated, decision-oriented measurement procedures (Deno, 1985; Förster and Kuhn, 2023; Fuchs and Fuchs, 2004). In this review, progress monitoring serves as the umbrella term, with CBM as its most extensively researched operationalization; the major terminological variants were reflected in the search strategy (see Method). The two central terms are related but not interchangeable: progress monitoring denotes the practice of repeatedly measuring student learning to inform instructional decisions, regardless of the instrument used, whereas CBM denotes a specific standardized measurement technology within this practice, namely brief, curriculum-anchored indicator tasks with parallel forms developed for repeated administration (Deno, 1985; Fuchs and Fuchs, 2004). Progress monitoring therefore also encompasses non-CBM procedures such as direct behavior ratings or language probes. Throughout this manuscript, CBM is used when a statement refers specifically to CBM instruments or CBM research, and progress monitoring when it refers to the practice or to the whole family of instruments.

In special education, CBM operates within multi-tiered systems of support (MTSS) and response to intervention (RTI), where progress monitoring distinguishes responders from nonresponders and informs tier movement (Fuchs and Fuchs, 2009; Jimerson et al., 2016). One-time universal screening can produce high false-positive rates: in the first-grade research programs reviewed by Fuchs and Deshler (2007), approximately half of the students initially flagged as at risk recovered within the school year. Repeated progress-monitoring data reduced false positives by capturing growth across the year (Fuchs et al., 2012).

The link between progress-monitoring data and student outcomes is mediated by data-based decision making. Teachers who use CBM data to adjust instruction outperform those relying on their own assessment practices (Stecker et al., 2008), and meta-analytic evidence indicates moderate effects of data-based individualization (g = 0.37), which grow substantially with weekly coaching (g = 0.66) or collaborative support (g = 0.86; Jung et al., 2018). A recent meta-analysis confirms this implementation dependence at scale: learning progress monitoring shows a small positive overall effect on academic performance (g = 0.30) that increases substantially with weekly assessment frequency, performance feedback to students, and ongoing teacher consultation (A. Fuchs et al., 2025). However, even experienced teachers may struggle to translate CBM graphs into instructional change (Stecker, 2017), and scaling such practice requires ongoing technical support (Fuchs et al., 2021).

Progress monitoring beyond academic curriculum-based measurement

Progress monitoring extends beyond academic CBM, that is, CBM focusing on academic achievement outcomes such as reading, mathematics, and writing. Direct behavior rating (DBR) and screening tools such as the Social, Academic, and Emotional Behavior Risk Screener (SAEBRS) operationalize formative assessment for social-emotional and behavioral risk (Christ et al., 2009), and language-focused measures provide analogous functions for students with developmental language disorder. These instrument families share the same formative logic as academic CBM, but their psychometric evidence has developed in research programs of their own (for direct behavior rating, see Chafouleas, 2011) and is less well synthesized. A review that addresses progress monitoring in special education must therefore span academic, language, and behavioral constructs rather than focus on reading and math alone.

The digital transformation of progress monitoring

The digital transformation of progress monitoring extends beyond administration efficiency. Computer-adaptive testing tailors item difficulty in real time and can outperform fixed-form CBM in predicting state-test performance (Shapiro and Gebhardt, 2012), and rule-based item generation enables large-scale parallel forms required for frequent monitoring (Förster and Kuhn, 2023). Beyond computer-adaptive testing, ecological momentary assessment (EMA) and wearable sensors extend repeated measurement into ambulatory, high-frequency data collection (Shiffman et al., 2008). EMA methods are established in clinical and health research with children and adolescents (Heron et al., 2017), and their sampling logic, brief and repeated in situ assessments of behavior and affect, matches the logic of progress monitoring; whether they have been validated as formative progress-monitoring tools in school-based special education has not been examined by prior syntheses. Modality is not binary: instruments and platforms vary along separable dimensions of administration, scoring, and data integration, and many contemporary tools combine paper administration with computerized scoring or use platform-mediated reporting on top of traditional CBM probes. These dimensions define the four administration-modality categories used in this review (digital, computer-adaptive, mixed, and paper administration with computerized scoring; see Method). These innovations do not change the measurement properties that progress-monitoring tools must demonstrate, but they change the conditions under which those properties are established: adaptive item selection replaces fixed parallel forms, so that reliability and form equivalence must be shown at the level of the item bank and the selection algorithm rather than the printed form; automatically generated or automatically scored items require evidence that item parameters and scoring are stable across generated forms and devices; and evidence obtained for paper-based versions does not transfer automatically to digital administration. Digital instruments therefore require validation evidence of their own against the same measurement-property standards.

Prior evidence and research gaps

Prior reviews have established the technical adequacy of CBM-reading, with strong reliability and validity (Wayman et al., 2007; Yeo, 2010, 2011) and moderate diagnostic accuracy (sensitivity .80–.83, specificity .71–.73; Kilgus et al., 2014). For CBM-math, reliability is largely adequate, but criterion validity lags behind reading (Foegen et al., 2007). Decision rules for individual progress remain largely expert-based rather than empirically validated (Ardoin et al., 2013), and the most recent cross-domain meta-analysis on early-grade screening reports moderate pooled validity (r = 0.51) with very high heterogeneity (Keles et al., 2025). The effectiveness evidence summarized above (A. Fuchs et al., 2025; Jung et al., 2018) does not by itself establish the measurement quality of the instruments involved: the effects are observed whether or not the instruments are sound. The mechanism assumed to produce these effects, however, runs through measurement. Progress monitoring can improve outcomes only if the data reflect student learning precisely enough, and are sufficiently sensitive to change, to support the instructional decisions involved. Observed effects therefore suggest that the instruments studied cannot all be inadequate, but they do not show which instruments are adequate, for which decisions, and in which digital form. Whether the available digital instruments meet these measurement requirements is a psychometric question, and it is the question of the present review.

The prior evidence is also uneven in its linguistic accessibility. Digital progress monitoring has a national tradition that predates the digital transformation of CBM: the Dutch pupil-monitoring system (leerlingvolgsysteem) has provided standardized, repeated assessment of pupils' learning with computer-supported implementation since the 1990s (Vlug, 1997), and its instruments are evaluated within a national test-review system of its own, whose test reviews are published in Dutch (Evers et al., 2010). Whether validation evidence on the digital generation of these instruments has reached the English- or German-language literature, and is thus accessible to international syntheses, has not been examined.

Three limitations of the existing evidence base motivate the present review. First, prior reviews focus almost exclusively on CBM-reading, with limited synthesis of CBM-math, language, or behavior progress monitoring. Second, digital and computer-adaptive modalities are rarely separated from paper-based administration in existing syntheses. Third, classification accuracy and responsiveness are systematically underrepresented relative to reliability and validity, despite their decision-relevance for tiered intervention systems. A further open question, whether psychometric properties hold for special-educational-needs subgroups, can only be answered by studies that report such subgroups separately; the present review records their presence descriptively (see Data Items) but does not synthesize subgroup-specific evidence.

The COSMIN framework

The synthesis is structured by the COSMIN framework (COnsensus-based Standards for the selection of health Measurement INstruments), a methodology for systematic reviews of measurement instruments that provides three elements: a taxonomy of measurement properties, consensus-based standards for judging the methodological quality of validation studies (the COSMIN Risk of Bias checklist, organized in boxes, that is, one set of standards per measurement property), and criteria for good measurement properties (Mokkink et al., 2010; Mokkink et al., 2018; Prinsen et al., 2018). The taxonomy distinguishes three domains: reliability (comprising internal consistency, reliability in the narrow sense, and measurement error), validity (comprising content validity, criterion validity, and construct validity, the latter including structural validity, hypotheses testing, and cross-cultural validity or measurement invariance), and responsiveness. Among these properties, responsiveness (the capacity to detect change in response to intervention) is the one most directly tied to formative decision-making in tiered systems. A second body of decision-relevant evidence, classification accuracy (sensitivity, specificity, and related indices as the basis for screening decisions), is not a COSMIN measurement property in the strict sense: COSMIN subsumes comparisons against a reference standard under criterion validity, and classification evidence is conventionally appraised with the dedicated QUADAS-2 tool (Whiting et al., 2011). How the present review combines the COSMIN taxonomy with classification accuracy, and at which level of the taxonomy it synthesizes, is specified in the Method. Other potentially relevant measurement aspects, such as differential item functioning or the linguistic accessibility of instruments, are captured only insofar as included studies report them within the validity facets; this restriction is noted in the Limitations.

The present review

The present systematic review addresses these three limitations. Building on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020; Page et al., 2021) and on the COSMIN methodology for measurement-property reviews (Prinsen et al., 2018), it synthesizes the contemporary evidence on digital progress monitoring across seven pre-specified synthesis groups, whose rationale is outlined next.

The seven synthesis groups of this review follow the structure of the progress-monitoring literature rather than a strict factorial scheme: four groups are defined by construct domain, two by measurement paradigm, and one by assessment tradition. The construct-domain groups mirror the domains in which formative progress monitoring has an established research base. Academic CBM has developed along two major strands with separate instrument families and review traditions, reading (Wayman et al., 2007; Yeo, 2010, 2011) and mathematics (Foegen et al., 2007), which define G1 and G2; G1 is labeled CBM-reading after its dominant domain but also comprises the smaller written-expression and content-area CBM literatures that emerged from the same literacy-oriented tradition. Language progress monitoring (G3) is kept separate from reading because it targets oral language and speech-language constructs in students with developmental language disorder and draws on instrument families of its own (e.g., individual growth and development indicators), whereas reading is an academic skill domain with a CBM tradition of its own. Behavior progress monitoring (G4) covers the social-emotional and behavioral domain, referred to as behavior for brevity; its instruments, direct behavior ratings and behavior screeners, were developed to assess social, emotional, and behavioral risk jointly (Christ et al., 2009; Kilgus et al., 2016) and are therefore not divided further.

Two further groups are defined by measurement paradigm rather than construct domain, reflecting the digital transformation described above. Ambulatory assessment through ecological momentary assessment and wearables (G5) and computer-adaptive testing and learning analytics (G6) are measurement technologies that cut across construct domains; they were specified as separate groups so that the validation evidence for the technologies themselves remains visible instead of being scattered across the domain groups (domain-anchored validations of adaptive instruments remain in their domain groups; see Method). Finally, the Dutch pupil-monitoring tradition (G7; Cito LVS, Boom LVS) was pre-specified as a group of its own because, as outlined above, it represents a long-standing, nationally established infrastructure for pupil monitoring with computer-supported implementation from early on (Vlug, 1997), whose validation evidence is documented largely outside the Anglo-American and German literatures. For G5 and G7, the background reviewed above made it uncertain whether eligible evidence existed at all: the EMA-and-wearables paradigm is established in adjacent fields but has not been synthesized for school-based special education, and the Dutch tradition is mature but may not be accessible to the international literature. Pre-specifying both groups ensured that an absence of eligible evidence would be recorded as a finding about the evidence landscape rather than disappear among unclassified exclusions. The seven groups are not claimed to exhaust the domains in which progress monitoring exists; they cover the domains and paradigms for which validation literatures on digital instruments in special education could be expected. Because digital modality was an eligibility criterion, all groups comprise digital instruments, and the group labels omit this qualifier.

From the three limitations of the existing evidence base and from the structure of the synthesis groups, three research questions and a derived aim follow:

(1) How is psychometric quality distributed across the four synthesis domains (the three COSMIN measurement-property domains and classification accuracy) in the contemporary digital progress-monitoring literature for students with or at risk for special educational needs? (2) How does the evidence base vary across the seven pre-specified synthesis groups? (3) Which evidence gaps emerge when psychometric coverage and coefficient strength are considered jointly with the decision functions that the domains serve in tiered intervention systems? Building on the answers to these three questions, the review derives implications for assessment-driven intervention practice and for future validation research in special education.

Method

Protocol and registration

This review was registered retrospectively on OSF after completion of screening (https://osf.io/tzvs4). The full review protocol, including the pre-specified eligibility criteria structured by the PICOS scheme (population, intervention, comparison, outcomes, study design; Methley et al., 2014), the exclusion taxonomy, synthesis groups, and adjudication procedures, is available in the Supplementary Material. The review is reported in accordance with the PRISMA 2020 guidelines (Page et al., 2021).

Eligibility criteria

Studies were eligible if they met the following criteria, which follow the PICOS scheme with the measurement construct taking the place of the intervention component and without a comparator requirement. The population comprised students in grades 1–10 (ISCED levels 1 and 2) with or at risk for special educational needs in the areas of learning, language, or social-emotional development, including equivalent international categories such as learning disabilities, developmental language disorder, and emotional and behavioral disorders. The construct was a digital instrument designed for formative progress monitoring, that is, for repeated administration across at least two measurement time points in support of instructional decisions, covering academic skills (reading, math, writing) and/or behavior. This criterion refers to the instrument and its intended use, not to the design of the individual study: studies that evaluated an eligible progress-monitoring instrument on the basis of a single measurement occasion (e.g., concurrent-validity or classification studies) were retained because they contribute evidence on reliability, validity, and classification accuracy, but they cannot, by design, contribute responsiveness evidence (see Methodological Adequacy for Progress Monitoring and the longitudinal sensitivity analysis in the Results). The setting was a regular school, special school, or inclusive classroom, including school-adjacent contexts where school-anchored. Outcomes were at least one reported psychometric property aligned with the four synthesis domains of this review: reliability, validity, responsiveness, or classification accuracy (see Synthesis Methods). Eligible study types were empirical quantitative studies, including validation studies, classification studies, generalizability studies, item response theory analyses, and ROC analyses. Eligible publication formats were peer-reviewed journal articles, dissertations, and technical reports from established institutions (e.g., NCII Tools Charts, IES What Works Clearinghouse). Established institutions were operationalized as university-based research centers and federally funded centers with a formal, citable technical-report series (e.g., Behavioral Research and Teaching at the University of Oregon; the Florida Center for Reading Research); repositories such as the NCII tools charts and the What Works Clearinghouse served as sources for identifying candidate reports rather than as publishers. The number of technical reports retained under this criterion is reported in the Results. Studies were eligible in English or German and had to be published in or after 2000.

Digital modality was defined broadly to include computerized administration, computerized scoring, or platform-mediated data integration. Hybrid formats (paper administration with computerized scoring) were included as borderline cases and are tracked separately in the synthesis. Four primary administration-modality categories were coded: digital (fixed-form computerized administration), computer-adaptive (administration in which item selection adapts to student responses), mixed (digital and paper administration within the same study), and paper administration with computerized scoring. Modality was coded independently of synthesis-group assignment: a computer-adaptively administered mathematics instrument, for example, is coded computer-adaptive in modality but assigned to the mathematics group by construct domain (see Synthesis Methods).

Ten exclusion criteria were applied hierarchically. Duplicates (E10) were removed first via Zotero (DOI match and title fingerprint) and a retrospective first author-and-title cross-check. Four criteria were applied primarily at the title-and-abstract stage: wrong population (E1), wrong construct (E2), no digital modality (E3), and wrong setting (E4); clearly ineligible study types (E6, e.g., editorials and narrative reviews identifiable from the abstract) were also excluded at this stage. The remaining criteria were applied at the full-text stage: no psychometric data (E5), wrong study type (E6), wrong language (E7), outside the publication window (E8), and full text unavailable (E9).

Information sources

Seven databases were searched: ERIC, PsycInfo, PsycArticles, FIS Bildung, PSYNDEX, Scopus, and Web of Science. MEDLINE and ProQuest Dissertations were screened in a pilot search but excluded because of low topic-specific yield and substantial overlap with ERIC and PsycInfo. An initial scoping search was conducted in February 2026. The last search to update the data before publication was conducted on April 28, 2026.

Search strategy

The search strategy combined three concept blocks: (a) digital or computerized assessment, (b) formative progress monitoring or curriculum-based measurement, and (c) special educational needs or relevant SEN equivalents. Database-specific search strings, including controlled vocabulary terms (e.g., MeSH, APA Thesaurus) and field tags, are reported in full in the Supplementary Material. The search window was defined a priori as January 2000 to April 2026, with eligibility limited to publications in English or German. Both limits were applied as database filters in all seven databases (see Supplementary Material, Part B). The publication-year limit was additionally enforced after export (see Results), and the 23 records in other languages that had passed the database filters (1% of the 3,223 screened records) were all excluded at the title-and-abstract stage on substantive grounds, so that the language criterion E7 was not applied to any record at the full-text stage.

Selection process

Study selection followed a multi-rater pipeline with mandatory consensus and adjudication of all conflicts. Title-and-abstract screening was performed by three independent raters with a tiebreak consensus rule (at least two concordant votes; otherwise the record was flagged for adjudication). Full-text screening was performed by four independent raters with a 3-of-4 consensus rule and a mandatory citation clause. Inter-rater reliability was Fleiss’ κ = 0.856 in the title-and-abstract phase and Fleiss’ κ = 0.997 in the full-text phase. Reliability was computed across all records entering the respective stage, that is, 3,223 records in the title-and-abstract phase and 319 full texts in the full-text phase. All flagged conflicts were adjudicated manually across several rounds (February to April, 2026).

Data collection process

Data extraction was performed against a 30-field COSMIN-aligned schema. Each included study was independently extracted by two reviewers across 100% of records, working in separate files with no visibility of each other's entries until both passes were complete. Inter-extractor agreement was assessed through field-level percentage agreement for numerical fields (target ≥ 90%) and Cohen's kappa for categorical fields (modality, study type, synthesis cluster; target κ ≥ 0.70). Discrepancies between the two extractors were resolved through structured consensus discussion, with the external senior reviewer acting as arbiter for unresolved cases.

Data items

Extracted variables fell into four categories. Study identifiers included author, year, country, and publication type. Sample descriptors included sample size, grade level, special educational needs status, and presence of a separately analyzed SEN subgroup. Instrument descriptors included platform, subtest, language, content domain, administration modality (digital, computer-adaptive, mixed, paper with computerized scoring), number of measurement time points, and total observation span. Psychometric data were extracted across the four synthesis domains: reliability (up to four facets: internal consistency, test-retest, parallel forms, inter-rater), validity (up to four facets: content, construct, criterion, convergent or discriminant), responsiveness (up to two facets: growth sensitivity, slope reliability), and classification accuracy (up to five facets: sensitivity, specificity, PPV, NPV, AUC or ROC).

Risk of bias assessment

A formal risk-of-bias assessment was specified in the protocol but not applied to the full corpus of 106 studies, given the multidomain scope of the synthesis. As a structural alternative, the breadth of psychometric reporting per study (i.e., the number of facets reported within each synthesis domain; see Table 1) is used as a transparent indicator of methodological comprehensiveness.

Table 1. COSMIN evidence map: psychometric findings per study (N = 106).

Studies are sorted by synthesis group and then alphabetically by first author within group; ordering and instrument labels are identical to Table 2. The main score (★★★, ★★, ★) reflects the strongest reported numerical coefficient within each domain; the value in parentheses (n/facets) indicates the breadth of psychometric reporting. Reliability is scored across up to 4 facets (internal consistency, test-retest, parallel forms, inter-rater); validity across up to 4 facets (content, construct, criterion, convergent/discriminant); responsiveness across up to 2 facets (growth sensitivity, slope reliability); classification across up to 5 facets (sensitivity, specificity, PPV, NPV, AUC/ROC). Symbol coding: ★★★ = strong (coefficients ≥ .80; e.g., α, ICC, AUC, sensitivity, specificity); ★★ = moderate (.60 to .79); ★ = weak (< .60); ✓ = reported but not numerically quantifiable (e.g., “significant correlations” reported without coefficient); – = not reported. G, synthesis group (G1 = CBM-reading; G2 = CBM-math; G3 = language progress monitoring; G4 = direct behavior rating and digital behavioral observation; G6 = computer-adaptive testing and learning analytics). Across all 106 studies, reliability is reported in 71 studies (67%; 55% strong), validity in 91 studies (86%; 33% strong), responsiveness in 48 studies (45%; 8% strong), and classification accuracy in 38 studies (36%; 31% strong). Substantial gaps exist for classification accuracy in G2 (4 of 21, 19%) and G4 (2 of 9, 22%) and for responsiveness in G4 (2 of 9, 22%; 0% strong). These gaps represent core findings of the review. For instrument abbreviations, see Note for Table 2.

In addition, a reduced formal assessment was conducted for the 60 studies validating the five most frequently studied platforms (easyCBM, DIBELS/Acadience, Levumi, SAEBRS, STAR). Reliability, validity, and responsiveness evidence was rated at the box level with the COSMIN Risk of Bias checklist (Mokkink et al., 2018; very good/adequate/doubtful/inadequate, worst score counts), and classification-accuracy evidence was additionally appraised with QUADAS-2 (Whiting et al., 2011). For each study, two raters first identified the measurement properties reported and then rated the applicable COSMIN boxes (B4: internal consistency; B6: reliability; B7: measurement error; B8: criterion validity; B9: hypotheses testing for construct validity; B10: responsiveness) on the standard four-point scale, with the overall study rating following the worst-score-counts principle; for classification evidence, the four QUADAS-2 risk-of-bias domains (student selection, index test, reference standard, flow and timing) and the three applicability domains were rated as low, high, or unclear. Ratings were generated with the same multi-rater consensus procedure as the study-selection pipeline (see Selection Process and Supplementary Material A.6); every doubtful and inadequate rating was documented with a written justification, and the complete rating workbook is available in the OSF repository. Box-level results are reported in Supplementary Tables S4a,b. The ratings are used in two ways: descriptively in the Results section, where they qualify the domain-level findings, and analytically in a sensitivity analysis restricted to studies rated adequate or better (see Sensitivity Analyses).

Methodological adequacy for progress monitoring

Beyond the synthesis domains, two design features are critical for any study that aims to validate a progress-monitoring instrument: the number of measurement time points and the total observation span. These features determine whether a study can in principle estimate change scores, growth-curve parameters, or slope reliability. A single time point cannot inform responsiveness. Two time points can estimate change but not slope stability. Three or more time points are minimally required for slope-based analyses, and at least eight time points are commonly recommended for stable slope estimation in curriculum-based measurement (Christ et al., 2013).

Of the 106 included studies, 15 (14%) were single-time-point designs and could not contribute to responsiveness evidence by design. Sixteen studies (15%) used two time points. Sixty-five studies (61%) used three or more time points and were therefore slope-eligible in principle. Only 17 studies (16%) reached the eight-time-point threshold associated with stable slope estimation. Ten studies did not report the number of time points. The median across the corpus was 3, with a range from 1 to 80. These design features are reported per study in Table 2 and are interpreted in the Discussion as a structural caveat on the responsiveness findings.

Table 2. Characteristics of included studies (N = 106).

Studies are sorted by synthesis group and then alphabetically by first author within group. Instrument labels follow a unified “Platform: Subtest” convention; the platform name alone is used when the study did not validate a specific subtest. G, synthesis group (G1 = CBM-reading; G2 = CBM-math; G3 = language progress monitoring; G4 = direct behavior rating and digital behavioral observation; G6 = computer-adaptive testing and learning analytics). Mod., administration modality (D, digital; CAT, computer-adaptive testing; M, mixed; P, paper administration with computerized scoring). Lang., instrument language; Level, educational level (primary=grades 1–4; lower sec. = grades 5–10). SEN, special educational needs focus (LD, learning disabilities; SLCN, speech, language, and communication needs; SEBD, social, emotional, and behavioral difficulties; MTSS, multi-tiered system of support; mixed, multiple SEN categories; general, general student sample with at-risk subpopulation; other, other category). SEMH, social, emotional, and mental health. Asterisks (*) indicate studies that report a separately analyzed SEN subgroup. TPs, number of measurement time points; Span, total time span. Em dashes (—) indicate that the corresponding information was not reported. Instrument abbreviations: AIMSweb, Aimsweb (Pearson); ATLAS-PA, Assessment of Tier 2 Language and Speech – Phonological Awareness; aLPA, automated Learning Progress Assessment; aWE-CBM, automated Written Expression Curriculum-Based Measurement; CBM, curriculum-based measurement; CBM-R, CBM of oral reading; CCSS, Common Core State Standards; Com-PASMA, Computer-based Phonological Awareness and Spelling Monitoring Assessment; CORE, Computerized Oral Reading Evaluation; DBR, Direct Behavior Rating; DBR-MIS, DBR Multi-Item Scales; DBR-Self, DBR student self-report version; DIBELS, Dynamic Indicators of Basic Early Literacy Skills; DORF, DIBELS Oral Reading Fluency; ERI, Early Reading Intervention Mastery Checks; FABLe, Functional Assessment-Based Learning Evaluation; FAIR-FS, Florida Assessments for Instruction in Reading – Florida Standards; FAST, Formative Assessment System for Teachers; HD, Highly Decodable; IGDI, Individual Growth and Development Indicator; ISIP, Istation's Indicators of Progress; ISIP-ER, ISIP Early Reading; ITRF, Integrated Teacher Report Form; LNF, Letter Naming Fluency; LSF, Letter Sound Fluency; LVD, Lernverlaufsdiagnostik; LVD-M, LVD Mathematik; MCRC, Multiple-Choice Reading Comprehension; mySAEBRS, student self-report version of SAEBRS; NWF, Nonsense Word Fluency; ORF, Oral Reading Fluency; PRF, Passage Reading Fluency; PSF, Phoneme Segmentation Fluency; PSI, Phonics Screener for Intervention; SAEBRS, Social, Academic, and Emotional Behavior Risk Screener; SAEBRS-TRS, SAEBRS Teacher Rating Scale; SiL, Silben lesen Levumi; SinnL, Sinnverstehendes Lesen Levumi; SinnL-CAT/ZasM-CAT, computer-adaptive Levumi modules; SKeVA, Science Key Vocabulary Assessment; STAR, Renaissance STAR; TPRI, Texas Primary Reading Inventory; WE-CBM, Written Expression CBM; WIF, Word Identification Fluency; WoL, Wörter lesen Levumi.

Synthesis methods

Studies were grouped into pre-specified synthesis groups based on instrument category and content domain. Group G1 covered CBM-reading. Group G2 covered CBM-math. Group G3 covered language progress monitoring. Group G4 covered direct behavior rating and digital behavioral observation. Group G5 was reserved for ecological momentary assessment and wearable-based progress monitoring. Group G6 covered computer-adaptive testing and learning analytics. Group G7 was reserved for instruments of the Dutch assessment tradition (e.g., Cito LVS, Boom LVS). Group assignment followed content domain: validation studies of domain-anchored instruments were assigned to their content group even when the instrument was computer-adaptive (e.g., ISIP and FAIR-FS in G1; STAR Math growth estimation in Van Norman et al., 2017, and the Levumi CAT simulation in G2), whereas G6 was reserved for studies whose primary object was the adaptive-testing or learning-analytics methodology itself (e.g., decision rules, growth estimation, platform diagnostics). Synthesis was narrative, structured by synthesis domain within each group, and supported by structured evidence tables (Table 1 at the study level and Table 3 at the group level). For research question 3, evidence gaps were operationalized descriptively as group-by-domain cells that combine low coverage (any-finding reporting below 50%) with low coefficient strength (strong-coefficient proportion below 20%), weighted by the decision function of the domain: gaps in responsiveness and classification accuracy were treated as more consequential than gaps in reliability and validity because these two domains directly underpin progress and screening decisions in tiered systems. For each synthesis domain, two indicators were reported per study: the strongest numerical coefficient (categorized as strong, moderate, or weak based on a priori thresholds) and the breadth of psychometric reporting (number of facets covered). For responsiveness, the protocol additionally listed minimal detectable change as a third facet; because it was not reported by any included study (k = 0 of 106), responsiveness was scored on two facets (growth sensitivity and slope reliability), and minimal detectable change was retained in the extraction schema.

Table 3. Synthesis aggregates per group (N = 106).

k = number of included studies per synthesis group; the pre-specified groups G5 (EMA/wearables) and G7 (Dutch-tradition instruments) are not represented in the included evidence and therefore omitted from this table; Sample n is reported as range across studies and as total sum (study-level; the Jungjohann/Schurig cluster contributes its sample of n = 761 three times because three independent tool validations were conducted on the same cohort). TPs=number of measurement occasions; median and range across studies in the cluster. COSMIN Coverage is reported per domain as the percentage of studies in the group reporting any finding (★ + ✓) and the percentage reporting strong coefficients (★★★, strongest coefficient ≥ .80). Reli, reliability; Val, validity; Resp, responsiveness; Class, classification accuracy. Values are computed directly from Tables 1, 2.

The COSMIN framework was applied as a structuring heuristic for psychometric domains, adapted to educational measurement contexts. Although COSMIN was originally developed for patient-reported outcome measures, its measurement-property domains (reliability, validity, responsiveness) translate to repeated educational assessments, where the same psychometric questions arise. Evidence was synthesized at the level of these three domains rather than the individual properties, because the included studies rarely report the finer-grained properties separately and because the synthesis spans instrument families with heterogeneous reporting standards; the finer-grained facets (internal consistency, test-retest, parallel forms, and inter-rater reliability; content, construct, criterion, and convergent or discriminant validity) were retained in the extraction schema and in the breadth indicator. Classification accuracy was added as a fourth synthesis domain. It is not a COSMIN domain, but screening decisions in tiered systems rest on sensitivity, specificity, and related indices, and treating it as a separate domain keeps this decision-relevant evidence visible; its methodological quality was appraised with QUADAS-2 rather than with the COSMIN checklist (see Risk of Bias Assessment). Throughout the manuscript, the term synthesis domains refers to these four domains.

Coefficient strength was categorized a priori using a uniform threshold across domains: strong coefficients were defined as values at or above .80 (e.g., Cronbach's α, ICC, AUC, sensitivity, specificity), moderate coefficients as values between .60 and .79, and weak coefficients as values below .60. Findings reported in qualitative form (e.g., “significant correlations” without coefficient) were marked as reported but not numerically quantifiable. Two properties of this categorization should be made explicit. First, it classifies the magnitude of the strongest reported coefficient, not the methodological quality of the study that produced it; we therefore speak of coefficient strength rather than strength of evidence throughout, and conclusions about the strength of the evidence base combine coefficient strength with the risk-of-bias appraisal. Second, a uniform threshold favors comparability across domains but is stricter than domain-specific conventions for some properties: COSMIN-based criteria for good measurement properties accept reliability and criterion-validity coefficients from .70 upward (Prinsen et al., 2018), and correlation-based validity coefficients rarely reach .80 when criterion measures differ in construct breadth or measurement occasion. The uniform threshold was retained as the primary, pre-specified analysis; its consequences are quantified in a threshold sensitivity analysis that brackets the proportions that would result from a .70 threshold (see Sensitivity Analyses).

Sensitivity analyses

Four sensitivity analyses probed the robustness of the domain-level findings against the main eligibility and synthesis decisions. First, a strict-modality analysis restricted the corpus to studies whose primary administration modality was digital or computer-adaptive, excluding mixed and paper-with-computerized-scoring modalities. Second, a longitudinal analysis excluded the 15 single-time-point studies, restricting the corpus to studies whose design comprised at least two measurement occasions. Third, a risk-of-bias analysis restricted the rated core-platform subset to studies with an overall COSMIN rating of adequate or very good. Fourth, a threshold analysis re-expressed coefficient strength using the interval bounds available from the extraction categories: because coefficients were categorized as at or above .80 and as .60 to .79, the proportion of studies that would qualify as strong under a .70 threshold lies between the ≥ .80 proportion and the ≥ .60 proportion, and both bounds are reported. All sensitivity analyses were computed from the study-level evidence map (Table 1) and the design and modality descriptors (Table 2); results are reported at the end of the Results section and in Supplementary Table S5.

Reporting bias and certainty assessment

Reporting bias was addressed through three mechanisms. First, gray literature was eligible (technical reports, dissertations) to reduce publication bias against null or partial findings. Second, studies that reported no quantifiable coefficient for a given facet were tracked as such (“reported but not quantifiable”) rather than discarded, preserving evidence on incomplete reporting as a finding of the review. Third, the breadth-of-reporting indicator (n facets per domain) explicitly captures selective reporting at the study level.

Use of AI tools

Large language models (Anthropic Claude; OpenAI ChatGPT; DeepSeek-chat V3.2-Exp) were used to support multi-rater consensus decisions in the screening, extraction, and risk-of-bias rating pipeline and to verify extracted values. All AI-assisted decisions were verified in both directions (independent re-extraction and independent verification) and were subject to manual human adjudication of all conflicts.

Results

The search across seven databases yielded 4,319 records; removing pre-2000 publications (n = 404) and duplicates (n = 692) left 3,223 records for title-and-abstract screening, which excluded 2,888 records and forwarded 335 for full-text retrieval. Sixteen reports could not be retrieved, and 319 full texts were assessed for eligibility (2 after OCR reconstruction). Full-text screening yielded 126 provisional inclusions and 193 exclusions, reported here in the hierarchical order in which the criteria were applied (wrong population, E1: n = 30; wrong construct, E2: n = 31; no digital modality, E3: n = 7; wrong setting, E4: n = 1; no psychometric data, E5: n = 112; wrong study type, E6: n = 12); because the criteria were applied hierarchically, each count is conditional on the criteria preceding it. Five adjudication rounds removed a further 20 records (9 paper-based instruments, E3; 11 duplicate or overlapping records, E10). The final corpus comprised 106 unique studies, of which 29 are flagged borderline cases combining paper administration with computerized scoring. The cumulative sample size was 774,420 students (median study-level n = 459; see the corpus-level caveat at the end of this section). Screening procedures and consensus parameters are reported in the Method section; the full flow is shown in Figure 1.

Characteristics of the included studies

The 106 studies were distributed across five of the seven pre-specified synthesis groups. Group G1 (CBM-reading) contained 68 studies (64%). Group G2 (CBM-math) contained 21 studies (20%). Group G3 (language progress monitoring) contained 4 studies (4%). Group G4 (direct behavior rating and digital behavioral observation) contained 9 studies (8%). Group G5 (ecological momentary assessment and wearables) was empty: two candidate records reached full-text assessment and were excluded, one for wrong population (E1; a community sample of 4- to 10-year-olds without special educational needs; Milojevich et al., 2023) and one for absence of psychometric data (E5; a feasibility study with children with ADHD in a special-education pre-K and first-grade setting that reported no reliability, validity, responsiveness, or classification evidence; Singh et al., 2025). Group G6 (computer-adaptive testing and learning analytics) contained 4 studies (4%). Group G7 (Dutch-tradition instruments) was likewise empty: no candidate record entered the screening pool, because the Dutch-tradition search terms retrieved no English- or German-language validation study of these instruments and Dutch-language publications lay outside the language limits applied at the search stage (see Search Strategy and Findings by Synthesis Group).

With respect to administration modality, 88 studies (83%) used digital administration as the primary modality (61 in G1, 16 in G2, three in G3, and eight in G4). Eleven studies (10%) used computer-adaptive administration as the primary modality (three in G1, three in G2, one in G3, and four in G6); because modality is coded independently of group assignment (see Method), computer-adaptively administered instruments appear in the domain groups as well as in G6. Four studies (4%) used a mixed digital-and-paper administration (three in G1 and one in G2). Three studies (3%) used paper administration with computerized scoring (one each in G1, G2, and G4). Independently of this primary-modality coding, a modality audit during adjudication flagged 29 studies (27%) as borderline cases in which test administration remained (partly) paper-based while scoring and/or data integration were computerized; these cases lie predominantly within the digitally coded corpus.

Sample sizes ranged from 5 to 358,032 students. The largest single sample was reported by Good et al. (2009) on DIBELS Nonsense Word Fluency (n = 358,032). The second largest was reported by Alonzo and Anderson (2018c) on easyCBM MCRC (n = 180,617). The median number of measurement time points was 3, with a range from 1 to 80. The 80-time-point study was Weber et al. (2020), a high-frequency self-monitoring design with daily ratings over four weeks.

The corpus was geographically bimodal. U.S. studies accounted for 75 of 106 (71%). German studies accounted for 24 of 106 (23%). The remaining 7 studies came from Canada (2), the Netherlands, Finland, India, Jordan, and New Zealand (1 each); in addition, two U.S.-based studies examined Spanish-language versions of established instruments (Basaraba et al., 2022; Klingbeil et al., 2021). Of the 106 studies, 82 (77%) reported a separately analyzed subgroup of students with or at risk for special educational needs. Educational level varied. Thirty-five studies focused on the primary level (grades 1 to 4), 16 on the lower-secondary level (grades 5 to 10), and 28 spanned both levels; the remaining 27 studies either did not report grade level or used cross-grade samples. With respect to publication format, 67 studies (63%) were peer-reviewed journal articles, 21 (20%) were technical reports from university-based research centers or federally archived grantee submissions (19 from Behavioral Research and Teaching, University of Oregon; one from the Florida Center for Reading Research; one ERIC grantee submission), and 18 (17%) were dissertations or other academic documents (15 doctoral dissertations, one university monograph, one conference proceedings paper, one archived research report). Full study-level characteristics are presented in Table 2.

Risk of bias (core-platform studies)

For the five most frequently studied platforms (easyCBM, DIBELS/Acadience, Levumi, SAEBRS, STAR; k = 60), a reduced formal risk-of-bias assessment complemented the breadth indicator. Fifty-six studies were rated at the COSMIN box level (four were not assessable at box level; see Supplementary Table S4a): overall methodological quality was adequate for 28 studies and doubtful for 26, with one very good and one inadequate rating. Concerns concentrated in responsiveness designs (11 of 21 rated studies doubtful) and reliability designs (8 of 19 doubtful or inadequate), whereas criterion-validity designs were predominantly adequate or better (34 of 39). All 21 studies with classification evidence were appraised with QUADAS-2: 15 were rated at high and 6 at low overall risk of bias, with concerns most frequent in student selection (7 high), index-test conduct (7 high), and flow and timing (7 high). Complete ratings, including written justifications for all doubtful and inadequate ratings, are provided in Supplementary Tables S4a and S4b. These ratings are taken up in the domain-specific findings below and in the risk-of-bias sensitivity analysis at the end of this section.

Psychometric findings across synthesis domains

Coverage was uneven across domains and across synthesis groups. Reliability was reported in 71 of 106 studies (67%). Validity was reported in 91 studies (86%). Responsiveness was reported in 48 studies (45%). Classification accuracy was reported in 38 studies (36%).

Strong coefficients were defined a priori as values at or above .80 (see Synthesis Methods). The strong-coefficient proportions were 55% for reliability, 33% for validity, 8% for responsiveness, and 31% for classification accuracy. Validity is therefore the most broadly reported domain in absolute terms. Reliability is the domain in which strong values are most attained. Responsiveness is the domain that is most directly relevant to formative progress monitoring. It is also the least frequently reported and the least frequently strong. The full study-by-study evidence map is presented in Table 1. Table 3 aggregates the same information at the synthesis-group level. Figure 2 summarizes coverage and strong-coefficient proportions by synthesis group and synthesis domain.

Reliability

In G1 (CBM-reading), 44 of 68 studies (65%) reported at least one reliability facet, and 35 of 68 (51%) reported coefficients at or above .80. Strong internal-consistency or alternate-forms estimates were reported, for example, by Souvignier and Förster (2011) for quop, by Foorman et al. (2015) for FAIR-FS, and by Clemens et al. (2019) for a multi-measure early-literacy battery. Clemens et al. reported all four reliability facets.

G2 (CBM-math) showed similar breadth (14 of 21, 67%) but a higher strong-coefficient proportion (13 of 21, 62%). Notable contributions came from Sikora and Voß (2017) on CBM Arithmetik (3 of 4 facets), Vanauer et al. (2022) on LVD-M 2–4, and Anderson et al. (2022) on Levumi mental computation.

G4 (direct behavior rating) showed the highest proportional coverage of reliability (7 of 9, 78%) but a lower proportion of strong values (4 of 9, 44%). This pattern is consistent with the well-documented sensitivity of single-occasion direct behavior ratings to rater-related variance.

G3 (language progress monitoring) and G6 (computer-adaptive testing) showed high proportions both of coverage (3 of 4, 75%) and of strong estimates (3 of 4, 75%). The small group sizes limit the generalizability of these proportions.

Validity

Validity was the most uniformly reported domain across groups. Coverage reached at least 56% in every group and 100% in G3 and G6. In G1, 59 of 68 studies (87%) reported at least one validity facet. The most commonly addressed facet was criterion-related (concurrent or predictive) validity against external reading achievement measures. Strong criterion validity was reported, for example, by Foorman et al. (2015) for FAIR-FS (4 of 4 facets), by Hautala et al. (2020) for a game-based reading assessment (4 of 4 facets), and by Misquitta et al. (2023) for FABLe in an Indian primary-school sample.

In G2, 19 of 21 studies (90%) reported a validity facet. Breadth-leading contributions came from Sikora and Voß (2017) and Vanauer et al. (2022), each reporting 3 to 4 facets.

The proportion of strong validity coefficients was substantially lower than the proportion of any-evidence reporting in every group. In G1, 87% reported any validity evidence, but only 34% reported strong coefficients. In G2, the corresponding proportions were 90% and 33%. This pattern reflects the well-known tendency of correlation-based criterion coefficients to fall in the .50 to .79 range. Coefficients at or above .80 are particularly hard to attain when external criteria differ in construct breadth or measurement occasion.

Responsiveness

Responsiveness was the least frequently reported domain. Coverage ranged from 22% in G4 to 62% in G2. The proportion of studies reporting strong responsiveness coefficients was 7% in G1, 14% in G2, and 0% in G3, G4, and G6.

Strong responsiveness coefficients were concentrated in a small number of studies that explicitly modeled growth-curve parameters or slope reliability. In G1, Tindal et al. (2009), Lai et al. (2010), Patarapichayatham et al. (2011), Christ et al. (2013), and Nese (2022) each reported both responsiveness facets at strong levels. In G2, Anderson et al. (2010), Gebhardt et al. (2015), and Nese et al. (2018a) reported strong responsiveness coefficients on both facets, using easyCBM Math and quop. Nese et al. (2010) reported moderate coefficients on both facets. The systematic underreporting of responsiveness is a core finding of this review. The risk-of-bias appraisal points in the same direction: among the rated core-platform studies, responsiveness designs showed the highest concentration of methodological concerns (11 of 21 rated studies doubtful; Supplementary Table S4a), so the small body of responsiveness evidence is, on average, also the least methodologically secure. The pattern is most pronounced in G4, where 22% of studies report any responsiveness evidence and 0% report strong coefficients. Responsiveness is the psychometric property most directly aligned with formative decision-making, the purpose for which the measures in this review were designed.

Classification accuracy

Classification accuracy is the basis for screening and risk-identification decisions in MTSS-aligned systems. It was reported in 38 of 106 studies (36%) overall.

Within G1 (reading), 29 of 68 studies (43%) reported classification facets, and 25 of 68 (37%) reported strong coefficients. Thirteen studies achieved the maximum of five facets (sensitivity, specificity, PPV, NPV, and AUC/ROC).

In G2 (math), only 4 of 21 studies (19%) reported any classification evidence. In G4 (behavior), only 2 of 9 studies (22%) did so. Both findings are striking given the central role of these measures in screening pipelines.

In contrast, G6 (computer-adaptive testing and learning analytics) showed the highest classification-accuracy reporting rate. Two of four studies (50%) reported classification facets, and both reported strong coefficients. This rate is consistent with the operational orientation of the included STAR-platform studies (e.g., Ochs et al., 2020; Van Norman and Forcht, 2024). The math and behavior gaps are substantive evidence gaps rather than methodological artifacts. The pattern is consistent across measurement modalities and platforms. The QUADAS-2 appraisal adds a quality caveat to the classification evidence overall: 15 of the 21 rated classification studies were at high overall risk of bias, most often for student selection, index-test conduct, or flow and timing. All six low-risk studies, in contrast, reported strong classification coefficients (Supplementary Table S4b).

Findings by synthesis group

G1: CBM-reading

G1 was the largest and most psychometrically mature synthesis group (k = 68; Σ N = 708,216). The dominant instruments were easyCBM (18 studies), DIBELS family measures (15 studies), generic CBM-reading variants (7 studies), and the German-language Levumi platform (7 studies). U.S. studies dominated (k = 53), followed by German (k = 9) and Canadian (k = 2) contributions.

Validity coverage reached 87%. Reliability coverage reached 65% and was strong in 51% of studies. Responsiveness coverage and classification-accuracy coverage were both 43%. Neither domain reached high proportions of strong coefficients (7% and 37%, respectively).

Notable comprehensive validations within G1 included Foorman et al. (2015), Merrill (2018), Hautala et al. (2020), and Matta et al. (2022, 2025). German-language contributions came from the Levumi and quop platforms. These included Souvignier and Förster (2011), Jungjohann et al. (2018a, 2018b, 2023), Schurig et al. (2021), and Jungjohann (2022). Three Levumi studies used the same n = 761 third-grade sample. They are reported as dependent samples.

G2: CBM-math

G2 (k = 21; Σ N = 21,869) was the second-largest group. It was almost evenly split between U.S. studies (k = 10) and German studies (k = 11). Reliability coverage reached 67% and validity coverage reached 90%. These proportions are comparable to G1. Responsiveness coverage reached 62%, the highest of all groups.

Classification-accuracy reporting was substantially lower than in G1. Only 4 of 21 studies (19%) reported any classification evidence. The four reporting studies were Nese et al. (2010), Clarke et al. (2011), Killen (2013), and Peltier et al. (2022). Three of them achieved high facet counts.

The most psychometrically comprehensive contributions in G2 were Sikora and Voß (2017) on CBM Arithmetik, Vanauer et al. (2022) on LVD-M 2–4, and Anderson et al. (2022) on Levumi mental computation. The under-reporting of classification accuracy in math is a substantive evidence gap with practical consequences for math screening in MTSS frameworks.

G3: language progress monitoring

G3 (k = 4; Σ N = 4,429) is the most evidence-thin group in this review. The four eligible studies were Tanabe (2013), Carson et al. (2014), Skibbe et al. (2020), and Besca et al. (2023). They span a heterogeneous instrument set (Picture Naming IGDI, Com-PASMA, ATLAS-PA, and a German-language progress-monitoring tool for sentence repetition and related constructs) and a heterogeneous SEN-subgroup composition. Most samples are speech-language-and-communication-needs samples.

Despite the small group size, validity coverage reached 100% and reliability coverage reached 75%. Only one study (Carson et al., 2014) reported classification-accuracy evidence at the 4-of-5-facet level. No study reported strong responsiveness coefficients. The G3 evidence gap is a research desideratum, particularly given the centrality of formative language progress monitoring to early intervention in students with developmental language disorder.

G4: direct behavior rating and digital behavioral observation

G4 (k = 9; Σ N = 29,214) was characterized by a small number of dominant instruments. SAEBRS and its variants accounted for five studies: Kilgus et al. (2016, 2018), Moulton et al. (2019), Whitley and Cuenca-Carlino (2020), and Renshaw et al. (2025). The Direct Behavior Rating family accounted for two studies: Casale et al. (2015, 2017). The remaining two studies were a German DBR instrument and the mySAEBRS large-scale validation.

Reliability was the best-reported domain (7 of 9, 78%; 4 of 9, 44% strong). This pattern is consistent with the methodological centrality of inter-rater and internal-consistency analyses for behavior-rating measures. Validity coverage was the lowest of any group (5 of 9, 56%). Both responsiveness (2 of 9, 22%; 0% strong) and classification accuracy (2 of 9, 22%; 1 of 9, 11% strong) showed substantial gaps.

The single G4 study reporting all five classification-accuracy facets was Kilgus et al. (2018) on SAEBRS in a middle-school sample (n = 1,243). The combination of low responsiveness reporting and low classification-accuracy reporting in G4 is a central finding of this review. G4 is the synthesis group most directly relevant to socio-emotional and behavioral risk identification.

G5: ecological momentary assessment and wearables

No study meeting the inclusion criteria was identified for G5. The eligibility criteria covered smartphone-based ecological momentary assessment, ambulatory assessment, and wearable-based progress monitoring within school-based formative progress-monitoring contexts for students with or at risk for special educational needs. Two candidate records reached full-text assessment. Milojevich et al. (2023) piloted an EMA smartphone application with 4- to 10-year-old children, but in a community sample that excluded children with developmental disorders (wrong population, E1); Singh et al. (2025) tested the feasibility of wearable- and EMA-based prediction of emotional outbursts among children with ADHD in a special-education pre-K and first-grade setting, but reported no reliability, validity, responsiveness, or classification evidence (no psychometric data, E5). A keyword audit of the screened records identified 288 further records referring to ecological momentary assessment, experience sampling, ambulatory assessment, or wearable sensors; all were excluded at the title-and-abstract stage, 251 of them for wrong population (E1; predominantly clinical or adult samples). The empty G5 group is reported here as a substantive null finding. At the time of the search (April 28, 2026), the EMA-and-wearables paradigm had not yet been operationalized as a formative progress-monitoring tool within the special-education school context covered by this review.

G6: computer-adaptive testing and learning analytics

G6 (k = 4; Σ N = 10,692) consisted of four STAR-platform studies: Mickelson (2017), Ochs et al. (2020), and Van Norman and Forcht (2023, 2024). Under the assignment rule described in the Method section, domain-anchored validations of adaptive measures remained in their content groups (e.g., Van Norman et al., 2017, on STAR Math growth estimation in G2), so G6 comprises the studies whose primary object was the adaptive-testing or learning-analytics methodology itself. The composition reflects the dominance of Renaissance STAR Reading and STAR Math in the operational learning-analytics literature for U.S. K to 12 contexts.

Reliability was reported in 75% of studies and validity in 100% of studies. Classification accuracy was reported in 50% of studies, the highest proportion across all groups. Both reporting studies achieved 4-of-5 or 5-of-5 facet coverage (Ochs et al., 2020; Van Norman and Forcht, 2024). Responsiveness was not reported at strong levels in any G6 study. This is striking given the platform-level capacity for fine-grained slope estimation that computer-adaptive testing should in principle afford.

The narrow instrument coverage of G6 is itself a finding. No included study validated an alternative computer-adaptive progress-monitoring platform within the inclusion period.

G7: Dutch-tradition instruments

No study was identified for G7. The search strategy contained the instrument names of the Dutch pupil-monitoring tradition (Cito, Leerling in beeld, Boom LVS, Snappet, Rekentuin, Taalzee, DigLin) in the EBSCOhost databases, Web of Science, and Scopus (see Supplementary Material, Part B), but no record with one of these instruments as its object of validation entered the screening pool. The only study that mentioned a Dutch-tradition instrument (two database records of one article) used a standardized Cito achievement test (Cito-VAS) as the criterion measure in a CBM study of foreign-language learning with grade 8 students without identified special educational needs and was excluded at the title-and-abstract stage for wrong population (E1). Because the database searches were limited to English and German, Dutch-language validation documentation, including the reviews of the national test-review system (Evers et al., 2010), could not be retrieved by design. The publication window (2000 onward) applies to all groups and does not explain the empty group: it excludes the foundational work on the Dutch system from the 1980s and 1990s, but several of the digital platforms named in the search terms (e.g., Snappet, Rekentuin) were launched after 2000, so that contemporary validation evidence would fall within the window if it were published in an eligible language. The empty G7 group therefore shows that, within the review period, validation evidence on the digital Dutch-tradition instruments has not reached the English- or German-language literature; it carries no information about the quality of these instruments.

Sensitivity analyses

The four sensitivity analyses left the domain-level pattern essentially unchanged (Supplementary Table S5). Restricting the corpus to studies with digital or computer-adaptive primary administration (k = 99) changed the coverage and strong-coefficient proportions only marginally (reliability 66% covered, 55% strong; validity 86%, 32%; responsiveness 46%, 8%; classification accuracy 36%, 31%). Excluding the 15 single-time-point studies (k = 91) likewise left the pattern intact (reliability 67%, 55%; validity 86%, 32%; classification accuracy 34%, 32%), while responsiveness coverage rose modestly (52%, 9% strong). The responsiveness gap is therefore not an artifact of retaining single-time-point validation studies.

Within the rated core-platform subset, restricting the corpus to the 29 studies with an overall COSMIN rating of adequate or very good reduced the strong-coefficient proportion for validity from 33% to 24% and left responsiveness (28% covered, 10% strong) and classification accuracy (38%, 31%) at comparable levels; conditioning on methodological quality therefore weakens rather than strengthens the apparent validity evidence. Under the threshold analysis, the proportion of studies that would qualify as strong under a .70 threshold lies between 55% and 59% for reliability, between 33% and 63% for validity, between 8% and 13% for responsiveness, and between 31% and 32% for classification accuracy. Even at its upper bound, responsiveness remains by far the least frequently strong domain, and the validity–responsiveness asymmetry persists under every specification.

Cross-cutting findings

Three cross-cutting patterns warrant explicit reporting before the discussion. First, fully paper-based progress monitoring was excluded by design (E3); within the corpus, paper administration with computerized scoring as the primary modality is rare (3 of 106 studies, 3%). A broader modality audit, however, flagged 29 of 106 studies (27%) as borderline cases in which administration remained (partly) paper-based while scoring and/or platform-based data integration were computerized. The CBM research base is shifting from paper to digital administration along a continuum. Scoring digitization typically precedes administration digitization.

Second, the geographic distribution is bimodal. U.S. and German studies together account for 93% of the corpus. The remaining 7% is distributed across six additional countries, with no more than two studies per country. Cross-linguistic and cross-system generalizability of digital progress-monitoring evidence remains thin, particularly for languages and educational systems outside the English- and German-language CBM traditions.

Third, only 38 of 106 studies (36%) report classification-accuracy facets, and only 33 of 106 (31%) report classification-accuracy coefficients at the strong (≥ .80) level. Classification accuracy is the psychometric basis for screening decisions in tiered intervention systems. Informing such decisions was one of this review's key motivations. This gap is the single most consequential reporting limitation in the contemporary digital progress-monitoring literature.

A final corpus-level caveat concerns the cumulative sample size (Σ N = 774,420). The distribution is heavily right-skewed. A small number of large-scale technical reports account for more than 80% of the cumulative sample. These include Good et al. (2009; n = 358,032), Alonzo and Anderson (2018c; n = 180,617), Foorman et al. (2015; n = 44,780), Roehrig et al. (2008; n = 35,207), and Moulton et al. (2019; n = 24,094). The median per-study sample size is 459. The interquartile range spans 145–1,353. The median provides a more representative description of the typical study-level sample than the cumulative sum. The cumulative N should not be misinterpreted as a typical study sample size. This caveat is relevant for the interpretation of the review itself: the synthesis conclusions rest on the many small and medium-sized studies rather than on the handful of very large technical reports, and the apparent breadth suggested by the cumulative N is concentrated in a few platforms, two countries, and, in the case of the three Levumi studies that share one sample, partially dependent samples.

Discussion

This systematic review examined the psychometric evidence base for digital progress monitoring in special education, drawing on 106 studies and a cumulative sample of Σ N = 774,420 students. Three principal findings stand out. First, validity is broadly reported (86%) but only modestly strong (33% of studies report coefficients at or above .80). Second, reliability is the domain in which strong coefficients are most often attained (55% strong). Third, the two domains most directly relevant to formative decision-making, responsiveness and classification accuracy, are the least frequently reported (45% and 36%, respectively) and the least frequently strong (8% and 31%). The corpus is large, but the evidence relevant to MTSS-aligned practice is thinner than the absolute counts suggest. These patterns were robust across all four sensitivity analyses.

The validity–responsiveness asymmetry

The clearest pattern across the corpus is the asymmetry between validity and responsiveness reporting. Validity is reported in 86% of studies and responsiveness in 45%. When restricted to strong coefficients, the gap remains: 33% of studies report strong validity, but only 8% report strong responsiveness. This asymmetry has a methodological explanation. Cross-sectional correlations against an external criterion are easier to estimate than slope reliability or growth-curve parameters, which require multiple measurement occasions, larger samples, and explicit longitudinal modeling. The asymmetry also has a conceptual cost. An instrument with strong concurrent validity but unknown responsiveness can rank students at one point in time, but its capacity to detect change in response to intervention remains untested. For formative decision-making in tiered systems, this is the wrong half of the evidence to be strong on.

This pattern is reinforced by the design features documented in the methodological-adequacy section. Of the 106 included studies, 31 (29%) used one or two measurement time points and could not, by design, contribute to slope-based responsiveness evidence. Only 17 studies (16%) reached the eight-time-point threshold associated with stable slope estimation in curriculum-based measurement (Christ et al., 2013; see Figure 3). The responsiveness gap is therefore not only a reporting gap but a design gap. Many of the included studies were not built to validate progress monitoring in the strict sense, even when the instruments are described as progress-monitoring tools. The risk-of-bias appraisal supports the same conclusion: responsiveness designs showed the highest concentration of doubtful ratings among the rated core-platform studies, and restricting the corpus to adequately rated studies lowered the strong-validity proportion without closing the responsiveness gap.

Domain-specific patterns

The five non-empty synthesis groups show distinct evidence profiles. G1 (CBM-reading) is the largest and most psychometrically mature group. The dominance of easyCBM, DIBELS, and Levumi reflects three decades of cumulative platform validation. Practitioners selecting reading progress-monitoring instruments can draw on a substantial evidence base, with the caveat that responsiveness evidence is concentrated in a small subset of studies.

G2 (CBM-math) is markedly different: despite reliability and validity coverage comparable to G1, most math instruments lack any classification-accuracy evidence. For practitioners, this means that a math instrument with strong validity for ranking students may have unknown sensitivity or specificity for screening decisions. The four math studies that did report classification accuracy (Clarke et al., 2011; Killen, 2013; Nese et al., 2010; Peltier et al., 2022) achieved high facet counts, which suggests that the gap is not technical but representational. Math instruments can be evaluated for screening, but most have not been.

G3 (language progress monitoring) is the most evidence-thin group of the corpus. Four studies are not enough to support strong claims about any single instrument, and the four eligible instruments do not overlap. Given the centrality of formative language assessment for early intervention in students with developmental language disorder, the G3 gap is a research priority. The corpus contains useful starting points (Picture Naming IGDI, Com-PASMA, ATLAS-PA, the German LVD Sprache), but the field is too small to triangulate evidence across instruments.

G4 (direct behavior rating and digital behavioral observation) shows a third distinct profile. Reliability is comparatively well documented, consistent with the methodological centrality of inter-rater and internal-consistency analyses for behavior-rating measures, but responsiveness and classification accuracy are both substantially under-reported. The combination is consequential. G4 instruments are widely used for screening and progress monitoring of social-emotional and behavioral risk, yet the evidence base for both screening accuracy and growth detection is thin. Kilgus et al. (2018) is the single G4 study reporting all five classification-accuracy facets, and the field would benefit from comparable validations of additional instruments.

G5 (ecological momentary assessment and wearables) was empty. No included study met the eligibility criteria for school-based, formative progress monitoring of students with or at risk for special educational needs using EMA or wearable technologies. The empty group is reported here as a substantive null finding rather than a screening artifact. The eligibility criteria were broad enough to include smartphone-based, ambulatory, and sensor-based formative assessment in school-anchored contexts, and the two candidate studies that reached full-text assessment were excluded on substantive grounds, one for its community sample without special educational needs (Milojevich et al., 2023) and one for the absence of psychometric evidence (Singh et al., 2025). The EMA-and-wearables paradigm has been applied extensively in clinical and adult research (Heron et al., 2017; Shiffman et al., 2008), but as of April 2026 it has not been operationalized as a formative progress-monitoring tool within special-education school settings. This is a research gap with substantial methodological potential, given that EMA and wearables can in principle capture behavior, attention, and emotion regulation at temporal resolutions far higher than any of the instruments in G1 through G4.

G6 (computer-adaptive testing and learning analytics) is operationally well-established but narrow. All four eligible studies validated Renaissance STAR products. Classification evidence, where reported, was strong; this is the highest classification-evidence rate of any group. Responsiveness, by contrast, was not strong in any G6 study, despite the platform-level capacity for fine-grained slope estimation that adaptive testing should in principle afford. This is a missed opportunity rather than a technical limitation.

Geographic and linguistic concentration

The corpus is dominated by U.S. and German studies, which together account for 93% of the included studies. The remaining 7% spans six countries, with no more than two studies per country. Three implications follow. First, the cross-linguistic generalizability of digital progress-monitoring evidence is weak outside the English- and German-language CBM traditions. Second, the dominance of two national systems narrows the policy contexts in which the instruments have been evaluated. MTSS in the United States and inclusive education in Germany are organizationally distinct, but the formative-decision-making logic underlying both systems is similar enough that instruments developed in one have often been adapted to the other. For systems with different referral pathways or different definitions of special educational needs, the evidence base is thinner. Third, the few non-U.S., non-German studies in the corpus (Carson et al., 2014, New Zealand; Chung et al., 2018, the Netherlands; Hautala et al., 2020, Finland; Misquitta et al., 2023, India; Al-Hmouz, 2013, Jordan; Mercer et al., 2021, and Mercer and Cannon, 2022, Canada; plus two U.S. studies of Spanish-language instrument versions, Basaraba et al., 2022, and Klingbeil et al., 2021) are individually informative but collectively too sparse to support cross-system synthesis. Programmatic validation of digital progress monitoring outside the English- and German-language traditions is a clear research priority. The empty G7 group illustrates the same point at the system level. The Dutch pupil-monitoring tradition is a mature, nationally established infrastructure (Vlug, 1997), but no validation study of its digital instruments entered the screening pool: Dutch-language documentation, including the reviews of the national test-review system (Evers et al., 2010), lay outside the language limits of the search, and the English- and German-language literature retrieved by the Dutch-tradition search terms contained no eligible validation study. The empty group thus reflects the eligibility boundaries of this review and the limited accessibility of a national evidence base rather than the state of that evidence, and a dedicated synthesis of the Dutch-language validation literature would be a worthwhile complement to this review.

The modality continuum

Fully paper-based progress monitoring was excluded by design; paper administration with computerized scoring as the primary modality is rare (3 of 106 studies, 3%), while the modality audit flagged 29 of 106 studies (27%) as borderline cases combining (partly) paper-based administration with computerized scoring or platform-based data integration. This pattern indicates that the field is shifting from paper to digital administration along a continuum, with scoring digitization typically preceding administration digitization. The implication for practice and research is that the binary distinction between digital and paper progress monitoring is no longer the right framing. Instruments and platforms vary along several dimensions: who or what administers the items (student, teacher, software), who or what scores the items (teacher, software, automated speech recognition), and how the resulting data are stored and visualized. A more useful taxonomy would treat each of these dimensions independently. The 29 borderline cases identified in this review (paper administration with computerized scoring) are not methodological hybrids but a distinct, operationally common category.

Implications for practice

Three practice-relevant conclusions follow from the synthesis. Because classroom teachers rarely access primary validation studies directly, these conclusions are addressed primarily to the actors through whom psychometric evidence reaches schools: school psychologists, district and school leadership, teacher educators, tool developers, and compilers of instrument repositories such as the NCII tools charts. First, those selecting reading progress-monitoring instruments can draw on a comparatively mature evidence base in G1, particularly for the dominant platforms (easyCBM, DIBELS, Levumi). Second, users of math instruments for screening decisions should be cautious. Most of the G2 instruments have not been formally evaluated for sensitivity, specificity, or predictive value, and the absence of classification-accuracy evidence should be interpreted as missing evidence rather than null evidence. Third, users of behavior-rating instruments for growth monitoring should be aware that responsiveness has not been established at strong levels in any G4 study. Behavior ratings can be reliably collected at single occasions, but their capacity to detect change over time is empirically less well-supported than their reliability would suggest. In all three domains, triangulation across instruments and across data sources remains the most defensible decision-making strategy. Finally, the effectiveness literature indicates that even psychometrically sound instruments improve learning only under supportive implementation conditions such as frequent measurement, feedback, and consultation (A. Fuchs et al., 2025; Jung et al., 2018), and these conditions depend on the very responsiveness evidence this review found to be scarce.

Implications for research

Five research priorities follow from the findings. First, responsiveness should become standard reporting for any instrument described as a progress-monitoring tool. The 8% strong-coefficient rate across the corpus is too low for a property that defines the instrument category. Reporting should include both growth-sensitivity and slope-reliability facets, and study designs should include enough measurement time points to support stable slope estimation. Second, classification-accuracy evidence is needed for math and behavior instruments. The 19% and 22% reporting rates in G2 and G4 are not commensurate with the role these instruments play in screening pipelines. Third, the EMA-and-wearables paradigm needs school-based operationalization. The empty G5 group is not evidence that the paradigm does not work in schools; it is evidence that the paradigm has not yet been evaluated in schools. Fourth, instrument validation outside the English- and German-language traditions should be supported as a programmatic research effort. Fifth, the field would benefit from greater transparency on dependent samples. Three Levumi studies in this corpus used the same n = 761 third-grade sample, and the cumulative N reported for the corpus is therefore not the cumulative N of independent participants.

Limitations

Seven limitations should be noted when interpreting the findings of this review. First, a formal risk-of-bias assessment using the COSMIN Risk of Bias Checklist (Mokkink et al., 2018) and QUADAS-2 (Whiting et al., 2011) was protocol-planned but not implemented for the full corpus; a reduced formal assessment covered the 60 core-platform studies (Supplementary Tables S4a, S4b), while the breadth indicator served as the transparency measure for the remaining studies. The breadth-of-reporting indicator (number of facets per synthesis domain) is reported as a structural alternative, but it does not substitute for instrument-level assessment of methodological adequacy. Future updates of this review should incorporate formal risk-of-bias assessment.

Second, the eligibility criteria included only English- and German-language publications; the corresponding language limits were applied at the search stage, and E7 served as a safeguard at the full-text stage. Studies in other languages, including the Dutch-language validation literature relevant to G7, did not enter the review. The cross-linguistic concentration of the corpus is therefore in part a product of the eligibility criteria and not only a finding about the field.

Third, the screening pipeline made use of large language models for multi-rater consensus decisions. All AI-assisted decisions were verified through independent re-extraction and human adjudication, and inter-rater reliability was high (Fleiss’ κ = 0.856 in title-and-abstract screening; κ = 0.997 in full-text screening). Nonetheless, the use of LLMs in screening introduces a layer of automation that was not standard practice in earlier systematic reviews.

Fourth, the synthesis is narrative rather than meta-analytic. Subgroup meta-analyses for homogeneous instrument clusters were specified as optional in the protocol but were not implemented because of substantial between-study heterogeneity in instruments, samples, and reporting standards. A meta-analytic synthesis of the easyCBM and DIBELS subsets within G1, or of the SAEBRS subset within G4, would be feasible in principle and is recommended as a follow-up project.

Fifth, despite the inclusion of gray literature, publication bias cannot be ruled out. Studies with null or partial psychometric findings may be underrepresented in the corpus, particularly for instruments that did not achieve commercial or institutional adoption. The breadth-of-reporting indicator captures incomplete reporting at the study level but does not capture studies that were never published.

Sixth, the a priori coefficient categorization applied a uniform .80 threshold across measurement properties. This choice favors cross-domain comparability but is stricter than property-specific conventions, particularly for correlation-based validity coefficients, and it classifies coefficient magnitude independently of study quality. The threshold sensitivity analysis brackets the resulting uncertainty, and the risk-of-bias sensitivity analysis addresses the quality dimension; the substantive conclusions, in particular the validity–responsiveness asymmetry, hold at both threshold bounds and under the quality restriction.

Seventh, measurement aspects beyond the four synthesis domains, most notably differential item functioning and measurement invariance across subgroups as well as the linguistic accessibility of instruments, were captured only where included studies reported them within the validity facets; they were not synthesized as separate evidence dimensions, although they are clearly relevant for instrument selection in linguistically and culturally diverse special-education populations.

Conclusion

The digital progress-monitoring evidence base in special education is large, internationally distributed, and methodologically uneven. Reliability and validity are broadly reported. Responsiveness and classification accuracy, the two domains most directly relevant to formative decision-making in tiered intervention systems, are systematically under-reported. These patterns persisted across all sensitivity analyses. The most consequential gaps are in math classification accuracy, in behavior responsiveness, and in the empty G5 paradigm of school-based ecological momentary assessment and wearables. The evidence base supports cautious use of established reading instruments, conditional use of math and behavior instruments, and active development of underserved domains and paradigms. The next research wave should treat responsiveness and classification accuracy as standard reporting requirements rather than optional add-ons.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI tools were used to support the extraction and processing of information from the included texts, to aid comprehension of the source material, and to prepare the extracted data for auditing. They were additionally used to assist with translations. All AI-supported outputs were reviewed and verified by the authors.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feduc.2026.1936281/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Rui Dai, Wichita State University, United States

Reviewed by: Gulzhaina K. Kassymova, Abai University, Kazakhstan

Daniel Sommerhoff, Ludwig Maximilian University of Munich, Germany