First-year students’ math anxiety predicts STEM avoidance and underperformance throughout university, independently of math ability

The study investigates whether math anxiety independently predicts STEM participation (course-taking) and achievement (grades) throughout university, beyond the influence of objective math ability. Prior literature widely claims that math anxiety leads to avoidance of math and underperformance, but evidence has been limited: (1) most studies did not control for math ability despite its robust negative correlation with math anxiety, raising confounding concerns; (2) outcomes often focused narrowly on math rather than the broader STEM spectrum; (3) many findings were retrodictive (measuring anxiety after choices/performance), and (4) avoidance and achievement have often been conflated rather than studied together. The authors measured math anxiety and math ability in first-semester undergraduates and tracked comprehensive STEM outcomes via transcripts over four years to test unique predictive effects, including controls for cognitive, affective, academic variables, and the other STEM outcome. They further examined mediation to quantify how much anxiety and ability account for each other’s links to STEM outcomes and moderation by gender and non-STEM performance to identify for whom math anxiety is most consequential. The work aims to inform theory on mechanisms (beyond ability) linking math anxiety to STEM outcomes and to guide interventions targeting STEM participation and achievement.

Prior research associates higher math anxiety with fewer math/STEM courses taken, lower likelihood of STEM majors/careers, and poorer performance in math-related classes. Individuals in non-STEM majors/careers report higher math anxiety than those in STEM. Longitudinal work links persistent or increasing math anxiety in adolescence to lower likelihood of STEM careers in adulthood. Math anxiety correlates negatively with math performance, with theorized mechanisms including avoidance limiting practice (depressing skills) and in-the-moment anxiety consuming working memory. However, much of this literature did not control for math ability (a strong correlate of math anxiety), often measured anxiety after the relevant STEM choices were made, focused on limited coursework (e.g., a statistics class), and rarely examined both STEM avoidance and achievement together. Conversely, a large literature links math ability to outcomes such as math grades, STEM career pursuit, income, and health, but typically does not control for math anxiety, raising the possibility that attitudes (anxiety) may partly account for those associations.

Design and participants: A prospective, correlational longitudinal study tracked university STEM outcomes over four years. Participants were 186 first-year undergraduates at the University of Western Ontario; 3 were excluded (not first-year or failed attention checks), yielding N=183 (117 female; mean age=18.55). Recruitment occurred in first semester via campus flyers and in-person. Power analyses indicated adequate power to detect small-to-moderate effects (correlations ≥ .205; multiple regression f ≥ .07 with up to 9 predictors). Ethics approval and written consent obtained; participants granted access to de-identified transcripts.

Procedure and measures (first semester):

• Math Anxiety: Short Math Anxiety Rating Scale (SMARS; 25 items; 0–4 scale; total 0–100; α=.96).
• Math Ability: Difficult mental arithmetic (ETS Kit of Factor-Referenced Cognitive Tests). Timed 3-min blocks for addition, subtraction, multiplication, division; mental calculation only; score = total correct across blocks (mean accuracy 81.2%, mean RT 9.91s; α=.89). Chosen to index broadly applicable foundational math ability.
• Trait Anxiety: State-Trait Anxiety Inventory–Trait (20–80; α=.93) to control non-math-specific anxiety.
• Verbal Working Memory: Automated reading span task (15 sequences of 3–7 sentences+letters); total score 0–75 (higher = greater capacity) as a general cognitive ability covariate.
• Gender: Self-reported (Male=0, Female=1).

Transcript-derived outcomes and covariates (over four years):

• % STEM Courses: Proportion of all courses designated as STEM, based on department classification. Traditional STEM departments (e.g., Physics, Mathematics, Engineering, Chemistry) were labeled STEM; ambiguous departments were classified via course content emphasis on math or scientific method (designation finalized prior to analysis; see Table S1). Using proportion accounts for different course totals.
• STEM Grades: Mean grade (0–100) across all STEM courses taken. Repeated courses were included (52 of 6211 courses; 0.8% repeats; sensitivity analyses indicated no substantive impact).
• Non-STEM Grades: Mean grade across non-STEM courses.
• Semesters Absent: Count of semesters missing transcript data (time off/left university) to adjust for differential exposure to course-taking.

Analytic approach:

• Relations among predictors/outcomes: Tested correlations between Math Anxiety and Math Ability; between % STEM Courses and STEM Grades (including partial correlation controlling for Semesters Absent).
• Zero-order associations: Correlated Math Anxiety and Math Ability with % STEM Courses and STEM Grades.
• Multiple regression (primary tests of unique prediction): Two models with DVs (a) % STEM Courses and (b) STEM Grades. Predictors of interest: Math Anxiety and Math Ability. Covariates: Trait Anxiety, Verbal Working Memory, Gender, Non-STEM Grades, Semesters Absent, plus the other STEM outcome (i.e., STEM Grades included in the % STEM Courses model; % STEM Courses included in the STEM Grades model). Predictors standardized; DVs in native units; effects reported as change in native units per 1 SD increase in predictor; df=174. Adjusted R²: .347 (% STEM Courses) and .554 (STEM Grades).
• Mediation (secondary analysis 1): Causal mediation models using R mediation package with 10,000 bootstrap iterations tested the extent to which (i) Math Ability mediates links from Math Anxiety to outcomes and (ii) Math Anxiety mediates links from Math Ability to outcomes. Same covariates as primary models (including the other STEM outcome as appropriate). Reported indirect effects and proportion mediated (%C) with bootstrapped 95% CIs.
• Moderation (secondary analysis 2): Interaction tests assessed whether the Math Anxiety–outcome links were moderated by Non-STEM Grades (proxy for general academic aptitude) and by Gender, controlling for the same covariates used in primary models.

Data availability: OSF https://osf.io/bctyg/?view_only=9ce7dd98a8464f1db2af1a40eb0336e7/

Descriptives and relations among measures:

• Math Anxiety and Math Ability were negatively correlated (r(181)≈−.346, p≈2e−6), replicating prior work.
• % STEM Courses and STEM Grades were weakly related overall (r(181)=.119, p=.107); after controlling for Semesters Absent, partial rp(180)=.191, p=.010; among those with no Semesters Absent, r(146)=.163, p=.047. These outcomes are largely distinct (<5% shared variance).

Zero-order predictions of STEM outcomes:

• Math Anxiety: Higher anxiety predicted fewer STEM courses (r(181)=−.390, p=5e−8; ≈13.1% decrease in % STEM per 1 SD increase) and lower STEM grades (r(181)=−.324, p=8e−6; ≈3.47-point drop per 1 SD increase).
• Math Ability: Higher ability predicted more STEM courses (r(181)=.241, p=.001; ≈8.1% increase per 1 SD increase) and higher STEM grades (r(181)=.280, p≈1e−4; ≈3.00-point increase). Results were similar controlling for Semesters Absent.

Unique predictive effects (multiple regression controlling for covariates and the other STEM outcome):

• % STEM Courses model (Adjusted R²=.347; df=174): • Math Anxiety significantly predicted fewer STEM courses: β=−.109 (native units; i.e., −10.9 percentage points), t=−4.20, p=4e−5, d=.637. Standardized unique effect ≈−.325. Interpreted as ≈3.93 fewer STEM courses over four years (≈1 per year), given 36.13 courses on average. • Math Ability did not uniquely predict % STEM Courses: β=.043, t=1.91, p=.058. • STEM Grades also predicted % STEM Courses (β=−.097, p=.001); Non-STEM Grades (β=.157, p=2e−7) and Semesters Absent (β=.103, p=2e−6) were significant covariates.
• STEM Grades model (Adjusted R²=.554; df=174): • Math Anxiety significantly predicted lower STEM grades: β=−2.406 points, t=−3.45, p=7e−4, d=.524. Standardized unique effect ≈−.225. Interpreted as a 2.41-point reduction per 1 SD increase in anxiety; high (＋1 SD) vs low (−1 SD) anxiety ≈4.82-point difference (nearly half a letter grade). • Math Ability did not uniquely predict STEM grades: β=.919, t=1.53, p=.129. • % STEM Courses (β=−2.121, p=.001) and Non-STEM Grades (β=7.579, p<2e−16) were significant covariates.

Mediation (proportion mediated, %C):

• Ability accounting for Anxiety→Outcome links: • Anxiety→% STEM Courses via Ability: %C=12.6% (95% CI [.5%, 30.8%]), indirect effect p=.040. • Anxiety→STEM Grades via Ability: %C=11.5% (95% CI [−.2%, 42.1%]), indirect effect p=.100 (ns).
• Anxiety accounting for Ability→Outcome links: • Ability→% STEM Courses via Anxiety: %C=40.9% (95% CI [17.2%, 95.0%]), indirect effect p=.001. • Ability→STEM Grades via Anxiety: %C=40.0% (95% CI [9.2%, 128.2%]), indirect effect p=.010. Thus, anxiety explains a substantial portion of ability–outcome associations, more so than ability explains anxiety–outcome associations.

Moderation:

• Non-STEM Grades moderated Math Anxiety→STEM Grades (β for interaction≈.140, t=−2.90, p=.004): the negative impact of anxiety on STEM grades was stronger among high non-STEM achievers. Predicted difference between high- vs low-anxiety students among high non-STEM achievers ≈8.48 points (nearly a letter grade), controlling covariates. Non-STEM Grades did not moderate Math Anxiety→% STEM Courses (p=.687).
• Gender trends: Anxiety effects tended to be more negative for men than women, but interactions did not reach α=.05 (for % STEM Courses p=.075; for STEM Grades p=.051). Women reported higher anxiety than men (t(181)=−5.35, p≈3e−7, d≈−.823).

Findings provide prospective, transcript-based evidence that math anxiety measured at university entry uniquely predicts two key and largely independent STEM outcomes—course participation and achievement—over four years, even after accounting for objective math ability, general anxiety, working memory, academic performance outside STEM, time away from school, and the other STEM outcome. Contrary to accounts attributing anxiety’s academic effects solely to depressed math ability (via less practice or working-memory interference), anxiety’s unique prediction of STEM grades suggests additional mechanisms. The authors propose ‘micro-avoidance’ behaviors (reduced attention, study time, engagement with math-heavy components within courses) as a plausible pathway distinct from ‘macro-avoidance’ (avoiding math/STEM courses altogether). The mediation analyses further indicate that math anxiety substantially accounts for previously reported links between math ability and real-world STEM outcomes, implying that attitudes toward math may confound ability–outcome relations when anxiety is not measured. Moderation results suggest that math anxiety may particularly undermine STEM performance among otherwise strong students (high non-STEM achievement), potentially contributing to the STEM pipeline leak. Overall, results emphasize that feelings toward math, not just ability, are critical determinants of STEM engagement and success, with implications for revising theoretical models and for designing interventions that target both avoidance and performance pathways.

This study demonstrates that first-semester math anxiety prospectively predicts taking fewer STEM courses and earning lower grades in STEM across university, independent of objective math ability and other covariates. Math ability did not uniquely predict these outcomes once anxiety was controlled, and math anxiety accounted for a sizable portion of ability–outcome associations. Theoretically, this calls for models in which anxiety impacts STEM via mechanisms beyond ability, including distinct pathways for macro-avoidance (course selection) and micro-avoidance (within-course engagement). Practically, interventions solely focused on boosting math skills may be insufficient; targeting math anxiety—potentially with distinct strategies for participation vs performance—may better improve STEM outcomes. Future research should include experimental and intervention studies to establish causality, test approaches that directly reduce math anxiety or mitigate its in-the-moment effects, examine micro-avoidance behaviors in authentic educational contexts, incorporate broader measures of math ability, and assess generalizability across institutions and populations.

• Correlational design: Although prospective and controlling for multiple covariates, the study cannot establish causality. Intervention studies are needed to determine whether reducing math anxiety improves STEM outcomes.
• Generalizability: Single large public Canadian university sample; findings may differ across institutions, cultures, or educational systems.
• Math ability measurement: A difficult mental arithmetic task may not capture all facets of math competence (e.g., higher-order reasoning, algebraic/geometry skills) relevant to specific STEM disciplines; other ability measures might explain additional variance.
• STEM classification: Department-level STEM/non-STEM designations, while defined a priori, involve some subjectivity.
• Missing semesters: Although modeled via a Semesters Absent covariate and robustness checks performed, varying exposure to coursework could still influence estimates.