The relationship between anxiety and mathematics has often been investigated in the literature. Different forms of anxiety have been evaluated, with math anxiety (MA) and test anxiety (TA) consistently being associated with various aspects of mathematics. In this meta-analysis, we have evaluated the impact of these forms of anxiety, distinguishing between different types of mathematical tasks. In investigating this relationship, we have also included potential moderators, such as age, gender, working memory, type of task, and type of material. One hundred seventy-seven studies met the inclusion criteria, providing an overall sample of 906,311 participants. Results showed that both MA and TA had a significant impact on mathematics. Sociodemographic factors had modest moderating effects. Working memory (WM) also mediated the relationship between MA and TA with mathematics; however, this indirect effect was weak. Theoretical and educational implications, as well as future directions for research in this field, are discussed.

The online version contains supplementary material available at

During school years, many students experience feelings of apprehension and fear brought on by testing situations or by facing a particularly challenging subject, such as math (Aiken,

Math anxiety (hereafter MA) refers to negative feelings of tension and apprehension experienced when thinking about and performing mathematical problems or number-related tasks in both academic situations and ordinary life (Ashcraft,

The other form of academic anxiety linked to a school context is test anxiety (TA). TA can be considered a form of anxiety in educational settings, encompassing affective, cognitive, and physiological reactions that are usually experienced during exams or similar evaluative situations (Sieber et al.,

Despite the uniqueness of MA and TA as specialized forms of anxiety, both share common features, including a perceived risk of failure and resultant disapproval by significant others who are evaluating the performance in comparison to a standard of achievement (Zeidner,

As previously mentioned, both forms of academic anxiety (MA and TA) are far-reaching and tend to have a detrimental impact on academic performance (Hembree,

MA, by its own definition, has a specific negative impact on an individual’s math achievement (Ashcraft & Moore,

An important factor to consider when looking at the relationship between MA and academic performance is the influence of gender and developmental changes. In terms of gender differences, some studies have reported the negative relationship between MA and math to be stronger in girls than in boys (Devine et al.,

Looking at MA from a developmental perspective, the negative relationship between MA and math performance may increase across years of schooling (Dowker,

TA is characterized to be (a) context specific (e.g. evaluative situations in educational settings) and (b) found in a variety of academic subjects (Hembree,

To date, only a few studies have looked specifically into the relationship between TA and math performance. In his meta-analysis, Hembree (

The literature identifies gender and age as important sociodemographic moderating variables in the relationship between TA and performance (Hembree,

As for developmental changes, as children move through the educational system, they tend to be tested more frequently, both for pragmatic reasons (e.g. increasing subject number complexity) but also for more legislative reasons (e.g. the use of test-based accountability policies) (Duckworth et al.,

Although the relationship between MA and math performance is well documented, the cognitive mechanisms underlining this relationship are not well understood. Some researchers argued that MA has a negative effect on math performance by impairing working memory (WM; Ashcraft & Kirk,

It has been proposed that individuals with high MA suffer from intrusive thoughts of failure when attempting to solve mathematical problems, and that this occurs more often than in individuals with low levels of MA. This indicates that cognitive factors, such as attention control and WM, could play a central role in explaining the anxiety-performance link (Eysenck & Calvo,

Regarding TA, only few studies investigated the role of WM in the relationship between TA and arithmetic task performance in children (Korhonen et al.,

It seems clear that the role of WM on the relationship between MA, TA, and math performance depends on a variety of factors. Some of these factors include the type of tasks being used to assess both math ability and WM, as well as the age and skill level of the participants (Raghubar et al.,

Given the potential impact of academic anxiety on math achievement, we decided to conduct a systematic meta-analytic review across the last 30 years of existing literature. Due to the recent publications of three meta-analyses separately targeting MA and TA (Barroso et al.,

No previous meta-analyses have assessed the specific role of both MA and TA on math performance, while also considering these different variables. In contrast with the meta-analysis of Namkung et al. (

Overall, the main objectives of the present meta-analysis were (i) estimating the impact of MA and TA on math performance; (ii) measuring the impact of potential moderators on the MA/TA and math relationship, by taking into account gender, age, type of math tasks, and type of WM tasks; and (iii) assessing the mediating role of WM when considering MA and math achievement relationship.

Primary research papers, in English, with quantitative data were reviewed. Only papers published from 1990 to 2018, in peer-reviewed journals, were included. Papers reviewed varied in their methodologies and studies were included if they directly measured math achievement or performance, WM, MA, and TA in typical developing individuals. The preferred reporting items for systematic reviews and meta-analyses (PRISMA) 2009 checklist were followed when writing this review. When assessing the risk of bias, the protocol described by Sirriyeh et al. (

We conducted an electronic survey of five online databases: Web of Science, Scopus, PubMed, Medline via Ovid (1996–present), and PsycINFO via Ovid (1996–present). When possible, a full text search was conducted, and in cases in which the full text search was not available, a search of the keyword (PsycINFO) or topic (Web of science) was conducted. The search terms used included “math* anxiety”, “test anxiety”, “performance”, “achievement”, and “working memory”. The wildcard * was used where variations like “math”/“maths”/“mathematics”/“mathematic”/“mathematical” might occur. Math* AND Anxiety AND (Achievement OR performance OR “working memory”). All search terms were combined in the same way on all databases, but some databases had different ways of inputting this. In order to include grey literature, we performed the same survey (using exactly the same procedure, same years, search terms, and Boolean operators) on ProQuest platform. We focused in particular on dissertations, theses, and conference proceedings.

After conducting systematic searches in each of the five databases, 3,223 articles were collected in total. Data were exported from each database into an EndNote library, and 1,360 duplicates were removed, leaving a total of 1,863 records. The remaining papers were exported into a Microsoft Excel spreadsheet. Two independent reviewers (S.C. and J.M.R.) independently evaluated each article and decided whether each specific paper was to be included in the meta-analysis. Disagreements were resolved by a third reviewer (I.C.M). As for the grey literature, the research was conducted through three different libraries (University of Padova, Genoa, and Leeds) resulting in 141 retrieved documents. After duplicate removal, a total of 109 abstract were screened by the two independent reviewers. To improve our strategy search, we also emailed leading researchers in the field requesting unpublished data, dissertations, as well as articles published on non-peer-reviewed journals that could be fit with the aims of the present meta-analysis.

Studies were entitled to be included in the meta-analysis if they met the following inclusion criteria. In the full-text analyses, we focused on three different study designs: (a) correlational studies, (b) longitudinal prediction studies, and (c) experimental studies. To prevent violating the independence of observations in longitudinal studies, data only from the first time point were coded. Similarly, for experimental or intervention studies, only data from control groups or pre-test period were coded. As for studies following an extreme-groups design (e.g. high vs low), only data from the initial screening were included, when available. Moreover, we only included studies reporting original empirical findings (i.e. not a review or re-analysis of findings that were reported previously). The inclusion of each study and its effect sizes in the database was based on the inclusion of at least one mathematical performance measure and one math/test anxiety assessment. Studies assessing perceived math competence or individuals’ self-reported beliefs about their math skills were excluded. Thus, studies had to either report zero-order correlation coefficients or report data made the calculation of correlation coefficients possible. When these were not reported, effect sizes were requested from authors via e-mail.

In a second phase, title and abstract screening were conducted and reviewers excluded 1,342 records that did not measure MA/TA or math performance, leaving 521 articles to be assessed for eligibility through full-text exploration. Similarly, for the grey literature, out of the 109 documents, only 65 were eligible for full-text screening. These documents were then subject to close review using the above outlined inclusion/exclusion criteria. Based on the full-text analyses, 370 peer-reviewed studies were removed, leaving a total of 151 studies. To these are added 26 documents from the full-text screening of the grey literature, reaching a total of 177 studies included in the present meta-analysis. Details concerning the literature search, inclusion criteria, and selected studies are shown in the flow chart below (Fig.

Flow diagram for the search and inclusion criteria for studies in this meta-analysis. Adapted from “Preferred Reporting Items for Systematic Reviews and Meta-Analyses: The PRISMA Statement,” by D. Moher, A. Liberati, J. Tetzlaff, D. G. Altman, and The PRISMA Group, 2009,

Extracted data included details on the study such as title, authors, year and type of publication, and author contact details. Details about the sample were also included such as sample size, mean age, and gender. It must be noted that each study could include more than one sample of participants (e.g. groups of children in different grades, or males and females separately), which we treated as independent samples (further details are given below in the “Meta-Analytic Model Fitting” section). Concerning the effect size, all correlations of interest (i.e. those involving MA, TA, GA, math achievement, and WM) were coded, along with the combination of constructs that they involved (e.g. “MA-Maths”, “MA-WM”). As the same study could report more than one correlation for the same combination or sample (or for different combinations or samples), the dataset included as many rows per study as were the effect sizes of interest that it reported. When outcomes involved response times or error counts, the signs of the effect size were inverted.

To investigate the moderators of interest, we coded (1) where effect sizes were reported separately for boys and girls, and (2) age group classifications (i.e. child or adult). For math performance, we coded math measures as tapping (a) school grades, (b) basic numerical knowledge, (c) arithmetical skills only, or (d) more advanced and complex skills. Where WM measures were available, we also coded the type of task as WM or STM, and the type of material as verbal or spatial.

Additional details on test characteristics were coded for each of the three main domains considered (i.e. emotional factors, achievement, and cognitive factors). Specifically, for each measure, we summarized test content (i.e. the skills measured by the test), complexity (i.e. for mathematical skills, low complexity would be basic math knowledge while more advanced arithmetic abilities would be classified as applied math), origin (i.e. standardized battery, national survey, or experimental task), and reliability of the measure(s) where available.

The presence of possible publication biases in the meta-analysis was tested and quantified using the “trim-and-fill” method (Duval,

Unlike other methods aimed at assessing the publication bias, the trim-and-fill method aims at directly quantifying the potential bias (Borenstein et al.,

The analytic strategy adopted in this meta-analysis followed the guidelines proposed by Borenstein et al. (

Correlational values were reported as Pearson’s

All eligible effect sizes were considered, and when multiple effects were reported for the same outcome combination in the same sample of individuals (e.g. a certain construct was measured by means of different indicators on the same group of participants), effects were combined using the formulas for non-independent outcomes reported by Borenstein et al. (

Heterogeneity across independent samples was quantified using τ^{2}, which is a measure of estimated variance, and by using the ^{2} index, which can be interpreted as the percentage of total variance of the estimated effect that is attributable to the variance across studies (Higgins et al., ^{2} (i.e. ^{2} < 50%) suggest limited heterogeneity across studies, indicating that the estimated effect sizes are generalizable. Higher values of ^{2} (i.e. ^{2} > 75%) suggest that there are large differences across studies (e.g. investigation of different constructs, or sampling from populations in which the true effects are very different; see Higgins et al. (

To test the robustness of the results, all meta-analyses that included at least three independent effects (which implies at least three independent samples, as effects were aggregated by sample) were tested using the “leave-one-out” method (Viechtbauer,

Moderating analyses were conducted using multilevel models, with each sample entered as a random effect. Multilevel modelling seemed the best choice in this case, as most studies reported effects simultaneously on different levels of the same moderating factor. For example, concerning the type of WM task, several samples of participants were tested both on verbal and on spatial tasks, meaning that the studies investigated correlated effects across different levels of the same moderating factor. Thus, multilevel modelling was the best way to account for the structure of dependencies among the effects. For consistency with the analyses described above, we combined the effects by independent sample using the formula suggested by Borenstein et al. (

Estimated coefficients were obtained using the “restricted maximum likelihood” method, which is set by default in the “metafor” package. Nonetheless, model comparisons, conducted by mean of likelihood ratio tests for nested models, were conducted on models fitted using “maximum likelihood”. This is due to the fact that models are not comparable when fitted with the “restricted” method; all of these specifications are set by default by the “metafor” package.

It should be noted that significant and non-significant effects of moderating factors were the same, even without multilevel models and without random effects of samples (or studies), even though the coefficient slightly changed when random effects are not included. This suggests that the moderating analyses were robust, with only minor variations in the estimated coefficients, which occur independently of the specific analytical choices.

An additional analysis of mediation was conducted to examine the role of WM as a mediator between MA and math achievement, and between TA and math achievement. To do so, multivariate regression models (or “meta-path analysis”, as shown in Fig.

We conducted this analysis using two different alternative strategies. First, we sought to take into account the random effects of samples on direct and indirect effects of interest. Therefore, we restricted this analysis to studies that reported complete correlation matrices for anxiety, WM, and math achievement. For each of these “complete” studies, we fitted a separate mediation model. Subsequently, we meta-analyzed all direct and indirect effects of interest using random effects models.

For the second strategy, we incorporated the maximum information available, at the cost of no longer considering random effects (see Yu et al. (

Summary of all meta-analyses conducted

Combination | N. of Studies | N. of indep. samples | N. of effects | Pooled N partic. | Range mean age (years) | Estimated correlation | 95% CI | τ |
| Leave-one-out range of variation | Trim-and-fill adjustment | ||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|

Added studies | Est. corr. | 95% CI | |||||||||||

MA - MATHS | 138 | 200 | 369 | 883,970 | 5.9−49.2 | −.30*** | [−.32, −.28] | 0.15 | 98% | [−.31, −.30] | 0 | ||

Math Type: arithmetic | 43 | 66 | 108 | 10,912 | 6.8−45.7 | −.25*** | [−.28, −.22] | 0.09 | 63% | [−.26, −.25] | 3 | −.24*** | [−.27, −.21] |

Math Type: basic know. | 10 | 11 | 14 | 2,144 | 6.8−23.6 | −.22*** | [−.34, −.09] | 0.20 | 89% | [−.25, −.16] | 0 | ||

Math Type: grade | 38 | 58 | 84 | 38,883 | 9.4−29.8 | −.31*** | [−.35, −.26] | 0.18 | 96% | [−.33, −.31] | 8 | −.27*** | [−.32, −.22] |

Math Type: math apply | 78 | 100 | 152 | 841,335 | 5.9−49.2 | −.31*** | [−.34, −.28] | 0.16 | 99% | [−.32, −.31] | 0 | ||

Gender: males | 16 | 21 | 28 | 3,962 | 8.2−45.7 | −.20*** | [−.27, −.13] | 0.14 | 79% | [−.22, −.19] | 0 | ||

Gender: females | 16 | 21 | 28 | 4,482 | 8.1−45.7 | −.29*** | [−.35, −.22] | 0.12 | 77% | [−.31, −.27] | 3 | −.26*** | [−.32, −.19] |

Age: children | 111 | 162 | 298 | 874,275 | 5.9−20.4 | −.31*** | [−.33, −.28] | 0.16 | 99% | [−.32, −.31] | 0 | ||

Age: adults | 29 | 38 | 71 | 9,695 | 19.1−49.2 | −.25*** | [−.29, −.21] | 0.10 | 75% | [−.27, −.25] | 8 | −.22*** | [−.26, −.19] |

MA – WM | 23 | 24 | 49 | 2,695 | 7.1−24.8 | −.18*** | [−.24, −.13] | 0.10 | 61% | [−.20, −.17] | 2 | −.17*** | [−.22, −.11] |

WM Type 1: verbal | 20 | 21 | 40 | 2,279 | 7.2−24.8 | −.19*** | [−.25, −.12] | 0.12 | 65% | [−.20, −.18] | 2 | −.17*** | [−.23, −.11] |

WM Type 1: spatial | 7 | 7 | 9 | 1018 | 7.1−23.3 | −.16*** | [−.24, −.09] | 0.06 | 33% | [−.19, −.15] | 0 | ||

WM Type 2: STM | 8 | 8 | 12 | 743 | 9.5−23.6 | −.14*** | [−.21, −.07] | 0.02 | 2% | [−.17, −.12] | 0 | ||

WM Type 2: WM | 18 | 19 | 37 | 2,240 | 7.1−24.8 | −.19*** | [−.25, −.12] | 0.12 | 68% | [−.20, −.17] | 4 | −.15*** | [−.22, −.08] |

TA – MATHS | 41 | 51 | 88 | 19,733 | 8.2−32.4 | −.23*** | [−.26, −.19] | 0.13 | 88% | [−.24, −.22] | 0 | ||

Math Type: arithmetic | 6 | 8 | 13 | 1,311 | 8.2−19.1 | −.12** | [−.21, −.03] | 0.11 | 70% | [−.15, −.10] | 0 | ||

Math Type: basic know. | 1 | 1 | 1 | 246 | 8.2 | .02 | [−.11, .14] | ||||||

Math Type: grade | 14 | 17 | 28 | 9,923 | 11.6−24.5 | −.23*** | [−.29, −.16] | 0.12 | 89% | [−.24, −.20] | 0 | ||

Math Type: math apply | 26 | 31 | 46 | 10,130 | 9.1−32.4 | −.26*** | [−.31, −.22] | 0.11 | 82% | [−.28, −.26] | 0 | ||

Gender: males | 4 | 4 | 6 | 666 | 13.4 | −.14*** | [−.21, −.07] | 0.03 | 13% | [−.17, −.11] | 0 | ||

Gender: females | 4 | 4 | 6 | 468 | 13.4 | −.22*** | [−.29, −.14] | 0.00 | 0% | [−.24, −.20] | 0 | ||

Age: children | 31 | 38 | 72 | 15232 | 8.2−24.5 | −.22*** | [−.26, −.17] | 0.12 | 87% | [−.23, −.21] | 0 | ||

Age: adults | 11 | 13 | 16 | 4501 | 18.4−32.4 | −.26*** | [−.34, −.17] | 0.15 | 87% | [−.28, −.24] | 0 | ||

GA – MATHS | 12 | 16 | 38 | 10,910 | 7.6−20.4 | −.19*** | [−.27, −.11] | 0.15 | 93% | [−.22, −.18] | 1 | −.18*** | [−.25, −.10] |

MA – TA | 11 | 15 | 19 | 4,002 | 8.2−22.1 | .46*** | [.33, .57] | 0.30 | 96% | [.46, .56] | 5 | .34*** | [.19, .47] |

MA – GA | 12 | 14 | 22 | 3,633 | 7.6−22.1 | .30*** | [.19, .40] | 0.21 | 93% | [.28, .34] | 0 | ||

TA – WM | 8 | 9 | 26 | 1,142 | 10.7−35.9 | −.18*** | [−.27, −.08] | 0.13 | 75% | [−.21, −.16] | 1 | −.16** | [−.26, −.06] |

WM Type 1: verbal | 8 | 9 | 24 | 1,142 | 10.7−35.9 | −.15** | [−.24, −.05] | 0.13 | 74% | [−.17, −.12] | 0 | ||

WM Type 1: spatial | 1 | 2 | 2 | 205 | 19.9−20.6 | −.40*** | [−.51, −.28] | [−.44, −.41] | |||||

WM Type 2: STM | 2 | 2 | 3 | 208 | 19.1 | −.15* | [−.26, −.03] | [−.22, −.14] | |||||

WM Type 2: WM | 8 | 9 | 23 | 1,142 | 10.7−35.9 | −.18** | [−.28, −.07] | 0.14 | 77% | [−.21, −.16] | 2 | −.14** | [−.24, −.04] |

MATHS - WM | 23 | 24 | 86 | 2,872 | 7.1−24.8 | .32*** | [.28, .36] | 0.08 | 52% | [.32, .34] | 3 | .30*** | [.26, .34] |

*

A total of 177 studies and 250 distinct independent samples (for a total of 906,311 participants) were included in the analyses presented below. Table

Funnel plots for the main meta-analyses summarized in 8 panels (

Moderation analyses were conducted for all combinations of outcomes, for which we examined subsamples of effects, as summarized in Table

Summary of the moderation analyses

Combination | Moderator | Comparison | χ | ||||
---|---|---|---|---|---|---|---|

MA - MATHS | Math Type | 36.42 | < .001 | ||||

basic know vs arithmetic | 0.07 | 0.03 | .005 | ||||

grade vs arithmetic | 0.02 | 0.02 | .443 | ||||

math apply vs arithmetic | −0.05 | 0.02 | .004 | ||||

grade vs basic know | −0.06 | 0.03 | .062 | ||||

math apply vs basic know | −0.12 | 0.03 | < .001 | ||||

math apply vs grade | −0.07 | 0.01 | < .001 | ||||

Gender | 3.20 | .074 | |||||

males vs females | 0.09 | 0.05 | .075 | ||||

Age group | 3.95 | .047 | |||||

Dev_Group vs Adult | −0.06 | 0.03 | .047 | ||||

MA – WM | WM type of material | 0.37 | .542 | ||||

verbal vs spatial | 0.03 | 0.04 | .472 | ||||

WM type of task | 0.10 | .756 | |||||

WM vs STM | −0.02 | 0.05 | .644 | ||||

TA – MATHS | Math Type | 6.16 | .104 | ||||

basic_know vs arithmetic | 0.00 | 0.09 | .756 | ||||

grade vs arithmetic | −0.10 | 0.05 | .072 | ||||

math_apply vs arithmetic | −0.12 | 0.05 | .016 | ||||

grade vs basic_know | −0.09 | 0.10 | .317 | ||||

math_apply vs basic_know | −0.12 | 0.09 | .202 | ||||

math_apply vs grade | −0.03 | 0.03 | .294 | ||||

Gender | 2.23 | .136 | |||||

males vs females | 0.08 | 0.05 | .130 | ||||

Age group | 0.84 | .360 | |||||

Dev_Group vs Adult | 0.04 | 0.05 | .353 | ||||

TA – WM | WM type of material | 11.21 | .001 | ||||

verbal vs spatial | 0.30 | 0.09 | .001 | ||||

WM type of task | 0.52 | .471 | |||||

WM vs STM | 0.06 | 0.08 | .429 |

The overall correlation between math anxiety and math performance, calculated on 169 independent samples within 121 different studies, was ^{2} = 98%, with a

Table

As shown in Table

Finally, age group emerged as a significant moderator of the relationship between math performance and math anxiety (Table

WM measures were coded according the type of content and degree of cognitive control involved. Regarding the former, most studies used verbal working memory tasks (20 studies totaling 21 independent samples) rather than spatial working memory tasks (7 studies totaling 7 independent samples). However, type of material had virtually no impact on the correlation of interest (Table

Similarly, involvement of cognitive control (short-term memory vs working memory) did not have a significant effect either on the relationships of interest (Table

Finally, a negative relationship emerged between MA and overall WM measures, although this link was relatively weak, ^{2} = 61%, with

The overall correlation between test anxiety and math achievement was notably weaker than the correlation between math anxiety and math achievement, ^{2} = 88%), with τ = 0.13, pointing to potential moderating factors behind the relationship between text anxiety and math achievement. Nonetheless, moderation analysis (Table

The overall correlation between test anxiety and working memory was

As explained above in the “Analytic Strategy” section, we examined the mediating role of WM using two alternative strategies. With the first, which was limited to studies reporting all variables of interest, we considered random effects. With the second, which was based on a Monte Carlo simulation approach, we sought to incorporate as much information as possible.

For the first strategy, our meta-analytic dataset had 15 independent samples reporting complete information for the mediation MA ➔ WM ➔ Math, and 6 independent samples reporting complete information for the mediation TA ➔ WM ➔ Math. The meta-analytic estimates, together with their 95% CIs, are reported in Fig.

Mediation analysis on the relationship between math anxiety (MA), working memory (WM), and mathematical achievement (Maths), and on the relationship between test anxiety (TA), WM, and Maths. Panel (A) shows the estimates obtained via meta-analytic multivariate regression based on samples with complete correlations. Panel (B) shows the estimates obtained via Monte Carlo simulation of multivariate regression based on meta-analyzed correlations. All estimates are standardized regression coefficients

Regarding the second strategy, the estimates are reported in Fig.

As reported above, the relationship between MA and math achievement appeared stronger than the relationship between TA and math achievement, even though both correlations were of medium strength. To formally test whether one correlation was stronger than the other, we conducted an additional meta-regression in which the correlations (involving MA or TA on one side, and math achievement on the other side) were examined in the same model. As in all previous cases, the correlations were combined by independent sample using the formula suggested by Borenstein et al. (^{2}(1) = 14.09,

The present meta-analysis examined the relationship between different forms of academic anxiety (MA and TA) and math performance, taking into account the impact of potential moderators, such as gender, age, and type of math and WM tasks.

We confirmed a moderate, negative relationship between MA and math performance (

In line with the previous literature (Devine et al.,

These results highlight how MA and TA share some risk factors (e.g. they both displayed effects trending in the same direction and were both negatively associated with academic success). At the same time, the results showed some specificity, as confidence intervals showed no overlap whatsoever, allowing us to make inferences about the stronger effect of MA on math performance. Similarly, Donolato et al. (

In contrast to TA, where the impact of this type of anxiety has been investigated mainly in relation to general academic examinations (see Putwin 2008), a large body of research has demonstrated that MA negatively impacts math performance specifically. For example, individuals with high MA perform worse than their less math-anxious peers on basic numerical abilities, such as counting and comparing numbers (Maloney et al.,

The effect of demographic variables, such as gender and age, have been more extensively considered as important source of variance in both forms of academic anxiety. In general, girls consistently reported significantly more anxiety than boys in terms of both MA (Ferguson et al.,

Further research is needed in order to clarify why female students frequently report higher level of anxiety compared to their male counterparts. Moreover, it is worth noting that in the present meta-analysis, we had a limited number of reviewed studies (

In general, it can be stated that almost all students are challenged with some form of academic anxiety during their school years, and age may affect the experience and expression of this anxiety (Mammarella et al.,

The influence of age on TA is less consistent across studies. Hembree’s (

Inconsistencies across studies are likely mirrored by cultural or educational aspects. When considering school performance, we should also take into the account the intrinsic variability of educational paths, especially in secondary education, where the frequency of math courses/lessons (or other STEM subjects, in general) can vary considerably. Another source of inconsistency could be represented by use of different tests/self-reporting in different populations. For example, self-reporting measures specifically developed for assessing primary school children could emphasize the emotionality (affective) dimension of anxiety (Lowe et al.,

Another core theme is related to the causal pattern between math performance and academic anxiety. As children gain more experience of mathematical success and failure (or testing situations in general), academic anxiety may increase in those students whose poor performance results in repeated experiences of failure, but not to the same extent in pupils who experience greater success in mathematics (Carey et al.,

There are theoretical reasons to believe that changes in the relationships between academic anxiety and math performance could reflect developmental changes in cognitive resources, and in the ways in which they are used in mathematical situations. Research on WM suggests that performance deficits caused by anxiety can be generally summarized by the extent to which individuals are able to use their WM capacity (Darke,

Our results revealed a weak but stable relationship (− .14 <

This work leaves another question open, namely how differences in the levels of WM may or may not influence the relationship between WM and MA/TA. There are indeed conflicting results on the negative impact of MA on math performance according to individual levels of WM resources. Some studies found that high-MA students with low working memory capacity are more predisposed to poor math performance (Ashcraft & Kirk,

The current meta-analysis provides an up-to-date synthesis on the relationship between MA/TA and math performance. Our results confirm that both forms of academic anxiety (MA and TA) are negatively related to math achievement not only in school age individuals, but also in adulthood. In terms of MA, the strength of this connection is strongly influenced by the type of math task and by gender differences; however, this influence is lower in TA studies. In addition, WM proved to be a stable, albeit weak, mediator in the relationship between both academic anxiety forms and math performance, confirming the key role of this cognitive construct. There was no substantial evidence on whether the type of WM components or the level of cognitive control can influence the relationship. Future studies are needed to clarify this point, as well as to map any potential developmental changes in the relationship of academic anxiety and math performance.

SC was supported by the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 700031.

The data and associated files for this meta-analysis are available at

Open access funding provided by Università degli Studi di Padova within the CRUI-CARE Agreement.

The authors declare no competing interests.

^{*}Henry, A. (2017).

(PDF 40 kb)

(PDF 907 kb)

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.