Family Socioeconomic Status and Child and Adolescent Developmental Outcomes: A Review and Meta-Analysis

1. Introduction

Family socioeconomic resources shape children's developmental contexts through material investment, parental stress, parenting practices, and neighbourhood conditions [1,2,3]. A prior meta-analysis of composite SES measures reported very small to small associations with aggression, internalizing behaviour and language/literacy, and concluded that family- and community-level processes may explain more variance than SES itself [4]. Primary studies since then have emphasised mediators and moderators, including parenting, temperament, community poverty and acculturation [5,6,7,8].

The way in which SES impacts development can be roughly described in one of two ways. The first (investment) account rests on the fact that families with higher income and education can invest in child learning: child-relevant experiences, learning materials, nutrition, health care, and more stable homes. The second (stress) and (parenting) account rests on the fact that economic hardship creates more parental stress and conflict, and that this stress and conflict can cascade to affect warm, consistent, sensitive parenting and child adjustment (e.g., [2,3]). Whether the dominant pathway is one or the other may vary across outcomes: investment resources are most directly relevant to learning and cognition, parenting processes to emotional and behavioural adjustment. Either way, SES refers to a distal index that is not determinative of outcomes, and this may partly explain why the relation is moderate.

The measurement of SES introduces yet another layer of variability. Income, education, and occupation are related but in different ways (families with similar income can differ in education and occupation, and vice versa). SES composites integrate various resources in different combinations [2]; for example, income may have a more direct influence through material resources than parent education does through information and behavior. Combining composites and other measures of risk then may produce estimates with a nuanced mix of correlations.

The developmental period is also likely to be important. Early childhood is a period of rapid development in language, social skills and regulation, with family life being the dominant influence on children (and to a lesser extent adolescents), while over adolescence other influences such as peers, school and neighbourhoods may be more influential. Therefore, associations between SES and child and adolescent developmental outcomes may well be different at early childhood and adolescent stages. Moreover, review studies often conflate large age ranges and have at best focused on negative aspects of psychosocial and socio-emotional development (such as problems or deficits), despite such competence and well-being valid and important domains in their own right.

Evidence on positive developmental trajectories, such as socio-emotional competence and psychological well-being in early childhood, is comparatively scarce, and most SES research has focused on problem behaviours [9]. The present study therefore had two aims: first, to evaluate whether the twelve primary studies identified through a systematic database search could inform a review of family income and positive early-childhood development; and second, because the corpus could not, to quantify what it does show about SES and the outcomes actually measured.

In this regard, the purpose of this review was to determine the magnitude and direction of the association between SES and children's development and whether this association varied according to outcome domain and age.

The models' intended scope was delineated by four protocol criteria: Population: children 0-10 years; Exposure: family income or an SES composite (parent education, occupation, income, or the alternative combinations); Outcomes: positive developmental outcomes, measured as socio-emotional competence or psychological well-being; Design: the primary quantitative observational or experimental studies eligible.

Hypotheses were generated in the expected direction of the extant literature [2,3,4]. For the protocol specified question, we hypothesised higher family income or SES would be related to higher socioemotional competence and psychological well-being; as below this could not be tested. For the post-hoc synthesis, we hypothesised (H1) higher SES was related to fewer externalising/antisocial problems (pooled r < 0); (H2) higher SES was related to fewer internalising/depressive symptoms (pooled r < 0); (H3) higher SES was related to better language/literacy outcomes (pooled r > 0); (H4) that the direction and magnitude of the association would differ by outcome domain; and (H5) that between-study heterogeneity would be partially attributable to child age and publication year. Small effects were expected based on the recent meta-analytic review [4]. Since the chosen outcome families are adopted after the protocol outcome was untestable, and this review was not registered, these hypotheses are considered exploratory.

2. Materials and Methods

2.1. Search Strategy

The synthesis followed PRISMA 2020 principles where applicable [10] (Supplementary Table 1).

Three databases were searched: Scopus (TITLE-ABS-KEY), Web of Science (Topic) and PubMed/MEDLINE (MeSH and title/abstract terms). Each search combined four concept blocks: 1) socioeconomic status, poverty or family income; 2) child, adolescent or youth; 3) outcome terms (externalizing, internalizing, antisocial behaviour, aggression, depression, anxiety, language, literacy); and 4) early-childhood terms (early childhood, preschool, kindergarten, young children, toddler, infant). The Scopus and Web of Science searches were limited to journal articles in English; the PubMed/MEDLINE search was limited to English-language human studies indexed in MEDLINE with free full text, published 2000–2026. The searches returned 7,662 (Scopus), 7,009 (Web of Science) and 2,946 (PubMed/MEDLINE) records, 17,617 in total. The full syntax is given in Supplementary Table 2. The outcome block covered problem behaviours and language/literacy; it contained no terms for positive developmental outcomes such as socio-emotional competence or well-being.

2.2. Eligibility Criteria

The pre-specified protocol required (a) children aged 0–10 years, (b) family income or composite SES as the exposure, (c) a positive developmental outcome (socio-emotional competence or psychological well-being), and (d) primary quantitative observational or experimental data. Reports were screened independently against these criteria by one reviewer (a second independent reviewer is required before submission). Records were de-duplicated and screened by title and abstract, and the full texts of potentially eligible reports were retrieved and assessed, with reasons for exclusion recorded. Because no report satisfied criterion (c), reports were retained for a post-hoc synthesis of three outcome families: externalizing/antisocial behaviour, internalizing/depressive symptoms, and language/literacy. This deviation from protocol is a principal limitation.

2.3. Data Extraction

For each report we extracted design, country, sample size, age, SES measure, outcome measure and informant, and the SES–outcome statistic with its source table. Minus signs and values were verified against page images. Effects were expressed as Pearson r with the sign defined so that positive values indicate that higher SES accompanies higher scores. Poverty and risk scores [6,11] were sign-reversed. Where only group means were available, Cohen's d was converted to r [12]; for ordinal consumption groups, r was derived from group means, standard deviations and sizes [13]. Where multiple SES indicators or outcomes within a family were reported, the study-level effect was the mean Fisher z. Subgroup correlations (girls/boys [8]) were combined by inverse-variance weighting. Korat and Haglili [14] reproduces demographic data identical to Korat et al. [12] and was treated as an overlapping sample. It is displayed in the forest plots and contributes to the overall, language/literacy, meta-regression and funnel-plot estimates, which should therefore be read as exploratory; it did not contribute to the externalizing or internalizing analyses.

2.4. Risk of Bias

All included reports were observational (cross-sectional or prospective cohort); Cochrane RoB 2 applies to randomised trials and was therefore not applicable. Quality was rated with the Newcastle–Ottawa Scale adapted for cross-sectional studies (maximum 10 points: selection 0–5, comparability 0–2, outcome 0–3). Comparability was scored 0 throughout because bivariate correlations were extracted. Ratings are provisional (single rater) and require independent duplicate rating.

2.5. Statistical Analysis and Software

Correlations were Fisher z-transformed (variance 1/(n − 3)) and pooled with random-effects models using restricted maximum likelihood (REML) for τ². Confidence intervals were additionally computed using the Hartung–Knapp–Sidik–Jonkman adjustment [15]. Heterogeneity was summarised by Q, I² and τ² [16], with 95% prediction intervals. Robustness was examined by leave-one-out analysis and by omitting studies with uncertain effect direction [17] or a mixed SES composite [11]. Pooling was performed overall across all effect estimates and by outcome domain (externalizing, internalizing, language/literacy), with a between-subgroup test of heterogeneity (Qb). Mixed-effects meta-regression (REML) examined publication year and mean/midpoint age as moderators. Small-study effects were examined with a funnel plot of all effect estimates and Egger's regression test [18], interpreted cautiously because the domain-specific analyses include fewer than ten estimates and the domains differ in effect direction. Stata version 17 was used for all analyses and meta-analytic figures.

3. Results

3.1. Study Evaluation and Characteristics

The database searches identified 17,617 records (Scopus 7,662; Web of Science 7,009; PubMed/MEDLINE 2,946). After removal of 11,583 duplicates, 6,034 records were screened by title and abstract and 5,791 were excluded. Of 243 reports sought for retrieval, 189 could not be retrieved and 54 were assessed for eligibility; reasons for exclusion are given in Figure 1. Twelve primary reports were included (Figure 1). Five reports had samples with mean age ≤ 10 years, but none measured positive developmental outcomes; seven concerned mainly adolescent or mixed 6–18-year samples. Nine reports contributed to the externalizing and/or internalizing analyses, and three reports (two independent samples) to the language/literacy outcomes. Table 1 gives study characteristics and extracted effects.

3.2. Risk of Bias

Provisional NOS scores ranged from 3 to 7 of 10 (median 5; Table 2). The population-based TRAILS cohort [8] and the national panel [6] scored highest; the small extreme-group literacy studies scored lowest. Full domain-level ratings and rationale are in the supplementary workbook.

3.3. Overall Association and Differences Between Outcome Domains

Across all 17 effect estimates from the 12 reports (Figure 2), the random-effects pooled correlation was r = 0.01 (95% CI [−0.06, 0.09]; z = 0.32, p = 0.75), with very high heterogeneity (Q(16) = 109.69, p < .001, I² = 91.35%, H² = 11.57, τ² = 0.02). This overall estimate is descriptive only: it combines problem-behaviour outcomes, for which a negative correlation indicates a benefit of higher SES, with language/literacy outcomes, for which a positive correlation indicates a benefit and it includes the overlapping Korat and Haglili [14] sample. Outcome domain accounted for a significant share of the heterogeneity (test of group differences Qb(2) = 35.20, p < 0.001; Figure 3), with the language/literacy subgroup clearly separated from the externalizing and internalizing subgroups.

3.4. SES and Externalizing/Antisocial Outcomes

Across eight studies (N = 6,245), the pooled correlation was r = −0.04 (95% CI [−0.11, 0.03], p = 0.22; HKSJ 95% CI [−0.12, 0.04]) indicating no statistically reliable association (Figure 3, EXT panel). Heterogeneity was substantial (Q = 42.8, df = 7, p < .001, I² = 84%, τ² = 0.0073); the 95% prediction interval [−0.26, 0.18] spans zero. Domain-specific leave-one-out estimates ranged from −0.06 to −0.02 (not plotted). Omitting Appleyard et al. [17], whose SES index direction is uncertain, gave r = −0.06 (95% CI [−0.12, 0.01]). Egger's test was non-significant (p = 0.17).

3.5. SES and Internalizing/Depressive Outcomes

Across six studies (N = 2,284), the pooled correlation was r = −0.04 (95% CI [−0.12, 0.05], p = 0.39; HKSJ 95% CI [−0.15, 0.08]) with moderate-to-substantial heterogeneity (Q = 16.9, df = 5, p = 0.005, I² =73%, τ² = 0.0076; Figure 3). The prediction interval was [−0.30, 0.23]. Omitting Dallaire et al. [11], whose risk composite includes sex and ethnicity, gave r = −0.00 (95% CI [−0.06, 0.06]). Egger's p = 0.16.

3.6. SES and Language/Literacy Outcomes

Two independent samples reported language or literacy outcomes: Korat et al. [12] (r = 0.36, d = 0.78) and Noble et al. [21] (PPVT r = 0.24; Word Attack r = 0.06; N = 38). Figure 4 also displays Korat and Haglili [14] (r = 0.46), which overlaps with Korat et al. [12]. The random-effects pooled correlation across the three displayed estimates was r = 0.38 (95% CI [0.26, 0.49]; z = 5.79, p < .001), with no heterogeneity (Q(2) = 1.56, p = 0.46, I² = 0%, τ² = 0.00). Because this estimate counts one sample twice and only two samples are independent, it is exploratory and likely overstates precision. Direction was consistently positive, in line with an SES gradient in language, but individual estimates were imprecise (for example, Noble et al. [21]: 95% CI [−0.09, 0.52]).

3.7. Small-Study Effects

The funnel plot of all 17 effect estimates (Figure 5) was not symmetric. Externalizing and internalizing estimates clustered around the pooled value (REML estimate ≈ 0.01 in Fisher's z), whereas the three language/literacy estimates (z ≈ 0.25–0.49), all with larger standard errors (> 0.10), lay to the right and outside the pseudo-95% confidence limits. This pattern is more plausibly explained by differences between outcome domains (Figure 3) than by selective publication, because the plot pools outcomes with different expected directions. Domain-specific Egger tests were non-significant; Egger's test result for all 17 estimates: insert from Stata. With few estimates per domain these checks are underpowered, and trim-and-fill was not performed.

3.8. Meta-Regression

Meta-regression of effect size on publication year (mean-centred; Figure 6) produced an almost flat fitted line, and the 95% confidence band was compatible with no trend across the 2001–2008 publications insert coefficient, 95% CI and p from Stata. Meta-regression on mean/midpoint age (mean-centred; Figure 7) showed a downward slope: effect estimates were most positive in the youngest samples (fitted Fisher's z ≈ 0.14 at about 6 years) and became negative in the oldest (fitted z ≈ −0.08 at about 16 years), with a confidence band that excluded zero only at the youngest end insert coefficient, 95% CI, p and R² from Stata. This age gradient is confounded with outcome domain, because the youngest samples are those that measured language/literacy and showed positive correlations, whereas older samples measured problem behaviours; with 17 estimates it cannot be interpreted as an independent age effect.

3.9. Sensitivity Analysis

Leave-one-out analysis of the overall model (Figure 8) gave pooled correlations between −0.02 and 0.03, with every 95% confidence interval including zero (p = 0.51 to 0.90). Omitting either Korat study turned the pooled estimate slightly negative (Korat and Haglili [14]: r = −0.02, 95% CI [−0.08, 0.05]; Korat, Klein and Segal-Drori [12]: r = −0.01, 95% CI [−0.08, 0.06]), whereas omitting Dallaire et al. [11] gave the largest positive estimate (r = 0.03, 95% CI [−0.05, 0.10]). No single study therefore changed the substantive conclusion, although the overall estimate depends on the positive language/literacy correlations offsetting the negative problem-behaviour correlations.

4. Discussion

In this corpus, composite SES showed small and statistically non-significant pooled associations with externalizing and internalizing outcomes, with considerable between-study heterogeneity. The direction of the point estimates (higher SES, fewer problems) is consistent with the broader literature, but the wide prediction intervals mean that a true association of either sign cannot be excluded for a new setting. These results are compatible with the conclusion that SES operates through family-, parenting- and community-level mechanisms rather than as a strong direct correlate [4,5,6,8]. Moderation findings in the primary studies, for example by temperament, gender, community poverty and age, further argue against a single uniform effect.

The pooled estimates here are smaller than the Hedges' g values reported by Letourneau et al. [4], which is expected given a different study set, different effect-size extraction (bivariate correlations versus converted group contrasts), and exclusion of studies that did not report an SES–outcome statistic. They should not be read as contradicting that review.

The language/literacy results are consistent with an SES gradient in early literacy, and the test of group differences shows that SES associations differ markedly by outcome domain. However, this estimate rests on two small independent samples, one of which is counted twice, and the near-zero overall correlation across domains largely reflects opposite-signed domain effects and should not be read as an absence of association. The age meta-regression suggests more positive correlations in younger samples, but this is confounded with outcome domain and cannot be separated from it with the present data; the flat publication-year trend gives no indication that associations changed across 2001–2008. The funnel plot was asymmetric, mainly because of the small language/literacy studies; this may reflect true domain differences or small-study effects, which cannot be distinguished with so few estimates.

The results for language/literacy require caution not only because the number of samples was small. The two independent samples used different approaches: Korat et al. [12] used a comparison of groups of low and high SES to produce an extreme group difference in mean, and the standardised mean difference may not be equivalent to r and likely to be over-estimated; the step from this to r over the SES range, in contrast, is likely to be larger; Noble et al. [21] used several SES indicators; in a very small sample (N = 38). Within Noble et al. [21], the r for vocabulary (PPVT r = 0.24) was over twice that for decoding (Word Attack r = 0.06), consistent with SES being more strongly related to oral language skills than to decoding, although again only suggestive because of the single small sample. Both Korat samples focused on maternal talk to children in book reading and the home environment - plausible proximal pathways of SES influences - but these are not tested as mediators here.

4.1. Implications for Future Research

We should pre-register a review protocol prior to searching (e.g. PROSPERO, Open Science Framework), screen and extract data in duplicate, include outcome terms for the positive development (e.g. Socio-emotional competence, social skills, self-regulation and well-being), implement the 0-10-year age restriction in the search itself [22] and obtain full-text articles through library access or interlibrary loan rather than limiting to free-full text records. When the number of studies permits, meta-analyses could disaggregate family income from education and occupation indicators, extract adjusted and unadjusted estimates, and model effect sizes within meta-analyses, within families, within studies (e.g. Multilevel models, robust variance estimation), rather than average effects across a set of outcomes. Primary studies would be more informative if authors provided statistics for continuous indicators of SES and outcome measures, and adopted longitudinal designs that track children from early ages. Since pooled associations were modest and prediction intervals largely included no association, these findings can hardly be accepted as evidence that family income or parental education has little influence on development, and certainly should not be used for policy. The main contribution of this review is its descriptive and methodological focus: highlighting evidence deficits, noting protocol deviations, and demonstrating how unstable and imprecise pooled estimates are, due to a small, diverse sample of studies.

4.2. Limitations

First, and most importantly, none of the twelve studies measured positive developmental outcomes, and most samples were adolescent; the corpus cannot answer questions about family income and positive developmental trajectories in early childhood. Second, the synthesis is post-hoc and not registered. Although studies were identified through searches of three databases, the search strategy contained no terms for positive developmental outcomes and included adolescent and youth terms despite the 0–10-year age criterion; only 12 of 6,034 screened records were included, 189 of 243 reports sought could not be retrieved, and the PubMed/MEDLINE search was restricted to free full-text MEDLINE records from 2000 onwards, so selection bias is likely. Third, extracted effects are bivariate and unadjusted, and several SES composites mix income, education, occupation and non-SES indicators. Fourth, effect direction for one SES index could not be confirmed, a risk composite included sex and ethnicity, and two effects were derived from grouped data, and one overlapping sample [14] is included in the overall, language/literacy, meta-regression and funnel-plot estimates. Fifth, all studies were cross-sectional or short-term observational, precluding causal inference. Sixth, quality ratings were single-rater and provisional. Seventh, k was small, limiting heterogeneity, meta-regression and publication-bias assessment; the funnel plot was asymmetric, the overall estimate pools outcomes with opposite benefit directions, and age and outcome domain are confounded in the meta-regression.

5. Conclusions

SES was weakly and inconsistently associated with externalizing and internalizing outcomes in this set of studies, and positively associated with language/literacy in two small independent samples. A registered systematic review with dual screening, a search strategy that includes terms for positive outcomes (socio-emotional competence, well-being), and a sample restricted to children aged 0–10 years is needed to evaluate family income and positive developmental trajectories.