Socioeconomic changes predict genome-wide DNA methylation in childhood

Abstract

Childhood socioeconomic position (SEP) is a major determinant of health and well-being across the entire life course. To effectively prevent and reduce health risks related to SEP, it is critical to better understand when and under what circumstances socioeconomic adversity shapes biological processes. DNA methylation (DNAm) is one such mechanism for how early life adversity ‘gets under the skin’. In this study, we evaluated the dynamic relationship between SEP and DNAm across childhood using data from 946 mother–child pairs in the Avon Longitudinal Study of Parents and Children. We assessed six SEP indicators spanning financial, occupational and residential domains during very early childhood (ages 0–2), early childhood (ages 3–5) and middle childhood (ages 6–7). Epigenome-wide DNAm was measured at 412 956 cytosine-guanines (CpGs) from peripheral blood at age 7. Using an innovative two-stage structured life-course modeling approach, we tested three life-course hypotheses for how SEP shapes DNAm profiles—accumulation, sensitive period and mobility. We showed that changes in the socioeconomic environment were associated with the greatest differences in DNAm, and that middle childhood may be a potential sensitive period when socioeconomic instability is especially important in shaping DNAm. Top SEP-related DNAm CpGs were overrepresented in genes involved in pathways important for neural development, immune function and metabolic processes. Our findings highlight the importance of socioeconomic stability during childhood and if replicated, may emphasize the need for public programs to help children and families experiencing socioeconomic instability and other forms of socioeconomic adversity.

Introduction

Socioeconomic position (SEP) is a fundamental determinant of health and disease across the lifespan (1). As defined by Krieger et al. (1997) (2), SEP is an ‘aggregate concept’ composed of diverse components of economic and social well-being across individual-, household- and neighborhood-level domains, including both resources (e.g. weekly income) and rank-based characteristics (e.g. occupational prestige). SEP therefore can be measured across time by various indicators, like job stability, ability to afford basic household needs and neighborhood quality, which are known to play related, yet distinct roles in health and life outcomes (3–5).

Dozens of observational and quasi-experimental studies examining these indicators have shown that children growing up in low SEP families have increased risk for both short- and long-term cognitive, socioemotional, behavioral and physical/mental health deficits compared to their high SEP counterparts (6–9). Some of these SEP-related disparities are evident very early in development, starting shortly after birth (10–13). Yet, the biological mechanisms that explain these well-established SEP and health relationships remain relatively unknown, limiting our ability to disentangle specific pathways of pathophysiology and design targeted interventions.

In the past two decades, epigenetic studies have exploded as a means of potentially unraveling the biological pathways through which SEP ‘gets under the skin’. Most epigenetic studies have focused on DNA methylation (DNAm) (14), which occurs when methyl groups are added to cytosines in the DNA sequence, typically within cytosine-guanine (CpG) dinucleotides (15). These DNA modifications do not alter the sequence of the genome, but can influence how genes are expressed in ways that can have important short and long-term health consequences (16).

Recent reviews summarizing the effects of SEP on epigenetic patterns suggest that SEP is linked to DNAm differences in childhood and adulthood (17–19). In fact, over 30 studies have found a relationship between childhood SEP and DNAm. However, less than a quarter of these studies were longitudinal by design (i.e. including repeated measures of SEP exposure across time). Further, less than half were epigenome-wide association studies (EWAS) analyzing SEP-related DNAm variations. In one recent comprehensive review of the SEP-DNAm literature, the number of significant, SEP-associated CpGs reported across prior EWAS studies ranged from 1 to 2546 (median = 10), yet relatively no consistent patterns in SEP-associated DNAm changes emerged between studies (see Cerutti et al. (19)). One possible explanation for these mixed results is that studies have conflated both the type of SEP indicator measured and the timing of SEP measurement (19). Indeed, few studies have investigated the effects of SEP type and/or timing on DNAm, even though it is well known that both features of SEP can influence the extent of its impact (20).

Prior studies that have analyzed the associations between multiple types of SEP indicators and DNAm have found little to no overlap in DNAm changes across SEP measures (21–23), suggesting that different SEP indicators may result in distinct biological signatures and subsequent cascading health risks. Yet, it remains relatively unknown whether exposure to distinct SEP indicators (e.g. low household income vs. neighborhood disadvantage) during childhood impacts later DNAm to a similar extent.

Even fewer studies have investigated the impact of SEP timing on DNAm, likely because it is difficult to collect multiple, repeated measures across time in large, epigenetic datasets. In some notable exceptions, studies comparing the time-dependent effects of childhood SEP (24–27) on DNAm have found timing differences with respect to SEP’s impact, consistent with the idea that there may be sensitive periods of elevated plasticity during childhood when adversity-induced biological changes are most likely to occur. However, whether different aspects of the socioeconomic environment across developmental stages differentially influence DNAm remains largely unexplored.

The current study aimed to address this gap by utilizing a large, longitudinal birth cohort with multiple, repeated measures of socioeconomic-related hardships assessed prospectively across childhood before epigenome-wide DNAm collection at age 7. We specifically sought to assess how different indicators of the socioeconomic environment (e.g. neighborhood quality, job loss, low household income) measured repeatedly across the first seven years of life associated with child epigenetic alterations. Given that different socioeconomic domains may impact health via related, but distinct pathways (4, 28), we analyzed exposure to seven distinct socioeconomic-related hardships. Additionally, because socioeconomic adversity could have multiple time-varying effects on DNAm, we tested three commonly examined hypotheses from the life-course epidemiology literature (29) to evaluate the circumstances under which childhood socioeconomic adversity associates with DNAm changes at age 7: 1) accumulation hypothesis, where the impact of low SEP increases with the number of time periods exposed, regardless of when it occurs; 2) sensitive period hypothesis, where the impact of low SEP is larger in magnitude during a certain developmental period compared to any other; and 3) mobility hypothesis, where the impact of SEP on DNAm is driven by an upward or downward change in SEP between adjacent developmental time periods.

Uncovering the dynamic relationships between SEP and DNAm across childhood will not only highlight the biological mechanisms driving the effects of SEP on long-term health, but also will offer clearer insights to guide targeted interventions aimed at reducing the negative consequences of socioeconomic-related adversity in childhood.

Results

Sample characteristics and prevalence of socioeconomic adversity

We analyzed data from 946 mother–child pairs from the Avon Longitudinal Study of Parents and Children (ALSPAC), a longitudinal birth-cohort in the United Kingdom (UK). Children included in our analytic sample were mostly White (97.1%) and from both sexes (49.9% female) (Supplementary Material, Table S1). Among the six SEP indicators analyzed (i.e. job loss, income reduction, low family income, financial hardship, major financial problem and neighborhood disadvantage), job loss was the least reported socioeconomic adversity (11.5% ever-exposed), and income reduction was the most common (73.8% ever-exposed) (Table 1). The prevalence of all adversities decreased over time (Table 1, Supplementary Material, Fig. S1). The six SEP indicators were moderately correlated with each other during all three childhood periods (Supplementary Material, Fig. S2), suggesting they captured distinct aspects of the socioeconomic environment.

Table 1.

Prevalence of exposure to socioeconomic adversity by developmental period in the ARIES analytic sample

	Job loss (N = 667)	Income reduction (N = 711)	Low family income (N = 619)	Financial hardship (N = 697)	Major financial problem (N = 710)	Neighborhood disadvantage (N = 687)
Very early childhood (0–2 years)	42 (6.3%)	458 (64.4%)	95 (15.4%)	127 (18.2%)	138 (19.4%)	83 (12.1%)
Early childhood (3–5 years)	32 (4.8%)	220 (30.9%)	79 (12.8%)	46 (6.6%)	69 (9.7%)	36 (5.2%)
Middle childhood (6–7 years)	18 (2.7%)	134 (18.9%)	55 (8.9%)	29 (4.2%)	60 (8.5%)	29 (4.2%)
Ever-exposeda	77 (11.5%)	525 (73.8%)	130 (21.0%)	147 (21.1%)	184 (25.9%)	98 (14.3%)
Average correlation over timeb	0.49	0.34	0.87	0.70	0.50	0.80

The first four rows present the number (%) of children who were exposed to the specific type of socioeconomic adversity at each developmental period or ever exposed throughout the three periods.

^aChildren who were exposed during at least one period were defined as ever-exposed for the specific type of socioeconomic adversity.

^bPolychoric correlations are presented, characterizing the average correlation over time within the given type of exposure. The average within-SEP correlations were moderate to high, suggesting these measures were variable across time, which allowed for detecting differences across periods. Exposures with correlations in excess of 0.90 typically cannot be used in the SLCMA.

Childhood socioeconomic adversities were associated with differential DNAm at 62 CpGs

We next examined possible time-dependent associations between each of the SEP indicators and DNAm at individual CpGs using a two-stage structured life-course modeling approach (SLCMA) (30–32), which identified the life-course hypothesis most supported in the observed data and estimated the magnitude of associations. In this and the following three sections, we summarize 1) the top CpGs associated with socioeconomic adversity, 2) the most selected life-course hypotheses, 3) the robustness of findings evaluated through a variety of sensitivity analyses and 4) the biological relevance of findings.

We identified 62 CpGs where exposure to socioeconomic adversity explained more than 3% of the variance in DNAm (R² > 3%, Supplementary Material, Table S2). Most of the 62 CpGs were linked to the two least commonly-reported adversities in this sample: neighborhood disadvantage (17 CpGs) and job loss (15 CpGs, Table 2). Only four of the 62 CpGs identified using the R² cutoff also passed a false discovery rate (FDR) < 0.05 significance threshold, all of which were associated with neighborhood disadvantage (Table 2).

Table 2.

Summary of the SLCMA results for the 62 CpGs with R² > 3%

Adversity	Number of R2 > 3% CpGs	Range of R2	Range of (P-values	Number of FDR < 0.05 CpGs
Neighborhood disadvantage	17	3.0–4.2%	1.3 × 10−7—7.1 × 10−6	4a
Job loss	15	3.1–3.7%	5.8 × 10−7—8.8 × 10−6	-
Low family income	13	3.0–3.8%	1.7 × 10−6– 2.5 × 10−5	-
Financial hardship	9	3.0–3.7%	5.9 × 10−7– 8.5 × 10−6	-
Major financial problem	5	3.0–3.8%	2.6 × 10−7– 4.7 × 10−6	-
Income reduction	3	3.0–3.3%	1.5 × 10−6– 4.5 × 10−6	-

The R² values reflect the increase in the variance of DNA methylation explained by the first hypothesis chosen after accounting for covariates. P-values were calculated using selective inference, which assesses the significance of the increase in R² explained. See Supplementary Material, Table S2 for the full list of the 62 CpGs. SLCMA = structured life-course modeling approach.

^aFour CpGs for neighborhood disadvantage passed an FDR < 0.05 significance threshold: cg20102336, cg08638097, cg23405172 and cg14212190.

Of note, 61 of these CpGs showed the same direction of effect as that reported in at least two prior EWASs examining SEP and DNAm. Furthermore, 17 out of 62 (27%) CpGs showed at least a nominal (P < 0.05) association in at least two prior EWASs. Of these 17 CpGs, two (cg23685969 and cg19260606) exceeded a statistical significance threshold of FDR < 0.05 in at least one prior EWAS (Supplementary Material, Table S3, Supplementary Material, Fig. S3).

Mobility and sensitive period hypotheses were most often selected

The SLCMA allowed us to determine which of the following three life-course hypotheses were most supported in the observed data: accumulation, sensitive period and mobility (Fig. 1). Of the life-course hypotheses we tested, mobility and sensitive period effects showed the strongest associations with DNAm (Fig. 2A).

Figure 1.

Study design and the conceptual life-course models used in the structured life-course modeling approach (SLCMA). (A) Measurement of childhood socioeconomic adversity (X) and DNA methylation (DNAm) over time (T). Exposure to socioeconomic adversities, or indicators of low socioeconomic position (SEP), was measured repeatedly across three childhood periods: very early (0–2 years, T₁), early (3–5 years, T₂) and middle childhood (6–7 years, T₃). DNAm was measured around age 7. (B) Illustration of the life-course hypotheses tested in the SLCMA, the least angle regression (LARS) variable selection procedure and selective inference test. Accumulation, sensitive period and mobility hypotheses were examined in this study. Accumulation assumes that the effect of low SEP increases with the number of exposed periods. Sensitive period assumes that low SEP is particularly impactful during one of the three time periods. Mobility assumes that changes in SEP across specific periods are particularly impactful. Early worsening and early improvement refer to adversity getting worse (↑SEP, i.e. increase in exposure) or better (↓SEP, i.e. decrease in exposure) from very early to early childhood, respectively; later worsening and later improvement refer to adversity getting worse or better from early to middle childhood, respectively. For each socioeconomic adversity, hypotheses were encoded into variables and then entered into the LARS variable selection procedure to identify the one explaining the most variability in DNAm at age 7 at each CpG site. We then performed post-selection inference to test the association between the selected variable and DNAm as well as estimate confidence intervals. See Supplemental Methods for more details about SLCMA.

Figure 2.

Mobility and sensitive period hypotheses were most often selected among the top 62 CpGs linked with socioeconomic adversity (or socioeconomic position, SEP) that explained > 3% variance in DNA methylation (DNAm). (A) Frequency at which each life-course hypothesis was selected among the 62 CpGs. For job loss and income reduction, we tested accumulation and sensitive period hypotheses, and middle childhood was the most selected hypothesis. For the other four socioeconomic adversities, we tested accumulation, sensitive period and mobility hypotheses. Mobility hypotheses, specifically worsening SEP, were most selected. Very early, Early and Middle refer to sensitive period hypotheses related to the three childhood periods: very early (0–2 years), early (3–5 years) and middle childhood (6–7 years). Early worsening/improvement refers to mobility hypotheses for changes between very early and early childhood, and later worsening/improvement refers to mobility hypotheses for changes between early and middle childhood. (B) For the four CpGs associated with neighborhood disadvantage at an FDR < 0.05, SEP mobility group implied by the selected mobility hypothesis showed the greatest shift in DNAm. The distribution of DNAm by SEP mobility group is shown in boxplots, where the center line indicates the median, box limits indicate the 25th and 75th percentiles, whiskers extend up to 1.5 inter-quartile range (IQR) from the box limits and individually plotted data points were values further than 1.5 IQR from the box limits. SEP mobility group was defined based on the exposure status at two consecutive childhood periods (very early and early, or early and middle) involved in the mobility hypothesis chosen for each CpG; persistently low was defined as being exposed to socioeconomic adversity during both periods; worsening SEP was defined as being unexposed during the former period but exposed during the later period; improving SEP was defined as being exposed during the former period but unexposed during the later period; persistently high was defined as being unexposed to socioeconomic adversity during both periods.

We first focused on the four socioeconomic adversities for which we tested all three life-course hypotheses (low family income, financial hardship, major financial problem and neighborhood disadvantage, Supplementary Material, Table S4). Here, 44 CpGs (R² > 3%) were identified, of which four passed an FDR < 0.05 threshold. The majority of CpGs reflected mobility (20 CpGs) or sensitive period (22 CpGs) relationships. The most selected life-course hypothesis varied by socioeconomic adversity. Sensitive period hypotheses were selected for all nine CpGs identified from financial hardship, with middle childhood selected for eight of them (Fig. 2A). By contrast, mobility (worsening SEP) explained more DNAm variability resulting from neighborhood disadvantage (11 of 17 CpGs) and major financial problem (4 of 5 CpGs). The time period when mobility had the greatest impact differed across SEP indicators, with very early to early childhood most often selected for neighborhood disadvantage, and early to middle childhood most selected for major financial problem (Fig. 2A). Accumulation was only selected for two CpGs, linked to low family income. Of note, mobility hypotheses were selected for all four FDR-significant CpGs, with a worsening hypothesis (meaning downward mobility) selected for three of them (Supplementary Material, Table S2). Fig. 2B shows at these three CpGs, children exposed to worsening SEP had the greatest shift in DNAm as compared to children with other types of SEP trajectories, including those who had persistently low SEP, worsening SEP, improved SEP or persistently high SEP.

For our instability indicators (job loss and income reduction), which innately capture the effects of socioeconomic mobility, we only tested accumulation and sensitive period hypotheses (Supplementary Material, Table S4). The strongest evidence was again for sensitive period effects, with middle childhood (age 3–5) most selected for job loss (9 of 15 CpGs) and very early childhood (age 0–2) most selected for income reduction (2 of 3 CpGs, Fig. 2A). Accumulation was only selected for one CpG linked to job loss.

Overall, exposure to socioeconomic changes (captured through instability indicators or mobility hypotheses) was associated with, on average, a 3.8% difference in DNAm levels, explaining 3.4% of the variance in DNAm across CpG sites after controlling for covariates (Supplementary Material, Table S2). The same patterns were found at the epigenome-wide level, with most CpGs showing most variability in response to adversity from mobility and sensitive periods, rather than the accumulation of exposure across development (Supplementary Material, Fig. S4).

SLCMA results were robust to sensitivity analyses

Additional covariate adjustment had minimal impact on results

To assess residual bias in the identified SEP-DNAm associations and further ensure the robustness of our findings, we additionally controlled for time-invariant SEP indicators, population substructure estimated from epigenetic data, cord blood DNAm, genetic variation and exposure to the other five time-varying SEP indicators. After additional covariate adjustments, the life-course hypothesis selected by Least Angle Regression (LARS) remained the same for all 62 CpGs with R² > 3% (Supplementary Material, Table S5, Supplementary Material, Table S6). Almost all CpGs remained significant at the nominal P < 0.05 threshold after adjusting for time-invariant SEP indicators (60 CpGs), population substructure (61 CpGs), cord blood DNAm (61 CpGs) and exposure to the other five SEP indicators (62 CpGs, Supplementary Material, Table S5). The associations between socioeconomic adversities and DNAm were also independent of genetic variation previously linked to significant CpGs (Supplementary Material, Table S6).

Mobility hypotheses improved our ability to identify CpGs related to SEP changes

SEP mobility during childhood had never been previously tested on childhood DNAm to our knowledge. Therefore, we assessed the insights gained from adding mobility hypotheses. We re-analyzed the CpGs with an R² > 3% for low family income, financial hardship, major financial problem and neighborhood disadvantage using only accumulation and sensitive period hypotheses. Considering only accumulation and sensitive period hypotheses, we were unable to fully detect shifts in DNAm patterns related to changes in socioeconomic environment. When mobility hypotheses were omitted from the SLCMA analyses, there were minimal changes to the main results showing effects of sensitive period on DNAm (n = 22 CpGs), as the same hypothesis was selected with similar effect estimates (Supplementary Material, Table S7). However, for CpGs originally linked to mobility (n = 20), there were substantial attenuations in the estimated SEP-DNAm associations: sensitive period hypotheses were selected instead, which in turn, showed smaller R² (ranging from 0.04% to 1.6%) and much larger P-values (ranging from 0.001 to 0.84, Supplementary Material, Table S7). These findings suggest that when the underlying association structure is misspecified, important DNAm signatures may not be identified.

EWAS of ever-exposed vs. never-exposed failed to identify time-dependent associations

To evaluate the loss (or gain) of information from the SLCMA compared to more conventional epigenetic approaches, we performed an epigenome-wide association study (EWAS) of any exposure to each type of SEP adversity before age 7 and DNAm, thus ignoring the timing or change of SEP over time. For 59 of the top 62 CpGs (including the 4 FDR-significant CpGs), the effect estimates from the SLCMA were larger in magnitude than those from EWAS (Supplementary Material, Fig. S5). In addition, no CpGs with an FDR < 0.05 were identified using EWAS of any exposure, meaning ever-exposed vs. never-exposed. These findings suggest the SLCMA was better able to identify developmentally sensitive effects of socioeconomic adversity on DNAm profiles, whereas EWAS might fail to detect signals if the true underlying hypothesis was time-dependent (24).

Biological significance of SLCMA findings

DNAm at top CpGs was weakly correlated across blood and brain

To examine the relevance of SEP-related DNAm pattern identified in peripheral blood tissues to brain health, we examined the correlation of DNAm at the top 62 CpGs in blood and brain samples, using data from the Blood Brain DNA Methylation Comparison Tool (http://epigenetics.essex.ac.uk/bloodbrain) (33). Overall, DNAm was weakly, but positively, correlated between blood and brain regions (Supplementary Material, Table S8) (prefrontal cortex: r_avg = 0.06; entorhinal cortex: r_avg = 0.10; superior temporal gyrus: r_avg = 0.08; cerebellum: r_avg = 0.09). Some top CpGs showed particularly strong correlations between blood and brain (e.g. cg24938210, r = 0.78 to 0.81 across brain regions).

Distinct biological pathways emerged across SEP indicators

The top 62 CpGs showed no significant differences in distributions of genomic features, CpG island locations or enhancers, as compared to all tested CpGs (Chi-squared tests P > 0.05, Supplementary Material, Fig. S6).

Gene set enrichment showed that SEP-related DNAm patterns were more likely to occur within or near genes involved in neural system regulation, developmental processes, immune functions, metabolic processes, substance localization and membrane transport (Supplementary Material, Fig. S7, Supplementary Material, Fig. S8). However, there was little overlap observed in the significant gene ontology (GO) terms across SEP indicators (Supplementary Material, Fig. S7), except for one GO term (morphogenesis of a branching epithelium), which emerged in the enrichment analysis for both financial hardship and major financial problem. These findings suggest different socioeconomic adversities may lead to shifts in distinct biological pathways.

Discussion

The main finding from this study was that changes in the socioeconomic environment may coincide with subsequent changes at a biological level as measured through DNAm signatures. Reports of a change in the socioeconomic environment, particularly worsening neighborhood quality (i.e. mobility) and parental job loss during middle childhood (i.e. sensitive period), were associated, on average, with a 3.8% difference in DNAm levels. These patterns were detected even after accounting for other dimensions of the socioeconomic environment, ancestry, DNAm levels at birth and genetic variation. To our knowledge, this study is the first to evaluate the role of socioeconomic changes in relation to epigenome-wide DNAm within childhood.

Our study extends prior literature on the effects of childhood SEP, providing new insights into the biological embedding of the socioeconomic environment. Only three studies to our knowledge have examined the relationship between socioeconomic mobility and DNAm (22, 34, 35). Each of these three studies included just two timepoints of SEP measures, one in childhood and another in adulthood, and only assessed DNAm in adulthood. Our results suggest that acute changes in children’s socioeconomic environment, compared to exposure to more stable socioeconomic adversity, might play a role in shaping DNAm profiles in childhood as early as age 7. Although our study is the first to measure the impact of exposure to socioeconomic changes on DNAm levels in childhood, our results parallel previous findings on SEP-related outcomes in the child development literature. For example, non-epigenetic studies focused on other SEP-related outcomes in childhood have shown that an episode of parental job loss may have a larger impact on child health and behavior than stable employment in low-income jobs (36–38). Indeed, the developmental literature largely suggests that children benefit from stable, predictable environments (39–41) and that changes in the socioeconomic environment can impact cognitive development and other mechanisms implicated in future risk of health and behavioral problems (36–38, 42, 43). Future studies are needed to replicate our findings and investigate how SEP-associated DNAm alterations may influence subsequent health and behavioral outcomes. Insights from such studies will be critical to discern whether SEP-related DNAm changes influence children’s vulnerability to disease and other negative health/behavioral outcomes.

We found more evidence for the importance of the developmental timing of SEP on DNAm rather than its accumulation. These results parallel previous findings from the ALSPAC cohort (24) and elsewhere (44), suggesting that sensitive period effects can be detected in the epigenome. Our results also specifically point to the importance of middle childhood as a potential sensitive period when the socioeconomic environment might be particularly impactful. SEP plays an important role during school-age years (39, 45), corresponding to our middle childhood time period findings, when children in the cohort began school. Socioeconomic disruptions during school-age years may lead to changes in parent–child interactions, afterschool care center attendance or extracurricular activities.

Consistent with prior epigenome-wide studies (21, 22), we found little overlap between the top CpGs across SEP domains, suggesting that various aspects of the SEP construct may trigger distinct mechanisms that lead to different alterations in DNAm patterns (19, 46). Across our six SEP indicators, the greatest number of detected CpGs (17 of 62) were related to neighborhood disadvantage, with 4 being the only CpGs to pass an FDR < 0.05 significance threshold. These findings point to the important role that neighborhood-level indicators, including more ubiquitous social and physical exposures experienced daily by larger segments of a population, may play in shaping the epigenome during child development. For example, we found that the DNAm alterations linked to neighborhood disadvantage were more likely to occur in genes related to peroxisomes, which are a key component of the biological response to various environmental pollutants (47). By contrast, we found that experiences of financial hardship (e.g. difficulty in affording common household necessities like food, clothing, heat and rent) and income reduction were linked to biological pathways related to diet quality, such as nutrient transport and metabolic processes. Overall, different clusters of biological pathways emerged across distinct DNAm-associated SEP domains, suggesting that socioeconomic adversities may affect child health through multiple mechanisms.

Many of the genes in which our top CpGs were located on or near have been linked to human health and diseases. For example, OAS3, in which our most significant CpG (cg20102336) resides, encodes an enzyme that plays a critical role in innate antiviral response (48) and has been linked with the incidence and severity of illness caused by coronavirus disease 2019 (COVID-19) (49, 50). TGFBR3, the nearest gene to another significant CpG (cg08638097), encodes a key receptor in the transforming growth factor-β (TGF-β) superfamily signaling pathways and has been implicated in various human cancers including prostate cancer and bladder cancer (51–54). Furthermore, one of the top CpGs showing strong evidence of replication across studies (cg24121967; same direction of effect and P < 0.05 in 8 and 3 other studies, respectively) was located in a putative oncogene MYEOV whose overexpression has been documented in many cancers such as gastric cancer (55), myeloma (55) and pancreatic cancer (56). These findings suggest that early life socioeconomic adversities are associated with biological disruptions that may ultimately lead to a wide constellation of health risks later in life.

While the current study uncovered many insights into SEP and DNAm associations, a major unanswered question is whether these DNAm changes are adaptive or maladaptive, in both the short- and long-term. Teicher and others have noted that early neurobehavioral changes that occur in response to experiences of childhood adversity often enhance immediate survival at the cost of long-term functioning (57). Thus, are specific epigenomic fluctuations in the face of family socioeconomic adversity reflective of increased risk, resilience or both? Although we found DNAm differences when comparing children who were exposed vs. unexposed to socioeconomic adversity, we do not know if these SEP-induced shifts represent systemic alterations of biological functions across tissue types, which may cause key impairments that lead to behavioral changes and increase disease risks. With existing publicly available data, we could only compare the potential implications of our findings to DNAm levels in brain tissue. Additional research comparing DNAm levels between different tissues is warranted to better understand the systemic effects of socioeconomic hardship.

Should these DNAm markers of socioeconomic adversity be replicated and identified as harmful (rather than adaptive) to health, our findings suggest at least two paths forward for prevention and intervention. First, our results suggest that children and families, especially lower-income families who may lack a safety-net to draw from during times of parental job loss or other socioeconomic transitions (58), might benefit from extending policies and social programs aimed at minimizing socioeconomic instability, such as the Supplemental Nutrition Assistance Program (59) and the American Families Plan (60). Second, prevention programs aimed at promoting socioeconomic stability during childhood might benefit from adopting a multisystemic approach that considers the social determinants of health (61) at multiple levels (62). In fact, interventions at the household-level (e.g. parenting-based) and neighborhood-level (e.g. community-based) have revealed measurable biological impacts on children’s DNAm profiles (63, 64) and on other biomarkers (65–67).

The current study should be interpreted in light of several limitations. First, like other epigenome-wide studies of this sample size, we identified few specific CpGs passing a stringent correction for multiple testing. However, following the recent movement to move beyond P-value thresholds alone (68, 69), we explored the patterns and implications of SEP-related DNAm profiles among top CpGs passing an effect-size-based threshold. The top CpGs passing this threshold were robust to various sensitivity analyses, and there was consistent evidence for the patterns of CpGs observed, with the majority showing effects in the same direction as previously published findings and two CpGs showing significance in other studies after correcting for multiple testing. Nevertheless, the results from individual CpGs should be interpreted with caution and validated in larger samples. Second, because this was a population-based sample, extreme cases of socioeconomic disadvantage were likely underrepresented in the ALSPAC cohort. Our results suggest that more severe forms of adversity may have more potent effects, as we identified most top DNAm CpGs (32 out of 62) from the two socioeconomic adversities that showed the lowest prevalence (job loss and neighborhood disadvantage). Future research in populations with more diverse SEP distributions capturing a wider gradient (i.e. extreme poverty) will help fully disentangle the impact of SEP on DNAm patterns. Third, the ALSPAC cohort is mostly White, which limits generalizability of these findings to other individuals and populations of non-European descent. Prior studies (see review (70)) show ancestry-related variation in DNA methylation that may lead to differences in gene regulation across populations. Thus, future replication efforts are needed in more diverse and representative populations. Finally, this study was observational and based on self-report measures of SEP, which could have been influenced by reporter bias, wherein participant responses may have been shaped by factors like social desirability or recall biases, leading to over- or underestimates of observed associations (71). Although self-reporting bias is common among survey/questionnaire data in observational studies, previous research has shown that individual-level SEP measures like education and income, compared to more objective measures assessed at the census tract-level, can more accurately capture the impact of SEP on a number of health outcomes, such as blood pressure and height (72). Future randomized experiments will help determine the causal effect of socioeconomic adversity on DNAm.

In summary, this study adds to a growing literature showing that early-life socioeconomic adversity can leave biological memories in the form of DNAm differences in childhood. Uniquely, our findings on socioeconomic mobility and instability suggest changes in the socioeconomic environment during childhood are especially impactful and associated with epigenetic disruptions related to various health outcomes. Ultimately, these findings will enable researchers to build toward better intervention and prevention efforts aimed at reducing socioeconomic disparities and promoting health across the life course.

Materials and Methods

Sample and procedures

Data came from the Accessible Resources for Integrated Epigenomics Studies (ARIES) (73), a subsample of 1018 mother–child pairs from the ALSPAC. ALSPAC is a prospective, longitudinal birth cohort in the UK designed to investigate genetic and environmental determinants of health across the lifespan (74–76). Women living in the county of Avon, UK, with estimated delivery dates between April 1991 and December 1992 were invited to participate. Mother–child pairs in the ARIES were randomly selected from ALSPAC based on availability of DNA samples across five waves of data collection (73). We analyzed data from 946 singletons in ARIES with blood-based DNAm profiles generated at age 7. Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committee. Note that the ALSPAC study website contains details of all the data that is available through a fully searchable data dictionary and variable search tool (http://www.bristol.ac.uk/alspac/researchers/our-data). See Supplemental Methods for full ALSPAC details.

Measures

Early-life socioeconomic position (SEP)

We analyzed six SEP indicators, spanning financial, occupational and residential domains: 1) job loss, 2) income reduction, 3) low family income, 4) financial hardship, 5) major financial problem and 6) neighborhood disadvantage. These were the only available, time-varying SEP indicators that were measured repeatedly via maternal report through mailed questionnaires during three developmental time periods (Fig. 1A): very early childhood (0–2 years), early childhood (3–5 years) and middle childhood (6–7 years).

For each SEP indicator, children were classified as exposed or unexposed at each period, using criteria described in Supplemental Methods (Supplementary Material). With these repeated, self-reported SEP indicators, we could identify changes occurring between time-periods for indicators capturing time-varying status of SEP. For job loss and income reduction, the measures inherently captured change within a certain developmental period, because they asked about socioeconomic mobility. To distinguish job loss and income reduction from other indicators, we refer to them throughout the manuscript as ‘instability indicators’.

DNA methylation (DNAm)

DNAm was measured from peripheral blood at age 7 using the Illumina Infinium HumanMethylation450 BeadChip microarray (Illumina, San Diego, CA). DNAm wet laboratory procedures, preprocessing analyses and quality control are described in Supplemental Methods (Supplementary Material). A total of 412 956 CpGs on autosomal chromosomes passed quality control and were included in this analysis. For each CpG, DNAm level is expressed as a ‘beta’ value (β-value) ranging from 0 to 1, which represents the proportion of cells methylated at each interrogated CpG.

Covariates

To adjust for baseline demographic differences in ARIES and technical variation in DNAm assessment, we controlled for the following variables measured at birth in all analyses: child age in months at blood draw, child race/ethnicity, child sex, child birthweight, maternal age, number of previous pregnancies, sustained maternal smoking during pregnancy and cell type proportions estimated using the Houseman method (77). Details can be found in the Supplemental Methods (Supplementary Material).

Data analysis

All analysis code is available through our GitHub page: https://github.com/thedunnlab/sep-dnam.

Structured life-course modeling approach

We used the two-stage structured life-course modeling approach (SLCMA) (30–32) to evaluate the time-dependent effects of socioeconomic adversity on DNAm. SLCMA is a method that leverages repeated exposure data to simultaneously investigate the relationship between exposure and outcome under multiple a priori-defined life-course hypotheses. In our analyses, we tested three life-course hypotheses, described previously, which were parameterized as follows (Fig. 1B).

First, to test the accumulation hypothesis, we created a sum score (ranging from 0 to 3), which captured the number of time periods across the three developmental stages that children were exposed. Second, to test the sensitive period hypothesis, we created three binary variables, one for each of the three developmental periods, to classify children’s exposure status (0 = unexposed during the period; 1 = exposed during that period). Third, to test the mobility hypothesis, we created a pair of indicator variables for change in SEP between very early and early childhood, and a pair of indicator variables for change in SEP between early and middle childhood. Each pair consisted of an indicator variable for worsening (1 = change from unexposed to exposed, 0 = other) and an indicator variable for improvement (1 = change from exposed to unexposed, 0 = other).

We tested all three hypotheses for low family income, financial hardship, major financial problem and neighborhood disadvantage. Only the accumulation and sensitive period hypotheses were tested for job loss and income reduction, as these two instability indicators inherently reflect SEP changes (Supplementary Material, Table S4).

We performed the SLCMA in two stages: 1) life-course hypothesis model selection followed by 2) post-selection inference (Fig. 1B, Supplemental Methods). In the first stage, we tested the variables described above using a Least Angle Regression (LARS) variable selection procedure (78) to identify the life-course hypothesis most supported in the observed data (i.e. explaining the most variation in DNAm). In the second stage, we used selective inference (30, 79) to test the association between the selected variable and DNAm and estimate confidence intervals.

Defining CpGs of interest

We used two thresholds to identify associations between SEP and CpG CpGs for further investigation. Given recent recommendations discouraging the use of P-values alone for statistical inference (68, 69), we used an effect-size-based threshold of R² > 3%, meaning that the SEP exposure explained more than 3% of the variance in DNAm. This cutoff was selected based on the effect sizes observed in previous epigenome-wide analyses of childhood adversity in ALSPAC (24, 26) and other well-established environmental exposures, including tobacco smoking (80). We also performed multiple-testing correction using the Benjamini-Hochberg method (81) at a 5% FDR to assess the significance of top CpGs.

Sensitivity analyses

We conducted three sensitivity analyses to evaluate the robustness of our SLCMA results. First, we additionally controlled for 1) time-invariant SEP indicators (e.g. maternal education at baseline), 2) population substructure estimated from epigenetic data, 3) cord blood DNAm (to account for differences in DNAm that might have been present at birth), 4) genetic variation (at methylation quantitative trait loci, or mQTL) or 5) exposure to the other five time-varying SEP indicators. Second, we reran the analyses of the CpGs with an R² > 3% for low family income, financial hardship, major financial problem and neighborhood disadvantage using only accumulation and sensitive period hypotheses and compared the results from analysis with and without mobility tested. Third, we performed an EWAS of any exposure to each type of SEP adversity before age 7 and DNAm and compared the findings with SLCMA results. See (Supplemental Methods, Supplementary Material) for details.

Secondary analyses

To interpret our findings and place them in the context of prior literature, we conducted two secondary analyses. First, we compared the effect estimates of R² > 3% CpGs to those reported in previous SEP-related EWAS studies (19) (Supplemental Methods, Supplementary Material). Second, we also evaluated the biological significance of our findings by examining the correlation between DNAm in blood and brain tissue for the R² > 3% CpGs and testing for the enrichment of genomic features, regulatory elements and Gene Ontology (GO) terms (Supplemental Methods, Supplementary Material).

Funding

Russell Sage Foundation and the Ford Foundation (# 96-17-05 to E.C.D., awarded July 2017; any opinions expressed are those of the author(s) alone and should not be construed as representing the opinions of either foundation); National Institute of Mental Health at the National Institutes of Health (grant number R01MH113930 to E.C.D); UK Medical Research Council and the Wellcome Trust (Grant ref: 217065/Z/19/Z); University of Bristol (to ALSPAC); BBSRC (BBI025751/1 and BB/I025263/1 to ARIES).

Ethical standards

All ethical guidelines were followed per research involving use of human subjects. Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committee. This study was approved with oversight by the Mass General Brigham Institutional Review Boards (IRB) (Protocol ID 2017P001110).