We included 17 701 women who had live singleton pregnancies, responded to the survey and had complete data. The study flow diagram is shown in figure 1. Altogether, 1189 women experienced PTB (1189/17 701, 6.7% (95% CI 6.2%, 7.1%)) and 1350 had a baby born SGA (1350/17 701, 7.0% (95% CI 6.5%, 7.5%)). Women who were Bangladeshi or Pakistani had the lowest mean ages, the highest proportions of women with no formal education and the lowest equivalised household incomes compared with other groups. They also had the highest proportion of marriage, alongside women who were Indian. Women who were Indian or White had the highest equivalised household incomes compared with others. Women who were Black African had the highest maternal age, a greater proportion of educational attainment at both extremes relative to the population total and along with women who were Black Caribbean, the highest proportions of single marital statuses, relative to other groups. Smoking was most common among women who were White or Black Caribbean. The ethnic composition and baseline characteristics of the sample are found in table 1. PTB was the most common among Indian (45/447, 9.6% (95% CI 7.0%, 13.0%)) and least common among Pakistani (59/833, 6.1% (95% CI 4.7%, 7.9%)) groups. On the international ‘Intergrowth’ chart, SGA was most common among Indian (98/447, 18.1% (95% CI 13.4%, 24.0%)) and least common among White (932/14 927, 6.2% (95% CI 5.7%, 6.6%)) groups. The prevalence of outcomes by ethnic group is shown in table 2.

Study flow diagram eligibility and inclusion criteria.

Baseline characteristics according to ethnic group

Prevalence of PTB and SGA according to ethnic group

For PTB, while the point estimates suggested a positive association for all groups except Pakistani women, there was only strong evidence for health inequality among Indian women who were at greater risk (RR 1.46 (95% CI 1.06, 2.02)), relative to White. Inequality increased once age, parity and SEP were adjusted for (RR 1.64 (95% CI 1.17, 2.29)). After adjustment, the point estimates for women who were Black African or Black Caribbean reduced but increased among other groups. The point estimates either did not change or increased after adjustment for smoking status, planned pregnancy and antenatal care.

Inequality for SGA using the Intergrowth growth chart was marked, with the greatest risk among women who were Bangladeshi (RR 2.65 (95% CI 2.10, 3.34)), Indian (RR 2.94 (95% CI 2.20, 3.95)) or Pakistani (RR 2.49 (95% CI 2.08, 2.97)), relative to white. Women who were Black African (RR 1.70 (95% CI 1.21, 2.40)), Black Caribbean (RR 1.59 (95% CI 1.06, 2.39)) or from ‘other ethnic groups’ (RR 1.75 (95% CI 1.28, 2.40)) were at moderately increased risk relative to White. After adjustment for maternal age, parity, education, income and marital status, the point estimates among all groups were attenuated except for Indian women, which increased. Inequality appeared to increase after accounting for smoking status, planned pregnancy and antenatal care. The multivariable analyses for PTB and SGA are found in table 3.

The association between ethnic groups and outcomes

In the mediation analyses of ethnic groups and PTB via SEP using multiple mediators (maternal education, income, marital status), significant total and direct ‘effects’ were similarly only observed among women who were Indian. Among this group, SEP may be protective, with weak evidence of a reduction in inequality (indirect effect RR 0.93 (95% CI 0.87, 1.00), p=0.066), relative to White women. This equated to a reduction in inequality of 15%, relative to White women. For SGA, there was evidence of mediation via SEP for women who were Black African and Black Caribbean (indirect effects RR 1.16 (95% CI 1.04, 1.30) and 1.12 (1.00, 1.26), respectively), equating to 29% and 10% of the inequality, respectively, relative to women who were White.

When we explored the latent class of SEP, a three-class solution was identified with entropy of 0.83. Based on the predicted probabilities of socioeconomic characteristics given class membership, we identified groups according to: ‘advantaged’ (A), more likely to be married homeowners in professional occupations with high income, ‘blue collar/white collar couples’ (BW), more likely to be cohabiting or married homeowners or private renters in routine or intermediate occupations with middle incomes and ‘disadvantaged’ (D), in routine or no occupation with low income. As the degree of entropy suggested a good degree of separation, women were then assigned their most probable latent class. The predicted probabilities given class membership are reported in online supplemental figure 1.

When we repeated the mediation analysis for latent class of SEP with similar adjustment for maternal age and parity, there was evidence of an indirect ‘effect’ on PTB for women who were Bangladeshi (RR 1.08 (95% CI 1.01, 1.16)), Black African (RR 1.07 (95% CI 1.01, 1.14)), Black Caribbean (RR 1.06 (95% CI 1.00, 1.12)) and Pakistani (RR 1.06 (1.00, 1.13)). This represented 60% of the difference between women who were Bangladeshi compared with White and 26% and 28% for women who women were Black African or Black Caribbean, respectively, compared with White. For SGA, there was evidence of an indirect ‘effect’ for all groups except for women who were Indian. The indirect ‘effects’ were RR 1.35 (95% CI 1.19, 1.52, 29% of inequality), 1.32 (1.17, 1.48, 48% of inequality), 1.26 (1.12, 1.41, 53% of inequality), 1.27 (1.13, 1.42, 25% of inequality) and 1.13 (1.02, 1.26, 20% of inequality) for women who were Bangladeshi, Black African, Black Caribbean, Pakistani or from an ‘other ethnic group’, respectively, compared with women were White. The full results for all mediation analyses are shown in table 4.

Relative risks for total, direct and indirect effects for PTB and SGA through SEP (multiple mediators) and SEP class

Overall, 2.4% of women had missing data for any variable except BMI, increasing to 10% including BMI. Data on BMI were missing for 1454 of the included women (1,454/17,701, 8.2%) who had otherwise complete data. The results following multiple imputations were insensitive to the inclusion of women with missing data on BMI. As the data were nearly complete other than BMI, the results following adjustment for the imputed BMI are reported as model four in table 2.

We identified health inequality according to the ethnic group for PTB under 37 weeks and SGA according to the Intergrowth chart that is not customised for ethnic group. The absolute risk of PTB was greatest among Indian, Black Caribbean or Black African women. The absolute risk of SGA was greatest among Indian, Pakistani and Bangladeshi women. We observed that adjustment for SEP attenuated the risk for both outcomes among women who were Black African or Black Caribbean and, within the assumptions of these casual models, we observed that SEP disadvantaged women who were Black Caribbean with respect to PTB and Black African or Black Caribbean with respect to SGA. We observed the opposite association for Indian women for both PTB and SGA, whose relative advantage through SEP may mask other sources of inequality, compared with women who were White. Importantly, women who were Indian were still at greater risk of PTB and SGA compared with those who were White. We identified multidimensional latent classes of SEP which were associated with both ethnic group and pregnancy outcomes. The latent SEP class explained a greater proportion of inequality for PTB and SGA than the simple SEP measures, with an indirect ‘effect’ across most groups. Class membership explained up to 60% of the health inequality for SGA (among Bangladeshi women).

The findings accord with the only other UK-based study to investigate the risk of PTB through socioeconomic measures (deprivation)5 which found 35% of inequality for PTB between ethnic groups was potentially mediated by small-area deprivation. In this large study using birth registration data, inequality in PTB was greater than that measured in the MCS. Socioeconomic classes identified in another diverse UK-based birth cohort showed that socioeconomic characteristics differed by ethnic group24 although these classes were not compared with health outcomes. Further evidence comes from birth certificate data from the USA.25–27 These data include a diverse range of socioeconomic information and the sample size to investigate more complex relationships between variables, although inequality may be qualitatively different in different environments and countries. Although we only observed evidence on the mediating role of SEP for PTB among women who were Indian, the direction of effect for the point estimates agreed with previous studies in larger data sets.28

We focused on the role of SEP because health inequalities are produced by systematic differences in the socioeconomic environments between groups.29 These differences may lead to acquired differences in health over the life course, including medical comorbidities that occur upstream of outcomes in pregnancy.30 Within a healthcare system, communication barriers and conscious and unconscious biases of healthcare providers may produce or exacerbate inequality too.31 As well as interpersonal racism and bias, education and training of health professionals, healthcare guidance and policies and workforce culture are other sources of systematic racism that may produce inequality closer to the time of birth.32 Other explanations for inequality include differences in access to or the quality of healthcare.33 34 Treatments that reduce the risk of PTB35 need to reach women from high-risk groups; for example, the level of antenatal care may contribute to health inequality.36 We observed differences in the level of antenatal care between ethnic groups but these did not appear to explain inequality in our multivariable analyses. Focussing on the provision of care through specialist clinics is a widely-used strategy but the evidence for this model, and therefore understanding how the model could be used to reduce inequality, is limited.37 There is randomised evidence that women who received care in either a midwifery-led continuity of care model or any form of alternative to routine antenatal care had a lower risk of PTB38, however. Tackling the underlying determinants of inequality is obviously challenging but in a healthcare context, there may be opportunities to reduce inequalities in pregnancy closer to the time of the outcome.

The strengths of this study include the combination of diverse measures of SEP to understand more complex socioeconomic distinctions between groups. Mediation analyses are limited by the assumptions that associations between ethnic groups, SEP and PTB/SGA are unbiased by confounding. We rely on the time-ordered nature of exposure, SEP and birth outcomes and the forward causal flow through potentially unmeasured mediating pathways. The unadjusted estimates are most informative because they do not make adjustments for what may be sequential mediators from SEP to the outcomes, such as antenatal care, smoking or BMI. The association between SEP and pregnancy outcomes may still be subject to unmeasured confounding.39 These models cannot be overinterpreted as supportive of causal inference. We did not include maternal age in our LCA as although this may be related to SEP, it may also represent other cultural influences. Maternal age may nevertheless contribute to health inequality.40 Although we investigated the interaction between income and education, ethnic group and income and ethnic group and education, the sample was likely insufficient to characterise these relationships. Instead, we identified latent socioeconomic classes. This may not only better understand how SEP affects health but may protect from the risk of bias when investigating multiple mediators and complex causal structures21–23 although we did not identify any studies of mediation via SEP which considered interaction.25–27 A limitation is that data were collected through maternal recall at 9 months. Recall of gestational age and birth weight in the MCS was reliable when compared with birth notification data.41 42 The use of the dichotomised outcomes reduced the risk of misclassification. We limited measures of SEP to antenatal measures; however, we included household income as reported at 9 months. We therefore assume that household economic conditions following birth represented those prior to and during pregnancy although this may not be the case for all women. Finally, the latent class solution may be subject to overfitting. In this exploratory analysis, we did not conduct an interval validation although did validate the classes with respect to health outcomes.

Healthcare research efforts could focus on investigating and exploiting modifiable risk factors that are proximal to health outcomes or deploying the available resources for PTB risk assessment and SGA detection most effectively. This does not necessarily mean developing services or recommending interventions solely based on ethnic groups32 but may include auditing interventions and services with a specific focus on ethnic groups and taking steps to support access and culturally-informed counselling. The underlying socioeconomic determinants require policy-level interdisciplinary approaches over a longer timeframe.