Key strengths of this study are that it uses standardised, routinely, prospectively collected healthcare data from all NHS neonatal units in England. This minimises biases in the data such as selection bias, observer bias, and recall bias. Findings are generalisable across the country. Data covered a 10-year period of 2014 to 2024, and the NPP intervention was implemented in 2018, which gives sufficient time for reliable trends to be observed. Statistical modelling estimated the total effects of each sociodemographic factor, adjusted for measured potential confounders, and as supplementary analyses we additionally estimated the direct effects of each of these key sociodemographic factors. All models were developed a priori using DAGs, to clarify assumptions and guide design of the statistical models to minimise confounding and best estimate causal effects. Results were robust to sensitivity analyses.

Observed differences in receipt of a treatment does not always indicate inequity in care. In some cases it could just reflect that different groups have different levels of need. However in this case, all mothers in this study are eligible for treatment, so are considered to have equal need. Medical contra-indication is rare and unlikely to be sociodemographically patterned. This indicates that the observed differences are likely to reflect a true inequity in care. Similarly, differences in treatment could be due to differences in patient behaviour – for example if certain patient groups tend to arrive later at hospital, so have a smaller window of opportunity for treatment. Arguably such cases are not ‘inequity in care’, or at least not inequity at the point of care, but just variations in patient behaviour that lead to differential treatment rates. Further exploration of upstream patient behaviour would be valuable here to better understand the causal pathways leading to differences in treatment. However, again we would argue that all of these mothers should be being treated, and if some groups are under-treated, that is an inequity to be addressed, regardless of whether the reason applies at the point of care or further upstream.

A limitation is that we did not have data on all possible covariates of interest, so residual confounding is plausible. For example maternal BMI, patterns and details of previous birth history (including assisted conception and stillbirths), how close the mother lives to their maternity unit, how early or late different groups of mothers tend to present at hospital, could all potentially affect receipt of MgSO₄. Similarly, there are other individual characteristics that as well as being potential confounders, could also be considered important dimensions of equity themselves – e.g., race, language, sexual orientation, disability status, chronic illness status, marital status, education, health literacy, and urban/rural living(46) – that were not measured in this data and so could not be evaluated here.

It is possible that changes in population characteristics over time could be explaining some of the differences observed here. The post-NPP population tended to have slightly fewer smokers, slightly fewer cases of fever in labour or receiving antibiotics, fewer cases of meconium-stained liquor, more cases of spontaneous onset of labour, and more caesarean sections. It is plausible that some of these factors are also associated with likelihood of treatment – e.g. non-smokers may be more likely to be present and available to treat, and fewer ‘complications’ such as fever and meconium-stained liquor could mean a bigger window of opportunity to treat between presentation and delivery. We tried to address this by adjusting for these factors wherever indicated in the DAGs, but residual confounding is possible, and of course there may be alternative views on the details of some of the relationships modelled in our DAGs.

Another limitation is that the neonatal dataset only includes data on live-born infants who were admitted to a neonatal unit, and their mothers. This is a sub-group of an underlying population of mothers and infants who may also have been treated with MgSO₄ for expected preterm birth, but who either did not go on to have a preterm birth, or where the infant died prior to NICU admission. If possible, it would be a valuable future research study to replicate this work on the wider underlying population.

More granular detail on individual-level deprivation factors may be useful in future work. It is possible that factors not captured by the area-level Index of Multiple Deprivation (IMD) (or captured, but different at the individual versus area-level) could also drive differences in receipt of healthcare. Similarly, more granular detail in the differences in treatment between the different minority ethnic groups would be valuable, as it is possible that combining them – as was indicated in this study to give sufficient power – may conceal variation in sub-group treatment rates.(10)

For the minority of eligible mothers not receiving antenatal MgSO₄, further exploration of the reasons why they did not – and whether or not this varies by sociodemographic factors – would be valuable. Previous research on this population indicates that the overwhelming reported reason for not receiving treatment (>75% of cases) is ‘imminent delivery’, with smaller proportions reported as ‘not offered’ (11%), ‘not appropriate’ (5%), and very rarely ‘contraindicated’ or ‘declined’.(21) In this study the sample size would not have been sufficient to power further analysis of this by sociodemographic sub-groups, but if numbers allowed in larger, international datasets (for example the Vermont Oxford Network dataset(47)), further analysis on this question would be useful.