Antenatal corticosteroids and newborn respiratory outcomes in twins: A regression discontinuity study

To estimate the effect of antenatal corticosteroids on newborn respiratory morbidity in twins.


| I N TRODUC TION
Administration of antenatal corticosteroids to singleton pregnancies at imminent risk of preterm birth reduces the risk of newborn morbidity and mortality. 1While the risk of preterm birth is a much as ten times higher for multiple gestations compared with singletons, 2 the effectiveness of antenatal corticosteroids in this population is less certain. 3,4A 2020 Cochrane review of randomised trials examining the effect of antenatal corticosteroids found only four trials that reported data separately for multiples and, although pooled risk ratios suggested that antenatal corticosteroids may be effective in multiples, the uncertainty around the risk ratios was also consistent with no effect. 1Two recent systematic reviews of non-randomised studies examined the effect of antenatal corticosteroids in twins 4 and multiples. 5Although both reviews reported that antenatal corticosteroids reduced the risk of adverse newborn outcomes, confidence in these findings is limited by the potential for confounding in observational studies of drug effectiveness. 4,6,7][10][11] We sought to generate such evidence using a regression discontinuity design, which provides better control for confounding than conventional non-randomised studies that typically rely on regression adjustment alone (e.g.3][14] Our objective was to estimate whether the incidence of neonatal respiratory morbidity was higher among twin pregnancies admitted for birth just after the gestational age cut-point for antenatal corticosteroid administration in Canadian clinical practice guidelines.

| Study design
Conventional non-randomised studies are at risk of bias due to confounding because pregnancies treated with antenatal corticosteroids have different clinical risk profiles from those that are not treated.Adjusting for measured confounders (e.g. using multivariable regression) can help to mitigate some of this bias, but is rarely able to eliminate all differences in risk profiles between groups.][14] Because the threshold is somewhat arbitrary, individuals on either side of this threshold are similar in most characteristics except their chance of being exposed to the treatment.
Through 2018, Canadian clinical practice guidelines recommended antenatal corticosteroids based on a gestational age cut-point of 34 +0 weeks (i.e.treatment was recommended through to 33 +6 weeks, but not at or after 34 +0 weeks). 9,15This cut-point is arbitrary in that the absolute risk of adverse respiratory outcomes and the absolute benefit of antenatal corticosteroids in singletons fall smoothly across gestational age (i.e.there is no sudden change in pregnancy or clinical characteristics between 33 +6 and 34 +0 weeks). 16Hence, assuming adherence to the guidelines, the proportion of pregnancies treated with antenatal corticosteroids would be expected to fall abruptly at 34 +0 weeks but other potential confounders (e.g.pre-eclampsia) would not change abruptly.If all characteristics except the probability of treatment are similar directly on either side of the cut-point, differences in newborn respiratory outcomes directly on either side of the cut-point can be attributed to differences in the proportion who were treated with antenatal corticosteroids.
In previous work we used a regression discontinuity design to estimate the effect of being above versus below the Canadian threshold for recommending antenatal corticosteroids on newborn respiratory outcome in singletons and found evidence consistent with results from randomised trials in singletons. 17Here, we extend this design to examine the effect of antenatal corticosteroids in twins.

| Study population
We used data from the British Columbia (BC) Perinatal Data Registry, a population-based registry containing detailed obstetric and neonatal medical records for >99% of all births in BC, Canada.Data are collected from pregnancies that end in a live birth or stillbirth and are at least 20 weeks of gestation or at least 500-g birthweight, in over 60 acute care facilities or from births occurring at home attended by BCregistered midwives.Data are also collected on infant transfers and readmissions through the first 28 days after birth.
We began our study period on 1 April 2008, because dayspecific estimates of gestational age at maternal admission for birth (necessary for the regression discontinuity design) were not abstracted into the BC Perinatal Data Registry before 2008.We ended our study period on 1 August 2018, as the Canadian clinical practice guidelines were updated to extend the treatment recommendation through 34 +6 weeks and to suggest that treatment could be considered through 36 +6 weeks. 9e included data from all twin pregnancies with an admission for birth from 31 +0 weeks through 36 +6 weeks of gestation that delivered one or more live births.Gestational age was estimated using the method that was in use clinically at the time.Before 2014, the recommended method for estimating gestational age was to confirm or revise the last menstrual period estimate with an early-ultrasound assessment and, from 2014, the recommended method was earlyultrasound assessment alone. 18We excluded individuals without an early-ultrasound estimate.

| Antenatal corticosteroids
The recommended course of antenatal corticosteroids during our study period was two 12-mg doses of betamethasone given 24 hours apart, or four 6-mg doses of dexamethasone given 12 hours apart. 19The BC Perinatal Data Registry contains information on whether corticosteroids were administered during the maternal admission for birth, but does not contain information on type of drug, number of doses, or if corticosteroids were administered at a previous admission.

| Exposure
Our exposure of interest was maternal admission for birth before versus after 34 +0 weeks of gestation.That is, we compared the probability of adverse respiratory outcomes if infants had been more likely to be exposed to antenatal corticosteroids (because they presented at a gestational age where administration was recommended by clinical practice guidelines) versus less likely to be exposed to antenatal corticosteroids (because they presented at a gestational age where antenatal corticosteroids were not recommended).

| Outcomes
Our primary outcome was a composite of any newborn respiratory distress diagnosis or in-hospital newborn mortality.We defined respiratory distress as a diagnosis of one or more of: respiratory distress syndrome (International Classification of Diseases 10th revision Canadian edition [ICD-10-CA] code P22.0), transient tachypnoea of the newborn (ICD-10-CA P22.1), other respiratory distress (ICD-10-CA P22.8) or unspecified respiratory distress (ICD-10-CA P22.9).Our secondary outcome was a composite of newborn receipt of respiratory intervention or in-hospital newborn mortality.We defined respiratory intervention as one or more of: intermittent positive pressure ventilation by mask or endotracheal tube immediately after birth (recorded in the Perinatal Data Registry), continuous positive airway pressure ventilation beginning at 60 minutes after birth or upon admission to the neonatal intensive care unit (recorded in the Perinatal Data Registry), non-invasive positive-pressure ventilation (Canadian Classification of Interventions [CCI] code 1.GZ.31.CB-ND), invasive per orifice positive-pressure ventilation via intubation (CCI 1.GZ.31.CA-ND), or invasive per orifice manual assisted ventilation via intubation (CCI 1.GZ.31.CA-EP).We included in-hospital newborn mortality in both composite outcomes because mortality can be caused by respiratory complications, and because mortality could be a competing event if it occurs before respiratory distress diagnosis or intervention.

| Demographic, pregnancy and newborn characteristics
We presented summary statistics for characteristics and interventions that may be associated with newborn respiratory outcomes.Pre-pregnancy characteristics were age (years), body mass index (kg/m 2 ) and parity (nulliparous versus multiparous).Pregnancy characteristics were smoking (any tobacco product during any part of the current pregnancy), gestational hypertension (ICD-10-CA O13), pre-eclampsia (ICD-10-CA O14), and gestational diabetes (as recorded in the BC Perinatal Data Registry), and admission-to-delivery interval (hours since admission and birth of first infant), mode of delivery (vaginal or caesarean), and whether the birth was iatrogenic (following prelabour caesarean or labour induction).Newborn characteristics were sex (male or female) and birthweight (grammes).

| Statistical analysis
We plotted the proportion of pregnancies treated with antenatal corticosteroids during the maternal admission for birth by the gestational age at the admission.
We used logistic regression to model the risk of our outcomes according to gestational age at maternal admission for birth.The model specified separate baseline levels and slopes for gestational ages above versus below the treatment cut-point (Equation S1).We compared the risks of our outcomes at 34 +0 weeks, estimated from the pre-cut-point trend in newborn outcomes (when treatment was more likely because guidelines recommended treatment) versus from the post-cut-point trend (when treatment was less likely because guidelines did not recommend treatment).This comparison is similar to an intention-to-treat (ITT) effect in a randomised trial, because we are estimating the effect of a guideline recommending antenatal corticosteroid administration, not the effect of antenatal corticosteroid administration (i.e. the effect takes into account that compliance with the guideline may be <100%).We used newborns as our unit of analysis and clustered standard errors to account for non-independent outcomes among twin pairs. 20,21We used triangular weights to down-weight gestational ages that were further from 34 +0 weeks to ensure that observations closest to the cut-point were more influential in our estimates.We estimated marginal risk ratios (RRs) and risk differences (RDs) with 95% confidence intervals (CI).We also estimated the magnitude of the discontinuity in treatment (i.e. the extent to which guidelines influenced the proportion of pregnancies treated during the birth admission) using the same models outlined above but using treatment with antenatal corticosteroids as the outcome.
In a regression discontinuity design, bias can be introduced if other causes of the outcome also change abruptly at the treatment cut-point.To verify the assumption that no other causes of the outcome were discontinuous at the cut-point, we graphed the prevalence of other demographic and pregnancy characteristics and interventions by gestational age at the maternal admission for birth.We assessed whether there was a discontinuity visually and by estimating the difference using the same regression discontinuity design described earlier but with each characteristic as the outcome (using an identity link for the continuous outcomes).

| Sensitivity analyses
We tested the robustness of our results by adding a quadratic term for gestational age at maternal admission for birth and by changing the bandwidth to restrict the cohort inclusion criteria to admission between 32 +0 and 35 +6 weeks of gestation (down from 31 +0 to 36 +6 weeks). 22We did not have data on antenatal corticosteroid exposure before the birth admission, so we further compared results when restricting to individuals without a previous hospital admission during pregnancy (i.e. who were unlikely to have been exposed before their birth admission).We did not have data on the number of doses, so we compared results when restricting to individuals with an antepartum stay of ≥24 hours (i.e.sufficient time to receive at least two doses of betamethasone).In the USA, late-preterm antenatal corticosteroid treatment in multiples increased soon after the 2016 Antenatal Late Preterm Steroids (ALPS) trial 23 showed that antenatal corticosteroids were effective in preventing respiratory complications in late-preterm singletons, 24 so we compared results when restricting to data from 2008 to 2016.Finally, the incidence of newborn mortality was very low in our study cohort (6 cases; 1 per 1000) and consequently would not be expected to introduce meaningful selection bias even if mortality occurred before a diagnosis of respiratory distress could be made.We therefore added an analysis that excluded newborn mortality from our composite outcomes.

| Patient and public involvement
There was no patient or public involvement in this study.

| R E SU LTS
Between 2008 and 2018, there were 6884 twin pregnancies (13 704 live births) in British Columbia, Canada.Our study population was restricted to 2524 pregnancies (5035 twin live births) with maternal admission for birth between 31 +0 and 36 +6 weeks of gestation.A study flowchart is shown in Figure S1.
Demographic and clinical characteristics of twin pregnancies admitted for birth before (20%) and after (80%) 34 +0 weeks of gestation (the Canadian guideline cut-point for recommended antenatal corticosteroid administration) are presented in Table 1.As expected, average characteristics were different among pregnancies admitted for birth in the 3 weeks before and the 3 weeks after the cut-point.However, we found no evidence of discontinuities for any of these characteristics at 34 +0 weeks (Figure S2; Table S1), suggesting that confounding by these characteristics is unlikely.The proportion of pregnancies treated with antenatal corticosteroids during the birth admission was much higher when treatment was recommended, with an average of 47% before 34 +0 weeks and 4.2% on or after 34 +0 weeks (Figure 1).The proportion treated during the birth admission was also discontinuous at the cut-point, falling sharply from an estimated 41% using the pre-cut-point trend to 12% using the post-cut-point trend (RD −29 per 100, 95% CI −19 to −39).
T A B L E 1 Demographic and clinical characteristics of our study population of all twin live births admitted for birth between 31 +0  The risk newborn respiratory distress or in-hospital mortality decreased with increasing gestational age at maternal admission for birth and then increased abruptly at 34 +0 weeks (Figure 2A).The risk of respiratory distress or in-hospital death at 34 +0 weeks was 27% estimated from the pre-cut-point trend, compared with 39% estimated from the post-cut-point trend (RR 0.69, 95% CI 0.53-0.90;RD −12 per 100, 95% CI: −20 to −4.1; Table 2).The risk of our secondary outcome (respiratory intervention or in-hospital death) also decreased with increasing gestational age at maternal admission for birth, but there was no evidence of an abrupt change at 34 +0 weeks (Figure 2B).The risk of respiratory intervention or in-hospital newborn death at 34 +0 weeks was 33% estimated from the pre-cut-point trend versus 38% estimated from the post-cut-point trend (RR 0.89, 95% CI 0.70-1.13;RD −4.2, 95% CI −13 to +4.2; Table 2).
Results from sensitivity analyses were compatible with the primary analysis for both outcomes (Table S2).As inhospital newborn mortality was very rare in our cohort (6 cases; 1 per 1000), results were almost identical when including only respiratory morbidities in our outcomes (Table S2).

| Main findings
We found that higher rates of antenatal corticosteroid administration lowered the risk of respiratory distress diagnosis in twins by 31%.Results for respiratory intervention were imprecise and did not show clear evidence for or against an effect.

| Strengths and limitations
We used a large, population-based cohort of all twin live births in BC, Canada.Our use of the regression discontinuity design to control for confounding is an important strength of this study over previous large, non-randomised studies on the effectiveness of antenatal corticosteroids in twins, which have relied on adjustment for measured confounders, risking bias from unmeasured confounding or poorly measured confounders. 4Another strength of our study was that we accounted for non-independent outcomes between twins born from the same pregnancy.Previous studies, including the 2020 Cochrane meta-analysis of randomised trials, 1 did not account for this clustering and therefore their estimates of uncertainty were likely underestimated (leading to artificially narrow 95% CIs).Potential limitations include that regression discontinuity results can be biased if other characteristics that affect our outcomes are discontinuous at 34 +0 weeks (our estimates would be an average of all these differences).However, we know of no clinical or biological reason why demographic or pregnancy characteristics that cause respiratory distress, intervention or mortality would be discontinuous at 34 +0 weeks.Consistent with these assumptions, we found no evidence of discontinuities in any of the characteristics or interventions considered.The estimates from the regression discontinuity can also be biased if the outcome trends before and after the cut-point were not modelled correctly.However, there were no obvious outliers or non-linearities (Figure 2) and estimates were similar when including a quadratic term for gestational age at admission for birth and when restricting to only 1 week either side of the cutpoint (Table S2).Missing data on gestational age at maternal admission for delivery could also bias our results, as we excluded 1837 pregnancies (27%) that did not have an earlyultrasound estimate of gestational age.In the BC Perinatal Data Registry, early-ultrasound estimates of gestational age were primarily missing for two reasons: (1) estimates were not recorded in the hospital chart for the birth admission (i.e.only in prenatal charts) or (2) the ultrasound was performed at 20 +0 weeks or later, the arbitrary cut-point for data extraction from patient charts into the BC Perinatal Data Registry.However, missingness is unlikely to be discontinuous at 34 +0 weeks and therefore excluding these individuals should not bias our estimates.Finally, the BC Perinatal Data Registry does not have data on the number of doses administered during the birth admission, administration of rescue doses, or on any treatments administered before the birth admission.Although this information would have not changed our analysis (the regression discontinuity model [Equation S1] only includes information on gestational age at admission for birth), it would have been useful for contextualising our results and considering the generalisability of our effect estimates.

| Interpretation
][10][11] For example, the most recent Green-top guidelines in the UK cited the 2020 Cochrane meta-analysis of trials 1 and a single non-randomised study in twins that reported null or harmful effects, 25 and consequently rated their guideline recommendation with the second lowest evidence quality score and lowest strength grade (D).We aimed to inform recommendations in twins by providing more robust evidence from a regression discontinuity design in a large population-based birth registry, which could increase precision relative to data from trials and decrease the risk of confounding relative to data from conventional nonrandomised studies.
Our findings are consistent with the results from the 2020 Cochrane meta-analysis of trials in multiples (e.g.RR for respiratory distress syndrome 0.85, 95% CI 0.61-1.20) 1 and the results from a 2022 meta-analysis of non-randomised studies in twins (e.g.odds ratio for respiratory distress syndrome 0.70, 95% CI 0.57-0.86). 4Together, these results provide strong evidence for the benefit of antenatal corticosteroids in twin pregnancies at imminent risk of preterm delivery and should be used to strengthen the current recommendations in clinical practice guidelines.
An important consideration is the generalisability of our effect estimates to earlier or later gestational ages, as regression discontinuity designs measure the effect at the cut-point (i.e.34 +0 weeks).Meta-analysis of randomised trials has suggested that the relative effect (risk ratio) of antenatal corticosteroids is generalisable across preterm gestational ages. 1 At present there is no evidence that the relative effect of antenatal corticosteroids would not generalise this way in twins, suggesting that it is reasonable that the risk ratios reported here would generalise to earlier and later preterm gestational ages.At the same time, the relative effect being generalisable means that the absolute effect would be greater at earlier gestational ages and smaller at later gestational ages. 26his suggests that risk differences for our outcomes in earlypreterm twins (before the cut-point) would be larger and risk differences for our outcomes in late-preterm twins (after the cut-point) would be smaller than at 34 +0 weeks of gestation.
More evidence is needed to inform discussions around the balance of benefits and harms of antenatal corticosteroid administration in late-preterm twin pregnancies, as current evidence relies heavily on studies in singletons. 16,23,27,28Our study provided evidence of neonatal benefits at gestational ages near our cut-point of 34 +0 weeks, but did not evaluate potential harms.One non-randomised study of twin pregnancies treated in the late-preterm reported a higher risk of neonatal hypoglycaemia (and lower risk of respiratory complications), 29 which is consistent with findings from the ALPS trial in late-preterm singletons. 23Two antenatal corticosteroid trials currently underway in late-preterm twin pregnancies, one enrolling between 34 +0 and 36 +5 weeks of gestation 30 and one restricting to planned twin births between 35 +0 and 38 +6 weeks of gestation, 31 and may help to provide higher-quality evidence on potential neonatal harms of administration in twins.The recent ALPS follow-up study did not find evidence for an adverse effect of antenatal corticosteroid administration on childhood neurodevelopmental outcomes and randomised trial data on longer-term neurodevelopmental outcomes in twins would be valuable. 28hile we found evidence for an effect of antenatal corticosteroids in twins, the magnitude of the effect may still be different in twins versus singletons.An important point when considering the magnitude of our effect estimate is that we are estimating the effect of recommendations to administer antenatal corticosteroids, not the effect of antenatal corticosteroid administration.This distinction is similar to the difference between an ITT and per-protocol analysis in a randomised trial.Our results are consistent with the ITT effect sizes in singletons (e.g.RR for respiratory distress syndrome from the Cochrane meta-analysis 0.65, 95% CI 0.57-0.74 1 ).However, in a placebo-controlled trial no individuals in the placebo group would receive antenatal corticosteroids, and in our data some pregnancies admitted on or after 34 +0 weeks were treated with antenatal corticosteroids.This suggests that our estimates can be viewed as a lower bound for the ITT effect of treatment in twins.However, direct comparisons with trials data in singletons are difficult because of imprecise estimates and differences in other study population characteristics that may affect the outcomes of interest.

| CONCLUSION
Our findings provide evidence on the effectiveness of antenatal corticosteroid treatment for preventing newborn respiratory morbidity in preterm twins.

AU T HOR C ON T R I BU T ION S
PMS conceived of the research question, with input on study design from ES, JAH and SH.PMS carried out the data analyses and wrote the first draft of the manuscript.All authors interpreted the results.All authors reviewed and approved final the manuscript.

AC K NO W L E D GE M E N T S
We thank Emma Branch at the British Columbia Women's Health Research Institute for preparing the study dataset.

C ON F L IC T OF I N T E R E S T S TAT E M E N T
None declared.

F I G U R E 1
Proportion of twin live births exposed to antenatal corticosteroids, by gestational age at maternal admission for birth, in British Columbia, Canada, between 2008 and 2018.F I G U R E 2 Proportion of twin live births diagnosed with respiratory distress or in-hospital mortality (A) and proportion with respiratory intervention or in-hospital mortality (B), by gestational age at maternal admission for birth in British Columbia, Canada, between 2008 and 2018.Lines are predicted probabilities with 95% confidence intervals.

F
U N DI NG I N FOR M AT ION This work was supported by a Canadian Institutes of Health Research project grant to ECS, JAH and SH; and Fonds de recherche du Québec -Santé Master's and Doctoral training awards to PMS.ECS is a McGill University William Dawson scholar.JAH holds a Canada Research Chair in Perinatal Population Health.The funders played no role in conducting the research or writing the manuscript.