Data Sources
This is a population-based birth cohort study using Hospital Episode Statistics (HES), a hospital administrative database containing data from all National Health Service (NHS) hospitals in England. HES captures data on all A&E attendances (HES A&E), admitted patient care (HES APC) admissions and outpatient appointments in NHS-funded hospitals across England. HES data include a unique pseudonymised patient identifier (the HESID) which enables longitudinal linkage of children’s hospital contacts over time (25, 26). We used a previously constructed nationally representative birth cohort generated using HES APC birth episodes linked to maternal delivery episodes, which includes all singleton children born in NHS hospitals to English resident mothers. The cohort includes 97% of all births in England; full details of the birth cohort derivation are described in elsewhere (27, 28).
We extracted data on all A&E attendances and emergency admissions for infants in the birth cohort from HES A&E and HES APC respectively. We used the methods proposed by NHS Digital for linking A&E attendances and emergency admissions (described in online supplementary materials S1) (29, 30).
Study population and follow-up
We included all children in the cohort born between 1st April 2012 and 31st March 2019. Infants were followed up from 1st April 2013 (start of follow-up) or discharge date of their birth admission (whichever occurred later), to date of first birthday, date of death, estimated date of migration or end of follow-up on 31st March 2019 (whichever came first). The study period ensured that all study years included follow-up for infants aged up to 12 months and excluded any time spent in hospital. Infants who moved out of England were identified by the presence of a non-English postcode for any A&E attendance or emergency admission during the study period. The date of emigration was set at mid-point between the infant’s start of follow-up date and the date at which the non-resident attendance or admission was recorded. Infants born < 24 weeks gestation and multiple births were excluded. Figure 1 shows the flow-chart of data linkage and final sample of cohort infants used in the analysis.
Outcomes
The study’s primary outcomes were rates of A&E attendances and rates of emergency admissions in infants. The secondary outcome was conversion the probability, i.e. the proportion of children admitted to hospital after attending A&E.
A&E attendances
A&E attendances were identified using the HES A&E dataset and defined as an unplanned attendance to a 24-hour consultant-led A&E department, consultant-led mono specialty A&E department (e.g. ophthalmology or dental A&E departments), and other types of A&E departments and minor injury units. Follow-up A&E attendances, attendances at NHS walk-in centres, and A&E attendances in private hospitals were excluded. For infants with more than one A&E attendance in one day, only the last attendance was included to avoid double counting and allow linkage with a potential subsequent emergency admission.
Emergency admissions
The HES APC dataset was used to identify emergency admissions. We identified two distinct pathways to an emergency admission within the APC dataset using the variable ‘admission method’. The first was via a hospital A&E attendance (indirect), where parents/carers would take infants to an A&E department in a hospital and the infant was then admitted as an emergency from the A&E department (also called conversion). The second was a direct hospital admission (circumventing an A&E attendance, often made by a general practitioner; GP). We included all emergency admissions (irrespective of admission pathway) for analysis. Admissions were classified as an emergency if the first episode of care within a multi-episode admission was recorded as an emergency episode.
Infant And Maternal Risk Factors
Infant birth characteristics included infant’s sex, gestational age at birth defined as severe prematurity (< 32 weeks), moderate prematurity (32–33 weeks), near term (34–36 weeks), and term (≥ 37 weeks) (31), and congenital anomalies, identified using an International Classification of Diseases version 10 (ICD10) code list (28, 32). Children were classified as having a congenital anomaly if a relevant code was recorded at birth, during any hospital admission before 2 years of age, or on a death certificate as any cause of death. Maternal age was categorised as under 20, 20–29, 30–39 and 40 + years at the time of delivery (33). Quintiles of the Index of Multiple Deprivation score (IMD, 2010 version) were used as an indicator of socio-economic status. The IMD is a composite measure of multiple deprivation at Lower Super Output Area level (covering 1500 people on average) across seven domains (34, 35). HES financial years (running April to March in the UK) were used to ensure comparison of annual rates published by Public Health England (PHE) (15, 16, 36).
We used information about infant’s local authority (LA) of residence at birth to examine variations in A&E attendances and emergency admissions across English local areas. There were 152 LAs (upper tier) across England during the study period. Due to small population sizes (< 1000 births across the study period), we combined the City of London with Hackney, and the Isles of Scilly with Cornwall. This resulted in 150 LAs included in the analyses. LAs were grouped into region of residence (North East, North West, Yorkshire and Humber, East Midlands, West Midlands, East of England, London, South East, and South West) to describe overall patterns of A&E attendances and emergency admissions at a regional level in England.
Statistical Methods
We described distribution of region, infant and maternal risk factors for children in the cohort (overall and in complete case cohort excluding children with missing data). Main analyses were restricted to complete case cohort. Rates of A&E attendances and emergency admissions were calculated per 1000 child-years, overall and stratified according to all risk factors (37).
We fitted mixed-effects negative binomial regression models to examine the association between infant and maternal risk factors and rates of A&E attendances and admissions and derive incidence rate ratios (IRRs) (38). The outcome variable was counts of A&E attendances or emergency admissions; separate models were fitted for each outcome. Year and month of birth (categorised as standard calendar quarters), infant’s sex and gestational age, presence of congenital anomaly, maternal age, and quintiles of IMD were included as exposure variables, and person-time at risk as the offset. Risk factors to be included in the models were selected a priori. Year and month of birth were included in the first model (model 1), then infant-related variables (infant sex, gestational age, presence of congenital anomaly) were added (model 2), and finally, maternal age, and quintiles of IMD were included (model 3). Heterogeneity due to unobserved variables within LAs was accounted for with a random effect term in the intercept of the models. Variance Partition Coefficients (VPC) defined as: \(\frac{{\sigma }_{u}^{2}}{{\sigma }_{u}^{2}+{\sigma }_{e}^{2}}\) (Eq. 1),
that is, as the ratio of the residual variance due to between-LA random effects (\({\sigma }_{u}^{2})\)and the total residual variance. Thus, VPC can be interpreted as the proportion of the total residual variance attributable to the random effects (38, 39). In other words, the proportion of variation that is beyond that explained by the fixed predictors that is due to between-LA variation.
We assessed goodness-of-fit for all models by changes in Bayesian information criterion (BIC) (40), with smaller values indicating better model fit. Normal probability plots of deviance residuals were used to determine deviations outside of the normal expected range, as only 5% of deviance residuals should lie outside ± 1.96. The assumption of normality for the models’ random effects was assessed with probability plots of the individual random effects predictions.
To visually examine LA level variation in rates of A&E attendances and emergency admissions, we used crude and adjusted rates to construct funnel plots with multiplicative over-dispersion adjusted control limits. Expected events were obtained from the final model’s predictions using both the fixed effects estimates and the LA random effect term (model 3). Adjusted rates were then calculated as the observed number of events divided by the expected number of events obtained using predictions from the final model for each outcome, multiplied by the overall crude rate for England (41). To further explore the contribution of infant and maternal factors within the model, we compared LA effect predictions from the final adjusted mixed-effects models (model 3) against those from model 1 (i.e., the null model). Maps were generated to show variations of adjusted rates (per 1000 child-years) of A&E attendances, emergency admissions and conversation probabilities by local authorities across England.
Models for conversion probabilities
Mixed-effects logistic regression models were fitted to determine the association between individual-level infant and maternal characteristics, and the probability of being admitted to hospital given that the infant had attended A&E (the conversion probability). These models were parametrised in terms of odds ratios. We included the same covariates and used the same model selection strategy as for the negative binomial regression models described above. LA was included as a random effect in the models’ intercept. For this analysis, we selected at random one A&E attendance per infant for those with multiple A&E attendances during infancy. VPC was calculated using:
\(\frac{{\sigma }_{u}^{2}}{{\sigma }_{u}^{2}+\frac{{\pi }^{2}}{3}}\) (Eq. 2),
where \({\sigma }_{u}^{2}\) is the variance of the within LA random effects (level 2), and the level 1 variance for a logistic regression model is obtained as the variance of the standard logistic distribution, \(\frac{{\pi }^{2}}{3}\) (42).
Sensitivity analyses
Logistic regression models were used to examine factors that were associated with missingness (outcome: complete data vs any missing data) showing that gestational age was associated with increased odds of missing information. To investigate the potential influence of missing data on the results, we ran all final models including an additional category of missing to the infant gestational age category. All statistical analysis were conducted in STATA v16 (43).