Biases arising from linked administrative data for epidemiological research: a conceptual framework from registration to analyses.

Linked administrative data offer a rich source of information that can be harnessed to describe patterns of disease, understand their causes and evaluate interventions. However, administrative data are primarily collected for operational reasons such as recording vital events for legal purposes, and planning, provision and monitoring of services. The processes involved in generating and linking administrative datasets may generate sources of bias that are often not adequately considered by researchers. We provide a framework describing these biases, drawing on our experiences of using the 100 Million Brazilian Cohort (100MCohort) which contains records of more than 131 million people whose families applied for social assistance between 2001 and 2018. Datasets for epidemiological research were derived by linking the 100MCohort to health-related databases such as the Mortality Information System and the Hospital Information System. Using the framework, we demonstrate how selection and misclassification biases may be introduced in three different stages: registering and recording of people's life events and use of services, linkage across administrative databases, and cleaning and coding of variables from derived datasets. Finally, we suggest eight recommendations which may reduce biases when analysing data from administrative sources.

Association of maternal risk factors with infant maltreatment: an administrative data cohort study.

OBJECTIVE: We aimed to evaluate the risk of infant maltreatment associated with commonly used criteria for home visiting programmes: young maternal age, maternal adversity (homelessness, substance abuse, intimate partner violence), newcomer status and mental health concerns in Ontario, Canada. DESIGN: This retrospective cohort study included infants born in hospital in Ontario from 1 April 2005 to 31 March 2017 captured in linked health administrative and demographic databases. Infants were followed from newborn hospitalisation until 1 year of age for child maltreatment captured in healthcare or death records. The association between type and number of maternal risk factors, and maltreatment, was analysed using multivariable logistic regression modelling, controlling for infant characteristics and material deprivation. Further modelling explored the association of each year of maternal age with maltreatment. RESULTS: Of 989 586 infants, 434 (0.04%) had recorded maltreatment. Maternal age <22 years conferred higher risk of infant maltreatment (adjusted OR (aOR) 5.5, 95% CI 4.5 to 6.8) compared with age ≥22 years. Maternal mental health diagnoses (aOR 2.0, 95% CI 1.6 to 2.5) were also associated with maltreatment, while refugee status appeared protective (aOR 0.6, 95% CI 0.4 to 1.0). The odds of maltreatment increased with higher numbers of maternal risk factors. Maternal age was associated with maltreatment until age 28 years. CONCLUSION: Infants born to young mothers are at greater risk of infant maltreatment, and this association remained until age 28 years. These findings are important for ensuring public health interventions are supporting populations experiencing structural vulnerabilities with the aim of preventing maltreatment.

Lessons learned from using linked administrative data to evaluate the Family Nurse Partnership in England and Scotland.

Introduction: "Big data" - including linked administrative data - can be exploited to evaluate interventions for maternal and child health, providing time- and cost-effective alternatives to randomised controlled trials. However, using these data to evaluate population-level interventions can be challenging. Objectives: We aimed to inform future evaluations of complex interventions by describing sources of bias, lessons learned, and suggestions for improvements, based on two observational studies using linked administrative data from health, education and social care sectors to evaluate the Family Nurse Partnership (FNP) in England and Scotland. Methods: We first considered how different sources of potential bias within the administrative data could affect results of the evaluations. We explored how each study design addressed these sources of bias using maternal confounders captured in the data. We then determined what additional information could be captured at each step of the complex intervention to enable analysts to minimise bias and maximise comparability between intervention and usual care groups, so that any observed differences can be attributed to the intervention. Results: Lessons learned include the need for i) detailed data on intervention activity (dates/geography) and usual care; ii) improved information on data linkage quality to accurately characterise control groups; iii) more efficient provision of linked data to ensure timeliness of results; iv) better measurement of confounding characteristics affecting who is eligible, approached and enrolled. Conclusions: Linked administrative data are a valuable resource for evaluations of the FNP national programme and other complex population-level interventions. However, information on local programme delivery and usual care are required to account for biases that characterise those who receive the intervention, and to inform understanding of mechanisms of effect. National, ongoing, robust evaluations of complex public health evaluations would be more achievable if programme implementation was integrated with improved national and local data collection, and robust quasi-experimental designs.

Identifying opportunities for upstream evaluations relevant to child and maternal health: a UK policy-mapping review.

OBJECTIVE: Interventions to tackle the social determinants of health can improve outcomes during pregnancy and early childhood, leading to better health across the life course. Variation in content, timing and implementation of policies across the 4 UK nations allows for evaluation. We conducted a policy-mapping review (1981-2021) to identify relevant UK early years policies across the social determinants of health framework, and determine suitable candidates for evaluation using administrative data. METHODS: We used open keyword and category searches of UK and devolved Government websites, and hand searched policy reviews. Policies were rated and included using five criteria: (1) Potential for policy to affect maternal and child health outcomes; (2) Implementation variation across the UK; (3) Population reach and expected effect size; (4) Ability to identify exposed/eligible group in administrative data; (5) Potential to affect health inequalities. An expert consensus workshop determined a final shortlist. RESULTS: 336 policies and 306 strategy documents were identified. Policies were mainly excluded due to criteria 2-4, leaving 88. The consensus workshop identified three policy areas as suitable candidates for natural experiment evaluation using administrative data: pregnancy grants, early years education and childcare, and Universal Credit. CONCLUSION: Our comprehensive policy review identifies valuable opportunities to evaluate sociostructural impacts on mother and child outcomes. However, many potentially impactful policies were excluded. This may lead to the inverse evidence law, where there is least evidence for policies believed to be most effective. This could be ameliorated by better access to administrative data, staged implementation of future policies or alternative evaluation methods.

Reductions in hospital care among clinically vulnerable children aged 0-4 years during the COVID-19 pandemic.

OBJECTIVE: To quantify reductions in hospital care for clinically vulnerable children during the COVID-19 pandemic. DESIGN: Birth cohort. SETTING: National Health Service hospitals in England. STUDY POPULATION: All children aged <5 years with a birth recorded in hospital administrative data (January 2010-March 2021). MAIN EXPOSURE: Clinical vulnerability defined by a chronic health condition, preterm birth (<37 weeks' gestation) or low birth weight (<2500 g). MAIN OUTCOMES: Reductions in care defined by predicted hospital contact rates for 2020, estimated from 2015 to 2019, minus observed rates per 1000 child years during the first year of the pandemic (March 2020-2021). RESULTS: Of 3 813 465 children, 17.7% (one in six) were clinically vulnerable (9.5% born preterm or low birth weight, 10.3% had a chronic condition). Reductions in hospital care during the pandemic were much higher for clinically vulnerable children than peers: respectively, outpatient attendances (314 vs 73 per 1000 child years), planned admissions (55 vs 10) and unplanned admissions (105 vs 79). Clinically vulnerable children accounted for 50.1% of the reduction in outpatient attendances, 55.0% in planned admissions and 32.8% in unplanned hospital admissions. During the pandemic, weekly rates of planned care returned to prepandemic levels for infants with chronic conditions but not older children. Reductions in care differed by ethnic group and level of deprivation. Virtual outpatient attendances increased from 3.2% to 24.8% during the pandemic. CONCLUSION: One in six clinically vulnerable children accounted for one-third to one half of the reduction in hospital care during the pandemic.

Changes in adolescents' planned hospital care during the COVID-19 pandemic: analysis of linked administrative data.

OBJECTIVE: To describe changes in planned hospital care during the pandemic for vulnerable adolescents receiving children's social care (CSC) services or special educational needs (SEN) support, relative to their peers. DESIGN: Observational cohort in the Education and Child Health Insights from Linked Data database (linked de-identified administrative health, education and social care records of all children in England). STUDY POPULATION: All secondary school pupils in years 7-11 in academic year 2019/2020 (N=3 030 235). MAIN EXPOSURE: Receiving SEN support or CSC services. MAIN OUTCOMES: Changes in outpatient attendances and planned hospital admissions during the first 9 months of the pandemic (23 March-31 December 2020), estimated by comparing predicted with observed numbers and rates per 1000 child-years. RESULTS: A fifth of pupils (20.5%) received some form of statutory support: 14.2% received SEN support only, 3.6% received CSC services only and 2.7% received both. Decreases in planned hospital care were greater for these vulnerable adolescents than their peers: -290 vs -225 per 1000 child-years for outpatient attendances and -36 vs -16 per 1000 child-years for planned admissions. Overall, 21% of adolescents who were vulnerable disproportionately bore 25% of the decrease in outpatient attendances and 37% of the decrease in planned hospital admissions. Vulnerable adolescents were less likely than their peers to have face-to-face outpatient care. CONCLUSION: These findings indicate that socially vulnerable groups of children have high health needs, which may need to be prioritised to ensure equitable provision, including for catch-up of planned care postpandemic.

National administrative record linkage between specialist community drug and alcohol treatment data (the National Drug Treatment Monitoring System (NDTMS)) and inpatient hospitalisation data (Hospital Episode Statistics (HES)) in England: design, method and evaluation.

OBJECTIVES: The creation and evaluation of a national record linkage between substance misuse treatment, and inpatient hospitalisation data in England. DESIGN: A deterministic record linkage using personal identifiers to link the National Drug Treatment Monitoring System (NDTMS) curated by Public Health England (PHE), and Hospital Episode Statistics (HES) Admitted Patient Care curated by National Health Service (NHS) Digital. SETTING AND PARTICIPANTS: Adults accessing substance misuse treatment in England between 1 April 2018 and 31 March 2019 (n=268 251) were linked to inpatient hospitalisation records available since 1 April 1997. OUTCOME MEASURES: Using a gold-standard subset, linked using NHS number, we report the overall linkage sensitivity and precision. Predictors for linkage error were identified, and inverse probability weighting was used to interrogate any potential impact on the analysis of length of hospital stay. RESULTS: 79.7% (n=213 814) people were linked to at least one HES record, with an estimated overall sensitivity of between 82.5% and 83.3%, and a precision of between 90.3% and 96.4%. Individuals were more likely to link if they were women, white and aged between 46 and 60. Linked individuals were more likely to have an average length of hospital stay ≥5 days if they were men, older, had no fixed residential address or had problematic opioid use. These associations did not change substantially after probability weighting, suggesting they were not affected by bias from linkage error. CONCLUSIONS: Linkage between substance misuse treatment and hospitalisation records offers a powerful new tool to evaluate the impact of treatment on substance related harm in England. While linkage error can produce misleading results, linkage bias appears to have little effect on the association between substance misuse treatment and length of hospital admission. As subsequent analyses are conducted, potential biases associated with the linkage process should be considered in the interpretation of any findings.

Cost-effectiveness of strategies preventing late-onset infection in preterm infants.

OBJECTIVE: Developing a model to analyse the cost-effectiveness of interventions preventing late-onset infection (LOI) in preterm infants and applying it to the evaluation of anti-microbial impregnated peripherally inserted central catheters (AM-PICCs) compared with standard PICCs (S-PICCs). DESIGN: Model-based cost-effectiveness analysis, using data from the Preventing infection using Antimicrobial Impregnated Long Lines (PREVAIL) randomised controlled trial linked to routine healthcare data, supplemented with published literature. The model assumes that LOI increases the risk of neurodevelopmental impairment (NDI). SETTING: Neonatal intensive care units in the UK National Health Service (NHS). PATIENTS: Infants born ≤32 weeks gestational age, requiring a 1 French gauge PICC. INTERVENTIONS: AM-PICC and S-PICC. MAIN OUTCOME MEASURES: Life expectancy, quality-adjusted life years (QALYs) and healthcare costs over the infants' expected lifetime. RESULTS: Severe NDI reduces life expectancy by 14.79 (95% CI 4.43 to 26.68; undiscounted) years, 10.63 (95% CI 7.74 to 14.02; discounted) QALYs and costs £19 057 (95% CI £14 197; £24697; discounted) to the NHS. If LOI causes NDI, the maximum acquisition price of an intervention reducing LOI risk by 5% is £120. AM-PICCs increase costs (£54.85 (95% CI £25.95 to £89.12)) but have negligible impact on health outcomes (-0.01 (95% CI -0.09 to 0.04) QALYs), compared with S-PICCs. The NHS can invest up to £2.4 million in research to confirm that AM-PICCs are not cost-effective. CONCLUSIONS: The model quantifies health losses and additional healthcare costs caused by NDI and LOI during neonatal care. Given these consequences, interventions preventing LOI, even by a small extent, can be cost-effective. AM-PICCs, being less effective and more costly than S-PICC, are not likely to be cost-effective. TRIAL REGISTRATION NUMBER: NCT03260517.

Reflections on modern methods: linkage error bias.

Linked data are increasingly being used for epidemiological research, to enhance primary research, and in planning, monitoring and evaluating public policy and services. Linkage error (missed links between records that relate to the same person or false links between unrelated records) can manifest in many ways: as missing data, measurement error and misclassification, unrepresentative sampling, or as a special combination of these that is specific to analysis of linked data: the merging and splitting of people that can occur when two hospital admission records are counted as one person admitted twice if linked and two people admitted once if not. Through these mechanisms, linkage error can ultimately lead to information bias and selection bias; so identifying relevant mechanisms is key in quantitative bias analysis. In this article we introduce five key concepts and a study classification system for identifying which mechanisms are relevant to any given analysis. We provide examples and discuss options for estimating parameters for bias analysis. This conceptual framework provides the 'links' between linkage error, information bias and selection bias, and lays the groundwork for quantitative bias analysis for linkage error.

Establishing a composite neonatal adverse outcome indicator using English hospital administrative data.

OBJECTIVE: We adapted a composite neonatal adverse outcome indicator (NAOI), originally derived in Australia, and assessed its feasibility and validity as an outcome indicator in English administrative hospital data. DESIGN: We used Hospital Episode Statistics (HES) data containing information infants born in the English National Health Service (NHS) between 1 April 2014 and 31 March 2015. The Australian NAOI was mapped to diagnoses and procedure codes used within HES and modified to reflect data quality and neonatal health concerns in England. To investigate the concurrent validity of the English NAOI (E-NAOI), rates of NAOI components were compared with population-based studies. To investigate the predictive validity of the E-NAOI, rates of readmission and death in the first year of life were calculated for infants with and without E-NAOI components. RESULTS: The analysis included 484 007 (81%) of the 600 963 eligible babies born during the timeframe. 114/148 NHS trusts passed data quality checks and were included in the analysis. The modified E-NAOI included 23 components (16 diagnoses and 7 procedures). Among liveborn infants, 5.4% had at least one E-NAOI component recorded before discharge. Among newborns discharged alive, the E-NAOI was associated with a significantly higher risk of death (0.81% vs 0.05%; p<0.001) and overnight hospital readmission (15.7% vs 7.1%; p<0.001) in the first year of life. CONCLUSIONS: A composite NAOI can be derived from English hospital administrative data. This E-NAOI demonstrates good concurrent and predictive validity in the first year of life. It is a cost-effective way to monitor neonatal outcomes.

A guide to evaluating linkage quality for the analysis of linked data.

Linked datasets are an important resource for epidemiological and clinical studies, but linkage error can lead to biased results. For data security reasons, linkage of personal identifiers is often performed by a third party, making it difficult for researchers to assess the quality of the linked dataset in the context of specific research questions. This is compounded by a lack of guidance on how to determine the potential impact of linkage error. We describe how linkage quality can be evaluated and provide widely applicable guidance for both data providers and researchers. Using an illustrative example of a linked dataset of maternal and baby hospital records, we demonstrate three approaches for evaluating linkage quality: applying the linkage algorithm to a subset of gold standard data to quantify linkage error; comparing characteristics of linked and unlinked data to identify potential sources of bias; and evaluating the sensitivity of results to changes in the linkage procedure. These approaches can inform our understanding of the potential impact of linkage error and provide an opportunity to select the most appropriate linkage procedure for a specific analysis. Evaluating linkage quality in this way will improve the quality and transparency of epidemiological and clinical research using linked data.

Resource and outcome in paediatric diabetes services.

The availability of resource (staffing and services) in all 21 paediatric diabetes services in Yorkshire and Humber Strategic Health Authority, UK was surveyed and this information was combined with demographic and clinical data on 2683 children and young people with diabetes (aged 0-23 years) to assess whether level of resource was associated with glycaemic control (mean HbA1c %). Multilevel modelling and graphical techniques were used to analyse the relationship between resource and outcome for paediatric diabetes services. No services achieved all resource recommendations based on National Institute for Health and Clinical Excellence guidelines, but there was no direct association between level of resource and glycaemic control after controlling for deprivation, age and duration of diabetes. Transitional care, nurse caseload and access to specialist services are not adequately resourced but variation in outcome between services is not accounted for by level of resource.

Rising rates of all types of diabetes in south Asian and non-south Asian children and young people aged 0-29 years in West Yorkshire, U.K., 1991-2006.

OBJECTIVE: To investigate incidence trends of all diabetes types in all children and young people and in the south Asian subpopulation. RESEARCH DESIGN AND METHODS: Annual incidence per 100,000 and time trends (1991-2006) were analyzed for 2,889 individuals aged 0-29 years diagnosed with diabetes while resident in West Yorkshire, U.K. RESULTS: Diagnoses comprised type 1 (83%), type 2 (12%), maturity-onset diabetes of the young (0.7%), "J"-type/other (0.1%), and uncertain/unclassified (4%). There was a lower incidence of type 1 and a threefold excess of type 2 in south Asians compared with non-south Asians. Type 1 incidence leveled out and type 2 increased after the first south Asian case of type 2 was diagnosed in 1999. Type 2 and unclassified diabetes incidence rose in all population subgroups. CONCLUSIONS: The burden of diabetes increased over time for both ethnic groups, with a significant excess of type 2 diabetes in south Asians. The rising incidence of type 1 diabetes in south Asians attenuated as type 2 diabetes increased after 1999.