Ethics approval and consent to participate

This project has been approved by the following ethics committees and data custodians, including a waiver to seek participant consent from The University of Queensland Human Research Ethics Committee (HREC) (approval number 2021/HE001925) and the Metro South Health HREC (approval number HREC/2022/QMS/83690). Access to the administrative health data-sets was approved by the Director-General of Queensland Health as Chief Executive under the Public Health Act 2005 (PHA 83690). The study is registered with the Australian and New Zealand Clinical Trials Registry (ACTRN12622000870752).

Study design

This is an observational, longitudinal birth cohort study linked to administrative health data. We will report findings per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (Supplementary Table 1)^{Reference von Elm, Altman, Egger, Pocock, Gøtzsche and Vandenbroucke18} and the Reporting of Studies Conducted using Observational Routinely-Collected Health Data (RECORD) Statement (Supplementary Table 2).^{Reference Benchimol, Smeeth, Guttmann, Harron, Moher and Petersen19}

Setting

The original study setting was a large hospital that accounted for around 50% of the births in Brisbane, the capital city of the state of Queensland, servicing both public and private patients.^{Reference Najman, Bor, O'Callaghan, Williams, Aird and Shuttlewood20}

Data sources/measurement

This study will link data to the MUSP cohort (Fig. 1) from the following Queensland-wide databases, up until 1 January 2020 (Figs 2 and 3): (a) the Queensland Hospital Admitted Patients’ Data Collection (QHAPDC) of all admissions for all diagnoses to all hospitals in the public (from 1 January 2000) and private sectors (from 1 July 2007), (b) the Emergency Data Collection (EDC) of all emergency department presentations state-wide (from 2008) and (c) the Consumer Integrated Mental Health Application (CIMHA) of all contacts with state-run community-based or out-patient mental health services in the State (from 2010).

Fig. 2

Contributing data-sets.

Fig. 3

Database linkages. CIMHA, Consumer Integrated Mental Health Application; EDC, Emergency Data Collection; MUSP, Mater-University of Queensland Study of Pregnancy; QHAPDC, Queensland Hospital Admitted Patients’ Data Collection.

All three databases have used diagnostic codes from the ICD-10²¹ for the entire period of the proposed study. We will investigate the possibility of using a cross-walk to convert these codes to ICD-11 diagnoses.

Participants

As noted previously, the cohort consisted of 7223 mother–child pairs recruited from the Mater Misericordiae Mothers’ Hospital, the main obstetrics unit for Brisbane. The recruitment started in January 1981, using a sampling frame involving all consecutive pregnant women attending a booking in clinic that usually occurred in the first trimester of pregnancy. The sampling frame excluded emergency transfers and admissions referred from other hospitals.^{Reference Najman, Bor, O'Callaghan, Williams, Aird and Shuttlewood20} This study will focus on the children's health outcomes. Data were collected from mothers at their first clinic visit, and from mother–child pairs when children were 6 months and 5, 14 and 21 years of age (Fig. 3).^{Reference Najman, Alati, Bor, Clavarino, Mamun and McGrath8}

Mothers’ and children's data were collected separately at the 27-year (mothers) and 30-year follow-ups (children). All follow-ups of the MUSP have received ethical clearance. In September 2000, cases of child abuse and neglect that were reported to the Department of Families, Youth and Community Care (DFYCC) were accessed and confidentially linked to the longitudinal database following approval from the ethical review committees of both the Mater Misericordiae Children's Hospital and The University of Queensland.^{Reference Strathearn, Mamun, Najman and O'Callaghan22} The DFYCC (renamed in 2001) was Queensland's statutory child protection agency. Information included the date of referral, whether the referral was substantiated on subsequent investigation, and the type of maltreatment (physical, emotional, sexual and neglect) up until the child was 15 years of age. Confidentiality was preserved by using an identification number to link the two databases anonymously.

Data linkage

There are two groups of researchers (Professor Najman and colleagues, as well as the MUSP data custodian) who are based at the School of Public Health, and the other team members are based in the Medical School at the Princess Alexandra Hospital. The Princess Alexandra Hospital researchers who will do the analysis will not have access to identifiable administrative health, MUSP or child protection data, and will only be able to view anonymised data extracts.

Confidentiality will be preserved by using an identification number to link the two person-level databases anonymously. The researchers analysing the maltreatment data will have no access to identifying information. Ethical approval for the anonymous database matching at the time was obtained from the ethical review committees of both the Mater Misericordiae Children's Hospital and The University of Queensland.

Data linkage will be done by the Statistical Services Branch (SSB) of Queensland Health. The MUSP data custodian will send a data-set to the SSB that is limited to the name, date of birth and gender of birth cohort participants (Fig. 3). This data-set will contain no other information so the SSB staff will only know that participants were born in Brisbane between 1981 and 1983. Within the SSB, these details will be matched to the Queensland Health master linkage file (MLF), using Oracle SQL Developer, version 22.2.1 for Windows (Oracle Corporation, Redwood Shores, CA, USA; www.oracle.com/database/sqldeveloper/technologies/download/). This involves a thorough review of all possible matches rather than using match thresholds, meaning the linkage is of very high accuracy. Every linkage process is cross-checked by the project linkage team manager. The MLF uses several custom ChoiceMaker (version 2.7 for Windows; ChoiceMaker LLC, Princeton, NJ, USA; www.choicemaker.com/Download) ‘models’ to build the linkage map. To determine linkage decisions for pairs of records, models use a series of weighted tests to calculate a match probability, apply decision thresholds (probability cut-offs) to assign a differ/review/match decision (all machine learning), and then in certain conditions apply rules to override the probability-assigned decision. To assure quality and efficiency, SSB's proprietary linkage models are developed for different types of data sources. Each model can comprise hundreds of tests, with test weights assigned through supervised machine learning, using robust training data with fine-tuned decision thresholds and deterministic rules. The error rates are approximately 99.5% for false positives and 98% for false negatives, with more detail about the processes provided online.²³

Each person is assigned a unique project identification number that is returned to the MUSP data custodian, as well as the custodians for each administrative data-set. The data custodian attaches the relevant clinical or service information to the unique project identification number, but no identifying information. These anonymised data extracts are then directly released to a third-party analyst (in this case, the research team at the Princess Alexandra Hospital), but not returned to the MUSP data custodian. The third-party analyst uses the project IDs from the anonymised data to link the four data-sets. The third-party analyst then does the analysis, but only releases summary tables, not individual-level data.

Data and linkage quality

Administrative data are not collected for research purposes and therefore subject to recording bias. However, linking several databases from the same population permits cross-checking and validation of information. As noted previously, there are several quality assurance checks.²³ Although weights assigned by the linkage software process can largely determine true positive and true negative links, there is an area of uncertainty that requires human judgement. Paired sets of identifying information are therefore examined manually to inform whether they are to be considered to be true matches. Following this, identifying variables are cleaned and standardised, and quality assurance checks are performed for missing or duplicate data. Previous work has reported that demographic features such as Indigenous status were correctly identified in 89% of cases when checked with other sources.^{Reference Kisely and Pais24}

In keeping with standard practice, researchers will also perform basic frequency analyses following receipt of the data to include checks for duplicate records before and after merging data-sets, and whether the gender distribution, or years covered, seems appropriate.²³

Variables

The following are predictor variables and possible covariates that will come from the MUSP data-set. We cross-checked these against all the confounders included in a 2021 systematic review and meta-analysis of prospective longitudinal cohort studies in this area.^{Reference McKay, Cannon, Chambers, Conroy, Coughlan and Dodd25}

1. (a) The number of notifications, if any, received by DFYCC in relation to each participant in the cohort before the collection in 2000, at which time the youngest children in the cohort were approximately 16.5 years of age.

2. (b) The number and proportion of these notifications that were substantiated by the DFYCC after investigations.

3. (c) The number and proportions of notifications and substantiations attributable to the recognised subtypes of abuse and neglect, i.e. physical abuse, sexual abuse, emotional abuse and neglect.

4. (d) Self-reported child maltreatment as measured by the Childhood Trauma Questionnaire Short-Form (CTQ)^{Reference Bernstein, Stein, Newcomb, Walker, Pogge and Ahluvalia26} at the 30-year follow-up. This has been used extensively and been found to have reliability, sensitivity and discriminant validity. There are five subscales with five items each: physical abuse, sexual abuse, emotional abuse, physical neglect and emotional neglect. Scores across the five items in each subscale are aggregated (producing an ordinal measure) and then recoded to four levels: none (or minimal), low (to moderate), moderate (to severe) and severe (to extreme). To ensure consistency with previous work on the same cohort, we will dichotomise the variables into none/low and moderate/severe, but also perform sensitivity analyses of the effect of splitting the scores into ‘severe’ versus all other categories.^{Reference Kisely, Strathearn, Mills and Najman16,Reference Najman, Kisely, Scott, Strathearn, Clavarino and Williams27} This approach will enable comparisons between self- and agency-reported maltreatment in line with our earlier work on which the CTQ designer was a co-author.^{Reference Najman, Kisely, Scott, Strathearn, Clavarino and Williams27} Self-reported data on sexual abuse were also collected at 21-year follow-up with a different questionnaire based on five non-exclusive abuse scenarios.^{Reference Mills, Kisely, Alati, Strathearn and Najman15}

5. (e) Pregnancy-related characteristics of both the mothers of the original MUSP participants and when they, in turn, had children. These include the following: prenatal intention of pregnancy, feelings about becoming pregnant, number of previous pregnancies, number of previous pregnancies and live births, maternal dyad and relationship status during pregnancy, perceptions of pregnancy (easy/hard), problems with delivery, prenatal stressful life events, attitudes toward their newborn and social support during or after pregnancy. Pregnancy-related characteristics have previously been shown to be associated with a range of adverse outcomes in the MUSP cohort.^{Reference Najman, Alati, Bor, Clavarino, Mamun and McGrath8}

6. (f) Sociodemographic variables on entry to the study, including year of birth, gender of the child, parental racial origin, maternal age, parental relationship status, family income at study entry (first prenatal visit), chronic socioeconomic disadvantage and parental education or employment.

7. (g) Sociodemographic variables up until the 30-year-old follow-up, including the offspring's employment status, income level at 30 years, educational achievement and marital status, maternal or paternal absence or both, and out of home placement. The administrative data will give employment and marital status at 40 years.

8. (h) Behavioural and environmental factors, including parental lifestyle, intimate partner conflict and violence, internalising or externalising behaviours, victim of bullying at 14 years, maternal mental health and family instability or dysfunction.

Outcomes from linked administrative data

These will be health service contacts as measured by QHAPDC (from 1 January 2000 for public hospitals and 1 July 2007 for private hospitals), the EDC (from 2008) and CIMHA (from 2010) up until 31 December 2019, in terms of numbers of visits and diagnoses as outlined by the relevant ICD-10 codes (Table 1). Participants will be classified as having a psychiatric disorder if they have ever had a hospital admission, emergency department presentation or community contact with a primary diagnosis of a psychiatric disorder (ICD-10 codes F00 to F99), suicidal ideation (code R45.8) or self-harm (codes X60 to X84). Psychiatric diagnoses will be divided into six broad categories, with each broad category further subdivided into more detailed subcategories (Table 1): severe mental disorders (non-affective psychoses and bipolar disorder), common mental disorders (anxiety and depressive disorders and post-traumatic stress disorder), personality disorders, substance use disorders (excludes substance-induced psychoses, which are included under severe mental disorders), other adult-onset disorders and other childhood onset disorders.

Table 1

ICD-10 codes

Diagnostic categories and subdivisions will be mutually exclusive sets of ICD-10 codes and coded as binary indicators (present/not present) for each individual. We expect that drug, alcohol and mental health problems will be common reasons for presentation, given that the cohort will be approximately 40 years of age, based on pilot data from the decision support team at the Princess Alexandra Hospital.

Analyses

We will compare people with all forms of self- and agency-reported child maltreatment to the rest of the cohort, adjusting for potential confounders with a series of logistic, Cox and multiple regressions as follows: (a) logistic regression analyses to explore the association with a discrete outcome (e.g. admission for a particular psychiatric diagnosis), and control for relevant demographic and clinical covariates – this will be the primary outcome for each psychiatric diagnosis; (b) survival analysis to identify a terminal event (i.e. admission) and the time to event; (c) multiple regression analyses to allow predictions of continuous outcomes (e.g. number of presentations, bed days) and (d) where appropriate, we will apply marginal structural models, inverse probability weighting of marginal structural cox models and the g-computation formula for causal inference.^{Reference Hernán, Brumback and Robins28,Reference Robins29}

Each form of childhood maltreatment will be sequentially entered into the models. We will run separate analyses for each administrative data-set (QHAPDC, CIMHA and EDC) and the psychiatric outcome of primary interest. Given our pilot data (see above), we will focus on the following diagnoses: self-harm, severe mental disorders (non-affective psychoses, bipolar disorder), common mental disorders (depression, anxiety or post-traumatic stress disorder), substance use (intoxication, hazardous/harmful use or dependence) and alcohol use (intoxication, hazardous/harmful use or dependence). Many of these outcomes were investigated in our earlier analyses of just the MUSP cohort, so allowing comparison with our previous work. Other psychiatric disorders will be explored if there are sufficient numbers.

In the case of self-reported child maltreatment, we will undertake propensity score analyses that considers baseline covariates across the entire at-risk cohort rather than multiple imputation, as we cannot assume the data will be missing at random. In addition, previous work has multiple imputation of data from this cohort changes neither the estimates nor their precision, and that rate of loss to follow-up, for both dependent and independent variables, has little effect on estimates of association.^{Reference Najman, Alati, Bor, Clavarino, Mamun and McGrath8}

Power calculations are based on dichotomous variables (our primary outcome), and used the WINdows Programs for EPIdemiologists (WINPEPI) program, version 11.65 for Windows (Joe Abramson, Jerusalem, Israel; www.brixtonhealth.com/pepi4windows.html), for logistic regression, titled LOGISTIC, which allows for adjustment for covariates, with specified multiple correlation coefficients between covariates and the tested predictor.^{Reference Abramson30} Data from studies that were restricted to the MUSP cohort suggest minimum differences in rates of adverse outcomes between agency- or self-reported child maltreatment of between 10 and 15%, with odds ratios ranging from 1.34 to 5.53. Power calculations therefore indicate we would require a minimum of 6500 participants to have an 80% chance of detecting a significant difference at P < 0.05, meaning that our current sample is sufficient for commonly expected outcomes. We will handle missing data following ‘Treatment And Reporting of Missing Data in Observational Studies’ framework, and clean data following the guidance to detect, diagnose and edit data abnormalities.^{Reference Lee, Tilling, Cornish, Little, Bell and Goetghebeur31,Reference Van den Broeck, Argeseanu Cunningham, Eeckels and Herbst32}

Feasibility

As outlined above, no new data will be needed. The MUSP data have already been collected and previously linked with the child protection data, whereas the administrative health data are available from Queensland Health. The QHAPDC, CIMHA and EDC data custodians have approved the project, and linkage and data cleaning are underway. In terms of notified maltreatment, 789 out of the 7223 participants (10.8%) had some type of notification by 16 years of age. In order of frequency, these were neglect (n = 500), emotional abuse (n = 446), physical abuse (n = 475) and sexual abuse (n = 259). At the 30-year follow-up, 600 out of the 2816 participants (24.7%) self-reported maltreatment of any type, with 326 (13.4%) rating this as severe. Neglect was the most frequent maltreatment (n = 382), followed by emotional (n = 225), sexual (n = 198) and physical abuse (n = 197).

Patient and public involvement

Patients and the public were not involved in the design of the study. However, as described in our dissemination activities, MUSP participants are regularly informed that information about studies arising from MUSP can be found on the MUSP website. This includes a full list of publications, as well as user-friendly descriptions of research. In addition, two of the authors are investigators in the National Research Translation Centre to Implement Mental Health Care at Scale in Primary Care and Community Settings (ALIVE) in Australia. This is a nationwide programme to co-design and evaluate approaches to improve mental health through evidence-based strategies for addressing physical, behavioural, psychological and other determinants, in partnership with people with lived experience. Early identification of risk factors and prevention of mental illness was identified in consultation with people with lived experience and their carers as a research priority, and this has informed the design of the present programme.

Plans for dissemination and publication of project outcomes

The immediate focus of the knowledge translation component of our work will be on those providing services to families with experiences of adversity. These service providers will primarily include nurses, social workers and other staff employed in hospitals or the Department of Child Safety of the Queensland Government. In addition, MUSP participants are also regularly informed of research arising out of the MUSP study (see above). Finally, findings will be submitted for presentation at academic conferences and publication in peer-reviewed journals.