Health and household environment factors linked with early alcohol use in adolescence: a record-linked, data-driven, longitudinal cohort study.

Introduction: Early alcohol use has significant association with poor health outcomes. Individual risk factors around early alcohol use have been identified, but a holistic, data-driven investigation into health and household environmental factors on early alcohol use is yet to be undertaken.
Objectives: This study aims to investigate the relationship between preceding health events, household exposures and early alcohol use during adolescence using a two-stage data-driven approach.
Methods: In stage one, a study population (N = 1,072) were derived from the Millennium Cohort Study (MCS) Wales (born between 2000-2002). MCS data were first linked with electronic-health records. Factors associated with early (<=eleven years old) alcohol use were identified using feature selection and stepwise logistic regression. In stage two, analogous risk factors from MCS were recreated for whole population (N = 59,231) of children (born between 1998-2002 in the Welsh Demographic Service Dataset) using routine data to predict the alcohol-related health events in hospital or GP records.
Results: Significant risk factors from stage two included poor maternal mental (adjusted odds ratio [aOR] = 1.31) and physical health (aOR = 1.25), living with someone with alcohol-related problem (aOR = 2.16), single-adult household (aOR = 1.45), ever in deprivation (aOR = 1.66), child's high hyperactivity (aOR = 3.57), and conduct disorder (aOR = 3.26). Children with health events, whose health needs are supported (e.g., are taken to the doctor), are at lower risk of early alcohol use.
Conclusion: Health events of the family members and the child can act as modifiable exposures and may therefore inform the development of prevention initiatives. Families with known alcohol problems, living in deprivation, experiencing child behavioural problems and those who are not taken to the doctor are at higher risk of early drinking behaviour and should be prioritised for early years support and interventions to target problem drinking in young people.

Introduction

Alcohol use in childhood is associated with the risk of later alcohol abuse, alcohol dependence [1] and several negative outcomes including poor educational achievement, death and disability [2-5]. Known factors that predict early alcohol use include a child’s hyperactivity and conduct disorder [6, 7], lack of family support, household dysfunction, parental alcohol drinking pattern, parental indifference towards young persons’ alcohol use [8-11] and adverse childhood experiences (ACEs) (e.g., child abuse and parental discord) [12]. Current research has largely focused on the family environment, individual level socio-demographic, neurocognitive, behavioural or emotional features, individually or in combination [13-15]. Although it is known that ACEs have a detrimental impact on a child’s health in early life [16, 17], it is not known whether a child’s own health status is associated with subsequent alcohol use and alcohol-related health outcomes.

Child health is a broad term that includes maintaining and protecting physical, mental and social health [18]. Broadly, there are two dominant methodological approaches in the investigation of child alcohol use that are increasingly regarded as complementary [19]. First, survey methodology allows researchers to focus on specific exposures and outcomes, such as volume of alcohol consumed, and to tailor validated [20] instruments to address preconceived study hypothesis [2]. Limitations include relatively small sample size, non-response, selection and volunteer bias [21]. Second, the analysis of routinely collected electronic health records (EHRs) facilitates the inclusion of a greater number of individuals, even entire populations, than is feasible using surveys. The analysis of whole population EHRs, however, imposes challenges relating to the processing and management of data, including addressing missing data on informative variables [22]. For example, EHRs are unlikely to capture occasional alcohol consumption but would be expected to capture health outcomes relating to hazardous alcohol use.

Existing literature on this topic has predominantly focused on preconceived study hypothesis [2], however this increases the chance of missing risk factors which have not already been identified. In contrast to this, a data-driven framework would avoid the limits of a pre-defined and hypothesis-bound investigation and significantly open up the exploration of the variable space. We anticipate that this will provide new insights and will ultimately help to develop a better understanding of the research problem under investigation. Hence, the current study does not focus on an explicit causal analysis, rather we aim to merge hypothesis-based knowledge with data-driven insights to investigate the risk factors associated with early alcohol use.

In this study we assess the relationship between childhood health factors, household environment and alcohol-related outcomes during adolescence using a two-stage data-driven approach. These broad categories of risk factors were based on hypothesis-based knowledge as discussed above. This method brings together a hypothesis-based study design followed by a data-driven approach which complements and minimises the limitation of both study designs.

Methods

A two-stage data-driven approach has been undertaken to investigate the association between the specific risk factors and the outcome in this study. In stage one, a machine learning feature selection algorithm and a classifier were used to identify the health conditions and socio-demographic factors associated with early alcohol use from linked EHRs and Millennium Cohort Study (MCS) survey data. In stage two, analogous risk factors identified from stage one were then sought in routine data and an analytic approach was used to determine the prediction model. The linked routinely collected EHRs and vast volume of administrative data from the whole population of Wales was analysed to determine the effect of the risk factors identified in the MCS data analysis as predictors to target alcohol-related health outcomes in the general adolescent population.

Stage one – Millennium Cohort Study (MCS)

Participants

The MCS is a longitudinal birth cohort of children born in the UK between the years 2000 and 2002 [23]. Parents of the original 18,819 singleton children were interviewed from all parts of UK when their child was nine months old, of those 1,951 were interviewed in Wales. Subsequent interviews took place at ages three, five, seven and eleven years of age. Written consent to link MCS children with their routine EHRs up to age fourteen years was obtained from their parents at the interview undertaken when children were seven years of age. Data of the 1,838 consented singleton children resident in Wales was subsequently linked with their EHRs. The study population included children who also participated in the interview at age eleven years, as the primary outcome data were collected at that point. The current study excluded participants who did not have a general practitioner (GP) record in the Welsh Longitudinal General Practice (WLGP) dataset before they were eleven years of age (Supplementary Figure 1).

Exposure

The study included parent reported socio-demographic and family-related variables for children from MCS interviews which took place between the age of nine months and seven years of the children. These include child’s sex, mother’s socio-economic classification (SEC), household poverty level (whether the household income was above/below 60% of national median using a modified Organisation for Economic Co-operation and Development scale), living area (based on 2005 Rural/Urban Area Classification), mother’s alcohol use during and post pregnancy, lone parent carer, and number of children. Based on lone parent status, the total number of siblings at household and total number of household members, the study derived a binary variable to identify whether the child was residing with any other additional household members. Using both parents’ responses on alcohol consumption, guardian alcohol use variables were derived. Children’s emotional and behavioural difficulties were measured using the parent completed Strength and Difficulty Questionnaire (SDQ) [24]. Since most of these variables are time varying (and collected from MCS at ages nine months until age eleven years) aggregated summary variables were derived based on average values. These variables include SDQ, mother’s SEC, lone parent status, guardians’ alcohol use, living area, poverty indicator, additional household member and mother’s alcohol use after their child was born. The exposure variables from MCS have been described in Table 1.

Table 1: Socio-demographic characteristics of the MCS population (following imputation) and whole population sample with descriptive statistics

MCS	Whole Population
Child Sex	n	%		n	%
Female	521	48.60	Female	28,770	48.57
Male	551	51.40	Male	30,461	51.43
Deprivation
Mother Socio economic classification (SEC)			Overall deprivation
Always managerial or intermediate	377	35.17	Low (WIMD quintile >=3)	29,102	49.13
Always semi-employed, self-employed, semi-routine or routine	280	26.12	High (WIMD quintile <3)	24,701	41.70
Unknown	415	38.71	Borderline (ever belong to high group but not always)	5,428	9.16
Poverty indicator			Employment deprivation
Above poverty level	539	50.28	Low (WIMD quintile >= 3)	29,394	49.63
Below poverty level	270	25.19	High (WIMD quintile < 3)	24,774	41.83
Ever been below poverty level	263	24.53	Borderline (ever belong to high group but not always)	5,063	8.55
Household alcohol use
Mother’s alcohol use during pregnancy			Mother’s alcohol-related health condition during pregnancy
Never	752	70.15	No	55,251	93.28
Low (less than once a month or 1–2 times a month)	218	20.34	Yes	3,980	6.72
High (more than 1–2 times a month)	102	9.51
Mother’s alcohol use after child was born
Never	82	7.65
Low	500	46.64
High	490	45.71
Guardian alcohol use			Household member identified with alcohol-related hospital admission
Low	247	23.04	No	57,799	97.58
Moderate	524	48.88	Yes	1,432	2.42
High	233	21.74
Variable	68	6.34
Living area
Rural	238	22.20		14,760	24.92
Urban	779	72.67		41,907	70.75
Ever been urban	55	5.13		2,564	4.33
Maternal age at child’s birth
Less than 20 years	102	9.51		7,111	12.01
20 to 24 years	202	18.84		9,266	15.64
25 to 29 years	305	28.45		17,389	29.36
30 to 34 years	324	30.22		17,005	28.71
35 years and over	139	12.97		8,460	14.28
Gestational age
Not term	52	4.85		1,317	2.22
Term	1,020	95.15		57,914	97.78
Household composition
Siblings at home			Living with single adult
No sibling	129	12.03	No	33,662	56.83
One sibling always or at some point	493	45.99	Yes	8,425	14.22
More than one sibling ever	450	41.98	Ever been	17,144	28.94
Lone parent
No	754	70.34
Yes	130	12.13
Ever been	188	17.54
Additional household member
No	792	73.88
Yes	118	11.01
Ever had	162	15.11
Mother’s health
Longstanding illness			Mother’s any comorbidity
No	589	54.94	No	46,170	77.95
Yes	170	15.86	Yes	13,061	22.05
Varies	313	29.20	Mother’s psychosis disorder
			No	58,924	99.48
			Yes	307	0.52
			Mother’s common mental health condition
			No	28,603	48.29
			Yes	30,628	51.71

The health records of the children were also considered as the exposures for risk of early alcohol use. EHRs of the MCS children obtained from hospital admission record and primary care events within the Patient Episode Database for Wales (PEDW) and the WLGP dataset. A broad list of explanatory health codes was constructed using the three-digit ICD-10 codes and Read Code Version 2 recorded in PEDW and WLGP from birth until age ten (one year before the alcohol data were collected). Wales Electronic Cohort for Children (WECC) [25] containing further details on child health in Wales, were used to obtain age and maternal age at birth.

Outcome

Alcohol data for MCS children were obtained from a self-report questionnaire at age eleven (Supplementary Table 1). Based on the responses to the questionnaire the children were classified into two groups: those who had consumed alcohol (case) and those who had not (non-case). Those who did not answer or provided contradictory responses were removed from analyses (Supplementary Figure 1).

Statistical analysis

In the cohort exposure dataset, the participants with more than 10 missing variables (out of 13) were removed from analyses to ensure the accuracy of the data. An explanatory variable with less than 10% missing data had been imputed using a predictive mean matching (PMM) imputation method [26, 27].

To identify the health codes that were associated with early alcohol use from the large volume of linked EHRs spanning 10 years, a chi-square (X2) feature selection method was applied [28]. A critical threshold value X2 ≥ 2.706 (one degree of freedom, p ≤ 0.1) was applied and health codes with a X2 above this threshold were retained in subsequent analyses. A multivariate stepwise logistic regression with bidirectional (forward and backward) search was then performed for the exposure variables to obtain the best-fit model [29]. In stepwise model the variables with least significance were removed at each iteration step and the final model was selected based on the minimum Akaike Information Criterion (AIC) value. From the final model, only the statistically significant (p ≤ 0.05) variables were selected as significant predictors associated with the risk of early alcohol use leading to a further reduction in variable space. This is justified due to the following reasons.

The variable selection process facilitates the choice of best model by incorporating the interdependence between the explanatory variables.

The approach only considers the statistically significant variables for the stage two analysis which reduces the variable space and optimises the time to recreate analogous variables.

Stage two – whole population

All children born between 1 January 1998 and 31 December 2002 and were resident in Wales during the first fourteen years of their life were included in the whole population dataset. The study population was selected from the Welsh Demographic Service Dataset (WDSD), which is an administrative dataset of individuals living in Wales registered with a GP. The participants without continuous record in the WLGP from age six months to fourteen years were excluded to ensure a complete follow-up period.

Analogous risk factors to those identified in the MCS analysis were created using the WDSD, WLGP and PEDW data. The study used an encrypted household identifier known as residential anonymised linking field (RALF) which enabled the participants to be linked with other household members and related records [30]. Each RALF is associated with the smallest geographical representation known as lower super output area (LSOA) which again is associated with a Welsh Index of Multiple Deprivation (WIMD) rank aggregated into a quintile or decile scale. Overall and employment WIMD scores were used as the measure of deprivation from routine data in the study. The main explanatory variables derived from routine data for the whole population analysis include child’s sex, employment deprivation and overall deprivation, living with single adult, mother’s alcohol-related condition during pregnancy, living with household member with alcohol-related condition, living area, maternal age, gestational age, and child mental and physical health. To be consistent with the MCS data, primary exposure data were collected for children up to age seven years. For time varying variables, the study used the same time points as MCS (birth to nine months, nine months to three years, three to five years, and five to seven years) and derived aggregated summary variables for the risk factors. Detailed descriptions of the variables are available in Supplementary Table 2.

Alcohol-related health events across the whole population cohort were obtained from ICD-10 codes in PEDW (Supplementary Table 3) and Read codes in WLGP (Supplementary Table 4) between the age seven and fourteen years [31].

As the case (alcohol-related EHRs) to non-case (no alcohol-related EHRs) ratio was 1:99 in the whole population cohort and unbalanced, to improve the efficiency and the sensitivity of model performance case-control selection was undertaken by randomly selecting 20 non-cases for each sex matched case [32]. The dataset was randomly split into a training (70%) and test set (30%). Logistic regression was used to obtain the best-fit model on the training data. Model prediction on the test data provided a predictive probability of the expected outcome associated with each individual. Model prediction was quantified by performance accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.

MCS and routine EHRs were anonymously linked and accessed within the Secure Anonymised Information Linkage (SAIL) Databank. Linkage was completed using an encrypted person-based identifier known as the anonymised linkage field (ALF), generated by the Digital Health and Care Wales (DHCW) [33, 34]. Data preparation (extraction, cleaning, and linkage) was performed in Structured Query Language (SQL) on an IBM DB2 platform, with subsequent analyses performed in R v3.3.2 [35].

Results

Stage one – MCS

Among the consented singleton children 1,838 were assigned an ALF, with 82% of the children having a GP registration record in SAIL before age eleven years (Supplementary Figure 1). Individual and household characteristics (following imputation) are described in Table 1. 7.6% of the MCS children were considered as ‘case’ based on their response. Health codes (256 ICD-10 and Read codes) were obtained after merging the first ten years of EHRs from PEDW and WLGP. Feature selection method reduced this to 13 health features (Table 2).

Table 2: Health codes identified as risk factors for early alcohol use by chi-square feature selection method in the MCS cohort and the percent of sample with these codes present in whole population (WP) following selection

Health code	Description of the code	Type of code	chi-square	MCS (%)	WP (%)
Read code H05%	Upper respiratory infections	Diagnosis	.60	62.50	59.95
Read code K2%	Male genital organ diseases	Diagnosis	7.77	12.41	8.46
Read code 919%	Child health surveillance related administrative code	Administrative	6.07	25.56	30.70
Read code 64N%	Child physical health examination	Administrative	4.63	17.35	15.56
Read code 656%	Tetanus vaccination	Administrative	4.11	28.26	34.21
ICD-10 code Z%	Factors influencing health status and contact with health services	Diagnosis	3.90	27.99	20.68
Read code 654%	Diphtheria vaccination	Administrative	3.69	27.71	-
Read code 655%	Pertussis vaccination	Administrative	3.35	29.94	-
Read code F%	Nervous system and/or sense organ diseases	Diagnosis	3.04	70.24	-
Read code F4%	Disorders of eye and adnexa	Diagnosis	3.00	46.27	-
Read code K27%	Disorders of penis	Diagnosis	2.99	9.42	–
Read code etc.%	Trimethoprim, an antibiotic used mainly in the treatment of bladder infections	Medication	2.93	16.70	–
Read code 4%	Laboratory test and procedures (e.g. urine culture, blood test)	Administrative	2.89	60.73	–

codes were not selected by the logistic regression models, hence were not selected for WP analysis

After merging health and socio-demographic variables, 31 main explanatory variables (13 health codes and 18 socio-demographic variables) were available for the two-way logistic model. The final 19 features with significant p values were considered to be significantly associated with the risk profile of early alcohol use (Table 3).

Table 3: The explanatory variables associated with higher and lower risk of early alcohol use for the MCS children (Stage one analysis) with the adjusted Odds Ratio (OR) and 95% confidence interval (CI)

Feature	Adjusted OR (95%CI)
Child’s sex
Female	1
Male	3.06 (2.35 to 3.99)***
Mother’s Socio-economic classification (SEC)
Always Managerial or intermediate	1
Always semi-employed, self-employed, semi-routine or routine	1.30 (0.93 to 1.81)
Unknown	1.94 (1.37 to 2.74)***
Lone parent
Never lone parent	1
Lone parent	1.68 (1.07 to 2.65)*
Ever been	1.77 (1.27 to 2.49)**
Mother alcohol use during pregnancy
Never	1
Low (less than once a month, 1–2 times a month)	2.48 (1.83 to 3.38)***
High	5.38 (3.58 to 8.15)***
Mother alcohol use after child was born
Never	1
Low	1.15 (0.70–1.92)
High	0.70 (0.04 to 1.24)
Guardian alcohol use
Low	1
Moderate	1.73 (1.22 to 2.25)**
High	1.07 (0.70 to 1.64)
Variable	0.91 (0.48 to 1.70)
Living area
Rural	1
Urban	1.61 (1.17 to 2.23)**
Ever been urban	4.54 (2.69 to 7.75)***
Poverty indicator
Above poverty level	1
Below poverty level	0.93 (0.60 to 1.45)
Ever been below poverty level	1.33 (0.95 to 1.86)
Maternal age at child’s birth
Less than 20 years	1
20 to 24 years	1.57 (0.97 to 2.58)
25 to 29 years	3.28 (2.03 to 5.36)***
30 to 34 years	2.68 (1.64 to 4.43)***
35 years or over	0.65 (0.35 to 1.21)
Gestational age
Not term	1
Term	9.42 (4.22 to 23.03)***
Additional household member
No	1
Yes	0.69 (0.45 to 1.06)
Ever had	0.57 (0.39 to 0.81)**
Hyperactivity
Always normal	1
Any mention of higher level of hyperactivity	1.84 (1.37 to 2.47)***
Conduct disorder
Always normal	1
Any mention of higher level of CP	2.10 (1.57 to 2.82)***
Emotional difficulty
Always normal	1
Any mention of higher level of ED	0.68 (0.48–0.97)*
Total Difficulty Score
Always normal	1
Any mention of higher level of TDS	0.45 (0.31 to 0.66)***
Mother longstanding illness
No	1
Yes	1.53 (1.09 to 2.16)*
Varies	1.25 (0.96 to 1.65)
Other acute upper respiratory infections (Read code H05%)
No	1
Yes	0.43 (0.34–0.55)***
Male genital organ diseases (Read code K2%)
No
Yes	2.77 (1.58–4.94)***
Child surveillance administration (Read code 919%)
No
Yes	1.38 (1.06 to 1.81)*
Child exam (Read code 64N%)
No
Yes	0.51 (0.35 to 0.75)**
Tetanus vaccination (Read code 656%)
No
Yes	0.60 (0.45 to 0.79)***
General examination (ICD10 code Z%)
No
Yes	0.73 (0.55 to 0.99)*
Disorders of penis (Read code K27%)
No
Yes	0.63 (0.33 to 1.19)

Note: * p < 0.05, ** p < 0.01, *** p < 0.001.

In Wales, 207,114 children were born in between 1 January 1998 and 31 December 2002, and their records were obtained from WDSD. After applying exclusion criteria there were 59,231 children as the study population (Supplementary Figure 2). Of the study population, 591 (0.99%) children had at least one alcohol-related event between seven and 14 years of age (Supplementary Figure 3) who were the cases from the whole population subset. After applying case control selection, the dataset had 591cases and 11,820 non-cases, which were further split into training and test set. There were 8,688 (417 cases and 8,271 non-cases) children in the training dataset. The variables identified as significantly associated with early alcohol use using MCS data were mapped into the whole population cohort (Supplementary Table 2). Table 1 presents descriptive statistics for this population. Mothers of 6.72% of the children had an alcohol-related event reported in PEDW or WLGP while pregnant. 2.42% children lived with a household member who had alcohol-related inpatient hospital admission. The adjusted odds ratio of the features with 95% confidence interval are presented in Table 4 (also see Supplementary Figure 4).

Table 4: The explanatory variables associated with higher and lower risk of early alcohol-related health outcomes for the whole population (Stage two analysis) with the adjusted Odds Ratio (OR) and 95% confidence interval (CI)

Feature	Adjusted OR (95% CI)
Child’s Sex
Female	1
Male	1.09 (1.02 to 1.17)**
Overall deprivation:
Low	1
High	1.11 (0.98 to 1.25)
Borderline	1.66 (1.41 to 1.95)***
Employment deprivation:
Low	1
High	0.84 (0.75 to 0.95)**
Borderline	0.82 (0.69 to 0.97)*
Living with single adult:
No	1
Yes	1.45 (1.32 to 1.59)***
Ever been	1.17 (1.08 to 1.26)***
Mother’s alcohol-related condition during pregnancy
No	1
Yes	0.88 (0.77 to 1.00)*
Household member with alcohol-related condition
No	1
Yes	2.16 (1.80 to 2.60)***
Living area
Rural	1
Urban	0.99 (0.92 to 1.08)
Ever in urban	2.42 (2.08 to 2.81)***
Maternal age at birth
Less than 20 years	1
20 to 24 years	0.88 (0.79 to 0.99)*
25 to 29 years	0.79 (0.71 to 0.87)***
30 to 34 years	0.68 (0.61 to 0.76)***
35 years or over	0.53 (0.46 to 0.60)***
Gestational age
Not-term	1
Term	1.11 (0.89 to 1.40)
Child – Attention deficit hyperactive disorder (ADHD)
No	1
Yes	3.57 (2.52 to 5.15)***
Child - Conduct disorder
No	1
Yes	3.26 (2.14 to 5.07)***
Child – Depression/Anxiety
No	1
Yes	0.75 (0.34 to 1.69)
Mother’s any comorbidity
No	1
Yes	1.25 (1.16 to 1.34)***
Mother’s common mental health condition
No	1
Yes	1.31 (1.23 to 1.40)***
Mother’s psychosis disorder
No	1
Yes	3.12 (2.04 to 4.90)***
Other acute upper respiratory infections (Read code H05%)
No	1
Yes	0.97 (0.91 to 1.04)
Male genital organ diseases (Read code K27%)
No	1
Yes	0.90 (0.79 to 1.02)
Child surveillance administration (Read code 919%)
No	1
Yes	0.80 (0.75 to 0.86)***
Tetanus vaccination (Read code 656%)
No	1
Yes	0.47 (0.44 to 0.51)***
Child exam (Read code 64N%)
No	1
Yes	0.59 (0.53 to 0.65)***
General examination (ICD10 code Z%)
No	1
Yes	0.84 (0.78 to 0.92)***

Note: * p < 0.05, ** p < 0.01, *** p < 0.001.

The model was run on the test dataset. The accuracy of the model was 61.32% with a sensitivity of 58.05% and specificity of 68.48% (additional details are provided in Supplementary Tables 5, 6). Out of 174 cases, the model was able to predict 101 (58%) children who had an alcohol-related health event recorded in the healthcare system between ages seven and fourteen.

Discussion

This study has developed a two-stage data-driven framework that can create a profile of the characteristics of children who end up with an alcohol problem in adolescence. The study undertook data linkage between a longitudinal survey data (MCS) and routine EHRs in stage one to select the significant risk factors associated with early alcohol use. Stage two built the analogous risk factors using only the linked routine data and based this, a prediction model was developed. Hybridisation of these two powerful data sources (routine and survey) enabled us to create a data-driven risk profile. The risk factors were significantly associated across both MCS and whole population analyses, but effect estimates varied. Children whose health needs are supported are at lower risk of early alcohol use, evidenced by protective effect of receiving vaccinations, attending routine health examinations with their GP, and contact with health services recorded in primary and secondary care were consistent across MCS and whole population analyses. Similarly, children with health codes relating to acute upper respiratory infections may have more protective guardians willing to consult medical professionals for mild conditions. Together, this suggests that the avoidance of regular healthcare contact is an indicator that increases the risk of early alcohol use. However, the trends relating to the two codes, the child surveillance administration code and the chapter heading linked to male genitals, differed between the whole population and the MCS analysis. The code linked to male genitals showed an association with higher risk of alcohol use in MCS but was statistically inconclusive for the whole population analysis. The child surveillance administration code was associated with higher risk for the MCS cohort in contrast to the whole population which can be attributed to the differential support received by two cohorts which was not captured by the data and hence this requires further investigation. Also, the proportion of cases obtained from MCS data (stage one) were higher than those obtained from the whole population data (stage two). This can be attributed to the fact that cases from stage one were based on the self-reported alcohol consumption data whereas the stage two routine data highlighted the most severe cases caused by alcohol among the adolescents and recorded on the healthcare system.

The overall risk profile obtained from MCS and whole population analyses were broadly consistent with each other and the research literature generally both in the UK and internationally. Similar risk factors include being male [13], ever living in an urban environment where there is a greater density of alcohol outlets [36], ever living in conditions of social deprivation, living in a household with higher level of alcohol use by household members [9]. Studies from USA highlighted that early onset of alcohol use was significantly associated with parental drinking pattern and living in a lone parent household [11], child’s attention deficit hyperactivity disorder (ADHD) and conduct disorder [6, 7]. The stage one MCS analysis in this study revealed that emotional difficulty and a higher level of behavioural difficulty (as assessed by parents) were associated with a reduced risk of alcohol use. However, diagnosis of clinically relevant behavioural/emotional problems was protective in the population model. Poor maternal mental health was linked with adverse outcomes, consistent with family-level risk factors that promote children’s alcohol use [12, 17]. A difference was observed in regards to the effect of maternal age at birth on the risk of a child’s early alcohol use. The protective effect of higher maternal age was observed for the whole population but the finding on MCS data differed and requires further investigation. Further, employment deprivation in the whole population analysis was associated with lower risk of a child’s early alcohol use after adjusting for overall deprivation. This finding is similar to the existing literature [15, 37], which found that early alcohol use is more common in higher income families. This suggests that reliance on employment indicators is not sufficient to understand the socio-economic factors influencing a child’s early alcohol use, the overall deprivation (also measured by education, health, access to the service, physical environment of living) plays an important role as well.

The result of this study needs to be interpreted in conjunction with a number of limitations. Firstly, mapping the MCS survey to the routine data was challenging, not all MCS variables were available in the routine data. In some instances, multiple variables had to be merged to derive summary variables. This may result in a degree of uncertainty about the information captured in the summary variables. Secondly, it was necessary to aggregate some time-varying variables into a single point estimate and, as such, the analyses are unable to capture how the recency of some events might influence results. Thirdly, due to unavailability of continuous GP records of some participants between six months and fourteen years (if the participants changed their GP and the their registered GP was not contributing to SAIL), they were removed from the whole population analysis. Similarly, the follow-up of children was not possible where they who moved out of the study area (Wales, UK), or died under age fourteen, because of which their exposure (sociodemographic and health related data) and outcome (alcohol data) data were not available. This resulted in a large reduction of the number of children in the study population. However, this did not contribute to selection bias as this happened randomly and the losses had no direct relationship with alcohol-related outcome. Fourthly, the EHRs did not include Emergency Department (ED) attendance data (but does include admissions into hospital via the ED) as there are no uniformly applicable codes for alcohol-related attendances in ED, and even when available, these are sparsely populated [38]. Lastly, in this study the model performance, measured by sensitivity and specificity, was moderate. However, even if we had a sensitivity and specificity of 90% the maximum positive predictive value, we can get is 31%, given the low prevalence of alcohol-related medical contact, as the prevalence influences the positive and negative predictive value of a model performance [39]. Machine learning approaches generally aim to achieve the best predictive models from the available data. The low positive predictive value, obtained here, suggests that the variables needed to improve model performance are not available in the data (e.g., genetic information, peer alcohol-related data).

Routine EHRs and administrative data are available to healthcare professionals and are used by policy makers and commissioners to determine how resources are best utilised to manage preventive interventions. However, the bulk of research considering early alcohol use and related outcomes has relied on self-report surveys. It has been shown that linking survey and routine data can offer new insights [40]. The results presented here are novel in that our approach generalised results from an established survey to a whole population analysis using predictive analytic techniques. This provides in-depth knowledge about the profile of the children susceptible to early alcohol use and can feasibly be used to inform population health strategies designed to reduce the prevalence of early alcohol use in children and related health outcomes.

Conclusions

The hybridisation of data of different nature, as carried out in this study, is a novel approach that combines the complementary advantages of EHRs with more personal insights from questionnaire-based cohort data. This provides a robust resource on which findings can be based and generalised to the wider population. The identified risk factors such as living with a single parent, alcohol problem in the household, social deprivation and children receiving poor support from the healthcare system indicate that involvement and support for the family is important in breaking cycles and improving children’s outcomes.