Study population and design
The HEalthy LIfe in an Urban Setting (HELIUS) study is a multiethnic cohort study conducted in Amsterdam, which focuses on cardiovascular disease (e.g. diabetes), mental health (e.g. depressive disorders), and infectious diseases [8, 9]. In brief, baseline data collection took place in 2011–2015 and included people aged 18 to 70 years of Dutch, Surinamese, Ghanaian, Moroccan, and Turkish origin. A random sample of participants, stratified by ethnic origin, was taken from the municipality register of Amsterdam. Participants filled in an extensive self-administered questionnaire (variables included in the questionnaire are described elsewhere) [9] and underwent a physical examination during which biological samples were obtained [9]. No information was provided regarding appropriate antibiotic use. Between 2011 and 2015, 24,789 persons were included. Data collection procedures have been previously described in detail [9]. Both questionnaire data and physical examination data were available for 22,165 participants. The HELIUS study was conducted in accordance with the Declaration of Helsinki and was approved by the AMC Ethical Review Board. All participants provided written informed consent.
Ethnicity was defined according to the country of birth of the participant as well as that of their parents [12]. Specifically, a participant is considered to be of non-Dutch ethnic origin if they fulfill either of the following criteria (i): they were born abroad and had at least one parent born abroad (first generation) or (ii) they were born in Netherlands but both their parents were born abroad (second generation). Dutch participants were born in the Netherlands and had both parents who were born in the Netherlands. After HELIUS data collection, the Surinamese group were further classified according to self-reported ethnic origin (obtained by questionnaire), into ‘African Surinamese’, ‘South-Asian Surinamese’, ‘Javanese Surinamese’ and ‘other/unknown Surinamese’.
Data linkage
Permission to link participants’ individual data to outside health registries was asked in the written informed consent form [8]. Of the 22,165 HELIUS participants, 19,895 agreed. HELIUS data of these individuals were linked to reimbursement data from the Achmea insurance company (Achmea Health Database, AHD) from 2010 until 2015. The AHD, obtained from the largest health insurance company in Amsterdam, contains all healthcare expenditures of every insured participant, including medications. A trusted third party linked data on reimbursed antibiotic prescriptions using an encrypted social security number and returned data without any identifying information. Procedures were in accordance with the General Data Protection Regulation [13].
Inclusion and exclusion criteria for present study
Of the 22,165 participants, we excluded those of Javanese Surinamese or other/unknown Surinamese origin and those with another/unknown ethnic origin because of small participant numbers. For analyses on antibiotic use, we included those who gave permission for data linkage and could be linked to the AHD. To reduce bias for individuals with short-term insurance, we excluded those who were insured with Achmea for less than 365 days in the year preceding their HELIUS study visit.
Outcome variables
The primary outcomes were level of antibiotic knowledge and antibiotic use during the year prior to the HELIUS visit. Level of antibiotic knowledge was based on five questions, used in other studies [4, 6, 14], which asked the perceived necessity (yes/no) for antibiotic treatment during influenza-like illness, pneumonia, fever, sore throat and bronchitis. Using these questions, we created an overall knowledge score of antibiotic use by summing the total number of correct responses, resulting in a score ranging from 0 to 5. A two-parameter logistic regression model was fitted to the five binary items based on the assumptions of item response theory (see Additional file 1). From this model, “higher” and “lower” levels of antibiotic knowledge were defined by a knowledge score of ≥4 and < 4, respectively.
Antibiotic use was obtained from linked AHD data and was based on the total number of reimbursed antibiotics (classified by ATC code J01; anti-infectives for systemic use) dispensed by community pharmacies from 2010 until 2015. We evaluated antibiotic use (yes/no) in the year prior to the HELIUS study visit, as well as the number of antibiotic prescriptions over the past year and during the entire insured period.
Other variables
Independent variables were obtained from the HELIUS study questionnaire (migration generation; sex; age; level of education; marital status; self-reported medical conditions; smoking; alcohol consumption; difficulty with the Dutch language and perceived health) and physical examination (body mass index (BMI, kg/m2)). Variables on antibiotic-related behavior were: not having finished antibiotic treatment; having saved antibiotics for later; and ever having asked the general practitioner (GP) for antibiotics. Definitions and grouping of variables are extensively described elsewhere [8].
Statistical analyses
Sociodemographics, health status, antibiotic knowledge level and questions on antibiotic use were presented by ethnicity. To assess selection bias resulting from AHD data linkage, the same variables were compared between participants who were successfully versus unsuccessfully linked. Comparisons between ethnic groups were made using a Pearson’s χ2 or Fisher exact test for categorical data and Kruskal-Wallis rank test for continuous variables.
Analysis on level of antibiotic knowledge included all HELIUS participants with available data. Odds ratios (OR) comparing levels of antibiotic knowledge across determinants and their 95% confidence intervals (CI) were estimated using logistic regression. All variables with an associated p-value < 0.2 in univariable analyses were included in a full multivariable model and variables with a p-value above this level were removed in backwards-stepwise fashion. Given that the research aim was to determine differences between ethnicity, ethnic groups were forced in all models. This multivariable approach was chosen to not only assess other variables associated with antibiotic knowledge, but also to understand the extent of confounding bias when assessing the relationship between ethnicity and outcome variables.
Analysis on antibiotic use in the year prior to HELIUS study visit included all HELIUS participants who were linked to the AHD and were insured for at least 365 days with Achmea in the year prior to their HELIUS study visit. Determinants for having received ≥1 antibiotic prescription were assessed using logistic regression. The same multivariable approach as above was used for this outcome. We also compared antibiotic use during the entire period insured at Achmea versus the year prior to HELIUS study visit to assess differences when considering longer time periods.
Determinants for the total number of antibiotic prescriptions were then evaluated. As this outcome contained a high proportion of zero values and was over-dispersed, we used a zero-inflated negative binomial regression model. This model contains two parts: one accounting for zero values in the count distribution (zero-inflated) and another accounting for the over-dispersed count distribution (negative binomial). Covariates for the zero-inflated part were determined a priori from the risk-factor analysis on ≥1 antibiotic prescription. Covariates for the negative binomial part were selected from covariates with a p-value < 0.2 in univariable analyses and variables above this p-value were removed in backwards-stepwise fashion. Incidence risk ratios (IRR) comparing the number of antibiotics prescribed over the past year across levels of determinants were estimated from this model.
Multicollinearity was verified using variance inflation factors, while any variable with an inflation factor of ≥4 was considered multicollinear and excluded from the model. To understand whether the association between ethnicity and outcome was modified by demographic variables, interaction between ethnicity and other demographic variables was also assessed in all multivariable models.
The three variables involving antibiotic-related behavior were not initially considered in the final multivariable models. To assess whether ethnic differences in antibiotic use could be explained by patterns of antibiotic-related behavior, additional multivariable models including these variables were constructed for the endpoints (i) having received ≥1 antibiotic prescription and (ii) total number of antibiotic prescriptions.
Figure 1 provides an overview of all descriptive analysis and modeling used in the study. Significance was determined using a p-value < 0.05. All analyses were conducted with Stata 13.1 (StataCorp., College Station, Texas, USA).