Setting
The study was conducted in Hubei province of central China. Hubei has a land size of 185,900 km2 and more than 59 million populations. With a gross domestic product at $8,915 per capita in 2017, its economic status is ranked in the middle range of all provinces in China. According to the World Bank [21], Hubei is considered as a middle-high income region.
We chose primary care facilities in Hubei as the study setting. In 2017, primary care facilities in Hubei served 205.08 million patient visits, accounting for 60.24% of total outpatient visits in the province [22]. About 44.28% of the patient visits to primary care were given an antibiotic prescription [23]. It was estimated that 60% of antibiotic prescriptions in primary care in China are inappropriate [24].
Sampling and participants
A stratified cluster random sampling strategy was adopted to select study participants. Hubei has 347 urban community health centres and 1137 rural township health centres. In proportion to the urban and rural numbers, 19 community healthcare centres from 3 urban districts and 48 township health centres from 6 rural districts were randomly selected, respectively. Details of the sampling methods have been published elsewhere [23].
Primary care physicians from the selected health centres who met the following criteria were invited to participate in the study: 1) having the authority to prescribe antibiotics independently; 2) having issued at least 100 prescriptions over the three-month study period, which contained at least one antibiotic prescription.
In total, 645 physicians met the inclusion criteria and 551 (85.58%) agreed to participate in the study. Of those who agreed, 458 (71.01%) returned a valid questionnaire.
Data collection
Eight field investigators were recruited and trained to conduct data collection. Each participating facility was visited by a pair of the trained investigators. The investigators had no servicing relationships with the facilities or their employees at the time. All eligible primary care physicians were approached and invited to participate in the study. Informed written consents were obtained prior to data collection.
Prescribing records issued by the 551 study participants over a three-month period (from 1 January to 31 March 2018) were extracted from the medical records system of the participating facilities, including the name, formulation, dosage, administration route, and price of the prescribed medicines, and information about the prescribers and facilities. This was followed by a questionnaire survey of prescribers (n=458) over the period from 23 April to 6 June 2018, tapping into their socioeconomic status and professional characteristics, and their knowledge and attitudes toward antibiotic prescribing. The respondents were asked to complete the questionnaire independently, which took roughly 15 minutes. A token gift ($1.65) was given to those who returned the questionnaire to the investigators. Missing items, if any, were re-filled by the investigators through an additional interview.
Measurements
Antibiotic prescribing patterns
Seven indicators were identified for measuring antibiotic prescribing patterns through a comprehensive literature review and expert consultations:
- average number of medicines per prescription;
- average number of antibiotics per prescription;
- percentage of prescriptions containing antibiotics;
- percentage of antibiotic prescriptions containing broad-spectrum antibiotics;
- percentage of antibiotic prescriptions containing parenteral administered antibiotics;
- percentage of antibiotic prescriptions containing restricted antibiotics imposed by the provincial government; and
- percentage of antibiotic prescriptions containing antibiotics included in the World Health Organisation (WHO) “Watch and Reserve List”.
The first three indicators were adapted from the prescribing indicators recommended by the WHO [25]. They measured the frequency and volume of antibiotics prescribed. Although we did not measure combined use of antibiotics directly because it was rare in primary care, the tendency of combined use of antibiotics was likely to be captured through the connection between the volume (indicator 2) and frequency (indicator 3) indicators [26]. Previous studies showed that higher number of medicines prescribed in general is also a significant predictor of higher antibiotic prescriptions [27, 28].
We added two additional indicators (indicator 4 and 5) in order to better assess irrational prescribing of antibiotics. Empirical evidence shows that broad-spectrum antibiotics is frequently used and is perhaps the most common form of antibiotic abuse in primary care [29, 30]. In addition, the high prevalence of parenteral administration of antibiotics has attracted increasing safety concerns in China. Studies showed that 36% to 60% of antibiotics were administered through parenteral injections in primary care settings in China [29-31].
Over the past two decades, China introduced some restrictive measures to reduce irrational antibiotic prescribing. These included a list of restricted antibiotics for primary care imposed by the regional governments [32]. Restricted access to certain antibiotics addresses the concerns of AMR [33-35]. The WHO also published an “Access, Watch and Reserve” (AWaRe) classification system [20]. All antibiotics were exclusively classified into three categories. The “Watch” list includes antibiotics that have higher resistance potential, while the “Reserve” list includes antibiotics that should be reserved for treatment of infections due to multi-drug-resistant organisms. We examined prescriptions of restricted antibiotics against the above two classification systems. Although the two share similar principles, they are not always consistent. In Hubei, antibiotics were classified into non-restricted, restricted, and special-restricted.
Factors associated with antibiotic prescribing patterns
Antibiotic prescribing behaviors can be influenced by the knowledge and attitudes of prescribers and their personal circumstances according to two systematic reviews [18, 19]. Prescribers with higher qualifications and better knowledge of antibiotics are less likely to prescribe antibiotics. However, their attitudes toward antibiotic prescribing are also influenced by patient expectations and collegial pressures.
This study used a 37-item questionnaire to measure the knowledge, attitudes and personal circumstances of prescribers. The questionnaire was developed based on some existing instruments [36, 37] with further consideration of the findings of the two systematic reviews [18, 19]. The reliability and validity of the questionnaire has been tested and confirmed in previous studies [36, 37].
The questionnaire respondents were asked to indicate whether they agreed to prescribe antibiotics for 11 common conditions such as upper respiratory tract infections and diarrhea [37]. A correct decision in line with the current clinical guidelines was given a score of 1, otherwise 0. The scores were summed up for each respondent.
The attitudes of the questionnaire respondents toward antibiotic prescribing were measured by 17 items, each being rated on a five-point Likert scale (0=strongly agree, 1=agree, 2=neutral, 3=disagree, 4=strongly disagree). The scores were summed up to measure the tendency of complacency to satisfied patients (0-8 measured by 4 items), fearful of adverse events (0-12 measured by 6 items), ignorance of AMR (0-16 measured by 8 items), indifference to changes (0-4 measured by 2 items), and responsibility avoidance by blaming others (0-28 measured by 7 items), respectively [36]. All item coding and summed scores were aligned into a unified direction, with a higher score indicating more positive attitudes toward reduction of irrational antibiotic prescribing.
The personal circumstances measured in this study included the demographic characteristics (age and gender) of the respondents, and their socioeconomic status (educational qualifications, and household income) and professional experiences (workplace, years of practice, sub-specialty, professional title, and continuing education on antibiotic prescribing). These factors have been proved to be significant determinants of antibiotic prescribing behaviors [18, 19].
Data analysis
Two datasets were prepared for data analyses. The first dataset contained 501,072 prescriptions made by 551 primary care physicians. For each physician, the seven prescription indicators were calculated. Antibiotics were defined based on the Anatomical Therapeutic Chemical (ATC) classification system and included only systemic use of antibiotics (ATC code J01) [38]. They were further divided into broad- and narrow-spectrum in line with the classification criteria used in the US national survey on antibiotic use [33]. Restricted antibiotics were defined based on the Hubei government’s antibiotic regulation policy and the WHO AWaRe list.
To determine the antibiotic prescribing patterns, latent profile analyses (LPA) were performed using the seven prescribing indicators at the physician level. LPA belong to finite mixture modelling which can identify and describe “hidden groups” within a population. Because the 551 physicians were clustered in 67 primary care facilities, a two-level LPA model was established. Differences at the facility level were treated as random effect. Maximum likelihood parameter estimates with standard errors (MLR) were applied. The model identification was checked using 1000 initial stage starts and 1000 final stage starts [39].
We tested different models that categorised antibiotic prescribing behaviors into one, two, three, four, or five groups. The best fit model was identified using the following model index: Bayesian Information Criterion (BIC), Sample-size Adjusted BIC (SABIC), Vuong-Lo-Mendell-Rubin Adjusted Likelihood Ratio Test (VLMR-LRT), Correct Model Probability (cmP) and Entropy. A lower value of BIC and SABIC indicates better fitness of data into the estimated model. VLMR-LRT compares the model fit between two neighboring models (for example, two groups vs three groups). A non-significant p value (>0.05) indicates a lack of statistical significance between the two compared models. cmP provides an overall assessment of all estimated models and a larger cmP value indicates a better model fit. Entropy assesses the accuracy of classification, with a higher value indicating better classification [39]. To avoid over-stratification, the smallest group should have a minimum of 5% of participants.
The second dataset contained the 458 returned questionnaires, as well as the classification of the antibiotic prescribing patterns of the 458 respondents. A three-group model was identified in the LPA. Each questionnaire respondent was assigned into one of the antibiotic prescribing pattern groups with the highest probability.
Differences in knowledge and attitudes scores and personal circumstances among the respondents in different antibiotic prescribing pattern groups were examined using Kruskal-Wallis rank tests, one-way analysis of variance (ANOVA), or chi-square tests. Post-hoc pairwise comparisons were performed using Dunn and Bonferroni tests. Multi-nominal logistic regression models were established to determine significant factors predicting the three groups of antibiotic prescribing patterns after adjustments for variations in other factors. In the regression analyses, knowledge and attitudes scores were transformed into dichotomous variables with mean scores serving as a cut-off point. An enter approach was adopted in the modelling.
The statistical analyses were performed using STATA (version 12.0) and Mplus (version 6.0). A p value <0.05 was considered statistically significant.