Trajectories of seasonal influenza vaccine uptake among French people with diabetes: a nationwide retrospective cohort study, 2006–2015

We conducted a retrospective cohort study in the Permanent Sample of Beneficiaries (Echantillon Généraliste des Bénéficiaires, EGB). The EGB, set up in 2005, is a permanent, representative, and open national random sample of 1/97th of persons currently affiliated with one of the three major national health insurance funds in France [13]. At the time of extraction (August 2017), it included 804,089 beneficiaries. For this study, we extracted data for salaried workers (including those who are retired) only (about 86% of the French population [13], covered by the French National Health Insurance Fund, NHIF) because people affiliated with the other insurance funds were only included in the EGB in 2011.

Data include age, gender, district of residence, reimbursement claims for consultations with private healthcare professionals, medical procedures (e.g., laboratory tests), drugs purchased in the community (classified by Anatomical Therapeutic Chemical (ATC) codes), and long-term illness (LTI) status, recorded by expert physicians according to the International Classification of Diseases (ICD-10). LTI status is granted to beneficiaries with long-term and costly diseases and exempts them, regardless of income level, from copayments for related medical care [14]. Since 2006, data regarding diagnoses associated with admissions to French public or private hospitals are also available.

The NHIF granted us authorization to access the EGB, in accordance with French law.

Using our adaptation of an NHIF algorithm [8] based on LTI status, hospitalization diagnoses, and reimbursement claims for antidiabetic drugs or hemoglobin A1c (HbA1c) assays (Additional file 1: Table S1), we selected all individuals residing in mainland France identified as treated for diabetes in 2006. We followed them up over 10 seasonal vaccination campaigns. Those who died or withdrew from the NHIF during the follow-up period were censored at the start of the year of the event.

For each individual and each SIV season n/n + 1 (temporal statistical units in our analyses), we constructed a binary variable “SIV uptake” (yes, no), based on SIV deliveries (Additional file 2: Table S2) recorded between September 1 of year n and March 31 of year n + 1. Each SIV delivered in a community pharmacy is recorded in the NHIF database.

To describe the diabetes type and treatments, we constructed for each year of follow-up a 5-category variable based on LTI status and reimbursement claims for antidiabetic drugs recorded during the 6 months before the start of season n/n + 1. Using these annual variables, we built a variable of “diabetes treatment intensification” (yes/no) during follow-up. “Intensification” was defined by at least one of the following modifications: from no antidiabetic drug to at least one antidiabetic drug; from only one to at least two noninsulin antidiabetic drugs or insulin; from at least two noninsulin antidiabetic drugs to insulin.

To measure comorbidities, we calculated for each year of follow-up an individual chronic condition score (ICC) based on drug deliveries according to a previously published methodology [14]. Then we built a 3-category variable describing the course of the ICC score from the first to the last year of follow-up (decreasing, increasing, or stable) and included it in our analysis, as a time-stable variable [15].

Each cohort member’s number of hospital stays for each year for diabetes, diabetes complications, influenza, and influenza complications was extracted, as were the numbers of visits – separately – with general practitioners (GPs), endocrinologists, and cardiologists. GPs are responsible for the management of most patients with type 2 diabetes [9] and for referral to specialists. We also extracted changes of GP during follow-up.

The NHIF sends free vaccination vouchers each season to individuals aged 65 years or older and to those patients with diabetes with an LTI status: we constructed a 3-category variable to describe receipt of this voucher (Additional file 3: Table S3). The voucher enables these patients to obtain the vaccine free of charge at the pharmacy, without a doctor’s prescription. They must then make an appointment with either a doctor or a nurse for its administration.

We ran group-based trajectory (GBT) modelling to identify subgroups of individuals with similar patterns of SIV dispensing over time during the 10-year follow-up period. GBT modelling is a data-driven semiparametric method designed to analyze the evolution of an outcome over time and to identify, within a population with unobserved heterogeneity, distinct clusters of individuals following similar trajectories of behaviors related to this outcome [16, 17]. It makes it possible to select the model with an optimal number of distinct trajectories that most appropriately represent the heterogeneity in the population [15]. To compare the models’ goodness of fit, we used the Bayesian information criterion (BIC) and individual posterior class-membership probabilities (i.e., the probability of belonging to a trajectory given the information collected). Starting with a one-trajectory solution, we added one trajectory at a time, testing each model fit and balancing it with our objective of identifying distinct and interpretable trajectories. The prevalence of each trajectory and the relevance of the solutions were also considered, as recommended by Nagin and Odgers [18]. To determine the order of polynomials for all trajectories, we started with third-degree polynomials and used the standard operating procedure, i.e. stepwise elimination of non-significant polynomial higher orders [19]. Early applications of GBT modelling have assumed that all attrition (including both loss to follow-up and mortality) is randomly distributed among all trajectories. A recent enhancement of the GBT approach enables the joint modelling of the outcome of interest and non-random missingness [20, 21]. Using this methodology, we were able to model attrition probabilities (mortality represented the vast majority of attrition in our study) jointly with the estimation of SIV trajectories.

The demographic, clinical, and healthcare utilization factors were added to the model as predictors of trajectory group membership. This joint estimation of trajectories and predictors of the probability of group membership allowed us to take into account the uncertainty in participants’ trajectory group membership [15]. We used Zhang’s correction to estimate adjusted risk ratios from the estimated ratios [22].

All statistical analyses were performed with SAS statistical software, version 9.4 (SAS Institute Inc., Cary, NC). GBT analyses were conducted with the TRAJ procedure [16].

Of the 17,259 subjects with diabetes included in the cohort in 2006, 46% were women; mean age at inclusion was 65.0 ± 13.7 years. About 10% were identified with type 1 diabetes at inclusion; only 70% had LTI status for diabetes then. Over the 10-year follow-up, 31% of the initial cohort died (Additional file 4: Table S4), for a death rate of 36‰ person-years, and 3% were lost to follow-up.

Based on the BIC values, the fit of the models improved as the number of trajectories modeled increased. From a seven-trajectory solution and after, the prevalence of some trajectories was very low and results were difficult to interpret. Accordingly, the solution that offered the best compromise between parsimony, fit, and interpretability was a six-trajectory solution. Classification quality was good for all six (mean posterior class-membership probability > 0.82).

In trajectory 1 (“continuously vaccinated”, 28% of the cohort), the SIV-uptake rate started at 92% at inclusion, then exceeded 97% throughout follow-up, with numbers of SIV injections over the 10-season follow-up ranging between 8 and 10. In trajectory 2 (“progressively less vaccinated”, 18%), SIV uptake exceeded 95% at inclusion, finally dropping to 59% in 2015/16 (range of SIV injections: [6–9]). The uptake in trajectory 3 (“postpandemic decreasingly vaccinated”, 10%) was relatively high (63–73%) and stable until the 2009/10 influenza A(H1N1) pandemic season; it dropped by 33 percentage points in 2010/11 (range of SIV injections: [2–8]). Trajectory 4 (“early increasingly vaccinated”, 9%) began with a very low SIV-uptake rate at inclusion and then immediately and rapidly increased, stabilizing around 90% in 2011/12 (range of SIV injections: [4–9]). Trajectory 5 (“late increasingly vaccinated”, 5%) looked like trajectory 4 with the increasing phase shifted forward several years (range of SIV injections: [2–6]). The individuals with trajectory 6 (“never vaccinated”, 30%) had very low SIV-uptake rates throughout follow-up (range of SIV injections: [0–2]).

With the “continuously vaccinated” trajectory as the reference (Table 2), the probability of belonging to the “progressively less vaccinated” trajectory was higher for individuals aged 65 years or older at inclusion, those receiving no antidiabetic drug, with high comorbidity scores at inclusion and remaining stable during follow-up, hospitalized for influenza during follow-up, and seeing GPs frequently. It was lower among women, for those with intensified diabetes treatment, seeing endocrinologists frequently, and changing GPs during follow-up.

The remaining four trajectories (“postpandemic decreasingly vaccinated”, “early”/“late increasingly vaccinated”, and “never vaccinated”) shared several characteristics. The probability of belonging to these four trajectories was higher in patients receiving no antidiabetic drug at inclusion and lower in those aged 65 years or older, with more comorbidities at inclusion, and with frequent visits with specialists during follow-up. These trajectories also showed some specificities. The probability of belonging to the “postpandemic decreasingly vaccinated” trajectory was higher for women and individuals hospitalized for diabetes or influenza; it was lower for those with worsening comorbidities. The probability of belonging to the “early” or “late increasingly vaccinated” trajectories was higher for those with worsening diabetes and comorbidities during follow-up, and those hospitalized for influenza (for the “early” trajectory only); it was lower for individuals with type 1 diabetes. Finally, the probability of belonging to the “never vaccinated” trajectory was higher for women and for individuals with stable comorbidities, and lower for those with type 1 diabetes, with worsening comorbidities, frequent healthcare utilization, and changing GPs during follow-up.

Overall, this study shows remarkable inertia in behavioural patterns, with 28% of the subjects continuously vaccinated and 30% never vaccinated from 2006/07 to 2015/16. For two other trajectories, the SIV-uptake rate decreased during follow-up, either progressively (18%) or more sharply after the 2009/10 season (10%), while the SIV-uptake rate rose for the last two trajectories (accounting for only 14% of patients). Compared to “continuously vaccinated” people, those in the “progressively decreasing” trajectory were older; those in the “postpandemic decreasing”, “increasing”, and “never” vaccinated trajectories were younger than the reference category with fewer comorbidities at inclusion. The “increasing” trajectories were positively associated with the worsening of diabetes and comorbidities during follow-up.

The strengths of this study include its 10-year follow-up, the longest for any study examining SIV behaviours over time [6, 9, 10], and its large sample size. Moreover, our algorithm to identify patients with diabetes was more sensitive and allowed earlier identification than an algorithm based solely on LTI [8]. We used vaccine deliveries, which are more reliable than self-reported vaccination behaviour [23]. The dropout extension of the group-based trajectory modelling allowed us to control for potential selection biases due to non-random participant attrition (especially those due to mortality, Additional file 3: Table S3) [20].

We acknowledge some limitations. Vaccinations that took place during occupational medicine visits or at vaccination centres or some nursing homes that buy vaccines for their residents (fewer than 20% of all nursing homes [24]) are not recorded in the French NHIF databases. However, these limitations are unlikely to affect our results substantially as the vast majority of vaccinations in France are administered by private healthcare workers and are thus recorded in these databases [25]. As SIV behaviour varies by socioeconomic characteristics [26], our results cannot be extrapolated to population categories not covered by the NHIF (e.g., farmers, the self-employed) or the very few people without insurance; nonetheless, the NHIF covers 86% of the French population. Several socioeconomic (e.g., educational level) and clinical (e.g., diabetes complications) characteristics are not recorded in NHIF databases and therefore could not be studied. Specifically, no data about individuals’ knowledge, attitudes or perceptions towards SIV (e.g., beliefs about SIV efficacy, side effects) were available, although they are important drivers of SIV behaviours [27, 28] and thus probably differ according to trajectories.

Our finding that most people with diabetes had stable SIV behaviours is consistent with results from previous qualitative [28] and quantitative studies with shorter follow-ups [9]. When health protective behaviours must be regularly repeated in stable contexts, patients’ responses to their healthcare workers’ recommendations may be performed almost automatically, without either conscious decision-making or thinking [29]. This interpretation is in line with recent advances in behavioural sciences showing that “much human behaviour is automatic, cued by environmental stimuli” [30]. We may assume that subjects continuously vaccinated were aware of their vulnerability to influenza (due to age and/or comorbidities [9]) before our follow-up began. Another hypothesis is that receiving a free voucher each year at least as early as inclusion (Additional file 3: Table S3) and regular medical consultations may foster SIV behaviours because they act as reminders [31] and the vouchers may facilitate access to the vaccine [8]. Conversely, studies show that continuously refusing SIV is often associated with attitudes of risk neutralization (e.g., comparing SI with other infectious diseases, feeling “mentally and physically” able to resist SI) [28]. Opportunities might also have been missed: we estimated that, at inclusion, 30% of patients with diabetes did not receive free vouchers because they did not benefit from LTI status. These patients can obtain a voucher from their doctor but this makes their pathway to vaccination still more complex as it requires first a doctor’s appointment to get a free vaccine voucher, then a trip to the pharmacy to pick up the vaccine, and then a second appointment for the actual injection.

Nonetheless, the shifts in SIV behaviour among distinct groups of patients suggest different underlying mechanisms of behaviour change. The characteristics of individuals in the “progressively less vaccinated” trajectory (the oldest in our cohort) may imply a progressive phasing-out of SIV among the frail elderly. This may result from doubts among patients, their relatives and/or their doctors about the benefits of SIV in the oldest populations, due to the scientific debate and its media coverage regarding SIV effectiveness and immunosenescence [32, 33]. Our results that patients in the “progressively less vaccinated” trajectory had less frequent consultations with endocrinologists and antidiabetic treatment might also suggest that diabetes itself and prevention of its complications has become a lower priority among these patients.

The “postpandemic decreasingly vaccinated” trajectory strongly echoes the fall in SIV coverage observed in most at-risk groups in France after the 2009 A(H1N1) pandemic season [8]. This drop has been linked to the controversies about the safety and effectiveness of the A(H1N1) vaccine surrounding the French mass vaccination campaign against the pandemic [8]. The overrepresentation of women in this trajectory is consistent with gender differences in vaccine hesitancy found for other vaccines [26].

Finally, our results regarding the “early/late increasingly vaccinated” trajectories suggest that adverse health events (e.g., intensification of diabetes treatment, worsening comorbidities) may foster or trigger adoption of SIV, which is in line with previous findings [9]. Finally, the percentages of subjects receiving free vouchers for the first time during follow-up rather than at baseline were highest in these trajectories (Additional file 3: Table S3). This finding suggests that offering a voucher might foster positive behaviour change [8, 31].