INTRODUCTION
In most winters, influenza causes substantial morbidity and mortality [Reference Fleming and Elliot1, Reference Fleming2], with widely variable clinical manifestations ranging from subclinical infection to fatal illness [Reference Nicholson, Nicholson, Webster and Hay3]. Although many persons become infected, seroprevalence studies indicate that only a minority make use of health services [Reference Miller4]. Nevertheless, those that do create considerable pressure on both primary- and secondary-care services [Reference Fleming, Zambon and Bartelds5, Reference Fleming, Harcourt and Smith6]. The risk of serious illness involving hospital admission or death is thought to be higher in persons with chronic illness [Reference Glezen7], but the total impact of influenza on health services comes at least as much from otherwise healthy persons [Reference Molinari8]. In the UK this impact is concentrated in a comparatively short epidemic period of around 8 weeks [Reference Fleming9], creating surge pressures on health services, commonly in midwinter, when the impact of other respiratory viral infections is also highest [Reference Elliot, Cross and Fleming10].
Policies to prevent and treat influenza need to accommodate the varying threat from different virus strain types, changes in severity of epidemics over time, advances in vaccines, the ageing population, increasing population expectancies from healthcare and changes in healthcare delivery. These factors interact to influence the cost effectiveness of medical intervention, and create the need to update burden assessments on a continuing basis using measurements from several seasons.
Most estimates of the influenza burden employ modelling methods in which the identification of the influenza epidemic period is determined from available virology data and the impact measured in the commensurate changes in the chosen outcome [Reference Thompson11–Reference Zambon15]. In some modelling studies, influenza virus data have been used exclusively, but the confounding effect of other viruses has prompted the inclusion of data for other viruses, especially respiratory syncytial virus (RSV). In a few studies, meteorological data has been incorporated as well [Reference Pitman13, Reference Hardelid, Pebody and Andrews14]. This study aims to provide a comprehensive analysis of the burden of persons consulting a general practitioner (GP) for a cause attributable to influenza over many seasons, to inform cost-benefit analyses of vaccination and treatment interventions. We used consultation data from the Clinical Practice Research Datalink (CPRD) and virology data from Public Health England (PHE) in a linear regression model to estimate influenza-attributable consultations and antibiotic prescriptions over 14 seasons, while controlling for RSV and other seasonal variables, and stratifying the analyses by age, comorbidity risk and influenza vaccination status.
METHODS
Study design
Viral surveillance time-series regression models were used to estimate the number of persons consulting a GP for influenza-attributable respiratory disease in England in each season (July in the index year through June the following year), from 1995 to 2009, stratifying by age, risk status and vaccination status (www.clinicaltrials.gov: NCT01520935). Seasons after July 2009 were not included as we were concerned to establish the normal or average seasonal burden of influenza, which was disturbed by the pandemic experience in summer 2009.
Data sources
At the midpoint of the study (January 2001) the monitored population was around 3·7 million. Patients included in the study were registered in approximately 550 general practices [Reference Williams16]. Individual patients could be followed for varying durations during this period. CPRD diagnoses and interventions are summarized and stored as READ codes.
Weekly influenza and RSV counts were obtained from PHE. Influenza reports are based on regular returns of virus swab-positive tests to PHE, mainly from community-based investigation of persons with influenza-like illness. RSV reports are mainly confirmed infections in young children admitted to hospital with respiratory infections. We used only counts of positive nose/throat swabs or nasopharyngeal aspirates. Viruses were classified as A or B without further classification.
All analyses of CPRD were restricted to English practices to match the coverage of virology data.
Outcome definitions
A GP episode was counted as the first in a series of consultations for a particular diagnosis/diagnostic group. A minimum of 28 days was required following any previous consultation for that same diagnosis/diagnostic group before defining a new episode.
Diagnostic READ codes were combined into respiratory diagnoses consistent with recognized ICD-10 groupings (Table 1) (see Supplementary material). We undertook a preliminary search of records to find codes that increased during influenza active periods. Our ‘respiratory disease broadly defined’ outcome included all disorders in the ICD-10 respiratory disease chapter, otitis media, as well as selected presenting symptom codes we found were used by some GPs in lieu of specifying a clinical diagnosis. A negative control outcome (urinary tract infection) with no seasonal pattern and association with influenza was included to control for the possibility of bias in model attribution over the entire study period.
Table 1. Outcomes: general practitioner consultations (CPRD)*
* Any Clinical Practice Research Datalink (CPRD) GP episode, including office visits, home visits, telephone consults and other types, for subjects registered with research quality data in CPRD. CPRD diagnostic data are coded using READ codes selected to correspond to ICD-10 definitions.
Time-series of numbers of prescriptions of broad spectrum penicillin, tetracycline and macrolide antibiotics, which are widely prescribed for acute respiratory infections, were also examined as an outcome category.
Vaccination and risk status
The CPRD dataset contains unique patient identifiers, allowing patients to be followed longitudinally. Since almost all influenza vaccination during the study period was undertaken by GPs and recorded using specific codes, we were able to check the vaccination status for each person in each season. Vaccine coverage in the CPRD population was used to estimate the vaccination coverage in each age-risk stratum. We also externally validated the CPRD vaccination coverage estimates against national data on vaccination coverage [Reference Pebody17]. We classified each patient as ‘high’ or ‘low’ risk by the prior occurrence of diagnoses that would prioritize the patient to receive influenza vaccination by reference to the accumulated historical records in the database (Table 1) [18].
Denominators
The CPRD covers a population which is representative of England in age and gender distribution and is thus a suitable population-denominated database for the assessment of vaccination status and national patterns of comorbidity in age- and gender-specific groups [Reference Walley and Mantgani19]. We found excellent agreement between CPRD and independent data sources for the UK with respect to the population age structure, the prevalence of various risk factors, and influenza vaccination coverage; for example, CPRD data yielded an average vaccination rate of 71% for persons aged ⩾65 years for the years 2001–2008, in close agreement with PHE estimates [Reference Pebody17]. The 2001 UK population (Office of National Statistics [20]) was thus used to weight the English CPRD population by age to reflect the UK profile.
Statistical methods
Statistical analyses were performed using SAS v. 9.2 (SAS Institute Inc., USA). Weekly time-series of the number of specimens positive for influenza A, influenza B and RSV were generated using PHE virology surveillance data. Weekly time-series for multiple influenza-related health outcomes (Table 1) were generated from the CPRD for age groups <5, 5–17, 18–49, 50–64, and ⩾65 years (for selected analyses age groups 65–74 and ⩾75 were examined separately) and low-/high-risk groups. A multiple linear regression model was applied to each age group to associate GP episodes to influenza A or B, while controlling for RSV and unspecified seasonal factors (Fig. 1). Major changes in virus detection methodology were undertaken by PHE in 2001; thus, to avoid bias the model separated the periods before and after this date:
$$\eqalign{Y = & {\beta } _{0} + {\beta } _{s1} {t} + {\beta } _{s2} {t^2} + {\beta } _{s3} {t^3} + {\beta } _{s4} {\rm sin}\left( {\rm 2}\pi {t}/52\hskip-1pt\!\cdot\!\hskip-2pt18 \right){\rm} \cr & + {\beta } _{s5} {\rm cos}\left( {\rm 2}\pi {t}/52\hskip-1pt\!\cdot\!\hskip-2pt18 \right){\rm} \cr & + {\beta } _{\,p1a} {\rm InfluenzaA}\left( {\rm pre- July\,2001} \right){\rm} \cr & + {\beta } _{\,p1b} {\rm InfluenzaA}\,\left( {\rm post - July\,2001} \right){\rm} \cr & + {\beta } _{\,p2a} {\rm InfluenzaB}\left( {\rm pre - July\,2001} \right){\rm} \cr & + {\beta } _{\,p2b} {\rm InfluenzaB}\left( {\rm post - July\,2001} \right){\rm} \cr & + {\beta } _{\,p3a} {\rm RSV}\left( {\rm pre - July\,2001} \right){\rm} \cr & + {\beta } _{\,p3b} {\rm RSV}\left( {\rm post - July\,2001} \right),}$$
where, Y is the incidence (rate) of outcome definition for each time period t (weeks), RSV and influenza are the proportion of laboratory isolates during t, sin and cos are harmonic functions of t, and the remaining terms track other types of secular trends in the data. The time period t used in the model is a running index of weeks starting 1 October 1997 and ending 31 March 2009.
Fig. 1. Attribution modelling showing excess GP episodes attributable to influenza A and B and respiratory syncytial virus (RSV) in children aged 5–17 years. The pattern was observed to be the same for other age groups. ‘All other’ refers to GP episodes not attributable to influenza or RSV. The denominator is the UK population.
β p1a Influenza A(pre-July 2001) + β p1b Influenza A(post-July 2001) are influenza terms pre- and post-July, 2001. The RSV terms control for RSV epidemics. Influenza A, Influenza B and RSV are observed counts of positive tests from the PHE LabBase dataset.
Outcome and pathogen series were tri-mean-smoothed, i.e. the smoothed value of X at week t is (X t–1 + X t + X t + 1)/3. The week numbers (t) were calculated as ISO 8601 V-weeks. The weekly attribution to each virus was computed as the product of the observed value of the explanatory variable (i.e. number of positive virology samples) and the corresponding regression coefficient (β p1a –β p2b ), and the weekly estimates summed to produce the seasonal estimates. The variability of seasonal estimates was assessed using standard deviations but limited to all-age estimates. This standard deviation represents the variability of the attributable burden between seasons and not the uncertainty of the individual seasonal estimates. Results were expressed as numbers of persons consulting a GP (GP episodes) and as the mean seasonal (from September to mid-May) rates of persons consulting a GP per 100 000 population.
Outside pandemics, influenza is exclusively a winter illness and thus summertime data were excluded from the analysis. Previous evidence suggests that winter increases in the incidence of RSV-associated outcomes (e.g. acute bronchitis) in adults follow those in children by 2–3 weeks [Reference Fleming and Elliot1]. Therefore, for age groups aged ⩾18 years, we lagged the RSV time-series by 2 weeks and achieved an improved model fit; we did not lag the influenza time-series as doing so did not substantially alter model fit or influenza burden attributions. The time-series of prescriptions of antibiotics was modelled in the same way as the health outcome time-series to estimate antibiotic prescribing attributable to influenza. The model did not attribute any positive influenza burden to the control outcome (urinary tract infection).
The protocol was approved by the Independent Scientific Advisory Committee of the Clinical Practice Research Datalink (CPRD, formerly the General Practice Research Database). Informed consent was not required.
RESULTS
Time trends
The onset of the influenza epidemic period usually followed RSV, and influenza A usually preceded influenza B (illustrated for the 5–17 years age group in Fig. 1).
Model fit
The goodness-of-fit was assessed using adjusted R 2. The model fit in most strata was very good, with adjusted R 2 between 0·75 and 0·86 for the respiratory broad outcome (Supplementary Table S1). Moreover, we found a substantial ‘lift’ in the adjusted R 2 value when introducing the virology terms into the base secular model. We did not adjust the model form (by including or excluding terms) individually for each outcome, age, and risk stratum, but applied the given form across all strata.
Overall disease burden
In a mean season there were an estimated 857 996 GP episodes in persons of all ages (1·5% of the population) for influenza-attributable respiratory disease (broadly defined), of which around 75% (649 219) were for influenza A (Table 2). There was wide inter-seasonal variability in influenza A- and B-attributable GP episodes, compared to a more stable consulting rate for RSV (Fig. 2). In the seasons studied the number of GP episodes for an influenza-attributable respiratory disease ranged between 174 305 (2002–2003) and 1 340 937 (1995–1996) for influenza A, and between 2075 (1997–1998) and 646 457 (1996–1997) for influenza B. The highest rates of GP episodes for influenza A-attributable respiratory disease were observed in children aged <5 years (Table 2). The highest seasonal incidence rates for influenza B were observed in those aged 5–17 years.
Fig. 2. Seasonal rates (with 95% confidence intervals) of GP episodes for respiratory disease (broadly defined) attributable to influenza A, influenza B and respiratory syncytial virus (RSV).
Table 2. Persons consulting a general practitioner in a mean season (number and rate/100 000 population) attributable to influenza in the UK, by age (1995–2009)
n, Number of episodes; s.d., standard deviation.
The rate of GP episodes for influenza A-attributable respiratory disease tended to increase after age 50 years, whereas GP episodes for influenza B-attributable respiratory disease decreased with age (Table 2). In the ⩾75 years age group, the rate of GP episodes for respiratory disease was 1255/100 000 for influenza A and 93/100 000 for influenza B.
In a mean season, the rate of GP episodes for influenza A-attributable bronchitis was higher in adults than in children, whereas the rate of GP episodes for influenza-like illness was higher in children than adults (Table 2). The rates of influenza A-attributable GP episodes for ‘pneumonia and influenza’ were higher in adults than in children. In those aged ⩾75 years, the mean seasonal rate of GP episodes for ‘pneumonia and influenza’ was 309/100 000 for influenza A, and 24/100 000 for influenza B.
Otitis media and antibiotic use
Influenza A and B contributed to 68 777 GP episodes for otitis media per season, with the highest number in children (Table 2). Around one third of the GP episodes for otitis media were attributed to influenza B.
There were 604 042 prescriptions for antibiotics attributed to influenza-associated respiratory disease (as measured in the broad respiratory outcome) per season across all age groups (Table 2).
Comorbid risk status and vaccination
Individuals with comorbid conditions that put them at high risk of severe influenza infection were more likely to consult a GP for an influenza-attributable respiratory disease than individuals at low risk (Table 3). The exceptions were children aged <5 years for influenza A, and adults aged 50–64 years for influenza B. The average seasonal rates of influenza A-attributable GP episodes in the unvaccinated high comorbid risk group were 1·79-, 1·56- and 1·83-fold higher than in the low comorbid risk group for older adults aged 50–64, 65–74 and ⩾75 years, respectively. Trends for influenza B were similar for the 65–74 and ⩾75 years age groups (1·74- and 2·97-fold higher rates in high comorbid risk group, respectively).
Table 3. Persons consulting a general practitioner in a mean season (rate/100 000 population) for outcomes attributed to influenza A and B by comorbid risk status
s.d., Standard deviation.
Higher rates of GP episodes were observed in all age groups when comorbid risk factors were present than when they were absent for the other outcomes studied.
Between 2000 and 2008 influenza vaccine coverage in the CPRD population aged ⩾65 years ranged between 64% and 76%, which was similar to coverage estimated by PHE [Reference Pebody17]. In each season studied the rate of influenza A-attributable GP episodes in the high comorbid risk group aged 18–49 years was 0·6- to 0·89-fold lower in vaccinated compared to unvaccinated individuals. The rate ratio of influenza A-attributable GP episodes in the high comorbid risk group in vaccinated over unvaccinated individuals was between 0·74 and 0·76 in those aged 50–64 years, and 0·81–0·85 in those aged ⩾65 years.
Similarly, in these age groups the rate of GP episodes was lower in vaccinated vs. unvaccinated persons at low risk: the rate ratio of influenza-A attributable GP episodes in the low comorbid risk group in vaccinated over unvaccinated was 0·78–1 in those aged 18–49 years, 0·79–0·92 in those aged 50–64 years, and 0·76–0·98 in those aged ⩾65 years.
Clinical syndromes
A wide spectrum of diseases that resulted in a GP episode was attributable to influenza A and B (Fig. 3). The relative proportions attributable to influenza A and B varied with the GP-reported diagnoses, highlighting the importance of influenza B in the 5–17 years age group; especially so since influenza B seasons are less frequent than influenza A.
Fig. 3. The percentage of the estimated burdens for several outcomes attributed to influenza A and influenza B by five age groups (mean across seasons). The total number of GP episodes for each outcome was: respiratory disease (broadly defined) (n = 857 996), bronchitis and bronchiolitis (n = 248 727), influenza-like illness (n = 265 123), pneumonia and influenza (n = 197 769), otitis media (n = 68 777), antibiotic use (n = 604 042).
DISCUSSION
We used the CPRD in the UK, a promising but as yet underutilized nationally representative database for influenza burden assessment, to estimate the age-specific influenza burden in primary care [Reference Williams16]. We employed traditional methods used for hospitalization and mortality studies, and applied them to time-series of GP episodes for a diverse range of outcomes forming part of the burden of influenza, such as comorbidity risk status and antibiotic prescriptions, after taking account of the potentially confounding effects of RSV. In a mean season, around 1·5% of the UK population were estimated to present to GPs with diverse respiratory syndromes attributable to influenza A or B. The estimated mean seasonal burden of influenza A was much larger than influenza B in all age groups and for most respiratory outcomes. Approximately 0·7% of children aged <5 years were estimated to consult a GP for influenza-related otitis media in a mean season, and nearly 2% of all children received antibiotic prescriptions which were attributable to influenza. Compared to the relatively stable disease burden attributable to RSV in each season, the influenza burden was characterized by marked inter-seasonal variability according to epidemic activity.
The single previous UK study by Pitman et al. that used the CPRD data source to study GP office visits examined only a single season (2002–2003) [Reference Pitman13]. In that study, Pitman et al. estimated approximately one million influenza-attributable consultations. This is higher than our estimate for that season of 330 763 GP episodes, but closer to our mean seasonal estimate encompassing 14 seasons. As well as the inherent instability of estimates produced using a single season, the model used by Pitman et al. [Reference Pitman13] reported two alternative estimates obtained by models with and without a bacterial pathogen series included in the explanatory variables. The inclusion in the model of Pitman et al. of the bacterial series, characterized by a dominant sinusoidal pattern, likely affected the influenza burden estimate similarly to our use of cyclic components.
Estimates of the number of persons infected with influenza in an average season vary widely, but when measured as persons consulting with clinical illness they are mostly less than 5% of the population, depending on the diagnostic outcomes measured [Reference Molinari8, Reference Fleming9, Reference Sullivan, Monto and Longini21, Reference Schanzer and Schwartz22]. Our findings are consistent with a simpler analysis conducted in a database of the Royal College of General Practitioners, which estimated an influenza burden affecting 2·1% of the population of England and Wales in seasons 1989–1999 [Reference Fleming, Zambon and Bartelds5].
We observed that in persons with comorbid conditions that placed them in a high risk group and in all persons aged <65 years, there were fewer GP episodes for influenza-attributable respiratory disease in vaccinated than non-vaccinated persons. This finding is difficult to interpret because our study was not designed to measure vaccine effectiveness, which would require a different study design. Such studies would require an analysis of the propensity to consult, which differs between vaccinated and non-vaccinated persons, and between those with high and those with low comorbidity risk.
Population-based studies of the influenza burden relying on individual laboratory-confirmed cases alone are not practical, due to the lack of standard testing, the expense of large-scale testing using sensitive PCR-based assays, and the fact that infections are often resolved by the time patients seek medical care (if they seek such care at all). Instead, indirect modelling approaches have been used for many years to estimate the burden of influenza [Reference Thompson12, Reference Serfling23]. We used a previously described linear regression model [Reference Matias24], which adopted a recent refinement of regression techniques guided by virus surveillance data as described in studies by Pitman et al. [Reference Pitman13] and Zhou et al. [Reference Zhou25]. Other models used to estimate the burden of respiratory infections, such as Poisson regression with a natural logarithm for the link function [Reference Thompson11, Reference Thompson26], imply multiplicative effects of respiratory viruses (i.e. the effect of multiple simultaneously circulating respiratory viruses is different from the sum of their individual effects). This is likely an unrealistic assumption for influenza strains [Reference Yang27, Reference Gay28]. The linear regression model has the advantage over other methods of being an additive model, which corresponds to the view that the total burden in the population is due to the sum of outcomes due to different causes, including viral infections [Reference Yang27, Reference Goldstein29].
The strengths of our model include the use of large nationally representative databases, which limited the potential for sampling error; inclusion of data over a long period that covered 14 seasons; and the use of control outcomes with no inherent seasonality and no association to influenza. By defining a new outcome category (‘respiratory disease broadly defined’), we aimed to improve the sensitivity in capturing the full burden of influenza-attributable disease, while retaining sufficient specificity; for example, the additional codes in the ‘respiratory broad’ outcome increased the average seasonal all-age estimate by 14% over the stricter ‘respiratory’ outcome, but left the confidence intervals and the model fit virtually unchanged (data not shown). Age stratification allowed detailed estimation of the age-specific burden, controlled for the observed age effect on consultation rates and overcame possible confounding due to higher frequency of viral testing in some age groups. Finally, weekly occurrences of outcomes were assumed to be determined by the circulation of influenza, RSV, and other causes which follow a seasonality that was estimated using a combination of sine and cosine terms. The inclusion of both cyclical terms (sine and cosine functions) controlled for confounders for which data are not available. Tri-mean smoothing of pathogens and outcomes series mitigated short-term effects such as national holidays and extreme weather conditions on viral testing and GP consultations.
A potential limitation of the study is the exclusive analysis of influenza and RSV virus time-series with no consideration of other respiratory pathogens; if such pathogens consistently circulate at the same time as influenza, the influenza burden could be overestimated. We did not generate time-series for influenza subtypes as we did not have access to subtyped data. We did not evaluate the effect of pneumococcal vaccination in children or older age groups [Reference Christenson30], or consider changes in herd immunity during the period of study, nor did we consider the effects of obesity [Reference Cocoros31, Reference Mertz32]. The study is also dependent on the quality and consistency of routine recording in the practice network. The burden of illness as encountered in primary care, which is the primary objective of this study, includes all consulting persons regardless of their propensity to consult. We did not study the indirect burden of illness in individuals with influenza who did not consult. While our study provides up-to-date and detailed information on health service utilization, the study does not provide information for evaluation of indirect medical costs associated with influenza. None of the potential limitations of the study are likely to have influenced the model estimates in any major way.
In addition to influenza-attributable respiratory disease, we estimated the burden of influenza-attributable otitis media, which has often been regarded as a common complication of influenza [Reference Heikkinen33]. Our study found almost as many children (aged <5 years) were diagnosed at presentation with influenza-attributable otitis media as with influenza-like illness, suggesting that otitis media is part of the primary symptomatology of influenza. The diverse range of presenting illnesses identified as attributable to influenza in our study questions the relevance of diagnostic criteria by which influenza is defined clinically. Like other studies that have tried unsuccessfully to correlate specific symptoms with a diagnosis of influenza [Reference Mulpuru34, Reference Michiels35], our study suggests that collections of symptoms with/without fever are largely inadequate for measuring the burden of illness.
Finally, the model estimated that more than 600 000 antibiotic courses were prescribed for influenza-associated respiratory disease per season. The potential misuse of antibiotics is a well-recognized cause for concern, particularly in relation to common respiratory infections [Reference Hawker36]. A simple test that is able to discriminate between bacterial and viral infection would be a great step forward, as would timely local influenza surveillance data that would allow physicians to assess the probability that an acute respiratory infection is caused by influenza.
Health economic models can inform policy choices, but they depend on accurate and age-specific parameter estimates of the disease burden. Furthermore, it is critical to use estimates based on recent assessments of the burden: earlier publications could overestimate the current burden. Because the effectiveness of influenza vaccination in preventing otitis media appears similar to that of pneumococcal conjugate vaccines [Reference Heikkinen37], it is important that estimates of the otitis media burden and the potentially preventable fraction be included in cost-benefit estimates of influenza interventions in this age group. Our study provides a source for understanding the national influenza burden and undertaking cost evaluations.
SUPPLEMENTARY MATERIAL
For supplementary material accompanying this paper visit http://dx.doi.org/10.1017/S0950268815001119.
ACKNOWLEDGEMENTS
The authors thank the Sage Analytica team, in particular Lone Simonsen and Lewis Kim, for their contribution to the study. GlaxoSmithKline Biologicals SA was the funding source and was involved in all stages of the study conduct and analysis. GlaxoSmithKline Biologicals SA also took in charge all costs associated with the development and publication of the present manuscript. The authors are grateful to all teams of GSK Vaccines for their contribution to this study, especially Robert Gardner (scientific writer), Archana Jastorff (scientific writer; XPE Pharma and Science, on behalf of GSK Vaccines), Els Tassenoy (global study manager; Harrison Clinical Research, on behalf of GSK Vaccines), and Dominique Rosillon (Biometrics). Finally, the authors thank Joanne Wolter (independent medical writer, on behalf of GSK Vaccines) for providing writing services and Bruno Dumont (Business and Decision Life Sciences, on behalf of GSK Vaccines) for editorial assistance and manuscript coordination. Virology data were sourced from Public Health England (formerly Health Protection Agency). The opinions expressed in this report are solely those of the authors. [www.clinicaltrials.gov identifier: NCT01520935]
DECLARATION OF INTEREST
Gonçalo Matias, Dave J. Webb, John Logie, and François Haguinet are employed by the GSK Group of Companies. Dave J. Webb and John Logie report ownership of stock options and/or restricted shares. Robert J. Taylor, Roger L. Lustig, and Cynthia Schuck-Paim report having received consulting fees from Sage Analytica LLC to perform this study, paid for by the GSK Group of Companies. Douglas M. Fleming reports personal fees from Sage Analytica, personal fees and non-financial support from the GSK Group of Companies during the conduct of the study.
Open access
