Vaccine 32 (2014) 6884–6901
Contents lists available at ScienceDirect
Vaccine journal homepage: www.elsevier.com/locate/vaccine
Review
Are influenza-associated morbidity and mortality estimates for those ≥65 in statistical databases accurate, and an appropriate test of influenza vaccine effectiveness? Roger E. Thomas ∗ Department of Family Medicine, University of Calgary, Health Sciences Centre, 3330 Hospital Drive NW, Calgary, AB, Canada T2N 4N1
a r t i c l e
i n f o
Article history: Received 5 January 2014 Received in revised form 14 July 2014 Accepted 27 August 2014 Available online 27 October 2014 Keywords: Influenza Morbidity and mortality Influenza vaccine effectiveness Statistical models Databases RCTs
a b s t r a c t Purposes: To assess the accuracy of estimates using statistical databases of influenza-associated morbidity and mortality, and precisely measure influenza vaccine effectiveness. Principal results: Laboratory testing of influenza is incomplete. Death certificates under-report influenza. Statistical database models are used as an alternative to randomised controlled trials (RCTs) to assess influenza vaccine effectiveness. Evidence of the accuracy of influenza morbidity and mortality estimates was sought from: (1) Studies comparing statistical models. For four studies Poisson and ARIMA models produced higher estimates than Serfling, and Serfling higher than GLM. Which model is more accurate is unknown. (2) Studies controlling confounders. Fourteen studies mostly controlled one confounder (one controlled comorbidities), and limited control of confounders limits accuracy. Evidence for vaccine effectiveness was sought from: (1) Studies of regions with increasing vaccination rates. Of five studies two controlled for confounders and one found a positive vaccination effect. Three studies did not control confounders and two found no effect of vaccination. (2) Studies controlling multiple confounders. Of thirteen studies only two found a positive vaccine effect and no mortality differences between vaccinees and non-vaccinees in non-influenza seasons, showing confounders were controlled. Key problems are insufficient testing for influenza, using influenza-like illness, heterogeneity of seasonal and pandemic influenza, population aging, and incomplete confounder control (co-morbidities, frailty, vaccination history) and failure to demonstrate control of confounders by proving no mortality differences between vaccinees and non-vaccinees in non-influenza seasons. Major conclusions: Improving model accuracy requires proof of no mortality differences in pre-influenza periods between the vaccinated and non-vaccinated groups, and reduction in influenza morbidity and mortality in seasons with a good vaccine match, more virulent strains, in the younger elderly with less immune senescence, and specific outcomes (laboratory-confirmed outcomes, pneumonia deaths). Proving influenza vaccine effectiveness requires appropriately powered RCTs, testing participants with RT-PCR tests, and comprehensively monitoring morbidity and mortality. © 2014 Elsevier Ltd. All rights reserved.
1. Introduction Outcome measures in reviews of influenza vaccine usually include laboratory proof of influenza, respiratory and cardiovascular illness attributed to influenza, hospitalisation for influenza or pneumonia, death from influenza or pneumonia/influenza, influenza-like illness (ILI), ILI consultations, and all-cause mortality.
∗ Corresponding author at: Department of Family Medicine, Faculty of Medicine, University of Calgary, 3330 Hospital Drive NW, Calgary, AB, Canada T2M1M1. Tel.: +1 403 220 0157; fax: +1 403 270 4329. E-mail address: [email protected] http://dx.doi.org/10.1016/j.vaccine.2014.08.090 0264-410X/© 2014 Elsevier Ltd. All rights reserved.
Influenza trends are estimated by statistical modelling of databases and are used as an alternative to RCTs to ascertain vaccine effectiveness but problems are: “To estimate the numbers of deaths attributable to seasonal and pandemic influenza is difficult, as influenza infections generally are not laboratory confirmed, are not often recognized and mostly not specified on hospital discharge forms or death certificates. Additionally many influenza associated deaths occur weeks after the initial infection from secondary complications or from exacerbations of chronic illnesses and in both cases influenza viruses are no longer detectable. . . . the influenza virus may predispose to . . . bacterial super infections and cardiovascular complications. . .The inability of diagnoses on death
R.E. Thomas / Vaccine 32 (2014) 6884–6901
certificates to give a reliable and consistent account of the burden of death due to influenza has been understood for many decades and this understanding led to the development of statistical models to estimate influenza associated deaths” [1]. The purpose of this review is to assess whether estimates of influenza-associated morbidity and mortality are accurate, and are an appropriate test of influenza vaccine effectiveness. 2. Material and methods Medline, Embase and Cochrane reviews were searched from inception to 28 December 2013 using “influenza” and (“death” or “mortality”) with no language limitations. Relevant citations were also followed up with the Pub Med single citation search utility. A final search was performed on 1 July 2014 with “influenza” and (“death” or “mortality)” and (“model” or “statistic”) 3. Results 3.1. Search In total, 8740 abstracts/titles were retrieved (Medline: 897, Embase: 7721, Cochrane reviews: 122); 1487 entries were duplicates. Of the remaining 7253 entries, 278 were read in full-text. All relevant abstracts were read in full-text and followed up in Pub Med with the single citation search. A final search on 1 July 2014 identified 266 in Medline, 991 in Embase and 117 in the Cochrane database, yielding 4 additional articles. 3.2. Completeness of death certificates and laboratory reports Death certificates markedly under-record influenza, especially older people with co-morbidities. In England and Wales 2005–8, 131 death certificates were coded influenza yet models estimated 12,700 premature deaths [2]. Autopsy rates are low: only 14% of Western Australian deaths received a coroner’s autopsy: 22% had unknown or suspected infectious causes of death, of these 8.3% had a respiratory virus (0–9 years 28%), 3% influenza and 2% RSV [3]. Laboratory testing also underestimates influenza deaths as physicians test few cases. During the Danish 2009 pandemic regression models estimated deaths ten times higher than laboratory-confirmed influenza [4]. 3.3. Influenza in nursing homes Nursing homes are appropriate to assess vaccine effectiveness but there are few prospective studies with comprehensive laboratory testing. In 578 French nursing homes (44,869 residents ≥65; 93.4% received influenza, 13% pneumococcal vaccine) patients were assessed for infection by questionnaire and clinical examination [5]. The annual infection rate was 11.23% (95%CI 10.50, 11.97), with 4.60% (4.04, 5.54) classified definite; probable pneumonia 0.89% (0.69, 1.04) and definite 0.39% (0.32, 0.52); and probable upper + lower respiratory tract infections 3.34% (2.88, 3.87) and definite 1.31% (1.09, 1.68) [6]. A study of 98 French long-term care institutions 2004–5 (8041 patients; 93% residents received influenza vaccine, 35% staff) found 64% of residents exposed to ILI but only 3.5% tested for influenza. ILI-exposed patients had slightly higher hospitalisation risks (9.2% vs. 7.4%; RR 1.24; 1.05, 1.47) and all-cause mortality (5.8% vs. 4.3%; RR 1.36; 1.10, 1.70) [7]. Neither study assessed influenza vaccination. Because few prospective studies assess all patients for respiratory illness, collect virological data and assess all patient outcomes, statistical models assess influenza-attributed morbidity and mortality. Evidence was sought from two sources for influenza morbidity and mortality: studies
6885
comparing different statistical models, and models controlling for confounders. For vaccine effectiveness evidence was also sought from two sources: studies of the same region with low and higher rates of influenza vaccination (usually a policy change), and studies controlling for multiple confounders. 3.4. Do Serfling, GLM, Poisson and ARIMA models provide similar influenza mortality estimates? Farr’s Serfling model subtracted deaths in London during a winter with little influenza from an epidemic winter [8]. Four studies comparing different models were identified (Table 1). Studies of single models are widely reported in the literature and not commented on here. Lemaitre for France 1969–2010 compared Serfling and Poisson models only for the A/H1N1 2009–10 pandemic. For those >65 years-of-age estimates differed substantially. For pneumonia and influenza excess deaths the Poisson model estimated 3.5 (95%CI 1.2, 5.6)/100,000 and the Serfling 1.8 (−0.16, 3.7) and for respiratory excess deaths 4.1 (−0.94, 8.9) and 3.2 (−1.1, 7.8), respectively. Estimates were in the opposite direction with wide 95%CIs for cardio-respiratory −3.1 (−18, 11) and 7.4 (−4.6, 19), and allcause deaths −6.59 (−41, 27) and 22 (−1.9, 45) [9]. López-Cuadrado found for Spain 1999–2005 that the Serfling model’s estimates were higher than the GLM model (which assumed influenza deaths in any week were proportional to isolates). The Serfling model for those >64 years-of-age estimated a higher influenza and pneumonia excess annual death rate (15.25/100,000; 6413 deaths) than the GLM model (6/100,000; 2473 deaths); and also a higher all-cause annual death rate (164.1/100,000; 68,977 deaths) than the GLM model (57.05/100,000; 23,560 deaths). López-Cuadrado explained the differences by the GLM model assuming deaths/week due to influenza are directly proportional to viral laboratory detections, and the Poisson log-link models exponentially increase deaths with the number of isolates [10]. Newall for Australia 1997–2004 found the Poisson model consistently estimated higher death rates for respiratory, circulatory and all-cause mortality than the Serfling model. For those 75–84 the Serfling model estimated respiratory deaths at 31.18/100,000 and the Poisson model 50.63, circulatory deaths as 52.93/100,000 (Poisson 63.80), and all-cause deaths as 87.92/100,000 (Poisson 138.06). Newall attributed the difference to fewer influenza months in the Serfling model [11]. Thompson for the US 1972–2003 for those ≥65 years-of-age found the Serfling model estimated annual influenza-associated deaths as 20,161 (95%CI 14,907 to 25,415), the Serfling-Poisson model 22,790 (95%CI 17,565 to 28,033) and the ARIMA model 24,856 (95%CI 19,576 to 30,136). The rate-difference model estimated the periseason lasted an average of 7.4 epidemic weeks (involving 19,954 excess deaths) and the summer season experienced 36,430 excess deaths. Thompson’s Poisson model using virological data produces estimates between Serfling and ARIMA models which do not [12]. Thompson commented that: “A strength of the Serfling least squares regression model is that it provides estimates of influenzaassociated deaths without the need for influenza virus surveillance data” [12]. This approach vitiates a key step in medicine of obtaining pathological data for confirmation. Thompson noted that: “Disadvantages of Poisson models as used by CDC include requirements for consistent, robust weekly viral surveillance data and for at least 5 years of mortality data before stable estimates . . .can be made.” and “The ARIMA method . . . suffers from some of the same weaknesses . . .including defining influenza seasons solely by the use of statistical models” [12]. For children 1–23 months a prospective study of virus-attributable hospitalisations was compared with estimates by six models (Serfling, peri-season differences, Poisson regression with log-link, negative binomial regression with identity link, and Box–Jenkins transfer function). No model provided accurate or consistent estimates. The authors concluded the problems
6886
R.E. Thomas / Vaccine 32 (2014) 6884–6901
Table 1 Studies which compared Serfling, Poisson or ARIMA models of influenza mortality. Author, date, country; Years included; Age groups
Data sources
Statistical methods
Lemaitre 2012; France; 1969–2010; 0–4, 5–24, 25–44, 45–64, >64
Inserm CépiDc; FluNet weekly positive influenza specimens database; weekly ILI from French Sentinel system
Poisson model with log-link: “We selected the indicator of influenza that provided the best fit to the mortality data . . .(ILI, weekly influenza virus percent positive, or their combination).” “We selected the best influenza indicator and associated lag based on the Akaike Information criterion of respiratory death models.” Serfling model: (no indicators of influenza activity) [“A good fit was achieved for most age groups and causes of death (Pearson correlation >0.56.”]
López-Cuadrado 2012; Spain; 1999–2005; 45–64, >64
National Statistics Institute; ICD-10; Spanish Influenza Surveillance System (covers 2% of Spanish population), Microbiological Information System (covers 25% of Spanish population)
Serfling model: Removed trend and applied tempoiral regression model, excluding data December to April to account for baseline mortality in absence of influenza activity. Assumes normal distribution of errors. Generalised linear model: weekly deaths, influenza and RSV isolates includes virological data. Assumes Poisson distribution of errors and “that the number of deaths due to a specific virus is directly proportional to number of virus detections in that week.”
Newall 2009, Australia, 1997–2004; 50–64,65–74,75–84, ≥85
Australian National Mortality Database; monthly viral Influenza A and B and RSV Australian Laboratory Virology and Serology Reporting Scheme; Australian Bureau of Statistics
Serfling model: Only mortality in “epidemic months” counted towards influenza-attributable mortality. Regression on each age group and cause-specific monthly time series of death rates, excluding June-October, to produce seasonal non-influenza baseline. Poisson model: Monthly mortality counts regressed on influenza and RSV counts; harmonics adjusted for seasonal fluctuations in deaths (independent of influenza and RSV) and linear and squared terms for temporal trends;; Poisson baseline = number of deaths predicted assuming zero laboratory influenza counts Peri- and summer-season rate-difference model: summer season was defined as May-September. Two definitions of the “peri-season” were tested (64 years: 15.25/100,000 (6413 deaths) GLM Model: 1.08/100,000 (2682 deaths); >64 years: 6/100,000 (2473 deaths) All-cause annual deaths: Serfling model: 30.42/100,000 (76,321 deaths); >64 years: 164.1/100,000 (68,977 deaths) GLM Model: 10.97/100,000 (27,307 deaths); >64 years: 57.05/100,000 (23,560 deaths) Deaths/100,000 [R = respiratory deaths; C = Circulatory deaths; A = all-cause deaths] Serfling model: 50–64: R 2.30; C 2.87; A 5.63; 65–74: R 13.17, C 6.36, A 25.01; 75–84: R 31.18, C 52.93; A 87.92; ≥85: R 136.90, C 223.39, A 461.07 Poisson model: 50–64: R 2.85; C 1.79; A 7.91; 65–74: R 14.10, C 9.69, A 34.56; 75–84: R 50.63, C 64.80; A 138.06; ≥85: R 192.66, C 301.40, A 663.94
Rate-differences model: 4 models (estimates for weeks when ≥10% and ≥15% specimens tested positive for influenza; and peri-season when 96% of all influenza-associated deaths occurred among those ≥65. [The correlation of summer rate influenza-associated deaths to the Serfling rate was 0.63, to Poisson regression 0.86 and to ARIMA 0.78; Linear to Poisson regression 0.54; and Poisson regression to ARIMA 0.68]
Table 2 Studies which assessed the effects of confounders on the relationship of influenza and mortality. Data sources
Statistical methods
Other confounders controlled
Results
Charu 2013; USA; 1990–2010; 1 year occur in hospital); CDC RSV and influenza surveillance data
Hebert 2005; USA; 1987–99; Medicare patients (mostly ≥65)
Medicare Provider Analysis and Review files (provides hospital stays and date of death if within two years of admission); ICD-9-CM
Prediction of hospitalisation, and death within 30 days of admission
Age, gender, race-adjusted rates for hospitalisation; Charlson comorbidity index; indicators highly predictive of mortality for patients with pneumonia (respiratory neoplasms, metastatic neoplasms, chronic lung diseases, septicemia); evidence of hospital upcoding to diagnoses that generate more revenue
Mytton 2012; England;
Contents lists available at ScienceDirect
Vaccine journal homepage: www.elsevier.com/locate/vaccine
Review
Are influenza-associated morbidity and mortality estimates for those ≥65 in statistical databases accurate, and an appropriate test of influenza vaccine effectiveness? Roger E. Thomas ∗ Department of Family Medicine, University of Calgary, Health Sciences Centre, 3330 Hospital Drive NW, Calgary, AB, Canada T2N 4N1
a r t i c l e
i n f o
Article history: Received 5 January 2014 Received in revised form 14 July 2014 Accepted 27 August 2014 Available online 27 October 2014 Keywords: Influenza Morbidity and mortality Influenza vaccine effectiveness Statistical models Databases RCTs
a b s t r a c t Purposes: To assess the accuracy of estimates using statistical databases of influenza-associated morbidity and mortality, and precisely measure influenza vaccine effectiveness. Principal results: Laboratory testing of influenza is incomplete. Death certificates under-report influenza. Statistical database models are used as an alternative to randomised controlled trials (RCTs) to assess influenza vaccine effectiveness. Evidence of the accuracy of influenza morbidity and mortality estimates was sought from: (1) Studies comparing statistical models. For four studies Poisson and ARIMA models produced higher estimates than Serfling, and Serfling higher than GLM. Which model is more accurate is unknown. (2) Studies controlling confounders. Fourteen studies mostly controlled one confounder (one controlled comorbidities), and limited control of confounders limits accuracy. Evidence for vaccine effectiveness was sought from: (1) Studies of regions with increasing vaccination rates. Of five studies two controlled for confounders and one found a positive vaccination effect. Three studies did not control confounders and two found no effect of vaccination. (2) Studies controlling multiple confounders. Of thirteen studies only two found a positive vaccine effect and no mortality differences between vaccinees and non-vaccinees in non-influenza seasons, showing confounders were controlled. Key problems are insufficient testing for influenza, using influenza-like illness, heterogeneity of seasonal and pandemic influenza, population aging, and incomplete confounder control (co-morbidities, frailty, vaccination history) and failure to demonstrate control of confounders by proving no mortality differences between vaccinees and non-vaccinees in non-influenza seasons. Major conclusions: Improving model accuracy requires proof of no mortality differences in pre-influenza periods between the vaccinated and non-vaccinated groups, and reduction in influenza morbidity and mortality in seasons with a good vaccine match, more virulent strains, in the younger elderly with less immune senescence, and specific outcomes (laboratory-confirmed outcomes, pneumonia deaths). Proving influenza vaccine effectiveness requires appropriately powered RCTs, testing participants with RT-PCR tests, and comprehensively monitoring morbidity and mortality. © 2014 Elsevier Ltd. All rights reserved.
1. Introduction Outcome measures in reviews of influenza vaccine usually include laboratory proof of influenza, respiratory and cardiovascular illness attributed to influenza, hospitalisation for influenza or pneumonia, death from influenza or pneumonia/influenza, influenza-like illness (ILI), ILI consultations, and all-cause mortality.
∗ Corresponding author at: Department of Family Medicine, Faculty of Medicine, University of Calgary, 3330 Hospital Drive NW, Calgary, AB, Canada T2M1M1. Tel.: +1 403 220 0157; fax: +1 403 270 4329. E-mail address: [email protected] http://dx.doi.org/10.1016/j.vaccine.2014.08.090 0264-410X/© 2014 Elsevier Ltd. All rights reserved.
Influenza trends are estimated by statistical modelling of databases and are used as an alternative to RCTs to ascertain vaccine effectiveness but problems are: “To estimate the numbers of deaths attributable to seasonal and pandemic influenza is difficult, as influenza infections generally are not laboratory confirmed, are not often recognized and mostly not specified on hospital discharge forms or death certificates. Additionally many influenza associated deaths occur weeks after the initial infection from secondary complications or from exacerbations of chronic illnesses and in both cases influenza viruses are no longer detectable. . . . the influenza virus may predispose to . . . bacterial super infections and cardiovascular complications. . .The inability of diagnoses on death
R.E. Thomas / Vaccine 32 (2014) 6884–6901
certificates to give a reliable and consistent account of the burden of death due to influenza has been understood for many decades and this understanding led to the development of statistical models to estimate influenza associated deaths” [1]. The purpose of this review is to assess whether estimates of influenza-associated morbidity and mortality are accurate, and are an appropriate test of influenza vaccine effectiveness. 2. Material and methods Medline, Embase and Cochrane reviews were searched from inception to 28 December 2013 using “influenza” and (“death” or “mortality”) with no language limitations. Relevant citations were also followed up with the Pub Med single citation search utility. A final search was performed on 1 July 2014 with “influenza” and (“death” or “mortality)” and (“model” or “statistic”) 3. Results 3.1. Search In total, 8740 abstracts/titles were retrieved (Medline: 897, Embase: 7721, Cochrane reviews: 122); 1487 entries were duplicates. Of the remaining 7253 entries, 278 were read in full-text. All relevant abstracts were read in full-text and followed up in Pub Med with the single citation search. A final search on 1 July 2014 identified 266 in Medline, 991 in Embase and 117 in the Cochrane database, yielding 4 additional articles. 3.2. Completeness of death certificates and laboratory reports Death certificates markedly under-record influenza, especially older people with co-morbidities. In England and Wales 2005–8, 131 death certificates were coded influenza yet models estimated 12,700 premature deaths [2]. Autopsy rates are low: only 14% of Western Australian deaths received a coroner’s autopsy: 22% had unknown or suspected infectious causes of death, of these 8.3% had a respiratory virus (0–9 years 28%), 3% influenza and 2% RSV [3]. Laboratory testing also underestimates influenza deaths as physicians test few cases. During the Danish 2009 pandemic regression models estimated deaths ten times higher than laboratory-confirmed influenza [4]. 3.3. Influenza in nursing homes Nursing homes are appropriate to assess vaccine effectiveness but there are few prospective studies with comprehensive laboratory testing. In 578 French nursing homes (44,869 residents ≥65; 93.4% received influenza, 13% pneumococcal vaccine) patients were assessed for infection by questionnaire and clinical examination [5]. The annual infection rate was 11.23% (95%CI 10.50, 11.97), with 4.60% (4.04, 5.54) classified definite; probable pneumonia 0.89% (0.69, 1.04) and definite 0.39% (0.32, 0.52); and probable upper + lower respiratory tract infections 3.34% (2.88, 3.87) and definite 1.31% (1.09, 1.68) [6]. A study of 98 French long-term care institutions 2004–5 (8041 patients; 93% residents received influenza vaccine, 35% staff) found 64% of residents exposed to ILI but only 3.5% tested for influenza. ILI-exposed patients had slightly higher hospitalisation risks (9.2% vs. 7.4%; RR 1.24; 1.05, 1.47) and all-cause mortality (5.8% vs. 4.3%; RR 1.36; 1.10, 1.70) [7]. Neither study assessed influenza vaccination. Because few prospective studies assess all patients for respiratory illness, collect virological data and assess all patient outcomes, statistical models assess influenza-attributed morbidity and mortality. Evidence was sought from two sources for influenza morbidity and mortality: studies
6885
comparing different statistical models, and models controlling for confounders. For vaccine effectiveness evidence was also sought from two sources: studies of the same region with low and higher rates of influenza vaccination (usually a policy change), and studies controlling for multiple confounders. 3.4. Do Serfling, GLM, Poisson and ARIMA models provide similar influenza mortality estimates? Farr’s Serfling model subtracted deaths in London during a winter with little influenza from an epidemic winter [8]. Four studies comparing different models were identified (Table 1). Studies of single models are widely reported in the literature and not commented on here. Lemaitre for France 1969–2010 compared Serfling and Poisson models only for the A/H1N1 2009–10 pandemic. For those >65 years-of-age estimates differed substantially. For pneumonia and influenza excess deaths the Poisson model estimated 3.5 (95%CI 1.2, 5.6)/100,000 and the Serfling 1.8 (−0.16, 3.7) and for respiratory excess deaths 4.1 (−0.94, 8.9) and 3.2 (−1.1, 7.8), respectively. Estimates were in the opposite direction with wide 95%CIs for cardio-respiratory −3.1 (−18, 11) and 7.4 (−4.6, 19), and allcause deaths −6.59 (−41, 27) and 22 (−1.9, 45) [9]. López-Cuadrado found for Spain 1999–2005 that the Serfling model’s estimates were higher than the GLM model (which assumed influenza deaths in any week were proportional to isolates). The Serfling model for those >64 years-of-age estimated a higher influenza and pneumonia excess annual death rate (15.25/100,000; 6413 deaths) than the GLM model (6/100,000; 2473 deaths); and also a higher all-cause annual death rate (164.1/100,000; 68,977 deaths) than the GLM model (57.05/100,000; 23,560 deaths). López-Cuadrado explained the differences by the GLM model assuming deaths/week due to influenza are directly proportional to viral laboratory detections, and the Poisson log-link models exponentially increase deaths with the number of isolates [10]. Newall for Australia 1997–2004 found the Poisson model consistently estimated higher death rates for respiratory, circulatory and all-cause mortality than the Serfling model. For those 75–84 the Serfling model estimated respiratory deaths at 31.18/100,000 and the Poisson model 50.63, circulatory deaths as 52.93/100,000 (Poisson 63.80), and all-cause deaths as 87.92/100,000 (Poisson 138.06). Newall attributed the difference to fewer influenza months in the Serfling model [11]. Thompson for the US 1972–2003 for those ≥65 years-of-age found the Serfling model estimated annual influenza-associated deaths as 20,161 (95%CI 14,907 to 25,415), the Serfling-Poisson model 22,790 (95%CI 17,565 to 28,033) and the ARIMA model 24,856 (95%CI 19,576 to 30,136). The rate-difference model estimated the periseason lasted an average of 7.4 epidemic weeks (involving 19,954 excess deaths) and the summer season experienced 36,430 excess deaths. Thompson’s Poisson model using virological data produces estimates between Serfling and ARIMA models which do not [12]. Thompson commented that: “A strength of the Serfling least squares regression model is that it provides estimates of influenzaassociated deaths without the need for influenza virus surveillance data” [12]. This approach vitiates a key step in medicine of obtaining pathological data for confirmation. Thompson noted that: “Disadvantages of Poisson models as used by CDC include requirements for consistent, robust weekly viral surveillance data and for at least 5 years of mortality data before stable estimates . . .can be made.” and “The ARIMA method . . . suffers from some of the same weaknesses . . .including defining influenza seasons solely by the use of statistical models” [12]. For children 1–23 months a prospective study of virus-attributable hospitalisations was compared with estimates by six models (Serfling, peri-season differences, Poisson regression with log-link, negative binomial regression with identity link, and Box–Jenkins transfer function). No model provided accurate or consistent estimates. The authors concluded the problems
6886
R.E. Thomas / Vaccine 32 (2014) 6884–6901
Table 1 Studies which compared Serfling, Poisson or ARIMA models of influenza mortality. Author, date, country; Years included; Age groups
Data sources
Statistical methods
Lemaitre 2012; France; 1969–2010; 0–4, 5–24, 25–44, 45–64, >64
Inserm CépiDc; FluNet weekly positive influenza specimens database; weekly ILI from French Sentinel system
Poisson model with log-link: “We selected the indicator of influenza that provided the best fit to the mortality data . . .(ILI, weekly influenza virus percent positive, or their combination).” “We selected the best influenza indicator and associated lag based on the Akaike Information criterion of respiratory death models.” Serfling model: (no indicators of influenza activity) [“A good fit was achieved for most age groups and causes of death (Pearson correlation >0.56.”]
López-Cuadrado 2012; Spain; 1999–2005; 45–64, >64
National Statistics Institute; ICD-10; Spanish Influenza Surveillance System (covers 2% of Spanish population), Microbiological Information System (covers 25% of Spanish population)
Serfling model: Removed trend and applied tempoiral regression model, excluding data December to April to account for baseline mortality in absence of influenza activity. Assumes normal distribution of errors. Generalised linear model: weekly deaths, influenza and RSV isolates includes virological data. Assumes Poisson distribution of errors and “that the number of deaths due to a specific virus is directly proportional to number of virus detections in that week.”
Newall 2009, Australia, 1997–2004; 50–64,65–74,75–84, ≥85
Australian National Mortality Database; monthly viral Influenza A and B and RSV Australian Laboratory Virology and Serology Reporting Scheme; Australian Bureau of Statistics
Serfling model: Only mortality in “epidemic months” counted towards influenza-attributable mortality. Regression on each age group and cause-specific monthly time series of death rates, excluding June-October, to produce seasonal non-influenza baseline. Poisson model: Monthly mortality counts regressed on influenza and RSV counts; harmonics adjusted for seasonal fluctuations in deaths (independent of influenza and RSV) and linear and squared terms for temporal trends;; Poisson baseline = number of deaths predicted assuming zero laboratory influenza counts Peri- and summer-season rate-difference model: summer season was defined as May-September. Two definitions of the “peri-season” were tested (64 years: 15.25/100,000 (6413 deaths) GLM Model: 1.08/100,000 (2682 deaths); >64 years: 6/100,000 (2473 deaths) All-cause annual deaths: Serfling model: 30.42/100,000 (76,321 deaths); >64 years: 164.1/100,000 (68,977 deaths) GLM Model: 10.97/100,000 (27,307 deaths); >64 years: 57.05/100,000 (23,560 deaths) Deaths/100,000 [R = respiratory deaths; C = Circulatory deaths; A = all-cause deaths] Serfling model: 50–64: R 2.30; C 2.87; A 5.63; 65–74: R 13.17, C 6.36, A 25.01; 75–84: R 31.18, C 52.93; A 87.92; ≥85: R 136.90, C 223.39, A 461.07 Poisson model: 50–64: R 2.85; C 1.79; A 7.91; 65–74: R 14.10, C 9.69, A 34.56; 75–84: R 50.63, C 64.80; A 138.06; ≥85: R 192.66, C 301.40, A 663.94
Rate-differences model: 4 models (estimates for weeks when ≥10% and ≥15% specimens tested positive for influenza; and peri-season when 96% of all influenza-associated deaths occurred among those ≥65. [The correlation of summer rate influenza-associated deaths to the Serfling rate was 0.63, to Poisson regression 0.86 and to ARIMA 0.78; Linear to Poisson regression 0.54; and Poisson regression to ARIMA 0.68]
Table 2 Studies which assessed the effects of confounders on the relationship of influenza and mortality. Data sources
Statistical methods
Other confounders controlled
Results
Charu 2013; USA; 1990–2010; 1 year occur in hospital); CDC RSV and influenza surveillance data
Hebert 2005; USA; 1987–99; Medicare patients (mostly ≥65)
Medicare Provider Analysis and Review files (provides hospital stays and date of death if within two years of admission); ICD-9-CM
Prediction of hospitalisation, and death within 30 days of admission
Age, gender, race-adjusted rates for hospitalisation; Charlson comorbidity index; indicators highly predictive of mortality for patients with pneumonia (respiratory neoplasms, metastatic neoplasms, chronic lung diseases, septicemia); evidence of hospital upcoding to diagnoses that generate more revenue
Mytton 2012; England;
