A Practical Guide for Designing and Conducting Influenza Disease Burden Studies

8/12/2019 A Practical Guide for Designing and Conducting Influenza Disease Burden Studies

1/52


2/52

Start1

iA Practical Guide for Designing and Conducting In uenza Disease Burden Studies

A Practical Guide forDesigning and Conducting

In uenza Disease Burden Studies

November 2008


3/52


4/52

TOC

iA Practical Guide for Designing and Conducting In uenza Disease Burden Studies

Table of Contents

Abbreviations and Acronyms .....................................................................................................................

Introduction ....................................................................................................................................................

Section 1: Studies Directly Measuring The Disease Burden Of In uenza ......................................1.1. Indicator-based Disease Burden Studies.................................................................................................... 1.1.1 Terms and De nitions ............................................................................................................... 1.1.2 Case De nitions ......................................................................................................................... 1.1.3 Study Population......................................................................................................................... 1.1.3.1 Hospitalized Acute Respiratory Infections (sARI)...................................................... 1.1.3.2 Outpatients With ILI.................................................................................................. 1.1.4 Specimen Collection and Handling............................................................................................. 1.1.5 Laboratory Diagnosis.................................................................................................................. 1.1.6 Data Analysis: Simple Proportions and Incidence Rates ........................................................... 1.1.6.1 Proportions ............................................................................................................... 1.1.6.2 Incidence Rates ........................................................................................................ 1.1.6.3 Individual Level Versus Population Level Data ......................................................... 1.1.7 Essential Indicators For In uenza Disease Burden Measurement............................................. 1.1.7.1 Hospitalized Sari Due To Laboratory-con rmed In uenza Infection.......................... 1.1.7.2 In uenza Managed In The Outpatient Setting...........................................................1.2. Absenteeism As A Measure Of Disease Burden........................................................................................ 1.2.1 Absenteeism In Schools ............................................................................................................ 1.2.1.1 Methods and Materials............................................................................................. 1.2.1.2 Data Storage and Analysis ....................................................................................... 1.2.1.3 Monitoring................................................................................................................. 1.2.1.4 Ethical Considerations ..............................................................................................

1.2.2 Absenteeism In Workplaces....................................................................................................... 1.2.2.1 Methods and Materials............................................................................................. 1.2.2.2 Data Storage and Analysis ....................................................................................... 1.2.2.3 Monitoring................................................................................................................. 1.2.2.4 Ethical Considerations ..............................................................................................

Section 2: Studies Indirectly Measuring The Disease Burden Of In uenza ....................................2.1. Estimating Disease Burden Of In uenza Through Mathematical Modelling............................................... 2.1.1 Introduction................................................................................................................................ 2.1.2 Data Preparations ...................................................................................................................... 2.1.2.1 Health Outcome Data............................................................................................... 2.1.2.2 In uenza Virus Laboratory Surveillance Data ........................................................... 2.1.2.3 Data On Other Covariates ........................................................................................

iii

1

333444456667777888891010

1010111212

13131314141415


5/52

TOC

ii

2.1.3 Statistical Methods In Disease Burden Of In uenza Estimation ................................................ 2.1.3.1 Summary For Different Approaches and

Data Requirement For Disease Burden Estimations.............................................

References ......................................................................................................................................................

Annexes ..........................................................................................................................................................Annex 1: Consultants For The Protocols Of In uenza Disease Burden Studies ...............................................Annex 2: Procedures For Specimen Collection..................................................................................................Annex 3: Key Clinical Data and Investigational Findings ...................................................................................Annex 4: Health Care Centre Recording Form (School Absenteeism)...............................................................Annex 5: Student Absentee Reporting Form ....................................................................................................Annex 6: Health Care Centre Recording Form (Workplace Absenteeism) .......................................................Annex 7: Employee Absentee Reporting Form.................................................................................................Annex 8: Approach 1 - Risk Difference Method................................................................................................Annex 9: Approaches 2 and 3 - Ser ings Method and Robust Regression Method........................................Annex 10: Approach 4 - Estimating In uenza Associated Excess Deaths Using

Autoregressive Integrated Moving Average (Arima) Method .........................................................Annex 11: Approach 5 - Stochastic Frontier Estimation Method .......................................................................Annex 12: Approach 6 - Poisson Regression Method........................................................................................Annex 13: Template For Data To Be Collected Using Approach 6 For Estimation

Of In uenza Disease Burden ..........................................................................................................

15

16

19

23232527282930313233

354143

45


6/52

TOC

iiiA Practical Guide for Designing and Conducting In uenza Disease Burden Studies

APACI Asia-Paci c Advisory Committee on In uenzaAPSED Asia Paci c Strategy for Emerging DiseasesCDC Centers for Disease Control and PreventionIHR International Health RegulationsILI In uenza-like IllnessMOH Ministry of HealthNIC National In uenza CenterNIID National Institute of Infectious Diseases

RSV Respiratory Syncytial VirusRT-PCR Reverse transcriptase polymerase chain reactionsARI Severe Acute Respiratory InfectionSEARO South-East Asia Regional Of ce of WHOWHO World Health OrganizationWPRO Western Paci c Regional Of ce of WHO

Abbreviations and Acronyms


7/52


8/52


9/52

Introduction

2

consultation were to: 1) develop generic protocols for in uenza disease burden studies; 2) reviewand introduce mathematical modelling methods for in uenza disease burden studies; and 3) draft aguide for data collection and preparation for in uenza disease burden studies through mathematicalmodelling. Eleven experts from eight countries participated in this consultation. The Asia-Paci cAdvisory Committee on In uenza (APACI) was also represented during this meeting. This document

was developed by the expert consultants of this meeting to provide speci c guidance for theplanning and operation of in uenza disease burden studies.


10/52

Section 1: Studies Directly Measuring The Disease Burden of In uenza

3A Practical Guide for Designing and Conducting In uenza Disease Burden Studies

1.1. INDICATOR-BASED DISEASE BURDEN STUDIES

1.1.1 Terms and De nitions

Disease burden: Morbidity and mortality due to in uenza infection and the related costs to thehealth care system and to society (e.g. loss of employment, missed schooling).

Eligible patient: A patient who meets the ILI or sARI case de nition and can be enrolled in thestudy.

Enrolled patient: A patient who meets the ILI or sARI case de nition and has provided informedconsent and respiratory specimens.

Informed consent: Agreement by the patient to participate in the study after receiving a thoroughexplanation of the studys purpose, risks and bene ts communicated in a manner understandableto the patient.

In uenza patient: A person with laboratory-con rmed in uenza infection.

Incidence rate: The number of new in uenza cases per unit of person-time at risk, often expressedas n cases / 100 000 persons/ year

Respiratory specimen: Respiratory secretions obtained speci cally for virus detection.

1.1.2 Case De nitions

In uenza-like Illness (ILI): A person with sudden onset of fever of > 38C and cough or sore throatin the absence of other diagnoses.

Severe Acute Respiratory Infection (sARI) 1 : Meet ILI case de nition (Sudden onset of fever over38C and cough or sore throat in the absence of other diagnosis),AND Shortness of breath or dif culty breathing, and Requiring hospital admission

Section 1:Studies Directly Measuring TheDisease Burden of In uenza

1 Also referred to as Acute Lower Respiratory Tract Infection (aLRTI)


11/52


4

1.1.3 Study Population

The study population should be representative of the wider population at risk for in uenzainfection. On balance, patients meeting the study case de nition who are enrolled in the study andprovide informed consent and respiratory specimens should re ect their communities in terms ofage, gender, ethnicity and socioeconomic status.

1.1.3.1 Hospitalized sARI patients

A retrospective analysis of patients admitted to hospitals in the previous year with severe acuterespiratory infection should be carried out in order to estimate the number of patients that couldmeet the case de nition during the prospective study. Depending on funding and staf ng, effortsshould be made to obtain informed consent and respiratory specimens from all patients who meetthe sARI case de nition. If this is not feasible, either due to large numbers of eligible patients or alimited study budget, a systematic sampling strategy should be instituted allowing for selection ofevery nth eligible patient for taking respiratory specimens, but all of the patients who meet the sARI

case de nition in a speci ed period (week or month) should be recorded. This sampling strategywill reduce the potential for bias in the selection of study patients.

Example:

Step 1: Review a random subset of hospital discharge diagnoses from various months during thepreceding year.

Step 2: Determine the proportion of hospital patients from the previous year that would have metthe sARI case de nition.

Step 3: Determine the number of patients and laboratory tests that the study budget will allow.Recognize that not all patients meeting the sARI case de nition will be enrolled for various reasonsincluding refusal to provide informed consent.

Step 4: If the budget does not allow for study participation from at least 50% of the estimatednumber of potentially eligible patients, design a random sampling plan that will attempt to enrolone out of every two eligible patients. Similarly, if the budget and staf ng will only allow for testingof specimens from 25% of the expected eligible patient population, plan to systematically enrolevery fourth hospitalized patient meeting the sARI case de nition.

1.1.3.2 Outpatients with ILI

For studies in outpatients, general practice or polyclinics are preferable as they are morelikely to represent the wider community in terms of age, gender and socioeconomic status thanpatients visiting emergency departments or other specialty clinics. Ideally, all patients who meetthe case de nition should be enrolled during randomly selected, complete/entire clinic days to allowfor the collection of the total number of outpatient visits (denominator) as well as the total numbermeeting the ILI case de nition. The proportion of patients meeting the WHO case de nition for ILIwill vary according to local seasonal patterns of in uenza activity.

Example:

Step 1: Begin by randomly selecting one (or more) polyclinic days per week as study budget andstaf ng permits.


12/52



Step 2: Assign a research nurse or assistant to monitor the outpatient department intake deskto record the age, gender and reason for visit of all patients attending the clinic from opening toclosing time.

Step 3: Self-reporting of the presence of a fever must be veri ed with measurement.

Step 4: If the patient has a measured fever over 38C, determine if the patient also has either acough or a sore throat. If yes, the patient meets the WHO case de nition for ILI. Attempt to obtaininformed consent and respiratory swab specimens from the patient.

Step 5: At the end of the clinic day, record: (1) the total number of all outpatient department patientvisits, (2) the number of patients that met the WHO ILI case de nition, (3) the number of patientsthat enrolled in the study. The laboratory results of this cohort should be followed, recording thenumber of patients with positive test results.

1.1.4 Specimen Collection and Handling

Diagnosis of infection with in uenza relies on the collection of high quality specimens, theirrapid transport to the laboratory and appropriate storage before laboratory testing (11). In uenzais best detected in specimens containing infected cells and secretions collected during the rstthree days after the onset of clinical symptoms. Therefore, every effort should be made to collectspecimens as close to the onset of illness as possible. The time between the onset of illness andspecimen collection should be recorded.

A variety of specimens are suitable, including: Nasal swab Throat swab Nasopharyngeal swab Nasopharyngeal aspirates or washes Nasal wash

Nasopharyngeal swabs, aspirates and washes are the best specimens for cell culture andPCR (12). However, these can be technically dif cult, require careful training to ensure propertechnique, and may be unpleasant for the patient. An acceptable alternative is to collect a nasal anda throat swab from the same patient and then combine these swabs into a single vial of transportmedium.

Failure to carefully handle respiratory swab specimens can signi cantly reduce their usefulness.It is essential that swab specimens be carefully and consistently managed. Respiratory specimens

should be collected, transported and stored in virus transport medium. Swab specimens should bepromptly transported to the laboratory and not exposed to heat or allowed to dry. The specimenscan be stored at 2-8C for up to 72 hours prior to processing by PCR or cell culture. Specimensfor direct detection of viral antigens by immuno uorescence staining of infected cells should berefrigerated and processed within two days. Specimens for use with commercial rapid in uenzatesting kits should be handled and stored in accordance with the manufacturers instructions.

Each specimen should be divided into aliquots for additional testing, re-testing or archivingprior to freezing at -70C for long-term storage. Heating and repeat freeze-thawing of specimensgreatly affect the quality and diagnostic yield, therefore should be avoided. Regular training andcareful supervision of staff responsible for specimen collection and storage is essential to ensure

high-quality specimens. Appropriate infection control and biosafety precautions (13, 14) must be


13/52


14/52



To calculate an incidence rate, the size of the population at risk must be known. This canbe dif cult when studying patient populations that can access health care services at multiplesites. Studies that are designed to examine speci c populations whose census is known arereferred to as population-based studies. In the absence of such information, certain adjustmentsand assumptions are necessary in order to estimate incidence rates. In such cases, advanced

epidemiological assistance should be sought and the methods used to estimate incidence shouldbe clearly and completely described in subsequent reports or publications.

1.1.6.3 Individual level versus population level data

Disease burden studies typically report data that describe the population at risk. For example,the age distribution, socioeconomic status, gender and ethnicities of the population under study.The WHO Global Agenda on In uenza Surveillance and Control (10) also calls for evaluation ofthe clinical aspects of in uenza infection. Reporting such patient-level clinical information can beuseful for clinicians providing care for in uenza patients and may also help to improve outcomesfor patients with serious complications. Therefore, wherever possible, disease burden researchers

should also collect and report individual level data describing the clinical presentation of in uenzainfection, underlying risk factors, medical interventions andthe outcome of in uenza infection. Examples of patient level data that could be collected areprovided in Annex 3.

1.1.7 Indicators for In uenza Disease Burden Measurement

1.1.7.1 Hospitalized sARI due to laboratory-con rmed in uenza infection

Disease burden:

Essential data: Number of sARI cases per week, month, year Proportion of all sARI cases caused by in uenza per week, month, year Distribution by age groups (0-2 years, 3-4 years, 5-17 years, 18-49 years,

50-64 years, 65 years) Demographic data of study population (gender, residence, ethnicity) Clinical data (history, symptom presentation, discharge status)

Desirable data: If catchment population is known or can be estimated, incidence per 100 000 persons per

year Indicators of severity and outcome such as need for oxygen therapy, treatment in Intensive

Care Unit and endotracheal intubation Mean, median and range of length of hospital stay Clinical status 21 days post-discharge form hospital (dead or alive)

Socioeconomic burden data: Direct treatment costs Lost work and lost school days Out-of-pocket costs to patient and families


15/52


8

1.1.7.2 In uenza managed in the outpatient setting

Disease burden:

Essential data:

Total number of outpatient visits per week, month, year Proportion of outpatients with ILI per week, month, year Proportion of ILI patients with laboratory-con rmed in uenza infection Age distribution of ILI patients and laboratory-con rmed in uenza infections Seasonality: proportion of ILI cases with laboratory-con rmed in uenza infection reported by

week and month

Desirable data: If catchment population is known or can be estimated, incidence per 100 000 persons

per year Distribution by age groups (0-2 years, 3-4 years, 5-17 years, 18-49 years,

50- 64 years, 65 years) Clinical data, including history, symptoms at presentation, medical intervention

Socioeconomic burden data: Direct treatment costs Lost work/school days due to laboratory-con rmed in uenza infection Out-of-pocket costs to patient and families

1.2. ABSENTEEISM AS A MEASURE OF DISEASE BURDEN

Absenteeism at schools and workplaces can provide important information on the social andeconomic costs of in uenza. Research has shown that trends in absenteeism often correspond toseasonal variations in in uenza activity (16).

1.2.1 Absenteeism in Schools

As school-aged children are at high risk for in uenza infection, data on absenteeism dueto in uenza-like illness (ILI) from schools may provide useful information on the seasonality,burden and social costs of in uenza. Individual schools may be selected to act as sentinels andfollowed for one or more school years to determine the frequency of ILI-associated absenteeism.If laboratory diagnosis is available, an improved understanding of the burden and seasonality oflaboratorycon rmed in uenza infection in the general community may also be gained by the studyof school-age populations.

1.2.1.1 Methods and Materials

Study design:Prospective, longitudinal, observational study

Study population:All children attending selected schools in the survey area.Selection of schools:Schools should be randomly selected from among all schools in a speci ed administrative area.

Data to be collected:Number of children under surveillance


16/52



Number of absence events/episodesNumber of missed school daysNumber of absence events associated with an ILINumber of missed school days associated with an ILINumber of absence events associated with laboratory-con rmed in uenza

Number of lost school days due to laboratory-con rmed in uenza infection

Survey procedure:1. Selected schools keep daily attendance books throughout the survey time period.2. All school absences are reported to the study coordinator on a daily basis

(Annex 4).3. The school nurse or designated study staff then contacts the childs family to inquire if the

absence is related to illness. If yes, the students parents will complete a reporting form (Annex5). The reporting form details symptoms of in uenza-like illness (ILI), such as fever, cough orsore throat as well as any visits to a medical provider. ILI-associated absences are recordedand tallied.

4. If the study has suf cient nancial resources and access to a quali ed diagnostic laboratory,children with an ILI-associated absence may be tested for the presence of an in uenza virusinfection. In this case, a health care worker from the study will visit the ill child in their home tocollect clinical swab specimens for laboratory diagnosis. If available, a rapid in uenza test maybe used to diagnose in uenza infection during the home visit.

1.2.1.2 Data Storage and Analysis

An ef cient system for storage and management of data should be established prior to thecommencement of the survey. Weekly tallies of the number of ILI-associated absences can beused as an indicator of increased in uenza activity.

There are three levels of potential data:1. Total (All-cause) Absence Episodes2. Total ILI-Associated Absence Episodes3. Total Laboratory-con rmed In uenza Infection Absence Episodes

Data analysis:

1. Calculate the proportion of absentees by:

Total number of students absent at the school during the week

Total number of students attending the school on the rst day of the week

If a student is absent across more than one week, he/she contributes to the numerator for bothweeks

2. Attempt to identify the commencement of the in uenza season by increases inabsenteeism.

3. Estimate in uenza burden by comparing the proportion of absentees during the in uenzaseason with those during the non-in uenza season (excess in uenza absenteeism).

4. Estimate the relative in uenza burden by calculating the proportion of absenteeism due to

in uenza, based on results of in uenza diagnostic laboratory tests.


17/52


10

1.2.1.3 Monitoring

An ongoing comparison of school ILI-associated absenteeism with the reported incidence ofin uenza from the national surveillance system will allow an assessment of the reliability of thesurvey results.

1.2.1.4 Ethical Considerations

This type of study will, in most cases, require a formal protocol and clearance from the relevantInstitutional Review Board(s).The coordinators of the survey must protect the privacy of individualswhen reporting and publishing data. Before commencing such a survey, the coordinators mustthoroughly explain the purpose, protocols, and material of the survey to the school administrationand to parents. Written consent should be obtained, and signed and dated by both the schoolcaretaker and the study coordinator. The titles, names, and contact information of the signatoriesshould be clearly provided. The original copies of the document are led at the school and at the

local health care centre. Informed consent must also be obtained from a parent or legal guardian ofthe absentee.

1.2.2 Absenteeism in Workplaces

Selected workplaces are identi ed to act as sentinels and monitored prospectively for absencesdue to ILI. People of working age are considered to be at lower risk for in uenza morbidity and itscomplications than other age groups. Workers may continue to work while ill or return to work beforethey are fully recovered. Consequently, workplace absenteeism survey data may underestimatethe true burden.

1.2.2.1 Methods and Materials

Study design:Prospective, longitudinal, observational study

Study population:All employees working in selected workplaces

Selection of workplaces:When possible, workplaces should be randomly selected from all workplaces within agiven region.

Data to be collected:Number of all-cause absentee episodesNumber of ILI-associated absentee episodesNumber of absentees episodes with laboratory-con rmed in uenzaCosts incurred due to medical assessment/intervention

Survey procedure:1. Selected workplaces keep daily attendance books throughout the survey year. The number

of absentees is recorded and reported to the study coordinating centre every day (Annex 6).


18/52



Employees absent from work provide a completed reporting form (Annex 7) at the time ofreturn. The reporting form details symptoms of in uenza-like illness (ILI), such as fever, coughor sore throat as well as any visits to a medical provider. ILI-associated absences are recordedand tallied.

2. If the study has suf cient nancial resources and access to a quali ed diagnostic laboratory,workers with an ILI-associated absence may be tested for the presence of an in uenza virusinfection. In this case, a health care worker from the study will visit the worker in their home tocollect clinical swab specimens for laboratory diagnosis. If available, a rapid in uenza test maybe used to diagnose in uenza infection.

3. Workplaces should calculate employees medical costs associated with these illnesses.

1.2.2.2 Data Storage and Analysis

An ef cient system for storage and management of data should be established prior to the

commencement of the survey. Weekly tallies of the number of ILI-associated absences can beused as an indicator of increased in uenza activity.

Data analysis:1. Calculate the proportion of absentees by:

Total number of employees absent at the workplace during the week

Total number of employees attending the workplace on the rst day of the week

If an employee is absent across more than one week, he/she contributes to the numerator forboth weeks.

2. Attempt to identify the commencement of the in uenza season by increases in absenteeism.

3. Estimate in uenza burden by comparing the proportion of absentees during the in uenzaseason with those during the non-in uenza season (excess in uenza absenteeism).

4. Estimate the relative in uenza burden by calculating the proportion of absenteeism due toin uenza, based on results of in uenza diagnostic tests.

5. Calculate excess medical expenses due to in uenza by comparing absentees medical expenseswith non-absentees medical expenses during in uenza season. It is important to recognize

that the proportion of people with underlying disease may be larger among absentees thanamong non-absentees. Thus, excess medical expenses due to in uenza can also be estimatedby comparing medical costs of absentees during the in uenza season with that during thenon-in uenza season.

1.2.2.3 Monitoring

An ongoing comparison of ILI-associated workplace absenteeism with the reported incidenceof in uenza from the national surveillance system will allow an assessment of the reliability of thesurvey results.


19/52


12

1.2.2.4 Ethical Considerations

This type of study will require a formal protocol and clearance from the relevant InstitutionalReview Board(s).The coordinators of the survey must protect the privacy of individuals whenreporting and publishing data. Before commencing such a survey, the coordinators must thoroughlyexplain the purpose, protocols, and material of the survey to the business administration and to theworkers. Written consent should be obtained, and signed and dated by both all workers. The titles,names, and contact information of the signatories should be clearly provided. The original copies ofthe documents are led at the study coordinating centre.


20/52


21/52

Section 2: Studies Indirectly Measuring The Disease Burden Of In uenza

14

regression method (Annex 9); robust regression method (Annex 9); ARIMA method (Annex 10);stochastic frontier estimation method (Annex 11); and Poisson regression method (Annexes 12- 13). Several of these methods can be applied in countries with both temperate and tropical/ subtropical climates, irrespective of the seasonality patterns.

Based on these approaches we propose a generic statistical approach in quantifying the healthburden of seasonal in uenza for member countries, and list out the epidemiological and viralsurveillance data which are normally required for such approaches.

2.1.2 Data Preparations

2.1.2.1 Health outcome data

The time series of the health outcomes are either weekly (more preferable) or monthly countsof all-cause deaths, or deaths from speci c underlying causes (e.g. due to cardio-respiratorydiseases, and pneumonia and in uenza), or the weekly or monthly counts of hospital discharges

with the main diagnoses of a certain underlying causes (e.g. due to cardio-respiratory diseases,and pneumonia and in uenza). The speci c causes are categorized using either the InternationalClassi cation of Disease, 9th Revision (ICD-9) or the International Classi cation of Disease, 10thRevision (ICD-10) as follows:

Underlying cause of death/hospitalization ICD-9 ICD-10All-cause All ICD codes All ICD codesCardio-respiratory 390-519 I00-I99, J00-J99Pneumonia and in uenza 480-487 J10J18

The sources of the mortality data are most likely derived from the national death registrationsystems for most member countries. There are usually less issues on completeness andcomparability of the mortality data among member countries.

However, for hospital data, representative and nation-wide data are usually not available orwould be problematic. In some member countries only government operated national registersare accessible but there may be substantial hospitalization undertaking in private healthcareinstitutions. In addition, the diagnosis may not be completely speci ed and coded for every hospitaldischarge.

The analysis is based on all ages. But if age information is available from the death/hospitalizationrecords, it would be desirable also to perform the analysis for strati ed sub-group who are at highrisk for in uenza infection, such as the 65 years of age or older group.

2.1.2.2 In uenza virus laboratory surveillance data

The best set of in uenza virus surveillance data would be obtained from national in uenzacentres or national epidemiology surveillance systems. It is because specimens from these centresor systems are usually better representative and good coverage of the countries or cities, and areconsistently collected and tested for in uenza viruses over time. The data aggregated into weekly


22/52



or monthly numbers of total respiratory specimens tested for in uenza virus, and numbers testedpositive for in uenza virus isolates are used to measure in uenza activities in de ning epidemicand non-epidemic periods as well as proportion of positive results (in uenza A or B types andsubtypes) in specimens tested for in uenza viruses.

2.1.2.3 Data on other covariates

Time series data on weekly or monthly mean temperature ( oC) and relative humidity (%) areused to control for the potential confounding effects of weather. Information of unusual events(e.g. heat or cold wave, strike in the health services, ood, earthquake, infectious disease outbreake.g. SARS, etc) will be collected whenever available. But for the control of other confoundingfactors such as seasonality and long-term time trends, and calendar and holiday effects, no specialcollection of the data is necessary. Since the respiratory syncytial virus (RSV) is known to co-circulate with in uenza viruses and affect morbidity and mortality among both young children andthe elderly (26-29). RSV data in numbers and proportion of specimens tested positive for RSV willalso be collected and included in the statistical models.

2.1.3 Statistical methods in disease burden of in uenza estimation

The quanti cation for the disease burden of seasonal in uenza has long been regarded as animportant issue but has not been fully developed particular for the tropical and subtropical regions.There are many approaches to disease burden modelling for in uenza. The rst method is the riskdifference method which relies on de nition of epidemic and non-epidemic (assumed to havelow or no in uenza viruses in the circulation as a baseline or control) periods so that the differencein observed mortality and morbidity between the two periods can be assessed to estimate theexcess impacts (27, 30). The second approach is the regression method, including the Ser ingmethod (20) and several adaptations of the Ser ing method (31-34). The third approach is the

robust regression method (35). The fourth approach is the classical ARIMA method (18, 36). Thefth approach is the stochastic frontier estimation method (37). Finally the sixth approach is thePoisson regression method (23-24, 28, 38-39).

The regression methods, robust regression method, ARIMA method, and stochastic frontierestimation method require well-de ned seasonal patterns of in uenza activities. However, in orderto overcome the problems arising from a lack of well-de ned seasonality patterns in the tropicsand subtropics, several alternative approaches have been identi ed. One approach is the simplerisk difference method which de nes periods based on available in uenza viral surveillance data(27, 40). Another approach is the Poisson regression model using viral surveillance data as thepredictors in the model and also including the seasonal terms in the model. In the temperateclimates, in uenza activities usually follow well-de nedseasonal patterns so that the epidemic periods with excess mortality and morbidity outcomes canbe recognized. But in the tropics and subtropics, estimation approaches become dif cult for thefollowing reasons:

i) There are no well-de ned seasonal in uenza patterns to allow estimation of the epidemic andnon-epidemic baseline periods because the in uenza virus can be circulating year round.

ii) A surveillance system with laboratory de ned in uenza activities data is either not available ornot complete.

iii) Population level mortality and morbidity databases are not available or not complete.


23/52


16

2.1.3.1 Summary for different approaches and data requirement for disease burdenestimations

The data requirement and the statistical methods for each of the approaches 16 are summarizedin the following table.

Table 1

Approach Data requirement Statistical methods

1. Risk DifferenceMethod

1) Weekly or monthly morbidityor mortality data2) Weekly or monthly viralsurveillance data to de neepidemic periods

De ned epidemic/ nonepidemicperiods.

2 Ser ing RegressionMethod

1) Weekly or monthly morbidityor mortality data2) Clear seasonal component

Ser ings method seasonal andexpected occurrence duringbaseline period and to identifythe threshold for in uenzaepidemic.

3. Robust RegressionMethod


Robust regression modelswhich do not require Gaussiandistribution and equal varianceassumptions, and are robust tothe presence of outliers in thehealth outcomes.

4. ARIMA Method 1) Weekly or monthly morbidityor mortality data2) Clear seasonal component

ARIMA models which areliable to over- tting the healthoutcomes.

5. Stochastic FrontierEstimationMethod


Stochastic frontier estimationmodel to identify excess healthoutcomes during periods withunexpected high level of healthoutcomes (in uenza data neededto validate the periods).

6. Poisson RegressionMethod

1) Weekly or monthly morbidityor mortality data2) Weekly or monthly viralsurveillance data

Poisson regression to modelthe time series patterns of thehealth outcomes due toseasonality and time varyingconfounding. Then the numbers

of health outcomes whenin uenza activities are assumedto be zero (E 0).The excess number =(observed number E 0).

The above Table 1 illustrates the statistical models and the underlying statistical methodsavailable to us for estimation of disease burden of in uenza. In Table 1 the data required in carryingout the estimation using different approaches are also speci ed. However in order to guide inthe selection of approach, a ow chart (Figure 1) is also provided. It is import to know whether

the required data for the estimation are available. If we have weekly or month data for the health


24/52



outcomes and for the virology activity measures, we can apply Poisson regression (Approach 6).If we only have virology data (to measure in uenza activities) and do not have long series of thehealth outcome data, we can de ne in uenza epidemics and baseline periods in each year andthen assess the rate difference between the two periods in each year, providing that we areaccessible to the disease incidence rates. But if we have long series of the health outcome data

but not virology data, we can apply either the Robust Regression model or the Stochastic FrontierEstimation model, depending on whether the data are from baseline years or not. If we knowwhich are the baseline years, we apply Robust Regression model or if not we apply the StochasticFrontier Estimation model.

Other approaches based on population surveys or based on intensive case nding for in uenzadiagnoses in health care institution can be applied. But this is beyond the scope of this guideline.Also to be stringent, validation of the Robust Regression model and Stochastic Frontier Estimationmodel could be performed rst before their recommendation for application in tropical and subtropicalregions. Finally disease burden may vary from place to place depending on the susceptibility ofthe population, and also from year to year depending on emerging of new or recurring strains.

Disease burden due to seasonal in uenza in the tropics and subtropics is largely unexplored, whichwarrants urgent investigation in order to raise the awareness and also preparedness of the public,the health professions and the vaccine production capacities during the inter-pandemic periods.

Figure 1: Flow chart of statistical approaches to be chosen for estimating disease burden ofin uenza

Health Data

5 years of weekly/ monthly data?

Viral data?

Denominator?

Approach 6

Approach 2-5

Viral Data?

Approach 6

Yes

Yes

Yes

Yes

Yes

No

No


25/52


18


26/52

References


References

1. WHO National In uenza Centres. http://www.who.int/csr/disease/in uenza/centres/en/

2. WHO Global In uenza Programme. Terms of Reference for National In uenza Centres.http://www.who.int/csr/disease/in uenza/en/TORNICs.pdf

3. Ryan J, et al., Establishing the health and economic impact of in uenza vaccination within theEuropean Union 25 countries. Vaccine , 2006. 24 (47-48): p. 6812-22

4. Poehling K, et al., The underrecognized burden of in uenza in young children.N Engl J Med , 2006. 355 (1): p. 31-40.

5. OBrien M, et al., Incidence of outpatient visits and hospitalizations related to in uenza ininfants and young children. Pediatrics , 2004. 113 (3 Pt 1): p. 585-93.

6. Harper, S.A., et al., Prevention and control of in uenza. Recommendations of the AdvisoryCommittee on Immunization Practices (ACIP). Morb Mortal Wkly Rep , 2005. 54 (RR-8):p. 1-40.

7. Fredrickson K., et al., In uenza vaccination coverage among children aged 6-23 months- siximmunization information system sentinel sites, United States, 2005-06 in uenza season.Morb Mortal Wkly Rep , 2006. 55 (49): p. 1329-30.

8. Naleway A, Smith W, Mullooly J. Delivering in uenza vaccine to pregnant women. EpidemiolRev, 2006. 28 : 47-53.

9. WHO Global Agenda on In uenza Surveillance and Control. Recommended SurveillanceStandardshttp://www.who.int/csr/resources/publications/surveillance/WHO_CDS_CSR_ISR_99_2_EN/ en/.

10. WHO Asia Paci c Strategy for Emerging Diseases South-East Asia Region, Western Paci cRegion, 2005.http://www.wpro.who.int/sites/csr/documents/APSED.htm

11. WHO guidelines for collecting, preserving and shipping specimens for the diagnosis of avianin uenza A(H5N1) virus infection

http://www.who.int/csr/resources/publications/surveillance/CDS_EPR_ARO_2006_1.pdf


27/52

References

20

12. Chana KH, Peiris JSM, Lim W, Nicholls JM, Chiu SS. Comparison of nasopharyngeal ockedswabs and aspirates for rapid diagnosis of respiratory viruses in children. J ournal of ClinicalVirology , 2008 in press

13. WHO laboratory biosafety guidelines for handling specimens suspected of containing avian

in uenza A virus, 2005. http://www.who.int/csr/disease/avian_in uenza/guidelines/handlingspecimens/en/index.html

14. WHO Laboratory Biosafety Manual 3rd Edition, 2004. http://www.who.int/csr/resources/publications/biosafety/Biosafety7.pdf

15. WHO recommendations on the use of rapid testing for in uenza diagnosis. http://www.who.int/csr/disease/avian_in uenza/guidelines/RapidTestIn uenza_web.pdf

16. Hirota Y, Takeshita S, Ide S, Kataoka K, Ohkubo A, Fukuyoshi S, Takahashi K, Hirohata T, KajiM. Various Factors Associated with the Manifestation of In uenza-like Illness.

Int J Epidemiol , 1992. 21(3): 574-782

17. Glezen WP. Emerging infections: pandemic in uenza. Epidemiol Rev 1996; 18(1):64-76.

18. Choi K, Thacker SB. Mortality during in uenza epidemics in the United States,1967-1978. Am J Public Health. 1982; 72(11):1280-3.

19. Nicholson KG. Impact of in uenza and respiratory syncytial virus on mortality in England andWales from January 1975 to December 1990. Epidemiol Infect 1996; 116: 5163.

20. Ser ing RE. Methods for current statistical analysis of excess pneumonia-in uenza deaths. Public Health Rep 1963; 78:494-506.

21. Simonsen L, Fukuda K, Schonberger LB, Cox NJ. The impact of in uenza epidemics onhospitalizations. J Infect Dis 2000; 181(3):831-837.

22. Wong CM, Chan KP, Hedley AJ, Peiris JSM. In uenza-associated mortality in Hong Kong. ClinInfect Dis 2004; 39(11):1611-1617.

23. Wong CM, Yang L, Chan KP, Leung GM, Chan KH, Guan Y et al. In uenza-associated hospitalization in a subtropical city. PLoS Med 2006; 3(4):e121.

24. Chow A, Ma S, Ling AE, Chew SK. In uenza-associated deaths in tropical Singapore.Emerg Infect Dis 2006; 12(1):114-121.

25. Viboud C, Alonso WJ, Simonsen L. In uenza in Tropical Regions. PLoS Med 2006; 3(4):e89.

26. Fleming DM, Cross KW. Respiratory syncytial virus or in uenza? Lancet. 1993; 342(8886-8887):1507-10.

27. Izurieta HS, Thompson WW, Kramarz P, et al. In uenza and the rates of hospitalization forrespiratory disease among infants and young children. N Engl J Med 2000;342:2329.


28/52

References


28. Thompson W, Shay DK, Weintraub E, Brammer L, Cox N, anderson LJ, Fukuda K. Mortalityassociated with in uenza and respiratory syncytial virus in the United States. JAMA 2003;289:17986.

29. Zambon MC, Stockton JD, Clewley JP, Fleming DM. Contribution of in uenza and respiratory

syncytial virus to community cases of in uenza-like illness: an observational study. Lancet2001; 358:14106.

30. Barker WH, Mullooly JP. Impact of epidemic type A in uenza in a de ned adult population. AmJ Epidemiol 1980; 112: 798-811.

31. Langmuir AD, Housworth J. A critical evaluation of in uenza surveillance. Bull World HealthOrgan. 1969;41(3):393-8.

32. Lui KJ, Kendal AP. Impact of in uenza epidemics on mortality in the United States from October1972 to May 1985. Am J Public Health. 1987;77(6):712-6.

33. Simonsen L, Clarke MJ, Williamson GD, Stroup DF, Arden NH, Schonberger LB. The impactof in uenza epidemics on mortality: introducing a severity index. Am J Public Health 1997;87:194450.

34. Simonsen L, Reichert TA, Viboud C, Blackwelder WC, Taylor RJ, Miller MA. Impact of in uenzavaccination on seasonal mortality in the US elderly population. Arch Intern Med 2005; 165:265-272.

35. andrew DF, Bickel PJ, Hampel FR, Huber PJ, Rogers WH and Tukey JW. 1972. RobustEstimates of Location: Survey and Advances. Princeton, N.J.: Princeton University Press.

36. Choi K, Thacker SB. An evaluation of in uenza mortality surveillance. 19621979. I. Time seriesforecasts of expected pneumonia and in uenza deaths. Am J Epidemiol 1981; 113:215-26.

37. Ohkusa Y, Shigematsu M, Taniguchi K, Okabe N. NIID model for estimation of excess mortalityin 2002/2003 season (in Japanese). Infectious Agent Surveillance Report 2003; 24(11):

288-289.

38. Thompson WW, Shay DK, Weintraub E, Brammer L, Bridges CB, Cox NJ, Fukuda K In uenza-associated hospitalizations in the United States. JAMA 2004; 292: 1333-1340.

39. Wong CM, Chan KP and Peiris J.S.M., The impact of in uenza on mortality in the Hong Kong

SAR, 1st European In uenza Conference, St Julians, Malta, 20-23 October 2002 . (PublicationNo. : 74910)

40. Chiu SS, Lau YL, Chan KH, Wong WH, Peiris JS. In uenza-related hospitalizations amongchildren in Hong Kong. N Engl J Med 2002; 347(26):2097-2103.


29/52


30/52

Annex 1


Professor Robert BOOYNational Centre for Immunization Researchand SurveillanceThe University of Sydney and The ChildrensHospital at WestmeadKerry Packer Institute for Child Health ResearchLocked Bag 4001Westmead, NSW 2145, Australia

Professor Paul CHANDepartment of MicrobiologyThe Chinese University of Hong Kong1/F Clinical Science BuildingPrince of Wales HospitalShatin, New TerritoriesHong Kong

Dr Yoshio HIROTAProfessor and ChairmanDepartment of Public HealthOsaka City UniversityGraduate School of MedicineOsaka, 545 8585 Japan

Dr Lance JENNINGSClinical VirologistCanterbury Health LaboratoriesP.O. Box 151Christchurch, New Zealand

Dr Stefan MAActing Deputy Director(Biostatistics and Research)Epidemiology and Disease Control DivisionMinistry of Health, Singapore

Dr Satoko OHFUJIResearch AssociateDepartment of Public HealthOsaka City UniversityGraduate School of MedicineOsaka, Japan

AnnexesANNEX 1

Informal Consultation on Protocols of In uenza Disease Burden StudiesManila, Philippines

12-13 November, 2007

Temporary Advisors


31/52

Annex 1

24

Dr Yasushi OHKUSAInfectious Disease Surveillance CenterNational Institute of Infectious Diseases1-23-1 Toyama ShinjukuTokyo, Japan

Dr David SHAYEpidemiologist and Team LeaderEpidemiolgy and surveillance BranchIn uenza DivisionCenters for Disease Control and Prevention1600 Clifton RoadAtlanta, GA 30333, USA

Dr Mark SIMMERMANEpidemiologist, In uenza Division

Coordinating Centre for Infectious DiseasesCenters for Disease Control and PreventionBangkok, [email protected]

Dr Sirenda VONGChief, Epidemiology and Public Health UnitPasteur InstituteP.O. Box 983, 5 Bldg MonivongPhnom Penh, Cambodia

Dr Chit-Ming WONGAssociate ProfessorThe University of Hong KongDepartment of Community Medicine5/F William MW Mong Block21 Sassoon RoadPokfulam, Hong Kong

Dr Takeshi KASAIRegional AdviserCommunicable Disease Surveillance and ResponseWorld Health Organization

Regional Of ce for the Western Paci cP.O. Box 29321000 Manila, Philippines

Dr Weigong ZHOUMedical Of cerCommunicable Disease Surveillance and ResponseWorld Health OrganizationRegional Of ce for the Western Paci cP.O. Box 29321000 Manila, Philippines

WHO SecretariatWHO/WPRO


32/52

Annex 2


ANNEX 2

Procedures for specimen collection

Materials required Sputum/mucus trap Polyester bre-tipped applicator (Swabs with a wooden shaft should not be used as they may

carry inhibitors for PCR and cell culture) Plastic vials Tongue depressor 15mL conical centrifuge tubes Specimen collection cup or Petri dishes Transfer pipettes Virus transport medium

Virus transportation medium for use in collecting throat and nasal swabs

1. Add 10g veal infusion broth and 2g bovine albumin fraction V to sterile distilled water (to makea total volume 400mL)

2. Add 0.8mL gentamicin sulfate solution (50mg/ml) and 3.2mL amphotericin B (250 g/ml) 3. Sterilize by ltration

Methods of collection

Clinical specimens should be collected as described below and added to transport medium. Nasal or

nasopharyngeal swabs can be combined in the same vial of virus transport medium. The followinginformation should be recorded on the Field Data Collection Form: general patient information, typeof specimen(s), date of collection, contact information of person completing the form.

Standard precautions including barrier protection should always be followed during sampling.

Nasal swab A dry polyester swab is inserted into the nostril, parallel to the palate, and left in place for a fewseconds. It is then slowly withdrawn with a rotating motion. Specimens from both nostrils areobtained with the same swab. The tip of the swab is put into a plastic vial containing 2-3mL virustransport medium and the applicator stick is broken off.

Nasopharyngeal swab A exible, ne-shafted polyester swab is inserted via the nostril into the nasopharynx and left inplace for a few seconds. It is then slowly withdrawn with a rotating motion. A second swab shouldbe used for the second nostril. The tip of the swab is put into a vial containing 2-3mL virus transportmedium and the shaft cut.

Nasopharyngeal aspirate Nasopharyngeal secretions are aspirated through a catheter connected to a mucus trap and ttedto a vacuum source. The catheter is inserted into the nostril parallel to the palate. The vacuum isapplied and the catheter is slowly withdrawn with a rotating motion. Mucus from the second nostrilis collected with the same catheter in a similar manner. After mucus has been collected from both

nostrils, the catheter is ushed with 3mL virus transport medium.


33/52

Annex 2

26

Nasal wash The patient sits in a comfortable position with the head slightly tilted backward and is advisedto keep the pharynx closed by saying K while the washing uid (usually 0.9% sterile saline) isapplied into the nostril. With a transfer pipette, 1-1.5mL washing uid is instilled into one nostril ata time. The patient then tilts the head forward and lets the washing uid ow into a specimen cupor a Petri dish. The process is repeated with alternate nostrils until a total of 10-15mL washing uidhas been used. Dilute approximately 3mL washing uid 1:2 in virus transport medium.

Throat swab The posterior pharynx is swabbed vigorously, and the swab is placed in virus transport medium asdescribed for nasal swabs.


34/52


35/52

Annex 4

28

A n n e x

4 :

H e a

l t h C a r e

C e n t e r

R e c o r d

i n g

F o r m

( S c

h o o

l A b s e n t e e i s m

)

D a t e o f I n f o r m a t i o n

:___

/___

/_______

( d d / m m

/ y y )

N a m

e o f S c h o o l :

S c h o o l C l o s e

: Y e s

_____

N o

_____

G r a d e 1

G r a d e 2

G r a d e 3

G r a d e 4

G r a d e 5

G r a d e 6

N u m

b e r o f a b s e n t e e s

M

a l e

F e m

a l e

M a l e

F e m a l

e

M a l e

F e m

a l e

M a l e

F e m

a l e

M a l e

F e m

a l e

M a l e

F e m

a l e

N u m

b e r o f a b s e n t e e s f r o m

f e b r i l e d i s e a s e

N u m

b e r o f l a b o r a t o r y

c o n r m

e d i n u e n z a

N u m b e r o f A H 1 i n u e n z a

N u m b e r o f A H 3 i n u e n z a

N u m b e r o f B i n u e n z a


36/52

Annex 5


Annex 5:

Student Absentee Reporting Form

Name of school :

Grade :

Name :

Periods of absenteeism : from __/__/__ (dd/mm/yy) to __/__/__ (dd/mm/yy)

Consultation in a clinic ( yes no )

Symptoms during the period of absenteeism.

1. Max temperature : degrees

2. Nasal discharge ; ( presence absence )

3. Cough : ( presence absence )

4. Sputum : ( presence absence )

5. Sore throat : ( presence absence )

6. Joint pain : ( presence absence )

Diagnosis : In uenza : others ( )

Rapid diagnostic test : done ( A B negative ) not done

Pneumonia : presence absence

Encephalopathy : presence absence

Otitis Media : presence absence


37/52

Annex 6

30

Annex 6:

Health Care Center Recording Form (Workplace Absenteeism)

Date of Information : ___/___/_______ (dd/mm/yy)

Name of work place :

Age < 20 years old 20 - 29 30 - 39 40 - 49 50 - 64 65+

Sex Male Female Male Female Male Female Male Female Male Female Male Female

Number of

absentees


38/52

Annex 7


Annex 7:

Employee Absentee Reporting Form

Name of workplace :

Unit / Department :

Name :

Periods of absenteeism : from __/__/__ (dd/mm/yy) to __/__/__ (dd/mm/yy)

Consultation in a clinic ( yes no )

Symptoms during the period of absenteeism.

1. Max temperature : degrees

2. Nasal discharge : ( presence absence )

3. Cough : ( presence absence )

4. Sputum : ( presence absence )

5. Sore throat : ( presence absence )

6. Joint pain : ( presence absence )

Diagnosis : In uenza : others ( )

Rapid diagnostic test : done ( A B negative ) not done

Pneumonia : presence absence


39/52

Annex 8

32

Annex 8:

Approach 1 - Risk Difference Method

Risk Difference method which is simple and require fewer assumptions, has been used formany years to estimate in uenza-associated morbidity and mortality. In this model, the averagenumbers of deaths or hospital admissions during the months assumed to have low or no in uenzavirus circulation (baseline periods) are de ned, followed by calculation of the excess mortality orhospitalization by subtracting this the numbers during the baseline periods from the observednumbers of deaths or hospital admissions during in uenza epidemics. These methods oftenmake use of viral surveillance data, but only to establish periods of in uenza epidemic or baselineperiods.

In one innovative application of this method in the subtropical region (Hong Kong) has beenadopted by Chiu et al (26), in estimation for the burden of in uenza on hospitalizations among

children 15 years of age or younger. In her method she de ned in uenza predominance periods tobe the in uenza epidemic periods in which RSV activities are low, and special baseline to be periodswhich both in uenza and RSV activities are low. She then demonstrated that there was a signi cantexcess rate of hospitalization for acute respiratory disease during the weeks when in uenza waspredominant, as compared with weeks when neither in uenza nor RSV had substantial activities.


40/52


41/52

Annex 9

34

The method is well implemented in temperate countries where there are well-established andclear seasonal patterns of in uenza, however it is not the case in sub-tropic and tropic countrieswhere the seasonality of in uenza is not that well de ned.

Nevertheless, this model has been extended by US CDC to a robust estimation method (namelyrobust regression method) using iteration of weighted residuals as follows: at rst, it applies

for all data by OLS where f (.) stands for some transformation if as necessary. By using theestimated residual , the weighted function in the next step is de ned as

where c is some positive constant. The second step and following step are weighted regressionfor

using weights Wt i . The weight for the next step i +1 is

where is the estimated residual in the ith step. Then we continue this process until convergence.The procedure is known as andrews weighting (see reference 35 for detail).

represents clear regular seasonality, and thus it may be inappropriateor tropical and subtropical areas. Additional higher term of trigonometric curve,

i.e. or some other weekly dummies like those speci ed under stochasticfrontier estimation model as in the Annex 11 below may be necessary.

f(Y t ) = 00 + 1

0 + 20 + 1

0 sin( 2t ) + 20 cos( 2t ) + t

0 (1) 52

52

t 0

sin t 0 /(c ))

W t 1 = f t 0 / (c )


42/52

Annex 10


Annex 10

Approach 4 - Estimating In uenza Associated Excess Deathsusing Autoregressive Integrated Moving Average (ARIMA) Method

Authors: Dhankhar P, PhD, Meltzer MI, PhD

1.1. Autoregressive Integrated Moving Average (ARIMA) Model: The ARIMA model is used to forecast using a time series data. A time series is a group ofobservations indexed by time t . For example, weekly deaths from in uenza is a time series. TheARIMA model uses past values of a time series to forecast future values. This is the main bene tof using ARIMA as we do not need information on any other covariates. For example, to forecastdeaths from in uenza, we just need data on weekly deaths from in uenza. We do not need anyadditional data such as viral surveillance data. The main components of an ARIMA model are theautoregressive model, moving average model, and the level of differencing needed to make the

time series stationary (to be described in 1.1c).

1.1. a. Autoregressive (AR) Model: In the autoregressive model, the value of the time series in the present period is a proportionof values of the time series in the past. The simplest AR model is the rst order autoregressiveprocess, denoted as AR(1). In AR(1) model, number of deaths in the current period are modelled asa proportion of deaths in the immediately prior time period plus a random error. We t the modelon the weekly deaths data to estimate this proportion. Once we estimate this proportion, theforecasted deaths are a proportion of deaths in the previous time period.

We can also increase the number of past time periods used in the AR model. In the secondorder autoregressive model (AR(2) model), the number of deaths in the current time period aremodelled as a proportion of deaths in the two previous time periods plus a random error. We canestimate these proportions by tting the model on the data (the proportions can be different forthe two time periods). The forecasted deaths, from the AR(2) model, are a sum of proportion ofdeaths in the two last periods. The order of an AR model corresponds to the number of previoustime periods used in the model. For example, an autoregressive model of order p , written as AR(p),contains p previous time periods.

1.1. b. Moving Average (MA) Model: In a moving average model, we base the deaths in the current period on the random error(shock) in the past. In the rst order moving average model, denoted as MA(1), number of deathsin the current period are modeled as a proportion of random error in the last period plus the randomerror in the current period. The forecasted deaths in the MA(1) model are a proportion of therandom error in the last period. The order of a moving average model corresponds to the numberof previous random errors used in the model. For example, in a moving average of second order,denoted as MA(2), we use random errors in the last two periods to model deaths in the currentperiod. In a MA(2) model, deaths in the current period are modelled as a proportion of randomerrors in the last two time periods plus a random error in the current period. A moving averagemodel of order q , models deaths in the current period as a proportion of random errors in the qprevious time periods and the random error in the current period.

1.1. c. Autoregressive Moving Average (ARMA) Model: The autoregressive moving average model, written as ARMA ( p, q ), combines both the

autoregressive model and the moving average model, where p is the order of autoregressive


43/52

Annex 10

36

part and q is the order of moving average part. For example, in a ARMA(1,1) model deaths in thecurrent time period are modelled as a proportion of deaths and random error in the last time periodplus a random error in the current period. ARMA model can be used only for time series that arestationary. A time series is stationary if it has no trend and the variance is constant across time. Ifthe time series is not stationary we use the ARIMA model.

1.1. d. Autoregressive Integrated Moving Average (ARIMA) Model: If the time series is non-stationary, we can take rst differences to make it stationary.Differencing a series means subtracting each observation from the subsequent observation. Ifthe series is still not stationary we can difference it once more (take second differences). Oncewe have a stationary series we can apply the ARMA( p, q ) model. The ARIMA model is written asARIMA(p, d, q ), where p is the order of autoregressive part, d is the level of differencing needed tomake the series stationary, and q is the order of moving average part.

1.2. Seasonality: Some time series show a seasonal pattern. For example, a time series of weekly deaths from

in uenza might show a seasonal pattern, where every year the deaths peak in winter and are at aminimum in summer. There are two ways to model seasonality in the ARIMA process. If the timeseries shows a strong seasonal pattern, we can difference the series with respect to the seasonallag. For example, if we use the time series of weekly in uenza deaths, we can take a 52 weekdifference and then apply ARMA model. We can also model seasonality as a multiplicative term inthe ARIMA model if the series shows signi cant correlation at seasonal lags.

1.3. Box and Jenkins (1976) (1) four step procedure to forecast using the ARIMA process:1) Transform the data if necessary to meet the assumption of covariance stationarity. If the series

is non stationary, we can difference it to make it stationary. We may also need to take aseasonal difference.

2) Once we have a stationary time series we use the autocorrelation function (ACF), inverseautocorrelation function (IACF), and the partial autocorrelation function (PACF) to identify thetime series. Identi cation of a time series means looking at the sample ACF, IACF, and PACFto make an initial guess about the p and q values of the ARMA ( p, q ) model.

3) Once we have identi ed the model, we estimate the parameters. Mostly we use a maximumlikelihood estimation procedure.

4) After estimating the model, we should also perform diagnostic checking to con rm theappropriateness of the model selected. We should check that the parameters are signi cantand the errors do not have a pattern. We can either plot the errors to look for a pattern or usea test statistic. The test statistic we used was the Ljung modi cation of the Box-Pierce Q statistic (ref Ljung and Box, 1978). The model is a good t if the errors have no correlation andthe p value of the test statistic is large.

1.4. Calculating Excess Deaths: The analysis consists of two steps. We rst calculated excess deaths for the rst two seasonsby constructing a Fourier series equation. We used Fourier series for the rst two seasons becauseof two reasons. Firstly, to model a seasonal time series we need at least two seasons of data.Secondly, Choi and Thacker (2) also used a Fourier series to model rst two seasons of data.

1.4.1 Fourier Method Steps:1) To identify epidemic weeks for the rst in uenza season, we calculated mean and standard

deviation for the time series from week 27 of the rst season to week 42 of the next season.The epidemic weeks for the rst season were de ned as any two consecutive weeks with

deaths greater than two standard deviations from the mean.


44/52

Annex 10


2) Model tting: We used Fourier analysis to produce an equation statistically describing thenumber of deaths during non epidemic weeks of the rst year of data.

3) Forecasting: We then used the Fourier equation to predict deaths during the rst in uenzaepidemic.

4) Excess Deaths: The actual deaths during these epidemic weeks were then compared with the

forecasted deaths, with the difference between the actual recorded deaths and the predicteddeaths being the excess deaths attributed to in uenza.

5) The deaths during the epidemic weeks during the rst season were replaced with forecasteddeaths to predict excess deaths for the second season.

6) Second season: We calculated mean and standard deviation of deaths from week 42 of thesecond season until week 42 of the subsequent season. The epidemic weeks for the secondseason were de ned as any two consecutive weeks with deaths greater than two standarddeviations from the mean for week 42 of the second season until week 42 of the third season.The method required the Fourier series equation to be updated/ re-calibrated, using data fromthe rst season (which included the forecasted, not actual, deaths during the epidemic period)to the beginning of the epidemic period of the second season. The recalibrated equation was

then to be used to forecast deaths for the second season epidemic period. As before, thedifference between the actual recorded deaths and the predicted deaths would be the excessdeaths attributed to in uenza.

1.4.2. ARIMA Method Steps:1) Fitting the model on non-epidemic data from previous years: We used the data set from

week 27 of 1972 to week 42 of 1974, consisting of actual, non-epidemic deaths and predicted(from the Fourier series equations) epidemic period deaths (for 1972-73 season), as inputs intoan ARIMA model.

2) Checking for goodness of t: After tting the model, we checked for goodness-of- t by testingfor autocorrelation in the residuals. The test statistic we used was the Ljung modi cation of theBox-Pierce Q statistic (refer to Ljung and Box, 1978). The model is a good t if the errors haveno correlation and the p value of the test statistic is large. We changed the model if the p valueof the test statistic was smaller than 5% for many lags.

3) Forecast forward: We used ARIMA methodology to forecast the next 52 weeks of deaths.4) Epidemic Weeks: The epidemic weeks during the next season (3rd season) were de ned

as two or more consecutive weeks when actual mortality was greater than the 95 percentcon dence interval of the forecasted deaths.

5) Excess Deaths: As before, the difference between the actual recorded deaths and the forecasteddeaths during the de ned epidemic weeks was the excess deaths attributed to in uenza.

6) We then replaced the actual deaths for the epidemic weeks of the 3rd season with the predicteddeaths.

7) We repeated the process for each subsequent season. For each season we forecasted

deaths (for next 52 weeks) from week 42 onwards using the data for all previous weeks. Afterpredicting deaths for a given year, that year (with predicted deaths) was included in the dataset used to predict subsequent years (i.e, the data set used to predict grew). When predictingdeaths for each year, we re-estimated the coef cients of the ARIMA equation.

1.5. Additional Technical Details:

1.5.1. Lag Operator: The lag operator is a highly useful operator in time series analysis. Sometimes it is easierto write the time series model using the lag operator. Suppose we have a series x

t , then the

application of lag operator on x t will generate a new series y t where the value of y on date t is equalto the value x took on at date t -1:

yt = Lx t x t-1


45/52

Annex 10

38

The lag operator satis es the following properties:

L(Lxt ) = L(x t-1 ) = x t-2

L2 xt = L(x t-1 ) = x t-2

Lk xt = x t-k

L(x t ) = Lx t = x t-1 , where is a constant

L(xt +w t ) = Lx t + Lwt = x t-1 + w t-1

Following is an example of the use of lag operator

yt = ( + L)Lx t = (L + L2 ) xt = x t-1 + x t-2

1.5.2. Autoregressive Integrated Moving Average (ARIMA) Process:

The ARIMA model combines the moving average and the autoregressive models with adifferencing term to remove trend in the data. A moving average model of order q , MA(q ), iswritten as

X t = + t 1t-1 2t-2 ............... qt-q

We can also write the above MA( q) model using the lag operator

X t = + t 1 Lt 2 L2t ............... q L

qt

or q X t = + 1 i L

i t i=1 ( )

where is the mean of the series, 1 are the parameters of the moving average part, q is thenumber of moving average terms, and t is the error term (or random shock) with zero mean,constant variance and s are uncorrelated over time. We call it moving average since it is anaverage of random shocks in the past, where t is the random shock in period t, t-1 is the randomshock in period t-1 , and so on. Thus the value of series in any time period is expressed as a sum ofcurrent random shock and q previous period random shocks.

The autoregressive model of order p, AR( p), is written as

X t = c + 1 X t-1 + 2 X t-2 + .......... + p X t-p+ t

Using the lag operator, we can also write the above equation as

X t = c + 1 LX t + 2 L2 X t + .......... + P L

P X t + t

or

q

X t = c + i Li X t +

t i=1 ( )


46/52

Annex 10


where c is the mean of the series, 1 are the parameters of the autoregressive part, and p is thenumber of autoregressive terms. In the autoregressive model, past values of the series in uencethe current value of the series, where X is the value of series in time period X t-1 is the value ofseries in time period t-1, and so on. Thus, in the AR( p) model, the value of series in any time periodis modelled as a sum of the random shock in the current period and a portion of p previous values

of the series.

An autoregressive moving average model (ARMA) combines both moving average andautoregressive models. An ARMA ( p, q ) model contains p autoregressive terms and q movingaverage terms and is written as

X t = + 1 X t-1 + 2 X t-2 + .......... + p X t-p+ t 1t-1 2t-2 ......... qt-q

or in lag operator terms as

Thus, in an ARMA ( p, q ) model, the current value of the series is in uenced by the past values ofthe series and also by the past random shocks.

If the time series data is not stationary, i.e. it shows a trend, we can rst take a one perioddifference of the series to remove the trend before we do the analysis. After differencing we geta new series W t such that

W t = X t X t-1

Since the differenced series W t is now stationary (no trend), we can use the ARMA( p, q ) model

on W t for forecasting. If the rst differenced series W t is also not stationary, we can do a seconddifferencing to make the series stationary (we rarely go beyond second differencing) and useARMA( p, q ) model on the stationary Z t series Z t = W t W t-1

Thus if we need to difference a series to make it stationary, we use the ARIMA ( p, d, q ) model,where d is level of differencing needed to make it stationary. The ARIMA model is similar to anARMA model, except that we rst difference the series and then apply the ARMA model on thedifferenced series.

An ARIMA ( p, d, q ) model can be written as,

Where L is the lag operator, and d is level of differencing needed to make the series stationary.

1.5.3. Seasonality: Some time series show a seasonal pattern. For example, a time series of weekly deaths fromin uenza might show a seasonal pattern, where every year the deaths peak in winter and are ata minimum in summer. There are two ways to model seasonality in the ARIMA process. If thetime series shows a strong seasonal pattern, we can difference the series with respect to the

seasonal lag. For example, if we use the time series of weekly in uenza deaths and we nd that

q

X t = + 1 i Li t

i=1 ( )

q

1 i Li

i=1 ( )

q (1-L) d X t = 1 i L

i t i=1 ( )

q 1 i L

i i=1 ( )


47/52

Annex 10

40

the autocorrelations have strong wave like shape with spikes at 52 weeks interval, we can take a52 week difference and then apply ARMA model. An ARIMA model with 52 week difference canbe written as

Note that in the above ARIMA model we also need a d period differencing to make the timeseries stationary. We may also need to incorporate seasonality as multiplicative term in the ARIMAmodel if the autocorrelations of the differenced series have spikes at the seasonal lag. In thefollowing equations we have modelled seasonality as a multiplicative term in the AR and MA partsrespectively.

References;

1) Box G, Jenkins GM. Time-series analysis; forecasting and control. San Francisco: Holden-Day,1970.

2) Choi K, Thacker SB. An evaluation of in uenza mortality surveillance, 1962-1979. I. Timeseries forecasts of expected pneumonia and in uenza deaths. Am J Epidemiol 1981;113:215-26.

q

(1-L 52 )(1-L 52 )(1-L) d X t = 1 i Li t

i=1 ( )

q

1 i Li

i=1 ( )

q

(1-L 52 )(1-L) d X t = 1 i Li (1-L 52 ) t

i=1 ( )

q

1 i Li

i=1 ( )

q

(1-L52

)(1-L)d

X t = 1 i Li

t i=1 ( ) q

1 i Li

i=1 ( )


48/52

Annex 11


Annex 11: Approach 5 - Stochastic Frontier Estimation Method

The Stochastic Frontier Estimation method is the of cial procedure for the estimation of excessmortality in Japan. It applies the stochastic frontier estimation to mortality data. Excess mortalityis de ned as the difference in the total number of deaths and the hypothetical number of deaths

if there is no in uenza epidemic. Thus the baseline must indicate the number of deaths if there isno in uenza epidemic. This stochastic frontier estimation can represent the baseline which is theminimum number of deaths in the month. The estimation equation for mortality in all causes atyear t , Yt , is

log Y t = + 1T + 2T 2 + i M it + t (5)where T 1 is the linear time trend and M it are dummy variables for month. The third degree ofpolynomial equation is not signi cant, therefore time trend variables for the third and higher degreepolynomial are not included in the model.

A disturbance term is de ned as

t = v t + | t | (6)

where vt and t are mutually independent random variables, and vt ~ N(0, 2 ) and

t ~ N(0, 2 ). A random variable vt is a purely stochastic disturbance and t represents the

magnitude of the epidemic which is a non-negative deviation term from the expected number ofdeaths. The equation form (5) is selected by statistical inference. The probability density functionof this model is

2 t / f ( t ) = ( ) (t ) (7) 2 + 2 2 + 2 2 + 2where is the probability density function for standard normal distribution and is its cumulativedistribution function. The likelihood function is de ned as a product of the probability densityfunction over time: L( 1 , 2 ,....... t ) = f ( t ). Estimators for , 1, 2 , s , / and 2 + 2 are chosen so as to maximize the likelihoodfunction. The estimators are therefore maximum likelihood estimators. This model is known ineconomics as the stochastic frontier estimation. In economics, t is often de ned as inef ciency,and stochastic frontier estimation is used to measure this inef ciency. For example, deviation

from the most ef cient production is de ned as inef ciency in estimation of production function,and deviation from the least cost is de ned as inef ciency in estimation of cost function. In thiscontext, we de ned t as the magnitude of the epidemic. In this way we can estimate the excessmortality without making any arbitrary assumptions of what epidemic or non-epidemic weeksare. For example, the referred models arbitrarily de ne a non-epidemic season and estimate theexcess mortality during the epidemic season using this non-epidemic season as a baseline. Theepidemic threshold is usually de ned as an upper limit of 95% con dence bound for stochasticdisturbance terms, vt . Therefore, the epidemic threshold should be

+ 1T + 2T 2 + i M it + 1.96 (8)


49/52

Annex 11

42

where indicates the estimated parameters and is de ned as ____________

( 2 + 2 ) 1/2 1 + ( / ) 2

Comparing the stochastic frontier estimation model with the robust regression model, these aresimilar in the sense of all information other than mortality. However, robust regression modelhas less or no weight on the lower outlier as well as the upper outlier, which is the epidemic.Conversely, the stochastic frontier estimation model put some weight on the lower outlier becausethe estimated baseline in the model represents the minimum mortality in the month. In this sense,we may sometimes nd a lower outlier in the robust regression model, and it may contradict thede nition of baseline, which is the minimum number of deaths in the month if there is no in uenzaoutbreak.

The choice of such a seasonality or function f (.) may be different from country to country,based on statistical inferences.


50/52

Annex 12


Annex 12: Approach 6 - Poisson Regression Method

(a) Poisson regression (assumed weekly data to be used)

log E ( Y t ) = + 1 Z 1t + ... q Z qt + X 1 (9)

where for each week index t (t = 1,2,...,n) , observation Y t follows Poisson distribution, E(Y t ) isthe expected numbers of health outcome; Z it , Z 2t ,...,Z qt are the covariates for trends, seasonality,temperature, relative humidity and proportion of positive results tested for RSV; X t is the weeklyproportion of positive results tested for in uenza; 1 , 2 ,..., q and are the estimated parametersfor respective covariates and is an intercept of the model.

Note: Here we use dummy variable to de ne each year and all the year dummy variable in modelto control for trend We also use

2kt 2

Documents

A Practical Guide for Designing and Conducting Influenza Disease Burden Studies