Weekly electronic mortality data for years 2006 through 2011 were obtained from the Bureau of Policy and Strategy, Ministry of Public Health (MoPH), Thailand. Deaths were categorized into three groups based on codes from the International Classification of Diseases, Tenth Revision [ICD-10]: circulatory diseases (ICD-10 codes I00-99), respiratory diseases (ICD-10 codes J00-99), and pneumonia and influenza (P&I) (ICD-10 codes J10-18). Other variables included age at death and location of death (hospital or community). We defined non-pandemic years as every year except 2009.

In 2004, the National Institute of Health, MoPH launched the national influenza sentinel surveillance system to monitor virus circulation in patients with influenza-like illness presenting at outpatient clinics in 11 sites throughout Thailand.12 Between 2004 and 2009, each site was first instructed to enroll a convenience sample of up to five patients per week with ILI for a total of 20 patients per month. In September 2009, the sample size was increased to 10 patients per week (five from children aged <15 years and five from persons aged ≥ 15 years). These data, collected systematically throughout the year, represent an unbiased sample of the timing of influenza activity and are appropriate for analyzing seasonal trends. All specimens were tested by reserve-transcription polymerase chain reaction for influenza viruses. We obtained the weekly numbers of influenza viruses (by type and subtype) and the total number of specimens tested, from the National Institute of Health, and calculated weekly positivity using weekly number of specimens as the denominator.

Of all-cause deaths among person aged <65 and ≥65 years, ill-defined deaths (i.e., no cause listed) accounted for 25% (58% of these occurred in the community) and 54% (87% occurred in the community), respectively. To account for the high proportion of ill-defined cause of death, we reapportioned weekly ill-defined deaths by age (<65 and ≥65 years), and location (hospital and community) to the various categories (P&I, respiratory, circulatory). We first excluded the ill-defined deaths and calculated the proportion of deaths due to P&I, respiratory and circulatory among deaths recorded in-hospital by age (<65 and ≥65 years) under the assumption that the coding for an in-hospital death was more accurate than for a community death. This assumption is based on the knowledge that deaths occurring outside the hospital are recorded by non-medical, civil registrars and based on lay reports from relatives.13 We then reapportioned all ill-defined deaths (including both those that occurred in-hospital and those in the community) to the various categories (P&I, respiratory, circulatory) so that the final proportion in each category were equivalent to those found in the in-hospital deaths.

We created two separated databases, one without reapportioning the deaths and the other with reapportioning.

Meteorological parameters including weekly mean temperature and absolute humidity were obtained from the Thai Meteorological Department.

A negative binomial regression model was used to estimate deaths attributable to influenza in two age groups (<65 and ≥65 years) by incorporating influenza viral data as covariates in the model. To assume an additive relationship between the exposure to influenza and resulting mortality, we used an identity link function in the models. The models included independent variables that comprised weekly percentage of specimens testing of confirmed infection with seasonal influenza A, seasonal influenza B, and influenza A (H1N1) pdm09 during a given week. We tested various lags (no lag, 1 week, and 2 weeks) to account for delays between infection and death. To account for the baseline, the models include either harmonic terms for annual [sine (θ) and cosine (θ), where θ = 2*π*week/52·25] and semiannual [sine (2θ) and cosine (2θ)] periods or climate factors [weekly mean temperature and absolute humidity, using linear terms or nonlinear terms (natural cubic spline function)]. Selection of the most statistically meaningful proxy for influenza activity was based on fitting 13 models to ≥65 years respiratory deaths and comparing Akaike information criterion (AIC) values. We obtained the smallest AIC values for a model with a 2-week lag and smoothing spline of temperature and humidity and report those results. A detailed description of the model fitting procedure is provided in Table S1.

In this model, where Y_t is the number of deaths in week t and α is the offset term and is equal to the log of the population size. β₀ represented the intercept, β₁ and β₂ represented coefficients associated with secular trends, β₃ represented coefficients associated with a natural cubic spline (NCS) of weekly mean temperature, β₄ represented coefficients associated with a NCS of absolute humidity, β₅–β₉ represented coefficients associated with the percentages of specimens testing positive for each influenza virus type and subtype during a given week. All viral terms had a 2-week lag. The models were run separately by death category (P&I, respiratory, and circulatory), age group (<65 and ≥65 years), and location of death (hospital or community, only for the data without reapportioned deaths). The analysis was conducted both ways, with and without reapportioning these deaths with the latter included as supplementary tables and figures.

To estimate excess mortality associated with influenza, we used the following procedure:

E0–E1 was the estimate of mortality due to influenza.

Confidence intervals were calculated based on the standard errors of model coefficients. We note that these confidence intervals provide only a minimum estimate of uncertainty.

The statistical analyses were carried out using sas version 9.3 (SAS Institute, Cary, NC, USA).

As a sensitivity analysis, we modeled all-cause deaths. We compared the excess deaths obtained using the all-cause deaths as outcome in the final models with estimates using respiratory and circulatory deaths as outcomes in the same models.

From January 2006 to December 2011, an annual mean of 399 853 deaths (range, 389 696 in 2006 to 413 209 in 2011) occurred in Thailand (42% occurred in hospital). An average of 13 554 (3·4%) underlying P&I deaths, 27 268 (6·8%) underlying respiratory deaths, and 37 112 (9·3%) underlying circulatory deaths occurred each year. Table1 shows the annual mean number and proportion of deaths by cause of death. Forty-one percent of total deaths were coded as cause ill-defined and redistributed to the various categories (P&I, respiratory, circulatory) by location and age group.

During the 6-year period, 21 560 specimens were tested for influenza viruses. There were 2961 (14%) positive results for influenza viruses. The annual mean number of tests positive was 14% (range 10–18%) for influenza A viruses, and 6·5% (range 4·1–10%) for influenza B viruses (Table S2).

For all ages combined, there was year-to-year variability in the annual number of influenza-associated deaths; the mean for all years was 2511 (4·0 per 100 000) (Table2). The average influenza-associated mortality rate for non-pandemic years was 4·3 per 100 000, and for the pandemic year, 2009, it was 2·4 per 100 000. The rates in pandemic year for both age groups were lower compared with the other years (<65 years: 0·6 versus 0·7 per 100 000 and ≥65 years: 24 versus 26 per 100 000). The annual death rate attributable to influenza was the highest in the year 2008 (5·6). Eighty-three percent (2094/2511) of the average annual estimated influenza-associated deaths occurred among person age ≥65 years. The average annual rate of influenza-associated deaths was 0·7 per 100 000 population for person aged <65 years and 42 for person aged ≥65 years. Among adults aged <65 years and ≥65 years, influenza A (H3N2) and B were associated with the most deaths. The average annual number of influenza-associated deaths was 1401 (2·2 per 100 000) for underlying P&I causes, 2751 (4·3 per 100 000) for underlying respiratory causes, and 529 (0·8 per 100 000) for underlying circulatory causes (Table3). Deaths among persons aged ≥65 years accounted for 74% (1031/1401), 76% (2094/2751), and 100% (529/529) of the overall estimated average annual influenza-associated deaths with underlying P&I, respiratory, and circulatory causes, respectively.

The average annual rate of influenza-associated deaths for adults aged ≥65 years was 21 per 100 000 persons for P&I deaths, 42 per 100 000 for respiratory deaths, and 11 per 100 000 for circulatory deaths (Table3). Among adults aged <65 years, the average annual rate of influenza-associated deaths was 0·6 deaths per 100 000 persons for P&I causes and 1·1 per 100 000 for respiratory causes. The weekly number of deaths from 2006 to 2011, and excess deaths where mortality exceeded the baseline, is shown in Figure1; there are no discrete peaks indicating influenza epidemics. In the sensitivity analysis of all-cause deaths, the average annual rate of influenza-associated deaths was 2·0 (95%CI: −24 to 27) per 100 000 population for person aged <65 years and 99 (95%CI: −322 to 516) for person aged ≥65 years.

Estimated mean annual influenza-associated deaths without reapportioning ill-defined death are shown in Table S3. The average annual number of influenza-associated deaths from underlying P&I, respiratory, and circulatory diseases was 632 (1·0 per 100 000), 1308 (2·1 per 100 000), and 295 (0·5 per 100 000, Table S4), respectively. The estimated number of influenza deaths from the model with reapportioned ill-defined deaths and from the non-reapportion model for persons aged <65 was similar, while for those age ≥ 65 years, the estimates from reapportioned models were approximately 3 times higher than from the non-reapportion model. Compared with Figure1, similar patterns were observed with no obvious peak for the excess circulatory and respiratory deaths and overall lower magnitude (Figure S1).