Research in contextEvidence before this studyRegression-based ecological methods linking respiratory-virus circulation to fluctuations in all-cause mortality have long been used to estimate virus-attributable deaths that cause-of-death coding misses. We searched PubMed, Embase, and medRxiv from database inception to Feb 28, 2025, using terms including “population attributable fraction”, “ecological regression”, “excess mortality”, “seasonal baseline”, “influenza”, “RSV”, “SARS-CoV-2”, and “all-cause mortality”. We identified classic influenza and influenza/RSV attribution studies and later refinements to seasonal-baseline estimation, but most were pre-COVID, pathogen-specific, or focused on overall pandemic excess mortality rather than placing SARS-CoV-2 on a common footing with established seasonal respiratory viruses across a multi-decade series.Added value of this studyUsing Ontario mortality and laboratory surveillance data spanning 1993 to February 2025, we estimated the mortality burden of influenza A, influenza B, RSV, and SARS-CoV-2 within a single ecological framework across the pandemic transition.
SARS-CoV-2 was associated with about 6.5% of all-cause mortality, versus 1.8% for pre-pandemic influenza A, a 3–4-fold higher burden. Model-estimated SARS-CoV-2-attributable deaths closely matched independently adjudicated provincial COVID-19 deaths, providing external validation rarely available in this literature. Influenza A and RSV estimates were sensitive to seasonal adjustment, underscoring the inherent difficulty of separating strongly seasonal viral signals from shared seasonal variation in mortality.Implications of all the available evidenceRespiratory-virus-attributable mortality remains substantial after the formal pandemic emergency, with SARS-CoV-2 continuing to contribute materially to population mortality.
These findings support sustained surveillance and continued investment in vaccination and other prevention strategies across respiratory viruses, and the transparent use of all-cause, regression-based attribution, with explicit sensitivity analyses around seasonal baselines, to quantify burden that cause-specific coding underestimates.
Introduction
Respiratory viral illnesses impose a substantial burden on individuals and the health care system, often resulting in significant morbidity and mortality. Many of these viruses are now vaccine preventable or can have their severity attenuated through vaccination. This holds true in a Canadian context, where two of the leading causes of death are respiratory illnesses; influenza, which prior to the COVID-19 pandemic, caused approximately 3500 deaths annually,1 and severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), which as of 2023 had caused a reported 55,282 deaths in Canada.2, 3 Other respiratory viruses, such as respiratory syncytial virus (RSV), can also cause substantial morbidity and mortality, particularly in vulnerable age groups (e.g., children and older adults).4, 5, 6
To prioritise vaccination and other prevention programs against respiratory viruses, it is important to understand the potential for prevention of morbidity and mortality at the population level. One such method of doing so is by analysing the population attributable fraction (PAF) for a given respiratory virus. PAF quantifies the proportion of cases that could be prevented if the exposure (in this case, the viral infection), were eliminated, offering a valuable metric for public health planning.7 PAF can be calculated as the difference between the observed number of cases (O) and the expected number of cases if the exposure was removed (E), divided by the observed number of cases (PAF = (O-E)/O).7 “Shortcut formulae” allow PAF to be estimated if the relative risk among the exposed, and the prevalence of exposure, are known.7, 8
When mortality is the outcome of interest, researchers often rely on cause-specific mortality data to estimate the impact of respiratory viruses. However, these data are subject to well-known limitations, including misclassification and inaccuracies in death certificate coding.9 As an alternative, regression-based methods offer an indirect means of estimating the PAF.10 These approaches do not require explicit cause-of-death attribution and are agnostic to the magnitude of the exposure effect. By modelling the statistical association between viral circulation and mortality, often using smoothing techniques to detrend and smooth data, such models can estimate the fraction of mortality attributable to a given virus, based on the observed changes in outcomes corresponding to fluctuations in exposure.
In this study, we sought to quantify the population-level mortality burden attributable to key respiratory viruses in Ontario, Canada, using regression-based approaches. Specifically, we estimated the PAF for influenza A and B, and respiratory syncytial virus (RSV) over the period from 1993 to 2025, and for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) from 2020 to 2025. By applying statistical models that account for temporal variation in viral activity and mortality, our aim was to provide robust, population-based estimates of virus-attributable mortality to inform future vaccination strategies and respiratory virus prevention policies.
We also evaluated how modelling choices, specifically the inclusion or exclusion of seasonal adjustment term, affect estimates of virus-attributable mortality.
Methods
Data sources
We calculated percent positivity for influenza A and B, and RSV in Ontario using laboratory-based surveillance data from FluWatch, Canada's national surveillance system for influenza and other respiratory viruses, available from 1993 onward.1, 11, 12 FluWatch is the Public Health Agency of Canada's national respiratory virus surveillance system, integrating laboratory, syndromic, outbreak, and severe-outcome surveillance streams contributed by provincial and territorial laboratories, sentinel providers, and hospital networks.1, 11, 12 The percent-positivity measures used here derive from its laboratory surveillance stream (the Respiratory Virus Detection Surveillance System), a sentinel network of public-health and hospital laboratories that report, by surveillance week, the number of specimens tested and the number positive for influenza (by type and subtype) and RSV. As a surveillance system, FluWatch is designed for trend monitoring rather than complete case ascertainment, so percent positivity indexes viral circulation rather than true population infection incidence.
For SARS-CoV-2, we used test-adjusted case counts from Ontario's Case and Contact Management System (CCM)13 from March 2020 to August 2022 (using methodology described in Fisman et al.14 and Bosco et al.15) and percent positivity from FluWatch from September 2022 onward. Both measures were included in pandemic-period models to account for the surveillance system transition from universal to targeted testing; the counterfactual sets both to zero when estimating SARS-CoV-2 PAF. Mortality data were obtained from Statistics Canada (1991–1993) and the Ontario Deaths Registry (1994–February 2025).16, 17 Data were aggregated to monthly intervals; monthly aggregation reduces noise in mortality attribution models and aligns with established influenza mortality attribution literature.10 Population denominators were obtained from Statistics Canada.16 Individual-level CCM data are not publicly available; all other data sources are public, and aggregated monthly data used in this analysis are provided in a public repository (see Data Availability).
Analysis
All analyses were conducted using Stata version 18. To estimate the number of deaths attributable to seasonal respiratory viruses and SARS-CoV-2, we fit regression models to monthly all-cause mortality data from Ontario, incorporating indicators of viral circulation. We modelled death counts using both Poisson and negative binomial regression, including Fast Fourier Transforms (FFT; sine and cosine terms) to account for seasonal periodicity, and linear and quadratic terms to model secular trends.
Population denominators were included as a log-offset to estimate mortality rates. Influenza and RSV positivity data were unavailable for a period during the 2009–2010 H1N1 pandemic; the corresponding months were therefore excluded from regression models by listwise deletion (complete-case analysis), consistent with their exclusion from Fig. 1.Fig.
1Mortality and virological surveillance data for Ontario, 1993–2025. (a) Monthly deaths per 100,000 from Ontario Deaths Registry (red) and Statistics Canada (teal), demonstrating close concordance between data sources (mean difference 0.3%). (b) Monthly percent positivity for influenza A (red) and influenza B (teal) from FluWatch surveillance, showing regular seasonal epidemics and marked suppression during 2020–2021.
(c) Monthly RSV percent positivity from FluWatch surveillance, demonstrating seasonal patterns and pandemic-era disruption. (d) SARS-CoV-2 adjusted case counts per 10,000 population (red, left axis) and SARS-2 positivity percentage (teal, right axis) from March 2020 onward, showing multiple pandemic waves. Vertical dashed lines indicate emergence of the SARS-CoV-2 pandemic in Ontario in March 2020.
Panels a and c exclude a period from 2009 to 2010 during the influenza A H1N1 pandemic for which data were not available (grey rectangle). Data truncated at February 2025 to account for reporting delays in mortality time series.
Poisson model fit was assessed using Pearson and deviance goodness-of-fit tests (estat gof in Stata). Evidence of overdispersion (p < 0.001 for both tests) supported the use of negative binomial regression for all further analysis. Given that SARS-CoV-2 did not circulate widely before March 2020, analyses were stratified into pre-pandemic (January 1993–February 2020) and pandemic (March 2020–February 2025) periods.
The pandemic period was defined as extending through February 2025, the most recent month for which mortality data were assessed as substantially complete; mortality data for subsequent months showed evidence of reporting lag and were excluded. Secular trend terms were centred within each period to reduce collinearity. To characterise the overall shift in baseline all-cause mortality associated with the pandemic era, we fitted a single negative binomial model to the full 1993–2025 series, including Fourier seasonal terms, secular trend terms, a population offset, and a binary indicator for the pandemic period (March 2020 onward).
This whole-series model was used solely to quantify the pandemic-era shift in baseline mortality; virus-specific attribution was estimated separately using the period-restricted models described above.
As a sensitivity analysis, we re-estimated all models excluding Fourier terms to evaluate the robustness of viral effect estimates to the inclusion of seasonal periodicity. Seasonal adjustment terms may compete with viral predictors for explanatory power, particularly when viruses exhibit strong seasonal patterns. The reduced models were otherwise identical in structure, retaining trend terms, population offsets, and viral covariates.
To formally assess whether removal of Fourier transforms significantly altered viral effect estimates, we compared regression coefficients between primary and sensitivity models using Wald tests. Seemingly unrelated estimation (suest command in Stata) was used to jointly estimate both models and obtain the variance-covariance matrix needed for hypothesis testing. For each viral covariate, we tested the null hypothesis that coefficients were equal across models. p-values <0.05 were considered evidence of significant modification by seasonal adjustment.
For both primary and sensitivity models, we estimated the fraction of all-cause mortality attributable to each virus by generating model-based predictions under two scenarios: one with observed viral circulation, and a counterfactual in which the coefficients for each virus (or all viruses) were set to zero. PAF was then calculated as the difference between predicted deaths under observed and counterfactual scenarios, divided by the predicted deaths under observed conditions. Confidence intervals for each PAF estimate were obtained using the punaf command, specifying virus-specific terms to be set to zero in the counterfactual.
Model-predicted SARS-CoV-2 attributable deaths were externally validated by comparing them to adjudicated Ontario COVID-19 deaths during the same period (March 2020–March 2024).18 The provincially reported COVID-19 deaths used for this comparison comprise both deaths for which COVID-19 was recorded as the underlying cause and deaths in which COVID-19 was a contributing but not underlying cause18; this composite definition is concordant with the all-cause framing of our attributable-mortality estimates.
In addition to the Fourier-term sensitivity analysis described above, we conducted three further pre-specified analyses. First, in a subgroup analysis, we stratified the pandemic period into the SARS-CoV-2 Public Health Emergency of International Concern (PHEIC) period (March 2020–April 2023) and post-PHEIC period (May 2023–February 2025). Given the smaller sample size of each stratum (n = 38 and n = 22 months, respectively), these temporal comparisons are presented as exploratory; the combined pandemic model (n = 60) is treated as primary.
Second, as a sensitivity analysis, we excluded March–June 2020 from the pandemic period to assess the influence of the early pandemic phase, characterised by health system disruption, incomplete surveillance, and high case-fatality rates prior to vaccination (July 2020–February 2025; n = 56 months). Third, as a further sensitivity analysis, we compared model fit using adjusted versus unadjusted SARS-CoV-2 case counts using the Akaike and Bayesian Information Criteria (AIC/BIC), with the model using test-adjusted counts demonstrating better fit (AIC = 5145.2 versus 5162.1); adjusted counts were therefore retained for primary analyses. For each pathogen, PAF estimates and 95% confidence intervals from the counterfactual models were extracted for each combination of modelling approach (with or without Fourier terms) and time period using the punaf command in Stata, and displayed descriptively in a forest plot.
To place the pre-pandemic influenza A and pandemic SARS-CoV-2 estimates on a common footing, we additionally fitted a single joint negative binomial model to the full 1993–2025 series in which the pre-pandemic and pandemic eras each retained their own intercept, secular trend, and Fourier seasonal terms, so that the 2020 structural break was modelled explicitly; influenza A was estimated over the pre-pandemic era and SARS-CoV-2 over the pandemic era. Because both attributable fractions derive from a single model with a shared variance-covariance matrix, we estimated the ratio of the two, with its confidence interval, by parametric simulation from the model covariance (5000 draws). We did not include a pandemic-period indicator: over the pandemic era it is near-collinear with SARS-CoV-2 circulation, and the elevated pandemic baseline is itself substantially a downstream consequence of SARS-CoV-2, so conditioning on it would constitute adjustment for a mediator.
In a sensitivity analysis over the window for which vaccination data were available (March 2020–August 2022), we entered incident monthly COVID-19 vaccination coverage (new doses per capita, obtained by differencing the cumulative Ontario dose series) as a time-varying covariate in the pandemic model. We assessed residual autocorrelation using the autocorrelation function and Ljung–Box tests on Pearson residuals and, as a robustness check, recomputed coefficient standard errors using Newey–West heteroskedasticity- and autocorrelation-consistent estimators (12-month lag). Generative AI tools (Claude, Anthropic; ChatGPT, OpenAI) were used to assist with review of statistical code and editorial revision of the manuscript.
All AI-assisted outputs were reviewed and verified by the authors, who take full responsibility for the accuracy and integrity of the analysis and text.
Ethics and data availability
All data used in this study were aggregate level, with viral positivity data and monthly mortality being publicly available. All data and analysis code are available at https://github.com/fismanda/respiratory-virus-mortality-ontario (archived at https://zenodo.org/records/20981617). Approval for use of aggregate level, test-adjusted SARS-CoV-2 case counts was granted by the University of Toronto Research Ethics Board (Protocol # 41690).
Age-specific mortality attribution will be addressed in future work.
Role of the funding source
This research was supported by grants to DNF from the Canadian Institutes of Health Research (2019 COVID-19 Rapid Research Funding OV4-170360; CIHR Project Grant #518192). The funders had no role in study design, data collection, data analysis, data interpretation, or writing of the report.
Results
Model fit
Analyses were stratified into pre-pandemic (January 1993–February 2020) and pandemic (March 2020–February 2025) periods. Negative binomial regression demonstrated superior fit to Poisson models (p < 0.001 for overdispersion). In a whole-series model spanning 1993–2025 that included a binary pandemic-period indicator, the pandemic era was associated with elevated baseline all-cause mortality (IRR 1.052, 95% CI 1.027–1.078, p < 0.0001), after adjustment for secular trend and seasonality.
Virus-specific attribution was estimated separately using period-restricted negative binomial models, described below. Mortality and virological surveillance data are shown in Fig. 1.
Residual autocorrelation was present in the pre-pandemic model (principally an annual component; lag-12 autocorrelation ≈0.59), but Newey–West standard errors for the viral coefficients were comparable to or smaller than model-based standard errors, and adding a semi-annual Fourier harmonic left the attributable fractions essentially unchanged (influenza A 1.8%–2.1%; SARS-CoV-2 6.2%–6.0%).
Primary model results
Separate negative binomial models estimated mortality attributable to respiratory viruses in pre-pandemic and pandemic periods (Table 1). In pre-pandemic models (Table 1), only influenza A predicted mortality (IRR 1.0037 per one-percent increase in test positivity, 95% CI 1.0028–1.0046). Influenza B did not predict mortality, and RSV showed a non-significant inverse association (IRR 0.9992, 95% CI 0.9979–1.0005, p = 0.24).
During the combined pandemic period (March 2020–February 2025), only SARS-CoV-2 activity predicted mortality. Both SARS-CoV-2 percent positivity (IRR 1.0073, 95% CI 1.0039–1.0106) and adjusted case counts (IRR 1.0029 per 10,000 cases, 95% CI 1.0022–1.0037) predicted mortality. Influenza viruses and RSV did not show significant associations with mortality in the pandemic-period models.Table 1Mortality incidence rate ratios for respiratory viruses derived from negative binomial models.PeriodVirusMortality incidence rate ratio (IRR)aIRR 95% confidence intervalPopulation attributable fraction (PAF) (%)PAF 95% CIp-value for associationbPre-pandemic (1993 to February 2020)Influenza A1.00371.0028–1.00461.821.39–2.25<0.0001Influenza B1.00090.9992–1.00260.15−0.14 to 0.440.30RSV0.99920.9979–1.0005−0.45−1.21 to 0.300.24Combined respiratory viruses––1.530.46–2.300.005Pandemic (March 2020 to February 2025)SARS-CoV-2 (2020–2022) (per 10,000 increase in adjusted cases)1.00291.0022–1.00376.48c4.48–8.43<0.0001SARS-CoV-2 (2022–2025) (per 1% positivity)1.00731.0039–1.0106<0.0001Influenza A1.00250.9985–1.00650.22Influenza B1.00070.9789–1.02300.95RSV1.00320.9990–1.00730.13Combined Respiratory Viruses––7.615.15–10.02<0.0001aAll incidence rate ratios are per percent increase in test positivity except for SARS-CoV-2 (2020–2022), which is incidence rate ratio per 10,000 increase in adjusted cases.bp-values for multivariable model coefficients and PAF from punaf are identical.cSARS-CoV-2 PAF estimated by setting both early (adjusted cases) and later (percent positivity) model terms to zero.
Sensitivity analysis
Models without seasonal adjustment
To evaluate whether seasonal adjustment terms competed with viral predictors for explanatory power, we re-estimated all models excluding Fourier transforms (Table 2). Formal comparison of coefficients using Wald tests revealed substantial differences for viruses with strong seasonal patterns. In the pre-pandemic period, removal of Fourier transforms significantly altered estimates for influenza A and RSV.
The influenza A incidence rate ratio increased from 1.0037 to 1.0060 (Wald p < 0.001). Most notably, RSV shifted from a counterintuitive non-significant inverse association (IRR 0.9992) to a positive association (IRR 1.0033, 95% CI 1.0023–1.0042, p < 0.0001), with this reversal being highly significant (Wald p < 0.0001). Influenza B estimates were unchanged (Wald p = 0.50).
During the pandemic period, the early SARS-CoV-2 indicator (adjusted cases) increased from 1.0029 to 1.0035 (Wald p = 0.042). RSV again demonstrated a significant difference, with the IRR increasing from 1.0032 (p = 0.13) to 1.0073 (95% CI 1.0030–1.0116, p = 0.001; Wald p = 0.006). Other viral predictors did not differ significantly between models.Table 2Mortality incidence rate ratios for respiratory viruses derived from negative binomial models without fast fourier transforms.PeriodVirusMortality incidence rate ratio (IRR)aIRR 95% Confidence intervalPopulation attributable fraction (PAF) (%)PAF 95% CIp-value for associationbp-value for difference from primary model with FFTdPre-pandemic (1993 to February 2020)Influenza A1.00601.0051–1.00682.892.49–3.29<0.0001<0.0001Influenza B1.00130.9996–1.00290.21−0.07 to 0.500.140.50RSV1.00331.0023–1.00421.871.33–2.40<0.0001<0.0001Combined Respiratory Viruses––4.874.34–5.41<0.0001Pandemic (March 2020 to February 2025)SARS-CoV-2 (2020–2022) (per 10,000 increase in adjusted cases)1.00351.0026–1.00448.03c5.73–10.27<0.00010.042SARS-CoV-2 (2022–2025) (per 1% positivity)1.00951.0055–1.0135<0.00010.08Influenza A1.00430.9994–1.00920.75−0.11 to 1.590.080.13Influenza B1.01070.9862–1.03600.36−0.46 to 1.170.390.16RSV1.00731.0030–1.01161.770.72–2.810.0010.006Combined Respiratory Viruses––10.598.02–13.09<0.001aAll incidence rate ratios are per percent increase in test positivity except for SARS-CoV-2 (2020–2022), which is incidence rate ratio per 10,000 increase in adjusted cases.
FFT, Fast Fourier Transform.bp-values for multivariable model coefficients and PAF from punaf are identical.cSARS-CoV-2 PAF estimated by setting both early (adjusted cases) and later (percent positivity) model terms to zero.dP-value from Wald test comparing coefficient to corresponding estimate in primary model with Fourier transforms (Table 1).
In a sensitivity analysis using unadjusted SARS-CoV-2 case counts in place of test-adjusted counts, model fit worsened (AIC = 5162.1 versus 5145.2), supporting the use of test-adjusted data for estimating attributable mortality.
Population attributable fraction
Using the primary models with seasonal adjustment (Table 1), we calculated the population attributable fraction for each virus. In the pre-pandemic period, influenza A accounted for 1.8% (95% CI 1.4–2.3%) of all-cause mortality. Combined respiratory viruses accounted for 1.5% of mortality pre-pandemic (slightly lower than influenza A alone, owing to the negative RSV contribution in seasonally adjusted models), and 7.6% during the combined pandemic period (corresponding to approximately 46,000 deaths (about 9200 per year)).
During the combined pandemic period (March 2020–February 2025), SARS-CoV-2 accounted for 6.48% (4.5–8.4%) of all-cause mortality, approximately 3–4-fold the pre-pandemic influenza A burden.
In the joint model, the pre-pandemic influenza A attributable fraction was 1.8% (95% CI 1.4–2.3) and the pandemic SARS-CoV-2 attributable fraction was 6.2% (95% CI 4.1–8.1), a ratio of 3.4-fold (95% CI 2.2–4.9; probability that the ratio exceeded 1, >0.9999).
Applying the fitted pre-pandemic model, influenza A was associated with approximately 1600 attributable deaths per year in Ontario (about 41,400 over 1993–2020). For comparison, the roughly 3500 annual influenza deaths commonly cited for Canada correspond to about 1350 per year when scaled to Ontario's population, consistent with, and modestly below, our estimate, as expected for an all-cause method that recovers deaths under-attributed by cause-of-death coding.
Including incident SARS-CoV-2 vaccination coverage in the model left the SARS-CoV-2 attributable fraction unchanged (5.3% versus 5.3%), and coverage was not an independent predictor of mortality (p = 0.92); the raw cumulative series gave the same result (5.2%).
Model-predicted SARS-CoV-2–attributable deaths over March 2020–March 2024 totalled 18,052, within approximately 3% of the 18,594 adjudicated Ontario COVID-19 deaths over the same interval18 (Supplementary Table S1).
A sensitivity analysis excluding March–June 2020 yielded a conservative SARS-CoV-2 PAF of 5.7% (95% CI 3.3–8.1%), confirming that the primary estimate is not driven by the extraordinary early pandemic phase. In exploratory temporal stratification, the PHEIC period (March 2020–April 2023; n = 38) yielded a SARS-CoV-2 PAF of 7.3% (95% CI 5.0–9.6%) with Fourier adjustment. In the post-PHEIC period (May 2023–February 2025; n = 22), the SARS-CoV-2 attributable fraction (9.8%, 95% CI 1.1–17.7%) remained significantly greater than zero but did not differ significantly from the PHEIC-period estimate, consistent with ongoing endemic SARS-CoV-2 mortality burden rather than a true post-PHEIC increase.
Post-PHEIC influenza A and B attributable fractions did not differ significantly from pre-pandemic baselines (Flu A PAF 1.3%, 95% CI −0.7 to 3.3%; Flu B PAF 0.6%, 95% CI −1.5 to 2.7%). Primary inference rests on the combined pandemic model. Fig.
2a illustrates the temporal patterns of virus-attributable mortality, demonstrating that pre-pandemic combined viral mortality closely tracked influenza A, while pandemic-era combined viral mortality was dominated by SARS-CoV-2.Fig. 2Model-estimated virus-attributable mortality over time. (a) Monthly death rates per 100,000 population attributable to respiratory viruses from primary models with Fourier seasonal adjustment.
Blue line shows all respiratory viruses combined, brown line shows influenza A, and green line shows SARS-CoV-2. Pre-pandemic, the combined viral mortality (blue) nearly overlaps with influenza A (brown), demonstrating that influenza A accounted for essentially all virus-attributable mortality. During the pandemic, combined viral mortality (blue) tracks closely with SARS-CoV-2 (green), illustrating SARS-CoV-2's dominance.
(b) RSV-attributable death rates comparing models with Fourier adjustment (blue) versus without (black). With seasonal adjustment, RSV shows frequently negative values pre-pandemic (PAF -0.45%) and substantially attenuated positive associations during the pandemic. Without seasonal adjustment, RSV demonstrates consistently positive mortality associations of similar magnitude (∼1.8–1.9%) in both periods.
Both panels exclude a period from 2009 to 2010 during the influenza A H1N1 pandemic for which data were not available (grey rectangle). Vertical dashed lines indicate emergence of the SARS-CoV-2 pandemic in Ontario in March 2020.
In models without seasonal adjustment (Table 2), attributable fractions were substantially higher. Pre-pandemic, influenza A accounted for 2.9% (95% CI 2.5–3.3%) and RSV for 1.9% (95% CI 1.3–2.4%) of all-cause mortality; combined respiratory viruses accounted for 4.9% (95% CI 4.3–5.4%). The contrast between RSV estimates with and without seasonal adjustment is shown in Fig.
2b. During the combined pandemic period, SARS-CoV-2 accounted for 8.03% (95% CI 5.73–10.27%) and RSV for 1.77% (95% CI 0.72–2.81%) of mortality; combined respiratory viruses accounted for 10.59% (95% CI 8.02–13.09%) (Fig. 3).Fig.
3Population attributable fraction (PAF) of mortality for respiratory viruses, by modelling approach and time period. Forest plot of stratum-specific PAF estimates with 95% confidence intervals for all respiratory viruses combined, influenza A, influenza B, respiratory syncytial virus (RSV), and SARS-CoV-2. For each virus, estimates are shown with and without Fast Fourier Transform (FFT) seasonal adjustment, for the pre-pandemic (January 1993–February 2020) and pandemic (March 2020–February 2025) periods.
For SARS-CoV-2, estimates are shown for the combined pandemic period and, as exploratory analyses, for the PHEIC (March 2020–April 2023; n = 38 months) and post-PHEIC (May 2023–February 2025; n = 22 months) sub-periods; primary inference rests on the combined pandemic estimate. Points represent PAF estimates and horizontal lines indicate 95% confidence intervals. FFT, Fast Fourier Transform; PAF, population attributable fraction; PHEIC, Public Health Emergency of International Concern.
Discussion
In this study, we used more than three decades of mortality and virological surveillance data from Ontario to quantify the proportion of all-cause mortality attributable to major respiratory viruses, extending a previously reported analysis through February 2025. Prior to the emergence of SARS-CoV-2, influenza A was the only virus with a consistent positive association with all-cause mortality after adjusting for seasonal and secular trends, with a PAF of 1.8% (95% CI 1.4–2.3%), consistent with prior studies.19, 20 During the combined pandemic period (March 2020–February 2025), SARS-CoV-2 was associated with a substantially greater population mortality burden, with a PAF of 6.48% (4.5–8.4%) in primary models, approximately 3–4-fold the pre-pandemic influenza A burden, with model-predicted SARS-CoV-2-attributable deaths closely matching independently adjudicated Ontario COVID-19 deaths over March 2020–March 2024 (Supplementary Table S1), supporting the validity of our approach. A sensitivity analysis excluding March–June 2020 yielded a conservative estimate of 5.7% (95% CI 3.3–8.1%), demonstrating that the primary result is not driven by the extraordinary early phase of the pandemic.
In contrast, influenza A, influenza B, and RSV did not show significant associations with mortality during the PHEIC period, consistent with the well-documented suppression of seasonal respiratory virus circulation due to SARS-CoV-2 dominance and the implementation of non-pharmaceutical interventions (NPIs).21, 22
A key contribution of our study is the formal comparison of models with and without Fourier-based seasonal adjustment. Seasonal terms are commonly used in time-series mortality models to account for periodic fluctuations,10 but because many respiratory viruses also exhibit strong seasonality, these terms may compete with viral predictors for explanatory variance. The direction of any resulting bias is not fixed.
Removing seasonal terms may allow viral covariates to absorb mortality from other seasonal processes, inflating viral attribution; conversely, because a substantial share of seasonal mortality conventionally attributed to cardiovascular, cerebrovascular, and metabolic causes is itself downstream of influenza infection,19, 20 seasonal adjustment may remove genuine virus-associated signal and bias attribution downward. We therefore interpret the Fourier-adjusted influenza A estimate (1.8%) as a conservative lower bound and the unadjusted estimate (2.9%) as illustrating the influence of shared seasonality, rather than treating either estimate as definitively unbiased.
Similar patterns have been observed in excess mortality studies during the COVID-19 pandemic, where models adjusting for seasonal pathogen circulation produced different estimates than those not accounting for these factors.23 Our findings illustrate this tension: removal of Fourier terms significantly increased estimated effects for pathogens with strong seasonal signatures, particularly influenza A (PAF increased from 1.8% to 2.9%, Wald p < 0.001) and RSV (shifted from PAF –0.45% to PAF 1.9%, Wald p < 0.001).
SARS-CoV-2 attributable-fraction estimates were stable across modelling specifications (6.48% with versus 8.03% without seasonal adjustment), in contrast to influenza A and RSV, whose estimates varied substantially with seasonal adjustment. Models including Fourier terms provide conservative estimates reflecting mortality variation uniquely attributable to viral indicators after accounting for shared seasonal patterns, while models without such adjustment isolate virus-associated mortality more directly but may conflate viral effects with other seasonal phenomena. The negative RSV coefficient observed in seasonally adjusted models is unlikely to represent a true protective effect; rather, it most plausibly reflects the limited identifiability of a strongly and regularly seasonal exposure when modelled alongside flexible seasonal terms.24 The emergence of a positive RSV association in models without Fourier adjustment is consistent with this interpretation, and we regard RSV as the least identifiable of the three pathogens in this framework, interpreting its estimates with corresponding caution.
A key new contribution of the extended analysis is characterisation of respiratory virus mortality in the post-PHEIC period (May 2023–February 2025). Although the post-PHEIC SARS-CoV-2 point estimate was numerically higher than the PHEIC-period estimate, the two did not differ significantly, and the post-PHEIC confidence intervals were wide given the limited sample size (n = 22 months); primary inference is therefore drawn from the combined pandemic model. The post-PHEIC SARS-CoV-2 attributable fraction (9.8%, 95% CI 1.1–17.7%) nonetheless remained significantly greater than zero, indicating that SARS-CoV-2 continued to exert a measurable population mortality burden after the formal end of the PHEIC, albeit not at a level distinguishable from the PHEIC period itself.
In contrast, post-PHEIC attributable fractions for influenza A (PAF 1.3%, 95% CI −0.7 to 3.3%) and influenza B (PAF 0.6%, 95% CI −1.5 to 2.7%) did not differ significantly from pre-pandemic baselines, providing no support for a sustained post-pandemic influenza mortality rebound in this population. Influenza B showed little evidence of measurable mortality impact over the full study period, aligning with prior studies showing it causes milder illness with a smaller population-wide mortality footprint than influenza A.25 The magnitude of SARS-CoV-2's mortality burden is notable: over the combined pandemic period, SARS-CoV-2 was associated with approximately 3–4-fold higher attributable mortality than pre-pandemic influenza A. This exceeded the influenza A burden despite the rapid deployment of countermeasures during the pandemic (non-pharmaceutical interventions,26, 27 multiple vaccine formulations,28 antivirals,29 and substantial partial immunity from prior infection30) and notwithstanding that influenza is itself addressed by long-standing vaccines, antivirals, and substantial population immunity.
The close concordance between model-based and independently adjudicated COVID-19 deaths likely reflects the unique features of SARS-CoV-2 surveillance: unlike influenza and RSV, which are under-attributed in routine cause-of-death coding,10, 19 COVID-19 deaths were subject to intensive clinical, laboratory, and public health scrutiny during this period.
The persistence of SARS-CoV-2-attributable mortality into the post-PHEIC period is not paradoxical, even in the context of widespread vaccination, frequent reinfection, and substantially improved clinical management. Improvements in the acute management of COVID-19 act most directly on deaths coded as COVID-19; our regression-based approach, however, is deliberately agnostic to cause-of-death coding. A growing body of evidence indicates that SARS-CoV-2 infection precipitates downstream cardiovascular, cerebrovascular, and thrombotic events,31 as well as secondary and reactivated infections,32 that may prove fatal but are recorded under causes other than COVID-19 and are largely unaffected by improvements in acute COVID-19 care.
Deaths arising through these pathways are systematically under-captured by cause-of-death-specific approaches but are detectable in models relating all-cause mortality to viral circulation; our own analyses link SARS-CoV-2 activity to downstream events including pulmonary embolism and invasive group A streptococcal disease.33, 34 This mechanism parallels the influenza-attributable burden recorded under non-respiratory causes discussed above. Persistence of attributable mortality is further consistent with waning population protection: vaccine uptake in Canada fell substantially beyond the primary series,35 and protection against SARS-CoV-2 is transient, reflecting both antigenic change and immune waning, with severity of outcomes predicted by recency of vaccination as well as cumulative dose number.36 A population with low uptake of additional vaccine doses and increasingly remote from its most recent dose would retain limited vaccine-derived protection, providing no expectation that attributable mortality would decline through the post-PHEIC period.
Strengths of this analysis include the use of more than 30 years of mortality data, integration of laboratory surveillance indicators, formal evaluation of modelling assumptions through sensitivity analyses and Wald tests, and an extended time horizon now encompassing the post-PHEIC period. Limitations include the use of virological surveillance indicators as imperfect proxies for true infection incidence, and the ecological design, which cannot establish individual-level causation and assumes viral effects are independent and additive (potentially missing interactions such as interference or co-infections37).
Our use of all-cause mortality is a deliberate design choice intended to capture burden that cause-specific coding misses, but it precludes mechanistic or syndrome-specific inference, and it means that aggregate mortality masks age-specific patterns and likely underestimates mortality attributable to age-targeted pathogens such as RSV, whose burden is concentrated at the extremes of age; we therefore regard our RSV estimates as conservative, and identify age-stratified attribution as a priority for future work. For the same reason, deaths mediated by secondary bacterial infection following viral illness are captured within, rather than excluded from, the viral attributable estimate; because invasive bacterial disease is frequently downstream of antecedent viral infection.37, 38, 39 The antecedent virus is, in a sufficient-component cause sense,40, 41 a component cause of these deaths, and their inclusion reflects component-cause attribution rather than cause-of-death reassignment; because component causes are not mutually exclusive, attributable fractions for different exposures need not sum to 100%.40, 41 We therefore do not enter a bacterial co-circulation covariate, which, like conditioning on the cardiovascular and thromboembolic sequelae of infection, would adjust away a mediator of the viral effect, though our design likewise cannot partition the directly-viral and bacterially-mediated components. Our surveillance indicators do not distinguish the setting of acquisition, so estimates encompass both community-acquired and healthcare-associated infection, the elevated relative mortality of which we characterise directly elsewhere.42 Mortality captures only part of the societal burden of these pathogens; morbidity, quality-of-life loss, and economic costs, including those of post-acute SARS-CoV-2 sequelae and of RSV at the extremes of age, substantially extend the total burden beyond what we quantify here, as we and others have shown using methods designed for that purpose.36, 42, 43, 44, 45, 46, 47 This study used aggregate population-level mortality and viral surveillance data for the province of Ontario and did not include data on race or ethnicity; analyses were therefore not stratified by these characteristics.
We note that the population-level burden estimated here may mask important inequities in respiratory-virus mortality across racialised and socially marginalised groups, which warrant dedicated investigation using appropriately disaggregated data.
No independent benchmark of influenza- or RSV-attributable mortality exists against which to validate those specific estimates, as cause-of-death coding under-attributes these viruses and surveillance reports viral circulation rather than attributable deaths; the close agreement between our SARS-CoV-2 estimate and independently adjudicated COVID-19 deaths nonetheless supports the validity of the approach where such validation is possible. The transition in SARS-CoV-2 surveillance from case-based to test-based reporting in September 2022 introduces measurement heterogeneity, which is addressed by including both measures in the pandemic model and setting both to zero in the counterfactual, though residual heterogeneity may persist. Statistical power is limited in subperiod analyses given smaller sample sizes.
Additionally, mortality data for recent months are subject to reporting delays, which we mitigated by restricting analyses to February 2025. Because the September 2022 surveillance transition falls within the PHEIC period (March 2020–April 2023), the PHEIC-period estimate is itself derived from both exposure metrics (test-adjusted case counts before, and percent positivity after, the transition) rather than from a single consistent measure.
Despite these limitations, our findings have clear public health relevance. During 2020–2025, SARS-CoV-2 was associated with a population mortality burden 3–4 times higher than pre-pandemic influenza A, a pathogen that itself has historically caused an estimated 3500 Canadian deaths annually.1 Influenza A remains an important contributor warranting continued seasonal vaccination efforts, while RSV's consistent ∼2% burden in models without seasonal adjustment and its concentration in age groups at increased risk, including young children and older adults, highlights opportunities for emerging prevention strategies including vaccines for older adults and monoclonal antibodies for infants. The data suggest that SARS-CoV-2 continues to be associated with substantially greater population-level mortality burden than seasonal influenza, even after the formal end of the PHEIC.
The persistence of significant SARS-CoV-2-attributable mortality in the post-PHEIC period underscores the need to sustain surveillance and prevention strategies beyond the formal emergency phase. More broadly, this study illustrates the value of long-term respiratory virus surveillance and the inherent difficulty of separating the mortality burden of strongly seasonal viruses from shared seasonal variation in mortality, emphasising the need for continued investment in surveillance and analytic methods to monitor shifts in viral circulation, quantify preventable mortality burden, and guide immunization strategies.
Contributors
DNF conceived and designed the study and acquired the data. DNF and AAG accessed and verified the underlying data. DNF and AAG conducted the statistical analysis.
NJW, SM, ART, and CEL contributed to interpretation of the data. DNF drafted the manuscript, and all authors critically revised it for important intellectual content, approved the final version, and take responsibility for the decision to submit for publication.
Data sharing statement
All data used in this study are aggregate and publicly available. Monthly mortality and viral surveillance data and all analysis code are available at https://github.com/fismanda/respiratory-virus-mortality-ontario (archived at https://zenodo.org/records/20981617). Individual-level case data are not publicly available.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process
The authors used Claude (Anthropic) and ChatGPT (OpenAI) for statistical code review and editorial revision. The authors reviewed, edited, and take full responsibility for all content.
Declaration of interests
DNF has received consulting fees and honoraria for advisory boards and consulting from Seqirus and Pfizer within the past 36 months. ART was employed by the Public Health Agency of Canada when the research was conducted. The work does not represent the views of the Public Health Agency of Canada.
Other authors: no competing interests.