^{1}

^{2}

^{2}

^{3}

^{4}

^{1}

^{2}

^{5}

^{¶}

^{2}

^{5}

^{*}

¶ Membership of the Toronto Invasive Bacterial Diseases Network is provided in the Acknowledgments.

DNF has received research funding from GlaxoSmithKline, Novartis, and Sanofi Pasteur, all of which manufacture vaccines against respiratory pathogens. AM has received research funding from GlaxoSmithKline and Sanofi Pasteur. The other authors declare that they have no competing interests.

Using a combination of modeling and statistical analyses, David Fisman and colleagues show that influenza likely influences the incidence of invasive pneumococcal disease by enhancing risk of invasion in colonized individuals.

The wintertime co-occurrence of peaks in influenza and invasive pneumococcal disease (IPD) is well documented, but how and whether wintertime peaks caused by these two pathogens are causally related is still uncertain. We aimed to investigate the relationship between influenza infection and IPD in Ontario, Canada, using several complementary methodological tools.

We evaluated a total number of 38,501 positive influenza tests in Central Ontario and 6,191 episodes of IPD in the Toronto/Peel area, Ontario, Canada, between 1 January 1995 and 3 October 2009, reported through population-based surveillance. We assessed the relationship between the seasonal wave forms for influenza and IPD using fast Fourier transforms in order to examine the relationship between these two pathogens over yearly timescales. We also used three complementary statistical methods (time-series methods, negative binomial regression, and case-crossover methods) to evaluate the short-term effect of influenza dynamics on pneumococcal risk. Annual periodicity with wintertime peaks could be demonstrated for IPD, whereas periodicity for influenza was less regular. As for long-term effects, phase and amplitude terms of pneumococcal and influenza seasonal sine waves were not correlated and meta-analysis confirmed significant heterogeneity of influenza, but not pneumococcal phase terms. In contrast, influenza was shown to Granger-cause pneumococcal disease. A short-term association between IPD and influenza could be demonstrated for 1-week lags in both case-crossover (odds ratio [95% confidence interval] for one case of IPD per 100 influenza cases = 1.10 [1.02–1.18]) and negative binomial regression analysis (incidence rate ratio [95% confidence interval] for one case of IPD per 100 influenza cases = 1.09 [1.05–1.14]).

Our data support the hypothesis that influenza influences bacterial disease incidence by enhancing short-term risk of invasion in colonized individuals. The absence of correlation between seasonal waveforms, on the other hand, suggests that bacterial disease transmission is affected to a lesser extent.

Although some pathogens (disease-causing organisms) cause illness all year round, others are responsible for seasonal peaks of illness. These peaks occur because of a complex interplay of factors such as the loss of immunity to the pathogen over time and seasonal changes in the pathogen's ability to infect new individuals. Thus, in temperate countries in the northern hemisphere, illness caused by influenza viruses (pathogens that infect the nose, throat, and airways) usually peaks between December and March, perhaps because weather conditions during these months favor the survival of influenza virus in the environment and thus increase its chances of being transferred among people. Another illness that peaks during the winter months in temperate regions is pneumonia, a severe lung infection that is often caused by

Although the co-occurrence of seasonal peaks of influenza and IPD is well documented, it is unclear whether (or how) these peaks are causally related. For example, do the peaks of influenza and IPD both occur in the winter because influenza enhances person-to-person transmission of

Between January 1995 and October 2009, 38,501 positive influenza tests were recorded in Ontario by the Canadian national influenza surveillance network. Over the same time period, the Toronto Invasive Bacterial Diseases Network (a group of hospitals, laboratories, and doctors that undertakes population-based surveillance for serious bacterial infections in the Toronto and Peel Regions of Ontario) recorded 6,191 IPD episodes. The researchers used a mathematical method called fast Fourier transforms that compares the shape of wave forms to look for any relationship between infections with the two pathogens over yearly timescales (a test of hypothesis 1) and three statistical methods to evaluate the short-term effect of influenza dynamics on IPD risk (tests of hypothesis 2). Although they found wintertime peaks for infections with both pathogens, there was no correlation between the seasonal wave forms for influenza and IPD. That is, there was no relationship between the seasonal patterns of the two infections. By contrast, two of the statistical methods used to test hypothesis 2 revealed a short-term association between infections with influenza and with IPD. Moreover, the third statistical method (the Granger causality Wald test, a type of time-series analysis) provided evidence that data collected at intervals on influenza can be used to predict peaks in IPD infections.

These findings support (but do not prove) the hypothesis that influenza influences IPD incidence by enhancing the short-term risk of bacterial invasion in individuals already colonized with

Please access these Web sites via the online version of this summary at

A related research article by the same authors evaluating links between respiratory viruses and invasive meningococcal disease can be found in

The US Centers for Disease Control and Prevention provides information for patients and health professionals on all aspects of

The UK National Health Service Choices website also provides information for patients about

MedlinePlus has links to further information about

More information about the

The International Association for Ecology and Health provides information on the

The seasonal periodicity of pneumonia and influenza deaths in temperate countries is regarded as sufficiently commonplace that the term “flu season” is part of the English language vernacular. However, the widespread recognition of wintertime seasonality of severe respiratory disease obscures the fact that remarkably little is understood about the genesis of such seasonality

However, seasonal oscillation in disease incidence implies a complex system that includes such elements as loss of immunity and seasonally enhanced transmissibility with the latter potentially attributable to environmental, microbiological, or social forcing factors. Such a complex system cannot be adequately characterized by evaluating cross-sectional correlation in risk, particularly where causal links are concerned. Indeed, statistically significant correlation in risk when evaluating disease processes with shared seasonality is to be expected, and such correlation could be casual, as seen with many human respiratory pathogens with a peak incidence in winter, rather than causal.

We aimed to investigate the relationship between influenza infection and invasive pneumococcal disease (IPD) in Ontario, Canada, using several complementary methodological tools. We evaluated the relationship between the seasonal wave forms for influenza and IPD using fast Fourier transforms in order to examine the relationship between these epidemic waveforms over yearly timescales, i.e., to test the hypothesis that influenza influences IPD by enhancing person-to-person transmission. We also used three complementary statistical methods (time-series methods, negative binomial regression, and case-crossover methods) to evaluate the short-term effect of influenza dynamics on pneumococcal risk, i.e., to examine whether the risk of IPD is increased in previously colonized individuals.

The Toronto Invasive Bacterial Diseases Network (TIBDN) study was approved by Research Ethics Boards at all participating institutions. Ethical approval was not required for the remainder of this work.

The Toronto Invasive Bacterial Diseases Network is a collaboration of all hospitals, microbiology laboratories, infection control practitioners, physicians, and public health units serving the population of metropolitan Toronto and Peel Regions (population 3.7 million), performing population-based surveillance for selected serious bacterial and viral infections

A national network of hospital and provincial laboratories submit weekly reports of numbers of tests performed (using viral culture or direct antigen detection) and numbers of positive tests for influenza A and influenza B to the Public Health Agency of Canada. For the purpose of this study, the surveillance data of the province of Ontario from 1 January 1995 through 3 October 2009 were included in the analysis.

We obtained time series data on ultraviolet (UV) radiation and weather from Environment Canada monitoring stations in the Greater Toronto Area

The seasonality of disease occurrence was evaluated through calculation of autocorrelations for weekly case counts ^{α + β1•sin(2π(week/52))+β2•cos(2π(week/52))}. Accordingly, the phase-shift of the composite waveform generated by combining sine and cosine components can be approximated as tan^{−1}(β_{1}/β_{2}) and can be used to estimate the timing of peak disease incidence _{1}^{2} + β_{2}^{2}). Standard errors for phase-shift and amplitude were estimated as sums of standard errors for each model coefficient minus covariance. We evaluated the correlation between phase and amplitude terms for influenza A and B (combined), influenza A, and influenza B and pneumococcal sine waves, by year, by calculating Spearman correlation coefficients. Heterogeneity of phase and amplitude terms for ^{2} statistics. Differences in the magnitude of heterogeneity of amplitude or phase by pathogen was assessed using Knepp and Entwisle's method for assessing difference between chi-squared statistics

As both deviance and Pearson goodness-of-fit statistics suggested over-dispersion of pneumococcal case counts, we evaluated the relationship between IPD rates, influenza, and environmental exposures using a series of negative binomial regression models

Once models had been fitted, we attempted to enhance model parsimony through model reduction, removing terms that had no influence on pneumococcal case counts, in a manner that minimized Akaike's information criterion (AIC), with final models representing those that best balanced parsimony and fit

Granger causality Wald testing was used to determine whether influenza A and B (combined), influenza A, and influenza B seasons can be used to forecast peaks in pneumococcal infections

We used a case-crossover approach to evaluate short-term associations between influenza A and B (combined), influenza A, and influenza B infection and IPD occurrence during influenza seasons in order to provide a means for evaluating the association between brief exposures and rare outcomes. The design is characterized by “self matching,” in that cases serve as their own controls. A “case” thereby is a day on which a case occurred, whereas a “control” is an adequately selected day on which a case did not occur

Data were analyzed using Stata version 11.0 (Stata Corporation) and SAS version 9.1 (SAS Institute).

The period from 1 January 1995 through 3 October 2009 was available for analysis, including a total number of 38,501 positive influenza tests in Ontario and 6,191 episodes of pneumococcal disease in Metropolitan Toronto/Peel Region (

Weekly numbers of IPD (black bars) and positive tests for influenza A (thin curve) and B (thick curve). in the areas under surveillance from January 1995 to October 2009.

Annual periodicity (i.e., recurrence at annual intervals) could be demonstrated for pneumococcal disease by spectral decomposition and construction of autocorrelations, with a maximum autocorrelation coefficient (ac) = 0.4488 at week 52 (

Periodicity of (A) IPD, (B) influenza A and B (combined), (C) influenza A, and (D) influenza B in the Toronto-Peel area, Ontario, Canada as illustrated by autocorrelograms, for the time period from January 1995 to October 2009.

Negative binomial regression models revealed strong statistical evidence for annual oscillation for both pneumococcal disease and combined influenza A and B infections (^{2} = 62810.15, df = 15, ^{2} = 100.0%) and amplitude (Cochran Q statistic: χ^{2} = 5752.86, df = 15, ^{2} = 99.7%) terms for influenza A and B (combined), there was less heterogeneity of amplitude (Cochran Q statistic: χ^{2} = 40.06, df = 15, ^{2} = 62.6%) terms for ^{2} = 23.28, df = 15, ^{2} = 35.6%) terms. Differences in heterogeneity χ^{2} statistics between pathogens

Accordingly, no significant association between amplitude and phase of oscillatory waves of pneumococcal disease and influenza A and B (combined), influenza A, or influenza B could be detected using Spearman correlation (

Phase or Amplitude | PhaseInfluenza A and B | Amplitude |
Phase |

Amplitude influenza A and B | −0.30 (0.26) | 0.19 (0.49) | −0.35 (0.18) |

Phase influenza A and B | — | −0.06 (0.83) | 0.09 (0.75) |

Amplitude |
— | — | −0.49 (0.05) |

Phase or Amplitude | Phase Influenza A | Amplitude |
Phase |

Amplitude influenza A | −0.19 (0.48) | −0.06 (0.83) | −0.20 (0.05) |

Phase influenza A | — | −0.11 (0.70) | 0.12 (0.65) |

Amplitude |
— | — | −0.49 (0.05) |

Phase or Amplitude | Phase Influenza B | Amplitude |
Phase |

Amplitude influenza B | −0.38 (0.15) | 0.07 (0.80) | 0.06 (0.84) |

Phase influenza B | — | −0.29 (0.28) | 0.06 (0.81) |

Amplitude |
— | — | −0.49 (0.05) |

Results from negative binomial regression analysis suggest influenza A and B (combined) activity is associated with IPD (incidence rate ratio = 1.09 case of IPD per 100 influenza cases,

Negative Binomial Regression Analysis | Incidence Rate Ratio | 95% Confidence Interval | |

Influenza A and B (combined, per 100 infections), 1-wk lag | 1.092 | 1.047–1.138 | <0.001 |

Influenza A and B (combined, per 100 infections), 3-wk lag | 0.932 | 0.890–0.976 | 0.003 |

UVI average, 1-wk lag | 0.927 | 0.893–0.963 | <0.001 |

UVI average, 3-wk lag | 0.946 | 0.917–0.976 | <0.001 |

Relative humidity, 1-wk lag (%) | 0.995 | 0.991–1.000 | 0.03 |

Mean temperature, 2-wk lag (°C) | 0.993 | 0.984–1.002 | 0.14 |

Mean temperature, 4-wk lag (°C) | 1.004 | 0.995–1.012 | 0.39 |

Based on the Granger causality Wald test, there is evidence that influenza A and B (combined) Granger-cause pneumococcal disease (Chi-square = 23.28, df = 2,

We were able to identify a significant association between total influenza (A and B) and IPD with a 1-wk lag using a case-crossover approach (odds ratio [95% confidence interval] for one case of IPD per 100 influenza cases, 1.10 [1.02–1.18]) (

Lag (wk) | Odds Ratio | 95% Confidence Interval | |

0 | 1.05 | 0.97–1.13 | 0.20 |

−1 | 1.10 | 1.02–1.18 | 0.01 |

−2 | 1.00 | 0.93–1.09 | 0.90 |

−3 | 0.93 | 0.86–1.01 | 0.07 |

−4 | 1.02 | 0.94–1.11 | 0.65 |

Although this observational study cannot prove epidemiological causality, our analysis of seasonal oscillation in IPD and influenza A and B incidence in Ontario, Canada, is strongly suggestive of a causal relationship between influenza and IPD. The fact that we did not find a relationship between seasonal dynamics of influenza and IPD suggests that this relationship likely results from effects of influenza on risk of invasive disease in individuals with pneumococcal colonization. We found no epidemiological evidence to support the concept that influenza has a strong effect on IPD risk via changes in pneumococcal transmission dynamics that would be caused by enhanced susceptibility to pneumococcal colonization, increased infectiousness of pneumococcus-colonized individuals, or increased duration of carriage. All such changes would be expected to perturb the dynamics of pneumococcal seasonality in such a way that pneumococcal seasonal waves would “track” influenza waves. While we found significant and temporally directional relationships between influenza burden in the population and IPD risk, the seasonal “waves” of influenza and IPD were independent of one another, with substantial variability in influenza contrasted with the far more regular seasonal occurrence of IPD. The nonstereotyped nature of wintertime seasonality of viral respiratory pathogens has been noted previously

While prior work has evaluated the impact of influenza infection on pneumococcal risk, much of this work has been performed in experimental rodent models that may not be generalizable to human populations

The major clinical implications of our study are 2-fold: first, that dramatic increases in influenza incidence, as might be seen during a pandemic year, would not in and of themselves imply a marked surge in risk of IPD, but rather would do so only if there were overlap between flu and pneumococcal waves. This observation may help explain the absence of a marked increase in IPD risk in our jurisdiction during the (atypical) spring wave of the 2009 influenza A-H1N1 pandemic. Although public health agencies issued calls for stepped-up pneumococcal vaccination because of the pandemic

If our findings are correct, and influenza increases the risk of IPD without influencing pneumococcal transmission dynamics, the question remains: what are the drivers of the remarkably stereotyped seasonality of IPD? Crude correlations between IPD risk and mean weekly temperature and mean weekly minutes of darkness, but not precipitation, have been reported by Dowell et al. _{2}-vitamin-D metabolism, or direct mutagenic effects on vegetative bacteria. The current study, performed in a different jurisdiction, replicates this finding.

Like any observational study, ours has several limitations. As it is not possible to randomize exposure to influenza, or to meteorological exposures, observational studies like this one are necessary to evaluate the impact of such exposures on invasive bacterial disease risk in real-world human populations. Limitations of observational designs like the one we have used here include difficulties with residual confounding, ecological fallacy due to aggregation of exposures, and measurement issues. While it is not possible to control for all residual confounding by unmeasured confounders in regression models, our use of seasonal smoothers, multivariable techniques, and case-crossover design should largely have controlled for confounding of influenza effects by seasonally varying factors including seasonal behaviors and meteorological effects. Our use of case-crossover design implicitly matches for all seasonal effects that would be constant over the 3-wk time block used for each risk stratum. The case-crossover approach assumes that the distribution of exposure is constant across referent times, which should be controlled by careful referent selection. It can be associated with bias if referent periods are not chosen a priori or if referents are functions of the observed event times

Although implied by its name, Granger causality testing does not prove true causation in the epidemiologic sense. It originally is an econometric method to infer predictive value of one time series over another and thus, rather than prove causation, adds a valuable piece to the overall interpretation of our analyses towards potential causation. With respect to ecological effects, it should be noted that we effectively treat influenza as an aggregate “environmental” exposure, in that we don't have influenza status on the individuals who develop invasive bacterial disease, which creates the risk of ecological effects, including ecological fallacy, in our study. There are practical barriers to formulating an adequately powered study that actually captured data on the temporal relationship between influenza and invasive bacterial disease at the level of the individual (e.g., through ongoing evaluation of individuals in the population for influenza infection, with subsequent identification of the [rare] individuals who developed invasive bacterial disease).

Finally, with respect to measurement issues, for influenza A and B, we had to rely on public health surveillance data that are incomplete both because of under-reporting, and also because a majority of individuals with influenza never undergo virological testing

In conclusion, our data support the hypothesis that influenza influences IPD risk by enhancing pneumococcal invasion in colonized individuals, but has little effect on the transmission dynamics of pneumococcal infection. We suggest that the mechanism for such an effect might be influenza-related alterations at the level of the respiratory epithelium

Collaborating investigators in the Toronto Invasive Bacterial Disease Network are as follows: S. O'Grady, Bridgepoint Health (Toronto, Canada); I. Armstrong and B. Yaffe, City of Toronto Public Health (Toronto, Canada); A. Sarabia, Credit Valley Hospital (Mississauga, Canada); J. Kapala, Dynacare Laboratories (Brampton, Canada); M. Loeb, Hamilton Health Sciences Centre (Hamilton, Canada); A. Matlow, S. Richardson, and D. Tran, Hospital for Sick Children (Toronto, Canada); K. Lee, Humber River Regional Hospital (Toronto, Canada); J. Allard, Joseph Brant Memorial Hospital (Burlington, Canada); M. Silverman, Lakeridge Health (Oshawa, Canada); D. Yamamura, LifeLabsMedical Laboratory Services (Toronto, Canada); P. Shokry, Markham Stouffville Hospital (Markham, Canada); K. Green, A. Plevneshi, S. Pong-Porter, and B. Willey, Mount Sinai Hospital (Toronto, Canada); M. Lovgren and G. Tyrrell, National Center for Streptococcus (Edmonton, Canada); K. Katz and B. Mederski, North York General Hospital (North York, Canada); N. Rau, Halton Healthcare (Oakville, Canada); J. Gubbay, F. Jamieson, and D. Low, Ontario Agency for Health Protection and Promotion (Toronto, Canada); E. de Villa, Peel Public Health (Brampton, Canada); I. Kitai, Rouge Valley Health System (Toronto, Canada); J. Rodgers, Royal Victoria Hospital (Barrie, Canada); C. Wigston, Soldier's Memorial Hospital (Orillia, Canada); S. Krajden, St. Joseph's Health Centre (Toronto, Canada); M. Muller and R. Devlin, St. Michael's Hospital (Toronto, Canada); A. Simor and M. Vearncombe, Sunnybrook Health Sciences Centre (Toronto, Canada); R. Lovinsky and D. Rose, The Scarborough Hospital (Toronto, Canada); J. Downey, Toronto East General Hospital (Toronto, Canada) and Headwaters Healthcare Centre (Orangeville/Shelburne, Canada); J. Powis, Toronto East General Hospital (Toronto, Canada); K. Ostrowska, Trillium Health Centre (Mississauga, Canada); W.L. Gold and S. Walmsley, University Health Network (Toronto, Canada); H. Dick, Vita-Tech Canada (Toronto, Canada); M. Baqi and D. Richardson, William Osler Health Centre (Brampton, Canada); D. Chen, Southlake Regional Hospital (Newmarket, Canada) and York Central Hospital (Richmond Hill, Canada). All members of the Toronto Invasive Bacterial Diseases Network have acquired data for this study and revised the article critically for important intellectual content.

autocorrelation coefficient

invasive pneumococcal disease

ultraviolet