Air Travel and the Spread of Influenza: Important Caveats

While air travel contributes to the spread of influenza epidemics, the magnitude of impact is not clear compared to other factors—a crucial issue when considering a flight ban in the context of pandemic planning. Recent modeling efforts simulating the spread of pandemic influenza have concluded that such an intervention would matter little relative to other interventions [1–3]. But this assessment has now been challenged by an observational study of influenza in the winter following the post-9/11/2001 depression in air traffic. Brownstein and colleagues' study published in the September issue of PLoS Medicine [4] correlates variations in air traffic volume with patterns of timing and spread in influenza epidemics, based on United States mortality data from nine epidemic seasons between 1996 and 2005. While we find the study interesting, we have identified several important caveats and question the robustness of the conclusions. 
 
The core of this study's results lies in the observation that the 2001–2002 influenza epidemic immediately following 9/11 was late in the season and peaked in March (week of year 11), whereas the eight surrounding epidemics peaked between the end of December and the end of February (week of year 52 to 9). The authors attribute this delay to the 27% decline in air traffic that followed 9/11. 
 
Given the complexities of influenza virus subtype cycling and antigenic drift [5,6], it is essential to consider longer-term disease data spanning much more than nine years to interpret the “lateness” of the 2001–2002 epidemic. Using US national vital statistics data covering 30 winters from 1972 to 2002 [5], we identified four epidemics peaking in the month of March (13%), including the 2001–2002 epidemic following 9/11, but also two epidemics in the 1970s and the more recent 1991–1992 epidemic (Figure 1A). Furthermore, the average timing of influenza epidemics has not changed between 1972 and 2002—despite a concurrent and steady increase in air traffic volume by over 300% (Figure 1A) [7]. Indeed, during the earlier part of the last century when air traffic was minimal, influenza epidemics rapidly circulated around the world. Moreover, real-time influenza virus surveillance data from the US Centers for Disease Control and Prevention [8] show that last winter's (2005–2006) epidemic was even more delayed than the epidemic following 9/11, despite a 20% increase in air passenger traffic compared to the situation before 9/11 [7]. Clearly, late-season influenza epidemics have occurred and are still occurring even in the absence of restrictions on air travel. Hence a longer time perspective, with observations from both prior and more recent data, challenges this study's conclusions. 
 
 
 
Figure 1 
 
Patterns of Timing (A) and Spread (B) of 30 Influenza Epidemics in the US, Together with Trends in Air Travel Statistics 
 
 
 
In addition to comparing the timing of influenza epidemics across different seasons, Brownstein et al. analyzed the rate of disease spread among US administrative regions for their nine seasons of interest (1996–2005). In our previous work, we estimated the rate of influenza spread among all US states for 30 consecutive seasons (1972–2002) [5]. Our analysis shows that the epidemic following 9/11 spread at a rate comparable to other epidemics (Figure 1B), even after adjusting for the subtype of circulating viruses [5]. To increase our understanding of the spread of influenza, it is essential to quantify the relative importance of different modes of transportation. As an example, our recent study considered multiple modes of transportation (including air travel) and identified travel to and from work as a key determinant of the regional spread of epidemics [5]. 
 
In conclusion, Brownstein and colleagues' analysis of the “natural experiment” of the post-9/11 season is innovative and ingenious—but in and of itself could not demonstrate a robust association or a causal link between the decrease in air traffic and delayed timing of influenza epidemics. Even if there in fact had been a delay as hypothesized, the study lacked power to address the hypothesis, because this single “natural experiment” was set in a background of considerable variability in influenza epidemic patterns. Extrapolations from the study's findings predict that a flight ban could delay a pandemic by two months [9]—but we have shown here that this prediction is not supported by the analysis of more extensive disease data and transportation statistics. It is also unclear how a “natural experiment” conducted in the inter-pandemic period is applicable to a pandemic situation, where novel influenza viruses have higher transmissibility and circulate in fully susceptible populations, and may cause different age-patterns of transmission [10]. While Brownstein and colleagues' study represents an intriguing starting point, this study alone does not provide the critical quantitative evidence needed to evaluate the impact of travel restrictions on future pandemics.

While air travel contributes to the spread of infl uenza epidemics, the magnitude of impact is not clear compared to other factors-a crucial issue when considering a fl ight ban in the context of pandemic planning. Recent modeling efforts simulating the spread of pandemic infl uenza have concluded that such an intervention would matter little relative to other interventions [1][2][3]. But this assessment has now been challenged by an observational study of infl uenza in the winter following the post-9/11/2001 depression in air traffi c. Brownstein and colleagues' study published in the September issue of PLoS Medicine [4] correlates variations in air traffi c volume with patterns of timing and spread in infl uenza epidemics, based on United States mortality data from nine epidemic seasons between 1996 and 2005. While we fi nd the study interesting, we have identifi ed several important caveats and question the robustness of the conclusions.
The core of this study's results lies in the observation that the 2001-2002 infl uenza epidemic immediately following 9/11 was late in the season and peaked in March (week of year 11), whereas the eight surrounding epidemics peaked between the end of December and the end of February (week of year 52 to 9). The authors attribute this delay to the 27% decline in air traffi c that followed 9/11. Given the complexities of infl uenza virus subtype cycling and antigenic drift [5,6], it is essential to consider longerterm disease data spanning much more than nine years to interpret the "lateness" of the 2001-2002 epidemic. Using US national vital statistics data covering 30 winters from 1972 to 2002 [5], we identifi ed four epidemics peaking in the month of March (13%), including the 2001-2002 epidemic following 9/11, but also two epidemics in the 1970s and the more recent 1991-1992 epidemic ( Figure 1A). Furthermore, the average timing of infl uenza epidemics has not changed between 1972 and 2002-despite a concurrent and steady increase in air traffi c volume by over 300% ( Figure 1A) [7]. Indeed, during the earlier part of the last century when air traffi c was minimal, infl uenza epidemics rapidly circulated around the world. Moreover, real-time infl uenza virus surveillance data from the US Centers for Disease Control and Prevention [8] show that last winter's (2005)(2006)) epidemic was even more delayed than the epidemic following 9/11, despite a 20% increase in air passenger traffi c compared to the situation before 9/11 [7]. Clearly, late-season infl uenza epidemics have occurred and are still occurring even in the absence of restrictions on air travel. Hence a longer time perspective, with observations from both prior and more recent data, challenges this study's conclusions.
In addition to comparing the timing of infl uenza epidemics across different seasons, Brownstein et al. analyzed the rate of disease spread among US administrative regions for their nine seasons of interest (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005). In our previous work, we estimated the rate of infl uenza spread among all US states for 30 consecutive seasons   [5]. Our analysis shows that the epidemic following 9/11 spread at a rate comparable to other epidemics ( Figure 1B), even after adjusting for the subtype of circulating viruses [5]. To increase our understanding of the spread of infl uenza, it is essential to quantify the relative importance of different modes of transportation. As an example, our recent study considered multiple modes of transportation (including air travel) and identifi ed travel to and from work as a key determinant of the regional spread of epidemics [5].
In conclusion, Brownstein and colleagues' analysis of the "natural experiment" of the post-9/11 season is innovative and ingenious-but in and of itself could not demonstrate a robust association or a causal link between the decrease in air traffi c and delayed timing of infl uenza epidemics. Even if there in fact had been a delay as hypothesized, the study lacked power to address the hypothesis, because this single "natural experiment" was set in a background of considerable variability in infl uenza epidemic patterns. Extrapolations from the study's fi ndings predict that a fl ight ban could delay a pandemic by two months [9]-but we have shown here that this prediction is not supported by the analysis of more extensive disease data and transportation statistics. It is also unclear how a "natural experiment" conducted in the inter-pandemic period is applicable to a pandemic situation, where novel infl uenza viruses have higher transmissibility and circulate in fully susceptible populations, and may cause different age-patterns of transmission [10]. While Brownstein and colleagues' study represents an intriguing starting point, this study alone does not provide the critical quantitative evidence needed to evaluate the impact of travel restrictions on future pandemics.  Infl uenza patterns are based on weekly national vital statistics from 1972 to 2002 [5]. Air travel statistics represent the annual number of domestic and international passengers on US air carriers (scheduled fl ights, secondary y-axis) [7]. (A) Time series of timing of national peaks of infl uenza mortality. The 2001-2002 epidemic following 9/11 peaked in March, and so did two epidemics in the 1970s and one in the 1990s. (B) Time series of rate of spread between US states. The rate of spread of the 2001-2002 epidemic following 9/11 is comparable to that of other epidemics. Rate of spread as calculated in [5], based on the timing of epidemic peaks in each state.

Mark A. Miller
Viboud et al. [1] offer thoughtful commentary on our paper [2], opening a scientifi c exchange that we hope brings attention to a critical issue. We reported the fi rst empirical and quantitative evidence for the effect of airline travel on the rate of epidemic infl uenza spread. Though other investigators have also found this relationship [3][4][5][6], there is no consensus on effect size. We welcome scrutiny of our methods and results and believe the fi ndings stand for the following reasons.
Viboud et al., taking a longer historical perspective, suggest that the slower spread and late peaking of the 2001-2002 season is not unique and highlight three other late seasons dominated by infl uenza B (1992-1993, 1973-1974, and 1976-1977). Recent studies, including one by Viboud and colleagues, fi nd that B seasons have different epidemiological characteristics than A/H3N2 seasons, which may explain the late peaking in these years [7,8] We examined peaking during that season using mortality data from the US Centers for Disease Control and Prevention, in order to compare to our estimates from prior seasons [9]. We fi nd that mortality as well as morbidity was bimodal, with larger peaks in January and December, respectively. We thus reaffi rm that 2001-2002, being the latest peaking H3N2 season in over 30 years, is an aberrant season for which airline travel interruption remains the best explanation.
Viboud and colleagues' letter does not take into account that our results are not simply based on an outlier year, nor are they based on a single data source. Rather, we have revealed an important correlation across nine infl uenza seasons. The impact of airline volume on fl u spread does not depend on the 2001-2002 season and remains signifi cant even after its exclusion. Since considering longer time series may provide insight, we repeated our methods on the 30-year mortality time series which Viboud et al. also analyzed [10]. We employed the same spatial aggregation (nine geographic regions) and time series methods as described in [2]. We  The association between airline travel from September to December and time to transnational US spread is displayed. The numbers of traveling domestic passengers signifi cantly predicts transnational infl uenza spread (r 2 = 0.265; p = 0.004). Index spread of infl uenza (blue dots) is estimated from the 99% confi dence interval of the cross-correlation values for the nine major geographic regions of the United States against the national curve for a give year, as previously described [2].
found substantial long-term log-linear trends toward earlier peaking and faster spreading infl uenza epidemics that are correlated with air travel volume (r 2 = 0.460; p < 0.001 and r 2 = 0.265; p = 0.004, respectively, Figure 1). Thus, our new analyses of longer-term data support an effect of airline travel volume.
We strongly caution that other factors may infl uence these trends, including population density, air passenger demographics, ground transportation, and climate. Our design relies on a shorter, more recent time series to avoid confounding by longer term secular trends that may be evident in the 30-year time series. Given the three year backlog of the 30-year dataset, the data Viboud et al. use do not permit the interrupted time series analysis at the core of our investigation.
Reconciliation of our different time series methodologies and datasets should be considered in future research. Nonetheless, because our results were confi rmed with viral surveillance data, we remain confi dent in the robustness of our analysis. We agree that other modes of transportation are important infl uences; our paper makes no claim that air travel is the only mechanism of spread, and we explicitly report that our model explains a portion of the variation in yearly infl uenza spread and peak.
Finally, Viboud et al. emphasize the limited applicability of our fi ndings to pandemics. We agree and have highlighted this limitation in our paper. The decision to restrict travel should be multifactorial. We do hope that it will be evidence based. Our analyses (including those presented here) provide empirical insight into the previously uncharacterized effect of air travel fl uctuation on infl uenza spread. They are one contribution to a small body of investigations that are forming the basis of global policy on fl u preparedness. Though the effect we observe might be smaller under pandemic conditions, the benefi t of a delay is worthy of consideration by scientists and policy makers where lives are at stake and even a short lead time may be of enormous public health value.
We are pleased that Viboud et al. have engaged in a discourse that we hope will strengthen the scientifi c basis of pandemic preparedness. We call on governments, industry, and health care to create a more accessible, freely available, and well-documented data repository for geographically and temporally detailed data on infl uenza [11] and encourage empirical analyses of the dynamics and mechanisms of infl uenza spread. Competing Interests: The authors have declared that no competing interests exist.

Fluoxetine and Suicide Rates: Suicide and the Economy Carlos A. Camargo, Daniel A. Bloch
We wish to comment on the paper by Milane et al. [1] and also refer to the Perspective by Baune and Hay [2] in the June issue of PLoS Medicine on the effect of fl uoxetine prescriptions on the suicide rate in the United States. Milane et al. examined two sets of variables: the number of prescriptions for fl uoxetine in the United States, and the Census Bureau mortality tables with the age adjusted suicide rates for the years 1988 to 2002. The date 1988 is chosen because in that year fl uoxetine was introduced in the US. The authors report that the Spearman correlation coeffi cient between the two sets of variables equals -0.92 with a p-value of less than 0.001. The less suicides, the more tablets of fl uoxetine are prescribed, or vice versa. The least-squares regression line is displayed in Figure 1.
From this simple association they build an elaborate edifi ce, predicting what the suicide trends would have been had fl uoxetine not been prescribed, and they calculate fi gures for "the thousands of lives saved" for both men and women...even though it is not known how many of these prescriptions were for men or for women, whether the patients took the tablets or not, or for how long they took the medication. In addition, the baseline period used to calculate the suicide trend, and thus to predict the future, was arbitrary: from 1960 to 1987, when the suicide rates had a slight gradual increase in the 1970s. Had they used the period 1950 to 1987, a different "trend" would have been obtained, since the suicide rates decreased during the economic expansion of the 1950s [3].
It is widely known that one cannot infer causality simply based on statistical association. Baune and Hay pointed this out and wrote: "In a study like this, it is also important to consider other potential explanations for the fall of suicide rates, such as improvements in the economy..." In this letter we report on the association of other variables with the suicide rates, for we fi nd that the most glaring defect of the Milane et al. article is the total absence of analysis to address likely confounding by many other factors.
Suicide is the fi nal outcome of many conditions, and there have been, for many decades, scholarly articles indicating the many risk factors which increase the likelihood of suicide: poverty, loss of employment, and several other economic indicators have been shown to have a strong effect upon suicide rates. For example, during the Great Depression of the 1930s the rate of suicide rose signifi cantly, and fell when the economy improved and unemployment decreased in the 1940s. On this matter, the literature is quite clear and the references abundant [4][5][6][7][8]. In the 1990s there was a very substantial and prolonged improvement of the US economy [9], which could partially explain a lowering of the suicide rate. We have chosen three economic indicators for the period from 1988 to 2002 and correlated them with the suicide rate, using the Spearman correlation coeffi cient to quantify the strength of the association. The yearly data for the suicide rates and numbers of fl uoxetine prescriptions, for the three economic indicators (Dow Jones average, food stamp rate, and unemployment rate) and for the property crime and burglary rates are all contained in Table 1. The fi ndings are not surprising: The unemployment rate during those years has a strong positive correlation with the suicide rate: r = 0.62, p = 0.014.
The percentage of the US population eligible for the Food Stamp Program, a reasonable indicator of poverty rates, has a stronger positive correlation with the suicide rates: r = 0.84, p = 0.0002.
The Dow Jones industrial average for each of those years, when compared with the suicide rate of the US population, gives an even stronger (negative) correlation: r = -0.98, p < 0.0001 (see Figure 2).
We also calculated the correlation between fl uoxetine prescriptions and the Dow Jones average. Not surprisingly, there is a very strong positive correlation: r = 0.925, p < .0001 (see Figure 3).
We doubt that many will advance the thesis that the increasing sales of fl uoxetine were, somehow, one of the    Table 2.
Given these fi ndings, we decided to explore the relationship of suicide rates with both fl uoxetine prescriptions and Dow Jones averages as potentially associative factors in a single multivariate model. Results are displayed in Table 3. This allowed us to assess the association between fl uoxetine and suicide adjusting for the Dow Jones, an economic indicator. Statistically, the association is quantifi ed with a "partial" Spearman correlation coeffi cient. With this analysis the fl uoxetine association was not signifi cantly correlated with the suicide rate: fl uoxetine had an adjusted Spearman correlation of -0.18 (p = 0.54) whereas the adjusted Dow Jones correlation remained high at -0.88 (p < 0.0001).
In conclusion, we believe that there is little likelihood that the increasing sales of fl uoxetine from 1988 to 2002 were the cause of the modest decrease in the suicide rate during those years. It appears more likely that factors such as those connected with the sustained economic recovery of the 1990s were responsible.   Suicide is a complex outcome which cannot be attributed to a single factor. While we explored an association between suicide and antidepressant prescriptions, we fully acknowledge, as Camargo and Bloch suggest [1], that our work does not fully explain the observed trends. However, in other settings, the same association between increased antidepressant prescriptions and decreases in suicide have been observed. Please refer to the following article, published after ours came out in PLoS Medicine: "Increased antidepressant use and fewer suicides in Jamtland county, Sweden, after a primary care educational programme on the treatment of depression" [2]. I look forward to reading Camargo and Bloch's new article on the association of socioeconomic factors and suicide rates.

Funding:
The author received no specifi c funding for this article.

Competing Interests:
The author has declared that no competing interests exist.

Richard Stallman
The article "How Do Intellectual Property Law and International Trade Agreements Affect Access to Antiretroviral Therapy?" is very useful for its substance, but due to an unwise choice of terminology, it will tend to mislead the public in a way that the authors and editors probably are not aware of, which will promote the sorts of abuse that it seeks to criticize. This results from the use of the term "intellectual property." This article uses the terms "intellectual property law" and "patent law" interchangeably, which is like using "Asia" and "India" synonymously. However, most readers will recognize the latter as loose use of language, doi:10.1371/journal.pmed.0030501.g003 Figure 3. Correlation between Fluoxetine Prescriptions and the Dow Jones Average so they will not really be led astray. Only a few will realize that identifying patents with "intellectual property law" is just as mistaken, so real confusion will result. I ask the editors of PLoS Medicine, and the readers and writers of articles, to be on guard against confusing use of the term "intellectual property"-which means, nearly all use of the term. See http:⁄⁄www.gnu.org/philosophy/not-ipr.xhtml for more explanation of the problems of this term.
Richard Stallman (rms@gnu.org) Free Software Foundation Boston, Massachusetts, United States of America and migration, changing lifestyles, and democratic decentralization have an enormous impact on the above issues. Now is the time for governments and policy makers not only to put feasible plans in place with short-term goals, but also to ensure that we have adequate human resources, strong community ownership, improved information dissemination through the media, investment in health systems research, and lastly, a system that will monitor progress towards a healthy future.
Ramesh Vidavalur (rvidavalur@yahoo.com) University of Connecticut Health Center Farmington, Connecticut, United States of America ideas that back the illusion of aid in the Western civil society. They are rhetorical terms that could even sound radical but, paradoxically, due to their apolitical approaches, make international aid and cooperation harmless for governments and policies that generate injustice. Moreover, postmodernism appeals for individual attention and marginalizes collective political actions that could have greater impact. Solomon R. Benatar wrote in this journal that poor health refl ects systemic dysfunction in a complex world and calls the attention to address the complex system forces that sustain poverty and poor health [2]. We are far from addressing these complex system forces that themselves remain, unfortunately, in good health. At the end, we might need politics rather than the illusion of independent aid.

Funding:
The authors received no specifi c funding for this article.

Competing Interests:
The authors have declared that no competing interests exist.

Guy-André Pelouze
Médecins Sans Frontières is a respectable organisation which helps people in jeopardy and provides medical relief in diffi cult situations all over the world. Despite this recognition I must confess I found the last paper of G. Ooms in PLoS Medicine [1] to develop highly debatable concepts about health, development, and even sustainability. Ooms postulates that sustainable intervention creates a bias for development agencies to maintain the status quo. On the other hand, it could very easily be argued that conventional emergency "humanitarian aid" has failed to improve health on even a medium-term basis in the last four decades. Common sense indicates that populations benefi t more from improvements in wealth, water supply, and agriculture than from consuming free goods from international aid.
Several studies provide a basis for this conclusion [2]. On the contrary, a persistent status quo could result from an exclusive humanitarian approach, precluding the necessary changes to develop health care.
International agencies say that, based on numerous studies, improving health on a medium-or long-term basis is a matter of "sustainable" programmes. But what is meant by "sustainability"? It seems that Dr. Ooms interprets sustainability as durability, a confusion which is obvious when one reads the French version of his paper. Durability is a plain matter of time; pollution, totalitarian regimes, or poverty could as well, unfortunately, meet a single durability criterion. This is the main point, if one wants to assess the criticism which Ooms aims at development agencies. Their main objective is to assess the sustainability of the health programmes which means (in French the supportabilité and not the durabilité) whether these programs exceed the economic, organisational, and ecological possibilities of the country and its population. And I recognize that I am in keeping with that! To give rice to people who usually consume rice is helpful, but next time it will be even more helpful to give them the means to grow their own! Yet the reality is far more complex, something which requires that agencies carefully assess the different programmes [3]. I must add that such an approach does not preclude an increase in annual governments outlays. But before increasing expenditure, it is wise to assess whether the programmes is working, and for whom. I don't miss the point that sustainable aid is for certain governments synonymous with conditional aid, and such diffi culties must not be hidden, but every one of us is able to make the distinction between the principle and some penny-pinching, restrictive policies which can be amended and reversed.
Indeed, the two approaches are complementary. When an emergency situation arises, it is obvious that some of the critical health issues of the local populations could be addressed by emergency international aid. But after a few weeks, only structural and political changes (that is, peace, convenient water supply, agriculture revival, affordable energy, information, free trade, free enterprise through microrenting...) are crucial for maintaining health and eventually improving it. As a matter of fact it could be a more dangerous illusion for these endangered populations to give credit to the ideas that Dr. Ooms develops in order to justify the spending of more public funds to extend emergency humanitarian aid indefi nitely.

Spinal Delivery of p38: TNF-alpha Inhibitors Edward Tobinick
The new study by Boyle and colleagues provides important data on basic science mechanisms involved in pain and infl ammation [1]. Their data, along with that from previous studies, provides further basic scientifi c evidence documenting p38-TNF-alpha interactions, and suggests that spinal p38 or spinal TNF-alpha blockade may have clinical relevance [1,2]. The present study documents that p38 activation may be occurring predominantly in microglia. The present study, therefore, joins other recent work which suggests the importance of p38-glial-TNF-alpha interactions in neuroinfl ammation and synaptic signaling [3][4][5][6]. This increasing evidence may have clinical relevance not only to arthritis pain, but also to the pathogenesis of various forms of neuropathic pain and Alzheimer disease [1][2][3][4][5][6][7][8].
Because the present study suggests that spinal delivery may be more effective than systemic delivery when attempting to intervene in spinally-mediated infl ammatory mechanisms, the authors note the potential importance of developing compounds that may bypass the blood-brain barrier. The present author speculates that the rapid and signifi cant clinical effects noted following perispinal administration of etanercept in small pilot studies suggest that perispinal administration of p38 inhibitors may also allow these compounds to reach the spinal cord and dorsal root ganglia in therapeutically effective amounts [7,8]. It is hypothesized that this may be possible via passage through the vertebral portion of the cerebrospinal venous system, and this may explain the effi cacy of perispinal etanercept in the above cited studies [7][8][9]. Previous studies have documented that epidural administration of large molecules may result in delivery to the endoneurial space [10]. This evidence, along with the basic scientifi c evidence provided by the present study of the potential clinical importance of spinal delivery, supports consideration of investigation of this novel route of administration.