The main wave of the 2009 (H1N1) pandemic infected many more children than it did adults. These differences are not explained by baseline antibody titres to H1N1pdm, but could be explained partly by social mixing patterns of the population in these different age strata. However, given that social mixing patterns within the 20–60-y age range do not exhibit substantial variation 
, and that we have controlled for the presence of a child in the household, it is plausible that increasing age leads to decreased susceptibility independently of mixing and titres to H1N1pdm, possibly as a result of repeated seasonal influenza infections, but by a mechanism not detectable by assays for neutralizing antibody. Whether this reflects antibody that protects by mechanisms other than neutralization, such as antibody-dependent cell cytotoxicity or cell-mediated immunity, remains worthy of investigation.
While older adults had low infection rates, those individuals infected developed severe disease much more frequently. Further, our results suggest that individuals over 60 y experience very high absolute rates of severe outcomes, with approximately one reported and positively tested death for every 200 infections. If continuing waves of H1N1pdm infection are driven by antigenic drift, and if that drift decreases the efficiency of the cross-protection currently possessed by older adults, it is likely that future waves could have higher overall mortality than initial waves. Surveillance of clusters of severe disease in older adults should be prioritized because this may be the first clear signal of a significant antigenic evolutionary event. The efficacy of alternate vaccine formulations in preventing infection in older individuals should be assessed as a matter of priority 
The low rate of ILI reported by phone and symptom diary for seroconverters in this study is consistent with results from an independent parallel study of household contacts of children in Hong Kong 
. However, much higher rates of symptoms were reported in military personal in Singapore 
. These differences may be driven by the age distributions of the different cohorts. The Singapore cohort was much younger: our comparison of symptom reporting by age suggests that reduced level of symptom reporting observed in adults was likely due to a reduced rate of experiencing symptoms, rather than just to a reduced propensity to report. The proportion of uninfected children that reported febrile illness was similar to that of uninfected adults. However, our results do suggest that adults are slightly less likely to report symptoms during a follow-up interview than are children. In general, rates of reporting were much higher but less specific when participants were asked a direct question during the follow-up interview compared with more passive symptom diaries or participant call-in. Future studies attempting to address rates of illness associated with influenza infection should attempt more intensive prospective follow-up of participants to minimize potential recall bias.
Our study has a number of limitations. Firstly, we did not measure incidence in children aged 2 and lower, who are much more likely to be admitted to hospital for acute respiratory infection than other age groups, but less likely to be infected with pandemic influenza than older children 
. Given our focus on the use of paired sera, this shortcoming was unavoidable. The incorporation of data from cross-sectional samples from young children is the topic of ongoing investigation. Also, our cases are defined by an observed 4-fold rise in titre, which may not have occurred for all infections; and we will not have captured all severe cases as some H1N1pdm cases will have entered the private hospital system (although this is likely to have occurred much more rarely than the average rate of ~10% for all types of admission). We suggest that the combined effect of imperfect test sensitivity and imperfect severe case matching will have generated only minor biases in our estimates of infection rates and severe disease rates and that the two effects will have acted in opposite directions.
In order to extrapolate from the period of our study to the full period of the pandemic in Hong Kong, we made the assumption that the testing process for individuals who became hospitalized was consistent. This is almost certainly not the case for all hospitalized individuals. In particular, anecdotal evidence suggests that less severe hospitalized cases were less likely to be tested for H1N1pdm after the end of September. Analysis of the rate of admission to ICU per positive hospital admission supports the anecdotal evidence (unpublished data). Therefore, it is reassuring that estimates of the overall and age-specific attack rates based on the three different outcome measures (hospitalization, admission to ICU, and death) are largely consistent.
We cannot exclude the possibility of substantial sampling bias in our serological survey. We were only able to successfully obtain paired sera from an average of 1.6 individuals in ~2% of households initially identified by random telephone number selection. Although similar in many respects, after using common demographic characteristics to compare the study population with the wider Hong Kong population (age, sex, district, and education), we cannot exclude the possibility that individuals more likely to take part in our study had a different probability of infection than the population at large. However, we suggest that the potential impact of sampling bias in our results (and the value of evidence presented here in general) should be assessed on a result-by-result basis against the background of other reported community surveys of the 2009 influenza pandemic.
For the 2009 Hong Kong pandemic season, the current results add substantially to our earlier work 
in a number of ways. Using an entirely different sampling scheme, the current study confirms the general pattern of sharply decreasing age-specific rates of infection for a similar period of the epidemic, for the age ranges contained in both studies: thus strengthening the case for rapid cross-sectional serological studies on the basis of convenient samples 
at the same time as suggesting that serious sampling biases were not present in either study. Also, to a certain extent, multiple recruitment groups in the current study provided a proxy for propensity to take part: those recruited via the parallel study had already agreed to complete one telephone questionnaire and to be contacted again for other studies. Having agreed for a third time to take part, by enrolling in the main serological survey, it seems reasonable to assume that parallel study recruits are from a subset of the population more likely to take part in this type of study. Although our univariate estimate of the odds ratio for the parallel recruitment group was greater than one, compared with the direct group, the inclusion of recruitment source as a covariate did not improve the parsimony of our multivariate regression model. Also, the strength of the odds ratio for the parallel group in the univariate model was reduced substantially when adjusting for our epidemiological variables of interest. Therefore, although it is certainly possible that propensity to take part in the paired serological study was correlated with the risk of infection, comparison with our previously published cross-sectional study and comparison of our two recruitment groups suggests that sample bias was of considerably lesser influence on infection than variables we were able to measure directly, such as age and the presence of a child in the household.
The current study, by recruiting from a wide age range using the same sampling framework and a paired sera outcome, allows us to add to the available literature in a number of other ways. We present important data on infection rates and severity for those aged 60 y and older that were not reported in our previous study 
: even the single observed 4-fold rise in titre out of 131 paired samples is valuable. In an age group typically at high risk from influenza morbidity and mortality, these data allow an informative upper bound for the absolute risk of infection and, hence, an informative lower bound for the absolute risk of severe outcomes. Without paired samples, obtaining accurate bounds for estimates of low rates of incidence from cross-sectional samples is problematic. When trying to estimate low rates of incidence with cross-sectional data, statistical noise becomes significant in the numerator: to overcome this noise, large sample sizes are required. Good evidence for high absolute risk in a particular age group may be of substantial public health value for the prioritization of interventions.
With good data on other potential risk factors for infection, we were able to show how the presence of a child in the household could explain an apparent age plateau in risk of infection, while variables such as education and profession did not appear to be risk factors once adjusted for age. This type of traditional risk-factor analysis is not possible with unlinked samples for which only the following variables are usually available: age, sex, and clinic location. Similarly, our analysis of home district (using only a single clinic location) suggests that micro-scale spatial heterogeneities persisted for longer than might have been expected in a large well-connected population. For the period of our study, residents of one district (New Territories East) appeared to be at substantially greater risk of infection than were residents of other districts. It is possible that the overall level of transmission was higher in that one district than in other districts, or that the epidemic occurred sooner there than it did elsewhere. Further, it seems possible that, in Hong Kong, spatial decorrelation took a long time to occur or never did occur. Individual-based models of respiratory infections, parameterized with the commuting patterns of adults 
, and also those parameterized with explicit school locations 
, suggest much more rapid spread at small scales in large populations. Had the pandemic strain been more severe, good knowledge of small-scale spatiotemporal patterns could have been of value in optimizing the provision of key health care facilities and the timing of rolling school closures.
Our results can be compared with serology-based studies of influenza incidence in other populations during the 2009 (H1N1) pandemic 
. In England and Wales, a study of cross-sectional clinical samples found substantial increases in the proportion of younger children with titres 1
32 or greater between a 2008 baseline (n=
1,403) and sample taken in September 2009 (n=
1,954), thus giving valuable early evidence that the infection attack rate was high in some age groups and, hence, that the rate of severe cases per infection in the most affected age groups was likely to be low 
. As already mentioned above, the consistency of our results for Hong Kong between the current study and our previous cross-sectional study 
validates the use of convenient clinical samples during the early stages of a pandemic as a useful tool for the estimation of incidence in high-incidence groups. However, a high degree of cross-reactivity in hemagglutination inhibition assays in sera from adults in many studies introduced considerable statistical noise and prevented reliable estimates of attack rates in older age groups. In Singapore, sera were collected from four groups: an existing sample of healthy adults (n=
838), military personnel (n=
1,213), staff from an acute care hospital (n=
558), and staff and residents of a long-term care facility (n=
. Although an overall infection attack rate of 13% was observed in this study, it is difficult to generalize these results because no subgroup contained school-aged children and many of the infection events occurred in the military substudy.
It is more difficult to compare our community-wide results with historical studies such as the Tecumseh 
and Seattle 
study. Both were designed to efficiently obtain viral samples from households with children and, therefore, did not attempt to recruit from childless households. Also, information on their precise sampling framework for households with children is difficult to obtain. However, future analyses of household-level data from the current study and follow-up waves should permit a like-for-like comparison between the subgroups in the Hong Kong study and canonical historical studies of respiratory infection.
Our simulation results show that a larger paired-sera cohort study with a shorter follow-up period could have generated—more rapidly—similar data to those presented here. We suggest that this revised design would be a valuable addition to revised pandemic preparedness plans for a small subset of large well-connected global cities. Sentinel hospitals could be established in early-affected populations to help ensure that the testing process remains consistent for ICU cases throughout the epidemic curve. Given that (a) many believe the 2009 response to have been overzealous and (b) the severity of the next pandemic strain is not known, there appears to be a substantial risk that the public health impact of the next pandemic will be underestimated. Therefore, revised preparedness plans should prioritize reactive studies that can rapidly and reliably distinguish between 2009 (H1N1)-like strains (~1
10,000 infection fatality rate) and more severe pandemics. If the next pandemic strain were similar in all other respects, but had an infection fatality rate of ~1
1,000; we could reasonably expect peak demand on key health care services such as ICU to be ten times greater than that observed during 2009/2010