INTRODUCTION Pneumonia is not only the world’s leading cause of death among children (1), but it is also the leading infectious cause of death among adults (2). Immunization of infants using pneumococcal conjugate vaccines has reduced invasive pneumococcal disease (IPD) and hospitalization for pneumonia in children in randomized trials (3–5). Since its introduction in the United States in 2000, seven-valent conjugate pneumococcal vaccine (PCV7) has reduced IPD dramatically, including in unvaccinated age groups, through induction of herd immunity (6–10). Most of the burden of pneumococcal disease, however, is not from IPD but nonbacteremic pneumonia in adults 65 years old and older, that is, pneumococcal pneumonia without a positive culture from a sterile site. It is biologically plausible that interruption of transmission of vaccine-type Streptococcus pneumoniae can reduce this burden. A groundbreaking prospective observational study through 2004 on the Health Care Cost and Utilization Project (HCUP) Nationwide Inpatient Sample, a 20% sample of hospitalizations nationwide, found a significant reduction in ICD9-coded pneumococcal pneumonia in infants 1 year of age. We converted these annual coverage estimates to seasonal estimates for the 1999–2000 through 2005–2006 winter seasons by averaging coverage in each pair of years (e.g., we took the average of the 2002 and 2003 coverage to arrive at the coverage for the 2002–2003 season). ICD9-coded outcome definitions and rate comparisons. For each state and season, we extracted the numbers of hospitalizations for each outcome from among all of the listed diagnosis codes. We defined IPD as ICD9 code 320.1 or 038.2 or as codes 320.8, 790.7, or 038.9 and 041.2; pneumonia with diagnosed S. pneumoniae infection as ICD9 code 481; and all-cause pneumonia as ICD9 codes 480 to 486. Nonbacteremic pneumococcal pneumonia cases were defined as those with any mention of an ICD9 code 481 diagnosis but without mention of a diagnosis of IPD as defined above. We scanned across discharge diagnoses in each patient record for any mention of these disease codes. We used the disposition information at discharge to identify the subset of hospitalizations for each outcome that resulted in inpatient death. We constructed time series of each outcome for each state and for six age groups, <2 years old, 2 to 4 years old, 5 to 17 years old, 18 to 39 years old, 40 to 64 years old, and ≥65 years old. Baseline rates before introduction of PCV7 were defined as the average annualized rates during the 1996–1997 through 1998–1999 seasons; incidence RR estimates and 95% CIs were calculated using outcome-specific Poisson regression models. Attributing pneumococcal and influenza-associated pneumonia hospitalizations. Because influenza virus infection is rarely confirmed by laboratory testing and because a triggering influenza virus infection is often resolved by the time a patient presents with secondary complications such as bacterial pneumonia, it is not possible to directly assess the influenza disease burden. Modeling the pediatric burden of influenza is further complicated by the concurrent impact of RSV during winter months. Moreover, because most pneumonia hospitalizations are not linked to a specific pathogen, pneumonia due to infection with S. pneumoniae is frequently not recorded as such on hospital discharge forms. To overcome these limitations, we applied a Poisson regression modeling strategy to monthly time series of outcome incidences per 100,000 in order to estimate the seasonal influenza-related and pneumococcal pneumonia-associated burdens in our six age groups. Our strategy was similar to that used by Thompson et al. (16) to assess influenza-related pneumonia. However, instead of weekly laboratory virus surveillance data for influenza and RSV epidemic patterns, we used ICD9-coded counts of influenza (ICD9 487, any mention) and RSV infection (ICD9 480.1, any mention) as explanatory variables, similar to Pitman et al. (22); we also included S. pneumoniae pneumonia (ICD9 481, any mention) as a third explanatory respiratory pathogen variable. A linear trend and a sinusoidal wave component accounted for seasonality and secular trends not captured by the “pathogen” explanatory variables. The best-fitting model was of the form AC pneumonia = exp [β0 + β1(month) + β2(influenza) + β3(RSV) + β4(pneumococcus) + β5sin(2 Π t/12) + β6cos(2 Π t/12)] where AC pneumonia is all-cause pneumonia, defined as ICD codes 480 to 486 as the primary cause (removing records with first-listed ICD9 codes of 480.1 and 481); influenza (ICD code 487), RSV (ICD code 480.1), and pneumococcus (ICD code 481) are the monthly rates of hospitalizations specifically associated with each outcome in each state (all ages, any mention); and month is the running month variable. The cyclical terms track additional seasonality in the pneumonia data. We computed fractions of all-cause pneumonia attributed to influenza virus, RSV, and S. pneumoniae for all available state/age group time series. For most state/age group combinations, all of the variables in the model were significant. However, we did not change the model form to accommodate states or age groups in cases where not all explanatory variables were significant at the P < 0.05 level. If any parameter value was less than zero, we set the number of attributed cases to zero. And finally, we summed the model attribution and the ICD9-coded attribution to generate the total attributed fraction of all-cause pneumonia to S. pneumoniae, RSV, and influenza virus, respectively. Figure S3 in the supplemental material shows a typical model fit, here for data from children 2 to 4 years of age in New Jersey. Modeling reductions in pneumococcal disease burden associated with PCV7 use. We first constructed Poisson regression models to assess the effect of PCV7 coverage on hospitalization and in-hospital mortality. We analyzed time series extracted from SID data from 10 states spanning the 1996–1997 through 2005–2006 seasons, with the exception of Utah, for which data were unavailable for the 1996–1997 season. We constructed time series for six age groups, younger than 2 years old, 2 through 4 years old, 5 through 17 years old, 18 through 39 years old, 40 through 64 years old, and 65 years old or older. RRs and accompanying 95% CIs were calculated to represent the association between 10 percentage point increments in PCV7 coverage and outcome rates. For hospitalization rates, we ran age-specific Poisson regressions on monthly time series for each state and age group. The model form that best fit the data was Y / N = exp { β 0 + β 1 [ sin ( 2 t π / 12 ) ] + β 2 [ cos ( 2 t π / 12 ) ] + β 3 ( PCV ) } where Y is the number of hospitalizations during a particular month for a specific age group, N is the population offset, and t is the calendar month, β0 yielded the intercept while β1 and β2 accounted for seasonal changes in hospitalizations and β3 accounted for the effect of PCV7 coverage in children <5 years old. For in-hospital mortality rates, we applied the same model form to seasonal counts of cases in which a patient had both a diagnosis code associated with each outcome and a discharge status code indicating that the patient died in the hospital. We used the results of these analyses to estimate cumulative national reductions in disease and mortality burdens associated with PCV7 use in each age group. For each outcome, we first used U.S. census data and the aggregate seasonal rates for the 1996–1997 through the 1998–1999 seasons from 10 states to estimate the national burden. Then, for each season and age group, we used national PCV7 coverage, the RR per percentage point increment in coverage, and U.S. census data to estimate the PCV7-associated change in burden for each season and summed these reductions for the 1999–2000 through the 2005–2006 seasons. Single-season analysis of IPD and model-attributed influenza-related pneumonia. Using SID data from 10 states for each age group, we determined the ratio of the rates of IPD and model-attributed influenza-related pneumonia in each post-PCV7 season compared to a baseline determined by the average of rates in the 1996–1997 through the 1998–1999 seasons. We then used Poisson regression to model the relative rate reduction associated with a 10 percentage point increase in PCV7 coverage and calculated two-sided P values to test the hypothesis of no effect of vaccination. SUPPLEMENTAL MATERIAL FIG S1 Monthly rate of hospitalization with a pneumococcal pneumonia diagnosis (ICD9 code 481) in New Jersey among adults ≥65 years old, 1994–1995 through 2005–2006 seasons. Download FIG S2 Total PCV7 coverage estimates, including both normal-schedule and catch-up vaccinations, versus season for U.S. children less than 5 years of age in 10 states. Download FIG S3 Model fit for attribution of pneumococcal pneumonia, RSV, and influenza-related pneumonia attributions for infants <2 years old and adults ≥65 years old. Download FIG S4 Scatterplots of state hospitalization RRs (season/baseline) versus state PCV coverage for the 1999–2000 through the 2003–2004 seasons for (a) IPD among those <2 years old and (b) IPD among those ≥65 years old. The lines represent exponential fits to the data from each season, showing the trend. An asterisk next to the year in each panel indicates a significant trend. Download