The NIS data provided a total of 5,172 LD cases from 447,132 hospitalizations for bacterial pneumonia (Table 1). After applying the discharge weights to produce a national-level estimate and before applying any exclusion rules other than reporting month and year, the NIS series compares favorably to CDC’s reported monthly LD estimates (23), with a correlation of 0.74. After applying exclusion rules (age >18 years and provision of a complete set of predictor variables), we had data on 3,005 LD cases and 189,412 hospitalizations for pneumonia. The most common reasons for exclusion were lack of the AHA ID, admission month and year, and race, because some states elected not to report these variables.

We compared several demographic and severity measures between the portion of the data dropped because of missing values and the proportion retained (Table 2). Among the cases, the only substantial difference was in the percentage of patients not insured; for 10.8% of dropped records, no insurance was reported, compared with 7.4% of those used in the model. Many statistically significant differences existed between the controls retained and those lost; however, the large sample size (retained sample n = 189,412) meant many non–clinically relevant differences would be statistically significant.

In general, hospitalized persons with LD were younger than those with other bacterial pneumonia and more likely to be male (Table 3). We found a large unadjusted difference in the mean monthly environmental temperatures between cases (58.5°F) and controls (52.6°F), and monthly environmental relative humidity was higher on average for cases (70.0%) than for controls (67.3%). Additionally, there was nearly 20 mm more rain for cases (80.4 mm) than for controls (61.7 mm).

In the regression analysis, the primary variables of interest—mean temperature, mean relative humidity, and their interaction—were all significant (likelihood ratio test against model with only main effects, χ² test statistic = 350.42; p<0.0001) (Table 4). Although total rainfall was independently a risk factor for LD, the effects of temperature and humidity were still significant.

The combination of the temperature and humidity main effects and interactions can make understanding the combined effect difficult. For this reason, we separately reported the estimated composite odds ratios for each combination (Table 5). Additionally, because odds ratios can be difficult to interpret, we provide the expected probabilities for a 61–70-year-old white man on Medicare admitted to a hospital at 42°N in April 2011 at the 3 different rainfall levels (Table 6). The relationship between temperature and relative humidity exhibits the Goldilocks principle: when it is too hot (>80°F) or too cold (<60°F), the odds of LD do not vary with humidity, but when the temperature is “just right” (60°–80°F), the odds have a dose-response pattern with humidity. The largest effect of this relationship between temperature and humidity was evident for warm and very humid months across all 3 rainfall levels.

We determined the monthly percentage of bacterial pneumonia discharges for which an LD diagnosis had been given within HCUP by US Census region (Figure 2). Percentages were relatively high in the Northeast and somewhat lower in the Midwest and South. The frequencies in the West were the lowest of all 4 regions and appeared not to be seasonal. The changes around 2002–2003 in all of the series are present in other data sources (23). The exact cause is unknown but is thought to be related to increased vigilance, testing, and reporting of atypical pneumonia after the outbreak of severe acute respiratory syndrome (24).

We also determined the probability of a case of bacterial pneumonia being diagnosed as LD in 2011 using the local weather data. We restricted this prediction to the states used to estimate the model (Figure 3). We set the nonweather, nonlocation covariates to their modal values for patients hospitalized with bacterial pneumonia (white 61–70-year-old man on Medicare) and used the weather station latitude and monthly average temperature, humidity, and rainfall. The estimated probabilities of LD (Figure 3) are the fitted values from the model described by Table 4 and these covariate values. Since the ISD is national in scope, we extrapolated from the model estimated using HCUP data to the entire United States, including non-HCUP regions (Figure 4). The risk for LD varied considerably by location (low risk along the Gulf Coast and relatively low risk in the West) and calendar month (high–risk areas such as the Mid-Atlantic region are only actually at high risk during June–September and are at low risk during December–April) (Figure 4).

Our results suggest that the incidence of LD varies considerably by season and local weather patterns. Specifically, LD is more likely to occur in warm (60°–80°F) and very humid (>80.0%) months. For example, the odds of LD being diagnosed in a pneumonia patient during a month when the rainfall is <18 mm and the temperature is 60°–80°F was 3.1 (1.70/0.55) times higher when the relative humidity was >80.0% than when it was <50.0%. When rainfall amounts were greater, the risk also increased; however, regardless of rainfall, warm and humid weather was a major risk factor. Also, we found a dose-response relationship between relative humidity and the odds of an LD diagnosis during periods of warm weather. In contrast, hot, cool, or dry weather patterns produce no meaningful increase in LD.

Previous work has demonstrated seasonality and the effects of weather patterns on LD (1,11–13,15,16). However, much of this work was based on regional investigations where LD is common. Regional investigations are limited in their ability to more fully describe the relationship between weather patterns and LD incidence. For instance, in the Rocky Mountains or the US Southwest, community-associated LD is comparatively rare, and the rate for LD is much lower than what would be expected given patient factors. In contrast, the Mid-Atlantic has a higher-than-expected risk during certain months of the year.

Using the national scope of our data, we estimated the differences that weather has on LD risk. The predicted probabilities of LD varied from 0% to 1% of all cases of bacterial pneumonia in the Southwest to nearly 8% during warm, humid, and rainy summer months in the mid-Atlantic. Weather appears to drive these differences because our model had no geographic information other than latitude. This finding is consistent with prior regionally and seasonally limited studies (11–13,15,16). Areas previously studied (e.g., the mid-Atlantic) vary considerably in risk depending on recent weather. Our results are biologically plausible because L. pneumophilia thrives in warm, wet environments (25), which support not only the pathogen’s survival but also the existence of aerosolizations. In contrast, conditions are not as supportive for the pathogen in dry or excessively hot environments (1,18).

LD is difficult to diagnose on the basis of clinical manifestations alone (26). Furthermore, rapid diagnostic tests do not cover all strains (27), and some tests have relatively low sensitivity (28). More definitive culture results may take 3–5 days after therapeutic decisions are needed (1). Thus, incorporating local weather conditions into clinical decision-making ultimately might help increase or decrease clinical suspicion for LD, especially when combined with diagnostic testing. Current US CAP guidelines recommend empiric therapy routinely covering atypical pneumonias (29). Results of a recent noninferiority trial suggest that monotherapy with a β-lactam, aggressive diagnostic testing, and use of clinical judgment may safely avert the use of fluoroquinolones or dual therapy with a macrolide in patients with CAP (10). However, the same trial replicated elsewhere with a higher rate of LD might yield different results. Another study investigating the effect of a β-lactam alone versus a β-lactam with a macrolide found delays in clinical stability for persons treated with only 1 agent, but the authors failed to show that the β-lactam alone was not inferior (30). Our model suggests that warm, humid, and rainy summer months in the mid-Atlantic may exhibit predicted probabilities of LD of nearly 8%. Accordingly, abandoning initial empiric coverage for LD might yield a differential effect on outcomes depending on season, region, and weather, and treating all CAP cases for atypical pneumonia in areas and seasons when LD is relatively uncommon may result in the excessive use of antimicrobial agents.

The antimicrobial drugs most commonly used to treat LD include either a fluoroquinolone or a macrolide (with a β-lactam), and resistance has increased for both (10,31–34). Thus, treating for LD only when and where risk is higher, along with increased diagnostic testing and good clinical judgement, may help reduce antimicrobial drug use, providing a new antimicrobial drug stewardship target. The temporal, climatologic, and geographic variations in LD risk emphasize the potential importance of regionally relevant guidelines. Basing treatment guidelines on estimates in high- or low-risk areas will lead to overuse or underuse of LD treatment for CAP. However, future work with more detailed clinical information is required to determine when and where antimicrobial drug use may be initially restricted and in which patients, without significantly impairing the quality of care. Specifically, future work will need to determine under which conditions and in which patients initial monotherapy with a β-lactam will not be inferior to combination therapy. Furthermore, future work will need to consider and incorporate diagnostic testing for LD and the patient’s clinical status (e.g., intensive care unit admission).

Our work has several limitations. First, our study uses administrative data for inpatient hospitalizations. Some CAP and LD cases are treated on an outpatient basis. Furthermore, state participation in the NIS is voluntary, and some states omit key variables (e.g., AHA IDs). Although we found no meaningful differences between the parts of the sample we retained and parts we dropped, it is possible that participation/reporting varies nonrandomly. In addition, we do not have information about medications, test results, or microbiology data, so we cannot confirm an LD diagnosis or determine whether and what microbiological testing was performed. Thus, future work should incorporate alternative case-finding approaches, including more granular information about cases, specifically the CDC legionellosis database. Alternative data sources may also enable the generation of more granular geographic estimates and age-based estimates of disease risk. We also are unable to confirm without more clinical data the extent to which having a primary diagnosis of pneumonia correlates with CAP for the control patients in our analysis. Second, some of the geographic differences in the predicted probabilities may be due to differences in propensity to test for, and therefore diagnose, LD. Third, the NIS reports only the month of admission. Thus, we aggregated weather information by month. Future work needs to consider more granular (e.g., daily) data. Fourth, we consider the weather around a hospital, not the weather experienced by the patients admitted to the hospital. Fifth, meteorologic variables are interdependent: relative humidity depends on temperature because the maximum amount of water suspended in the air rises with the temperature, and our model may inadequately specify these relationships. Finally, our analysis did not contain possibly important geographic differences (e.g., use of monochloramine in municipal water).

The experience of warm and humid weather patterns common during summer resulting in substantial increases in LD might have driven the current view about the frequency of LD in the United States. Our results demonstrate the need to investigate the effects of incorporating recent weather patterns, particularly wet, warm, and humid weather, as an additional consideration in the clinical decision-making process for CAP. Our results suggest that the risk for LD is highly related to temperature and humidity regionally. We found locations where LD relatively rarely causes hospitalization for CAP, such as the Southwest and Rocky Mountains, but also the Mid-Atlantic region during the winter. Information about weather exposures for patients also should help inform the design and interpretation of CAP-treatment trials. Future work examining more granular environmental data may ultimately enable clinicians to safely limit initial empiric antimicrobial drug selection for CAP to monotherapy with a β-lactam in specific seasons and regions.