Associations between Legionellosis and meteorological variables in Ohio, 2001-2023

Abstract

Legionellosis, a respiratory disease caused by Legionella bacteria, has increased in incidence across the United States, yet the environmental factors influencing its rise remain unclear. We examined monthly Legionellosis case counts from 2001 to 2023 in four populous Ohio counties using Distributed Lag Nonlinear Models (DLNMs). The models assessed delayed and nonlinear associations between disease incidence and meteorological variables, including temperature extremes, precipitation, humidity, soil moisture, and large-scale climate indices. Significant lagged relationships were found for soil moisture, relative humidity, temperature, and oscillation indices, though associations and delayed effects varied based up geography. Elevated soil moisture and humidity were most consistently linked to increased risk. These results indicate that both local weather and broader climatic conditions influence Legionellosis patterns in Ohio. As climate change intensifies humidity and precipitation extremes, the environmental conditions favoring Legionella growth and transmission are likely to become more frequent, underscoring the importance of climate-informed public health surveillance.

1. Introduction

Legionellosis is a pulmonary infection caused by Legionella species, which are intracellular, Gram-negative bacteria. Two distinct clinical forms of the disease are recognized: Legionnaires' disease and Pontiac fever (Braeye et al., 2020; Han, 2021) both of which are characterized by specific symptoms. Legionnaires' disease is a form of pneumonia that cannot be clinically distinguished from other strains, and it poses a high risk to individuals aged 50 years or older or to those with a compromised immune system (Parr et al., 2015; Wade & Herbert, 2024). Pontiac fever, on the other hand, is a self-limiting extrapulmonary infection with symptoms similar to influenza, which is often underdiagnosed and underreported (Neil & Berkelman, 2008). Legionellosis primarily occurs through inhaling aerosols or soil dust containing the bacteria, and it was first identified in 1977 following a severe outbreak of pneumonia in the USA in 1976 (Yao et al., 2024). The bacteria are present in low numbers in freshwater sources such as lakes and streams, but they can proliferate in artificial water systems containing warm (25–45 °C), or stagnant water with inadequate disinfectant. Systems such as cooling towers, showers, hot tubs and plumbing have been identified as sources of infection (Barskey et al., 2022). Biofilms and free-living amoebae support their survival in these environments (Abdel-Nour et al., 2013; Pereira et al., 2021). Therefore, these engineered water systems are the main way that people are exposed to the bacterium and are closely linked to outbreaks of Legionellosis (Margot et al., 2024).

Legionellosis is an increasingly prevalent notifiable public health condition in the United States (Moffa et al., 2023). Reported cases of Legionnaires' disease have increased over the past two decades, peaking around 2018. Cases also spiked during lockdowns associated with COVID (Kunz et al., 2023; Rhoads & Hammes, 2021). Surveillance data indicates that 75–90% of cases occur sporadically, with sources rarely being identified, and outbreaks rarely being linked to an identifiable source (Moffa et al., 2023). The CDC estimates that around 6000 cases of Legionnaires' disease are reported in the US each year, with similar upward trends observed in other parts of the world, such as Europe and Australia (Capon et al., 2023; Kanarek et al., 2022). Given the global significance of the upward trend in outbreaks, closer attention is required at both national and regional levels, including in states such as Ohio, where the number of cases of Legionnaires' disease is increasing annually (Legionella).

Distributed lag nonlinear models (DLNMs) offer one approach for understanding increases in disease cases over time, especially as they relate to meteorological and climatic factors. DLNMs are a type of time series regression models that model non-linear, delayed effects between predictor variables and a response; these relationships are often seen in the health impacts of environmental exposures, such as air pollution or climatic factors, on disease incidence (Luo et al., 2022; Miley et al., 2020; Zhang et al., 2024). Researchers frequently use this in time series analysis of epidemiological research to understand the nonlinear lagged responses that are often overlooked in conventional linear models (Gasparrini, 2011). DLNMs have been used to study various diseases in different parts of the world. Examples include the impact of temperature variables on hemorrhagic fever with renal syndrome in Shandong Province, China (Xu et al., 2018); ambient air pollution in Wuxi, China (Luo et al., 2022); the impact of temperature and precipitation on hepatitis A in Seoul, Republic of Korea (Zhu et al., 2017); and the study of mortality risk during extreme temperature events in US metropolitan areas (Allen & Sheridan, 2018). However, the use of DLNMs in investigating legionellosis has been relatively limited and geographically restricted, with studies in Belgium (Braeye et al., 2020) and Switzerland, as well as a study of Medicare patients in the U.S. (Wade & Herbert, 2024). More studies are needed to identify associations with rapidly increasing cases in other locations. Extending the application of DLNMs to legionellosis can therefore aid in clarifying the ecological factors contributing to its increasing prevalence and reinforcing the evidence base for prevention and policy strategies.

In this study, we use DLNMs to explore how environmental conditions influence Legionellosis incidence across Ohio's four most populous counties. We hypothesize that both short- and long-term variations in meteorological and hydrological factors. Specifically, we aim to (1) identify environmental variables exhibiting significant delayed or nonlinear relationships with Legionellosis cases, (2) quantify the lagged influence of climatic and hydrologic variability on disease risk, and (3) evaluate how these associations differ across counties that vary in geography and climate.

2. Methods

2.1. Study area

We analyzed delayed and nonlinear associations between environmental factors and Legionellosis incidence across the four most populous counties in Ohio (Fig. 1). Franklin County, located in central Ohio, includes the state capital Columbus and features a mix of urban and agricultural land cover. Cuyahoga County, situated along Lake Erie, encompasses the city of Cleveland and is predominantly urbanized with substantial industrial and residential areas. Lucas County, also bordering Lake Erie, contains the city of Toledo and exhibits a mixture of urban, industrial, and agricultural land uses. Hamilton County, located in southwestern Ohio along the Ohio River, contains the city of Cincinnati and is characterized by an urban core surrounded by agricultural and industrial land uses. Franklin County's population increased by 24.1% between 2000 and 2023, while the other three counties experienced modest population declines over the same period (Table 1).

Fig. 1.

Study area map.

Table 1.

Population estimates for four counties in Ohio between 2000 and 2023 (source US Census Bureau).

County	2000	2023	Percent Change
Franklin	1068978	1326063	+24.1%
Cuyahoga	1393978	1233088	−11.5%
Hamilton	845303	827058	−2.16%
Lucas	455054	425484	−6.5%

2.2. Data

We obtained Legionellosis case data from 2001 to 2023 from the Ohio Department of Health. We summarized the case counts by month for each of the four counties (Fig. 1). A total of 213 occurred in Franklin County, 177 in Cuyahoga, 51 in Hamilton, and 24 in Lucas. Legionellosis cases increased in all four counties from 2012 to 2023 in comparison to 2001-1001 (Fig. 2). Monthly mean case counts during 2001-2011 were 4.5 (Franklin), 3.5 (Cuyahoga), 0.5 (Hamilton), and 0.5 (Lucas), while during 2012-2023 they were 9.2, 9.1, 1.9, and 1.1, respectively. Legionellosis incidence generally peaked during summer (June to August) and was lower during winter (December to February), although wintertime cases occurred at least occasionally in all counties and more frequently during the latter decade in comparison to the first.

Fig. 2.

Monthly cases of Legionellosis in four counties in Ohio.

Environmental predictors included climate indices (Southern Oscillation Index [SOI], North Atlantic Oscillation [NAO]), temperature variables (Days Below Freezing [DT32], Days Above 70 °F [DX70], Cooling Degree Days [CLDD], Heating Degree Days [HTDD]), precipitation (PRCP), relative humidity (RHAV), and soil moisture percentage (SM). The SOI measures fluctuations in air pressure between Tahiti and Darwin, Australia, reflecting the strength and phase of the El Niño–Southern Oscillation (ENSO); negative SOI values indicate El Niño (warmer tropical Pacific), while positive values indicate La Niña (cooler tropical Pacific). The NAO represents the difference in atmospheric pressure between the Azores High and the Icelandic Low and influences the strength and direction of westerly winds and storm tracks across the North Atlantic, thereby affecting winter temperature and precipitation patterns in eastern North America (Whan & Zwiers, 2017). DX70 records the number of days during a month when the daily maximum temperature exceeds 70 °F (21.1 °C). Conversely, DT32 measures the number of days during the month when the minimum temperature falls below 32 °F (0 °C). Cooling Degree Days (CLDD) and Heating Degree Days (HTDD) quantify cumulative thermal energy demand, calculated relative to a base temperature of 65 °F (18.3 °C), and serve as proxies for overall warm-season and cold-season intensity, respectively (Bhatnagar et al., 2018). PRCP represents total monthly precipitation (mm), RHAV measures average relative humidity (%), and SM is the volumetric soil moisture (%) in the top 0–10 cm layer.

We obtained monthly SOI and NAO data from the National Oceanic and Atmospheric Administration (NOAA) and meteorological data from NOAA's Climate Data Online portal. Soil moisture data were derived from the NASA Global Land Data Assimilation System (GLDAS) version 2.1, specifically the NOAH land surface model product (NASA/GLDAS/V021/NOAH/G025/T3H) at approximately 0.25° spatial resolution (Mitchell, 2005). We extracted the surface soil moisture variable representing the 0–10 cm depth layer as monthly averages spatially aggregated over county boundaries using Google Earth Engine. Each variable was summarized by month for each county for the same 22-year time period. An overall summary of the data is found in Table 2, while monthly visualizations are found in Supplementary Materials. Cuyahoga and Lucas counties were on average cooler than Franklin and Hamilton. Lucas had the lowest amount of precipitation, while Hamilton tended to be the least humid with lower SM.

Table 2.

Mean values for 7 climatic variables by county: Days Below Freezing (DT32), Days Above 70 °F (DX70), Cooling Degree Days (CLDD), Heating Degree Days (HTDD), precipitation (PRCP), relative humidity (RHAV), and soil moisture percentage (SM).

County	DT32	DX70	CLDD	HTDD	PRCP	RHAV	SM
Cuyahoga	8.9	12.2	41.7	260.2	90.3	72.0	35.5
Franklin	8.4	13.8	53.0	235.4	91.5	68.0	33.5
Hamilton	8.7	14.7	52.3	220.9	90.4	66.6	34.4
Lucas	10.0	12.7	41.7	272.8	77.3	69.6	37.4

2.3. Modelling

We explored associations between the environmental factors and Legionellosis case counts in each county using DLNMs using the dlnm R package (Gasparrini, 2011; Gasparrini et al., 2013). DLNMs offer a flexible time series approach that simultaneously estimates nonlinear exposure–response relationships and lagged effects, allowing for assessment of how environmental predictors influence disease risk across multiple time lags (Gasparrini, 2014). In other words, they describe how exposure influences health outcomes both nonlinearly (i.e. effects that increase sharply above a temperature threshold) and with delay (i.e. effects that occur several days, weeks, or months later). The key feature of a DLNM is the cross-basis function, which simultaneously represents the exposure–response relationship and its lag structure in a two-dimensional space. This allows estimation of complex temporal dependencies, such as how unusual meteorological conditions affect disease risk (Lüthi et al., 2023).

We modeled monthly Legionellosis case counts as county-specific time series over the period January 2001 to December 2023, yielding 276 monthly observations per county. We modeled the response using a quasi-Poisson distribution because the variance exceeded the mean, indicating overdispersion (Ver Hoef and Boveng, 2007). We did not model incidence rates based on population, because monthly and annual population estimates were not available prior to 2010. Instead, we accounted for apparent annual trends in the monthly case data using natural cubic splines (d.f. = 2). This allowed the models to account for changes in the overall trends over time that may be driven by changes in population, reporting, or similar factors (Howe et al., 2011). Likewise, we accounted for the distinct seasonal trend using sine/cosine terms for month, as is typical when modelling monthly data with DLNMs (Downs et al., 2025; Stolwijk et al., 1999). This allowed the models to capture for the observed cyclical pattern in cases driven by seasons.

Within this modelling framework, we represented exposure–response and lag–response relationships using low-degree polynomial functions and constrained maximum lag durations to 12 months to limit model complexity, consistent with recommended practice when applying DLNMs to sparse count data (Gasparrini & Leone, 2014). For each county, we first identified which environmental variables displayed lagged effects when used in isolation while still controlling for annual and seasonal trends. We systematically searched over all lag durations of (1-12 months) as well as polynomial degrees for both the exposure-response and lag-response bases (1-3°). If the DLNM cross-basis terms were collectively significant (all terms with p < 0.05) for a set of parameters, we considered the variable for possible inclusion with lagged nonlinear terms the next step. If not, we considered the variable as a simple term. Generically, the DLNM can be represented as:

$$\mathit{log}\left(E\left[Y_{c,t}\right]\right)={\alpha }_c+ns\left({Year}_t,2\right)+{\beta }_1\sin \left(\frac{2\pi t}{12}\right)+{\beta }_2\cos \left(\frac{2\pi t}{12}\right)+\sum _{k=1}^K{cb}_{poly}\left(X_{k,c,t},L_k\right)\text{,}$$

where:

$Y_{c,t}$ is the monthly number of Legionellosis cases in county $c$ at month $t$;

$E\left[Y_{c,t}\right]$ denotes the expected value of $Y_{c,t}$ under the quasi-Poisson model;

${\alpha }_c$ is the county-specific intercept;

$ns\left({\text{Year}}_t,2\right)$ is a natural cubic spline with 2 degrees of freedom applied to calendar year;

$\sin \left(\frac{2\pi t}{12}\right)$ and $\cos \left(\frac{2\pi t}{12}\right)$ represent harmonic seasonal terms with coefficients ${\beta }_1$ and ${\beta }_2$;

$X_{k,c,t}$ is the environmental predictor $k$ measured for county $c$ at time $t$;

$c{b}_{\text{poly}}\left(X_{k,c,t},L_k\right)$ is the polynomial cross-basis function describing both the nonlinear; exposure–response and lagged effects of $X_k$ up to a maximum lag of $L_k$ months;

$L_k$ is the maximum lag duration (1–12 months) considered for each environmental variable; and

$K$ is the number of environmental predictors included in the final model.

Because different DLNM specifications can produce significant results, we developed a composite scoring metric to heuristically select a significant parameter combination that balanced model fit and parsimony. The score included three metrics: the proportion of significant cross-basis terms (penalizes insignificant terms), highest p-value among cross-basis terms (penalizes isolated weak effects), and total degrees of freedom of the cross-basis (to penalize model complexity). The score is computed as:

$$score=w_a\times \left(1-\%significant\right)+w_b\times \max p+w_c\times totaldf\text{.}$$

We specified weight $w_a$ = 10 to heavily prioritize models where the majority of lagged terms are statistically significant, ensuring consistent association across the lag period. We specified $w_b=1$ in order to favor cross-basis terms with lower p values. We specified $w_c=0.01$ to apply a small penalty to discourage overly complex models. Note the magnitudes of the weights differ, because they are measured at different scales, and lower scores indicate more preferable models.

We constructed multivariable DLNMs by sequentially adding variables through forward selection, beginning with the lagged predictor that produced the lowest scoring metric in isolation. At each step, we tested the inclusion of a new variable and computed the scoring metric. Variables without significant lagged effects were included in multivariable models without lagged cross-basis terms. Additionally, due to strong collinearity among temperature predictors, we restricted final models to include only one temperature variable per county. The model producing the lowest scoring metric after running all iterations was selected as the final model. We confirmed absence of multicollinearity in final models using variance inflation factors (VIF), considering VIF <5 acceptable.

2.4. Sensitivity analysis

We performed a sensitivity analysis to evaluate the robustness of parameter selection and variable inclusion to the weights assigned using the scoring metric. We systematically varied weights $w_a$ (1, 2, …, 20), $w_b$ (0, 0.5, 1, 1.5, 2), and $w_c$ (0, 0.005, 0.01, 0.015) over a predefined grid and repeated the parameter search and forward selection procedure for each combination. From these repeated analyses for each county, we recorded the predominant lag structure for each variable as the combination of lag length, exposure degrees of freedom, and lag degrees of freedom most frequently selected across varying scoring parameterizations. For each county-variable pair, we calculated a predominance ratio, defined as the proportion of sensitivity scenarios in which the predominant parameter set was selected.

3. Results

The final models consistently identified significant effects for year, season, and environmental predictors across all counties (Table 2). Temporal trends modeled using natural splines for year and seasonal patterns modeled with harmonic sine and cosine term were also statistically significant (p < 0.05) in all final models. Across all counties, five environmental variables were significant in at least one final model: SM, RHAV, DT32, SOI, and NAO. SM and RHAV were the most frequently selected predictors, each retained in three counties. DT32 and SOI were included in two counties, while NAO was retained in one. These variables exhibited significant lagged associations with the outcome. None of the final models retained any variables with only simple, immediate effects. Polynomials with a single degree of freedom were selected for almost all variables except two: SOI for Cuyahoga and SM for Lucas. Additionally, all VIFs were below 5.0, indicating acceptable collinearity levels across retained predictors. The full model outputs are available in the Supplementary Materials.

SM was retained in the final models for Cuyahoga, Hamilton, and Lucas counties (Fig. 3a–c). In Cuyahoga, higher SM was associated with increased risk for 1 month. In Hamilton, extreme SM values produced elevated risk for a period of 8 months, with low values suppressing risk for the same period of time. In Lucas, risk peaked 3 to 9 months following extreme soil moisture, with low values producing the opposite effect. RHAV was retained in the final models for Cuyahoga, Franklin, and Lucas counties (Fig. 3d–f). For all three counties, higher RHAV was associated with increased risk for two months. DT32 was retained in the final models for Cuyahoga and Hamilton counties (Fig. 3g–h). In Cuyahoga, larger numbers of DT32 suppressed risk for 1 month, while in Hamilton this effect was observed for 4 months. An oscillation index was retained in models for Cuyahoga, Franklin, and Hamilton (Fig. 3i–k). In Cuyahoga, higher SOI was associated with reduced risk at a lag of one month, increased risk at 2 months, and reduced risk again at a lag of 3 months. In Franklin, high values of NAO produced elevated risk for a period of 7 months, while low values supressed risk for 6 months. In Hamilton, lower SOI values were associated with high risk for five months, while high values suppressed risk for 6 months.

Fig. 3.

Three-dimensional DLNM surfaces showing the estimated relative risk (RR) for four Oiho Counties (Cuyahoga, Franklin, Hamilton, Lucas) based upon soil moisture (SM), relative humidity (RHAV), Temperature (Days below freezing, DT32), and Southern (SOI) or North Atlantic (NAO) Oscillation indices.

We also used cumulative relative risk curves for each county to summarize the net effect of each predictor across the full lag period, providing an interpretable measure of total risk that complements lag-specific results. For example, Legionellosis risk showed a positive association with SM in both Cuyahoga and Hamilton Counties, increasing risk when above about 36% (Fig. 4). Legionellosis risk was positively associated with RHAV, showing similar trends in Cuyahoga, Franklin, and Lucas Counties with elevated risk occurring when RHAV exceeded 69-72%, depending on the county. Legionellosis risk was positively associated with extreme low values for DT32, translating to higher risk during summer seasons, in both Cuyahoga and Hamilton counties, though the relationship was less certain during months with consistent low winter temperatures. Cumulative relative response curves are shown for all variables in Supplementary Materials.

Fig. 4.

Cumulative relative risk response curves for Soil Moisture in three Ohio counties.

The sensitivity analysis evaluated the stability of selected lag lengths and polynomial degrees across variations in scoring weights (Table 3). Out of the eleven county-variable combinations, seven had a predominance ratio of 1.0, indicating that the same lag structure and polynomial degrees were selected regardless of the scoring weights applied (i.e. 100% of the time). These fully stable variables included SM, RHAV, DT32, and SOI in Cuyahoga County, and SM, DT32, and SOI in Hamilton County. For these variables, the selection of lag length and polynomial degrees were unaffected by changes in scoring emphasis. The remaining four variables showed moderate sensitivity to scoring. DT32 and NAO in Franklin County had predominance ratios of 0.80, as did DT32 in Cuyahoga and RHAV in Franklin, indicating that alternative lag lengths and polynomial degrees were selected under 20% of scoring schemes, though the final model parameters remained among the most frequently chosen. In Lucas County, RHAV and SM exhibited the lowest predominance ratios (0.72 and 0.65, respectively), suggesting greater variability in parameter selection across scoring variations. However, these differences were relatively minor, shifting only between a lag of 2 or 3 months and 1 or 2° for the polynomial function for any affected variables. Overall, the sensitivity analysis demonstrated that final model selections were robust to the range of scoring weights considered and had little impact on the associations identified.

Table 3.

Summary of DLNM parameters for the final selected model in comparison to the majority of models recorded in a sensitivity analysis. The predominance ratio records the proportion of models that utilized the dominant set of parameters, including lag (months), exposure degrees of freedom (d.f.), and lag degrees of freedom (d.f).

County	Variable	Final Model				Dominant Model
		Score	Lag	Exposure df	Lag df	Predominance Ratio	Lag	Exposure df	Lag df
Cuyahoga	SM	0.02	1	1	1	1.00	1	1	1
Cuyahoga	RHAV	0.02	3	1	1	1.00	3	1	1
Cuyahoga	DT32	0.02	1	1	1	0.80	1	1	1
Cuyahoga	SOI	0.04	12	1	2	1.00	12	1	2
Franklin	RHAV	0.02	3	1	1	0.80	3	1	1
Franklin	NAO	0.02	9	1	1	0.80	9	1	1
Hamilton	SM	0.02	12	1	1	0.96	12	1	1
Hamilton	DT32	0.04	7	1	1	1.00	7	1	1
Hamilton	SOI	0.06	10	1	1	1.00	10	1	1
Lucas	RHAV	0.02	2	1	1	0.72	3	1	1
Lucas	SM	0.05	10	1	2	0.65	10	1	2

4. Discussion

This study applied distributed lag nonlinear models (DLNMs) to investigate how climatic and environmental conditions influence Legionellosis incidence across four Ohio counties over a 23-year period. A central contribution of this work is the development and application of a transparent scoring framework to guide DLNM model selection in settings characterized by sparse outcomes and multiple plausible lag structures. Model specification in DLNM studies is inherently sensitive to choices regarding lag duration, functional form, and degrees of freedom, particularly when outcomes are rare. Rather than seeking a single optimal or predictive model, the scoring approach was designed to prioritize parsimony, stability, and interpretability by penalizing inconsistent statistical significance across lag periods, isolated weak effects, and unnecessary model complexity. This framework provides a reproducible strategy for navigating DLNM parameter uncertainty while avoiding overfitting, and it is well suited to epidemiologic applications where inference, rather than prediction, is the primary objective.

Using this approach, we identified county-specific sets of climatic predictors associated with Legionellosis risk in Ohio. Our findings reveal associations between Legionellosis risk and soil moisture, relative humidity, temperature, and large-scale climate indices, although the relevance and form of these relationships varied across counties and predictors. Although all four counties displayed pronounced seasonal patterns, with higher incidence during summer months, both the direction and timing of temperature-related effects differed by county. Moisture-related variables were retained in several county-specific models and generally exhibited positive associations where included, consistent with the ecological preference of Legionella bacteria for warm, humid environments where biofilms can persist and aerosolized transmission is more likely to occur (Abdel-Nour et al., 2013; Declerck, 2010; Reinares et al., 2024).

Lag durations varied substantially across predictors and counties, ranging from short-term effects of one to two months to extended lags of up to eight to twelve months. These longer lags are unlikely to represent biological incubation alone, which typically occurs over days to weeks (Cassell et al., 2025; Lehfeld et al., 2023). Instead, they likely reflect climatic pre-conditioning processes, whereby sustained antecedent moisture and temperature conditions promote Legionella persistence within biofilms and water systems, increasing the likelihood of amplification and exposure during subsequent favorable periods (Van der Kooij, 2014). While delays in diagnosis and reporting may contribute to apparent long lags, such delays would be expected to operate similarly across predictors. The variable- and county-specific lag structures observed here therefore suggest that extended lags primarily capture indirect environmental and infrastructural processes rather than reporting artifacts alone.

Temperature-related effects further exhibited marked geographic heterogeneity. For example, extreme cold temperatures (DT32) were positively associated with subsequent Legionellosis risk in Cuyahoga County but negatively associated in Hamilton County. Rather than representing contradictory findings, these opposing associations likely reflect differences in regional climate regimes and infrastructure-mediated exposure pathways. In northern counties such as Cuyahoga, where freezing temperatures are more frequent and prolonged, repeated freeze–thaw cycles may disrupt aging water infrastructure (Fan et al., 2023), facilitating Legionella release during subsequent warming periods. In contrast, southern counties such as Hamilton experience milder winters, where short cold spells may suppress bacterial growth and aerosolization. These temperature effects likely interact with baseline climate, building stock age, and water system design, producing distinct lagged responses across counties.

The inclusion of large-scale climate indices (SOI and NAO) in several county models suggests that broader climatic oscillations may modulate environmental conditions that favor Legionella proliferation and transmission. Both indices exhibited delayed, nonlinear associations with Legionellosis risk, implying that their influence operates indirectly through regional weather anomalies. The alternating positive and negative effects observed for SOI in Cuyahoga and Hamilton underscore the complexity of these lagged teleconnections, likely reflecting interactions between El Niño–Southern Oscillation (ENSO) phases and local moisture conditions. El Niño events (low SOI) tend to bring warmer and wetter winters to the Midwest (Mauget & Upchurch, 1999), potentially enhancing bacterial persistence and human exposure opportunities, whereas La Niña phases (high SOI) are often cooler and drier, which may suppress risk. Similarly, the significant role of the North Atlantic Oscillation (NAO) in Franklin County may indicate a teleconnection between positive NAO phases, characterized by mild, wet winters (Notaro et al., 2006), and increased Legionellosis risk several months later. These findings suggest that large-scale climate oscillations could act as upstream drivers influencing local environmental suitability for Legionella through lagged effects on temperature and hydrometeorological variables (Gasparrini et al., 2015).

Taken together, these findings have important implications under projected climate change (Walker, 2018). The state has already experienced measurable warming over the past century, and average annual temperatures are projected to increase by approximately 3–4 °C (5–7 °F) by mid-century, with the strongest increases during winter and spring (Wilson et al., 2024). Precipitation is also expected to become both heavier and more variable, characterized by wetter winters and springs with more frequent heavy rainfall events interspersed with short dry periods. This combination of warmer temperatures, higher atmospheric moisture, and greater hydrologic variability could intensify the lagged relationships observed here between Legionellosis risk and soil moisture, humidity, and large-scale climate oscillations. In addition to ecological effects, warmer summers will likely increase reliance on air conditioning and evaporative cooling systems, which, if inadequately maintained, can provide favorable environments for Legionella amplification and aerosolization (Lin et al., 2009; Wüthrich et al., 2019). Longer periods of elevated humidity and soil moisture, together with more frequent freeze–thaw and saturation cycles, may further enhance opportunities for Legionella persistence and transmission in both natural and engineered water systems. As Ohio's climate continues to warm and precipitation extremes increase, these environmental and behavioral changes may expand the temporal window and geographic range of elevated Legionellosis risk, underscoring the importance of adaptive surveillance, infrastructure maintenance, and water management strategies (Sciuto et al., 2021).

To assess the reliability of these findings in counties with lower incidence, we evaluated model robustness through an extensive sensitivity analysis. Lucas County exhibited greater variability in selected lag structures and polynomial degrees than the other counties, consistent with the increased statistical uncertainty expected under sparse outcome data. Nevertheless, dominant parameterizations remained stable across most sensitivity scenarios, and observed differences were limited to modest shifts in lag length or polynomial degree rather than changes in the direction of associations. These results suggest that the identified exposure–lag relationships are not artifacts of overfitting or arbitrary parameter choices. However, reduced statistical power in lower-incidence counties remains an important consideration, and county-specific estimates from these settings should be interpreted with appropriate caution.

While this study demonstrates robust associations across multiple counties, several limitations should be acknowledged when interpreting county-specific estimates. The analysis relied on county-level aggregated data, which may mask finer-scale environmental and infrastructural heterogeneity. Population change over the study period may also influence interpretation of relative risks across counties. Franklin County experienced substantial population growth, whereas Cuyahoga County experienced population decline, which could affect absolute case counts independent of environmental drivers. Although our models accounted for long-term temporal trends using spline terms, the absence of consistent monthly population estimates prior to 2010 precluded direct modeling of incidence rates. As a result, relative risks should be interpreted primarily as within-county associations over time rather than as direct quantitative comparisons of effect magnitude between counties. Additionally, the occurrence of Legionellosis has been closely linked to artificial water systems such as cooling towers and building plumbing (Schwake et al., 2021), and we were unable to explicitly account for changes in water system operation or maintenance. Periods of reduced building occupancy, such as during the COVID-19 pandemic, may have contributed to water stagnation and altered exposure risk, effects that could not be directly represented in the models (Chao & Lai, 2022; Dey & Ashbolt, 2020; Rhoads & Hammes, 2021). Although pandemic-related water stagnation may have influenced exposure risk, the highest monthly case counts in the study counties did not coincide with 2020, indicating that additional mechanisms likely contribute to the observed lagged associations. Estimated lag structures likely reflect a combination of bacterial ecology, exposure timing, and reporting delays, which could not be disentangled using available data. Future research integrating higher-resolution environmental data, building-level infrastructure information, and case-level reporting timelines may help clarify the mechanisms linking climate variability and disease emergence. Extending this framework to national or global datasets may also improve understanding of how large-scale climatic teleconnections modulate Legionella risk under changing climatic conditions.

Despite these limitations, the findings demonstrate that Legionellosis risk in Ohio is shaped by complex, lagged interactions between local environmental conditions and large-scale climatic oscillations. As climate change accelerates, similar mechanisms are likely to operate globally, with warming temperatures, increased humidity, and shifting precipitation regimes creating more favorable conditions for Legionella persistence and transmission. Regions historically too cool or dry for sustained bacterial growth may experience expanding seasonal windows of risk, while infrastructure stress from rising cooling demand could amplify outbreaks in urban areas. These results highlight the importance of integrating climate variability into infectious-disease forecasting models and underscore the need for proactive adaptation strategies.

CRediT authorship contribution statement

Joni Downs: Writing – original draft, Visualization, Project administration, Investigation, Formal analysis, Conceptualization. Hansamali Abeysinghe Mudiyanselage: Writing – review & editing, Visualization, Data curation, Conceptualization. Jaryd Hinch: Writing – review & editing, Formal analysis, Data curation, Conceptualization. Nji Mukwa Njoh: Writing – review & editing, Data curation, Conceptualization. Saurav Chakraborty: Writing – review & editing, Data curation, Conceptualization.

Conflict of interest

The authors have declared no conflict of interest.

Handling Editor: Dr Yijun Lou

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Data availability

Data and code for this study are archived at 10.5281/zenodo.18488353.