Introduction

Climate change is driving a rise in both the frequency and intensity of extreme weather events [1]. These events impact human health through multiple pathways [2], and notably by altering the transmission dynamics of respiratory pathogens. One driver of this is the disruption of human behavior and social mixing patterns during and after extreme events [3]. Studies showed that severe snowfall and rainfall were associated with a reduction in community contact rates and, through that, COVID-19 incidence [4,5], while heatwaves and wildfire had the opposite effect by increasing the amount of time spent indoors and thus the likelihood of airborne transmission [6,7]. Wildfires and COVID-19 were also proven to have a compound effect on excess mortality [8]. Also, severe floods were associated with a wider spatial spread of emergent respiratory pathogens [9]. More generally, the relationship between extreme weather and epidemic circulation is complex, reflecting the ways in which such events disrupt human mobility and contact patterns. Weather affects daily routines, reshapes social interactions [10–13] and alters the way individuals engage with points of interest, such as workplaces, recreational areas, healthcare facilities or shelters [3,7]. Crucially, these behavioral shifts can influence pathogen transmission both directly, by modifying contact networks, and indirectly, by undermining the public health response [14]. Disruptions to service access [15] and infrastructure, or risk perception, may delay the detection of cases, hamper access to healthcare [16] and hinder the implementation of preventive and containment measures [14,17]. In particular, reduced testing impairs situational awareness of epidemic dynamics, delaying or biasing the timely implementation of public health intervention [14,17]. When extreme weather coincides with epidemic emergencies, it can thus compromise the effectiveness of public health campaigns, on top of its direct effect on circulation [2]. It also exacerbates health inequalities, as extreme weather is known to disproportionately undermine access to healthcare among socioeconomically disadvantaged populations [18–21]. And as climate change increases the likelihood of such compound events, it becomes essential to understand their impact to inform preparedness and response strategies.

In this study, we examined whether heavy rainfall during the early COVID-19 epidemic was associated with reductions in testing rates in the state of New York and in France. We quantified the strength of this association and assessed its geographic variability. The COVID-19 pandemic generated an unprecedented volume of data, offering detailed insight into transmission dynamics, testing behavior, and public compliance with health measures [22–24]. These datasets have proven critical for evaluating the effectiveness of interventions [25–27], and, more recently, they have powered AI tools that have been improving public health preparedness and surveillance, through earlier and more accurate diagnosis, screening and predictions of epidemic impact [28,29]. A retrospective analysis of how rainfall affected the public health response during this large-scale respiratory epidemic will thus contribute to the broader effort of preparing for future outbreaks of emerging or re-emerging pathogens under a changing climate and improve the ability of available tools to account for climate-related effects.

Methods

Data

We selected Admin-2 as the geographic resolution of our analysis, corresponding to departments (France) and counties (New York State). In France, we included the 94 departments of mainland France excluding Corsica. Data on COVID-19 tests performed and incident cases were sourced from Santé Publique France, the French national public health agency [30] and from the New York State Department of Health open data repository [31] and was collected in both cases at Admin-2 level. We restricted our analysis to the time window from May 15, 2020, to November 1, 2021 (see Fig 1a–1b), to avoid the very first months of the pandemic, when testing was limited, and sharp change in incidence patterns brought about by the emergence of the Omicron variant [32,33]. Corresponding daily rainfall data were obtained from the French national weather agency Météo-France [34] and the US National Oceanic and Atmospheric Administration [35]. Data for New York State were provided at the county level. Data for France were provided for individual weather stations, and department-level values were obtained as population-weighted averages of weather station data. Local population data are available at: https://www.data.gouv.fr/datasets/population-municipale-des-communes-france-entiere/. Weather data were retrieved from January 1, 1998 to December 30, 2022 for France and from January 1, 2016 to September 30, 2023 for New York State.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. COVID-19 cases, tests, rainfall and extracted trends.

(a) Daily number COVID-19 test performed (orange, left y-axis) and recorded COVID-19 cases (gray, right y-axis) in France in the study period. (b) Daily number COVID-19 test performed (orange, left y-axis) and recorded COVID-19 cases (gray, right y-axis) in New York state. (c) Daily number COVID-19 test performed (orange, left y-axis), its trend extracted from the best-performing model (see Table 1) and daily rainfall (blue, right y-axis) for the French department with the highest income (Hauts-de-Seine). (d) Daily number COVID-19 test performed (orange, left y-axis), its trend extracted from the best-performing model (see Table 1) and daily rainfall (blue, right y-axis) for the French department with the lowest income (Seine-Saint-Denis). (e) Daily number COVID-19 test performed (orange, left y-axis), its trend extracted from the best-performing model (see Table 2) and daily rainfall (blue, right y-axis) for the New York State county with the highest income (Nassau). (f) Daily number COVID-19 test performed (orange, left y-axis), its trend extracted from the best-performing model (see Table 2) and daily rainfall (blue, right y-axis) for the New York State county with the lowest income (Bronx).

https://doi.org/10.1371/journal.pclm.0000894.g001

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Results of the Poisson regression models for France. Each row represents a specific model setup, including the Prophet setup used (additive or multiplicative) and the way of including the weekly component. The AIC column gives the Akaike Information Criterion for each model setup. The lowest value of AIC is in bold. The Effect of rain (Δ) column reports the percentage change in the number of tests associated with 25mm of rainfall, with 95% confidence intervals in parentheses; when not included is reported this means that the model setup excluded rainfall as predictor, using only trend and weekly periodicity. The last two columns report the estimated spatial autocorrelation parameter and the variance of the area-level random effect, respectively.

https://doi.org/10.1371/journal.pclm.0000894.t001

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Results of the Poisson regression models for New York. Each row represents a specific model setup, including the Prophet setup used (additive or multiplicative) and the way we included the weekly component. The AIC column gives the Akaike Information Criterion for each model setup. The lowest value of AIC is in bold. The Effect of rain (Δ) column reports the percentage change in the number of tests associated with 25mm of rainfall, with 95% confidence intervals in parentheses; when not included is reported this means that the model setup excluded rainfall as predictor, using only trend and weekly periodicity.The last two columns report the estimated spatial autocorrelation parameter and the variance of the area-level random effect, respectively.

https://doi.org/10.1371/journal.pclm.0000894.t002

Socioeconomic data were, for New York State, the county-level median household income in US dollars from the US Census Bureau (https://data.census.gov/). For France the indicator selected was the niveau de vie provided by the French statistical institute INSEE (https://www.insee.fr), which is defined as the disposable income in Euros of a household divided by the number of consumption units (approximately the household size). We referred to both as income indicators.

Statistical model of the effect of rainfall

We assumed that the likelihood of observing tests was a Poisson distribution with mean . For a given area i, we assumed that the expected number of tests performed on day had the following expectation value:

Where is the value of the trend of the area at time and possibly the periodic component from the time series model (see previous section), is a time-independent offset, is the daily rainfall (in mm), is the coefficient to be determined, and is a latent spatial random effect, the dots indicate other possible predictors. We tested four different choices for : i) the trend component of the additive model; ii) the trend component of the multiplicative model; iii) the trend plus weekly component of the additive model; iv) the trend times weekly component of the multiplicative model. For cases i), ii) (i.e., without the weekly component) we also tested whether a weekly effect should be incorporated as an additional one-hot-encoded categorical variable (day of the week) in the Poisson regression. To account for spatial correlation across areas, we fitted a spatial generalized linear mixed model using the fitme function from the spaMM R package. Spatial dependence across departments (France) or counties (New York State) was modeled as a Conditional Autoregressive (CAR) structure based on a binary adjacency matrix W. As customary, the spatial random effects u were assumed to follow a multivariate normal distribution:

where ρ controls the strength of spatial autocorrelation between neighboring areas and λ is a variance scaling parameter. See S1 Appendix for the treatment of possible correlations in time. All candidate models were initially fitted to the national (France) and state-level (New York) datasets, assuming a single common effect of rainfall across the entire country or state. Based on AIC (Akaike Information Criterion) model comparison, we then extended the best-performing specification by including an additional covariate for public holidays, which further improved model fit by accounting for systematic calendar-related variation in testing behavior. We also considered statistical models without rainfall as covariate, to check whether including rainfall was indeed meaningful in explaining the observed number of tests. Once the model was selected, it was extended by including a specific fixed effect for each department (France) or county (NY), allowing the rainfall effect to vary by area. The estimated coefficient α for rainfall was then translated into a more interpretable quantity, Δ, representing the percentage change in daily test counts associated with an additional 25 mm of rainfall: . All the models were coded and fitted in R (version 4.4.1).

Results

Fig 1 provides an overview of the COVID-19 epidemic and the testing efforts in France and New York State along the study period, displaying the daily number of recorded COVID-19 cases and of performed COVID-19 tests. It also shows the daily number of tests and the daily amount of rainfall in the French departments and NY counties with the lowest and highest values of the income indicator, together with the trend component extracted with Prophet (multiplicative model).

Table 1 reports the results of the different statistical models on French national testing data. The modeling setup whereby trend and the weekly periodicity were extracted with Prophet using a multiplicative model was the one which performed best according to AIC. This indicates that weekly periodicity (e.g., weekday-weekend differences) is best represented as a multiplicative modulation of the baseline trend, rather than an additive offset. This is indeed consistent with day-specific variations in testing availability. Importantly, rainfall indeed emerged as a meaningful predictor of the number of tests, as including it improved the lowered AIC. Adding the holiday predictor, which accounts for holidays except Sundays, further reduced the AIC, while slightly attenuating the estimated effect of rainfall.

The estimated effect of rain (Δ) was substantially and significantly lower than zero (Δ < 0), indicating that rainfall was associated with a reduction in the number of tests performed. Specifically, we estimated a Δ = -7.25%) CI=[-7.3, -7.2], reduction in the number of daily tests for every 25 mm of rainfall. The other models including rainfall, while performing worse, provided nonetheless similar estimates with reductions ranging from 7.2% to 10.5%. As expected, also the holiday predictor was associated with a reduction in the number of tests. The estimated spatial autocorrelation parameter was small (ρ = 0.12), with a variance scale of , indicating limited residual spatial dependence.

Table 2 reports the same analysis for New York State. The setup using the multiplicative model for Prophet, incorporating the weekly seasonality extracted and including rainfall as a predictor was again the best-performing. The effect of rain was again significantly less than zero (Δ < 0), indicating that rainfall was associated with fewer tests in New York State, too. The effect was, however, lower in magnitude than in France: Δ = -4.9% CI=[-5.0%,-4.8%], an almost 5% reduction in tests for every 25 mm of rain. Worse-performing models again gave comparable results, with reductions ranging from 5.2% to 5.4%. Again, adding the holiday predictor attenuated the estimated rainfall effect while improving model fit (lower AIC). Consistent with the results for France, the spatial CAR model for New York State showed modest spatial dependence (ρ = 0.10), with a variance scale of .The best-performing model at the national and state level was then fitted to each department and county. In France (Fig 2a), most departments exhibited a negative relationship between rainfall and tests, compatible with the national estimate. All estimated values of Δ along with their uncertainties are reported in S1 Appendix. The departments with the lowest Δ values were Indre (Δ = -23.12%), Moselle (Δ = -20.54%), Aude (Δ = -18.73%), Hérault (Δ = -18.27%) and Cher (-17.36%). The exception to this trend was a cluster of departments in the northern part of the country, from Ardennes (Δ = 12.81%) and Somme (Δ = 6.29.0%) in the East to Orne (Δ = 7.86%) in the West, and spanning from the coast of Normandy to the northwest of Paris (see for example Yvelines, Δ = 10.14% and Oise, Δ = 7.37%). In that area, additional rainfall was associated with an increase in the number of tests. It should be noted, however, that the estimate for some departments (e.g.,Orne) had a large uncertainty. Finally, Aisne, Calvados, Manche and Paris exhibited no significant association between rainfall and tests. It should be also noted that the confidence in the estimate of Δ varied from department to department: those with high uncertainty are hatched in Fig 2a. In New York State (Fig 2b), as in France, most counties exhibited a negative value of Δ, compatible with the state-level fit. All estimated values of Δ along with their uncertainties are reported in S1 Appendix. The strongest negative association between rainfall and testing was found in the counties of St. Lawrence (Δ = -16.81%), Schoharie (Δ = -13.06%), Franklin (Δ = -12.85%) and Clinton (Δ = -10.81.7%) in the north of the state, and Chemung (Δ = -11.72%) bordering Pennsylvania. The strongest positive association was found in the counties of Yates (Δ=+10.81%), Tompkins (Δ=+8.20), Rensselaer (Δ=+5.29%) and Columbia (Δ=+4.45%) along the eastern border with Vermont, and Niagara (Δ=+2.88%). The counties with Δ > 0 did not seem to cluster geographically, unlike in France. In Cattaraugus, Chautauqua, Essex, Fulton, Genesee, Greene, Hamilton, Jefferson, Lewis, Montgomery,Orleans, Seneca and Wyoming the estimate of Δ was not significantly different from zero.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Geographical patterns of the effect of rainfall (Δ) and relationship to rainfall and socioeconomic profiles.

Values of Δ come from the best-performing models (see Table 2). (a) Value of Δ in French departments. Hatched departments indicate that the estimated uncertainty interval on Δ is larger the third quartile of all uncertainties. Departments marked with stars indicate areas where the rainfall predictor is not statistically significantly different from zero.(b) Value of Δ in the counties of New York State. Hatched counties indicate that the estimated uncertainty interval on Δ is larger than the third quartile of all uncertainties. Counties marked with stars indicate areas where the rainfall predictor is not statistically significantly different from zero. (c-e-g) scatter plots of local (French department) Δ versus the yearly number of days with more than 0.5 mm of rain (c), number of days with more than 25 mm of rain (e), and the value of the income indicator (g). (d-f-h) scatter plots of local (NY counties) Δ versus the yearly number of days with more than 0.5 mm of rain (d), number of days with more than 25 mm of rain (f), and the value of the income indicator (h). In panels (c-e-g) and (d-f-h) the red dashed line displays the national value of Δ, the orange line is the linear fit from the weighted linear regression. Numerical values of Δ displayed in (a) and (b) are reported in S1 Appendix, along with their uncertainties. Base layer from geoBoundaries (https://www.geoboundaries.org) [40] available under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

https://doi.org/10.1371/journal.pclm.0000894.g002

We then investigated whether the local Δ was associated with the local weather profile. First, we considered the number of days in a year with more than 0.5 mm of rain. In France, the department with the largest number of such days was Finistère (197 days), the one with the smallest number was Bouches-du-Rhône (80 days). In NY, the county with the largest number of such days was Hamilton (191 days), the one with the smallest number was Richmond (136 days). We found a positive correlation between Δ and the number of rainy days in both France and New York State. In France, the correlation was marginally significant (Pearson = 0.13 CI=(-0.01, 0.28), p value = 0.075) while in New York State it was statistically significant (Pearson = 0.35, CI=[0.16, 0.53], p value = 0.00) – see Fig 2c and 2d. Then, we considered the number of days in a year with more than 25 mm of rain. In France, the department with the largest number of such days was Ardèche (13 days), the one with the smallest number was Aisne (<1 day). In NY, the county with the largest number of such days was Ulster (11 days), the one with the smallest number was Wayne (3 days). This time, the correlation was still similar in France and NY, but negative (Fig 2e and 2f). In France, however, it was not significantly statistically different from zero (Pearson = -0.14, CI=[-0.28, 0.01], p value = 0.064); while in New York State it was: (Pearson = -0.32, CI=[-0.46, -0.19], p value = 0.00).

Finally, we tested the correlation between the effect of rain (Δ) and the income indicator (Fig 2g and 2h). In France, a positive correlation was found: Pearson = 0.23 (CI=[0.10, 0.36], p value = 0.001). In NY, there was no statistically significant correlation: Pearson = 0.04 (CI=[-0.06,0.15], p value = 0.401).

The same analysis using the effect of rain (Δ) estimated from models other than the best-performing (see Table 1 and 2) gave similar results and is available in S1 Appendix. An alternative approach using Generalized Estimating Equations (GEE) is presented in S1 Appendix. results are consistent with those presented here.

Discussion

In this study, we investigated the effect of rainfall on COVID-19 testing across French departments and counties in the State of New York. We found that rainfall was significantly associated with a reduction in the number of COVID-19 tests performed, in the period from May 15, 2020, to November 1, 2021. This was true in both France and the US state of New York, but in France the effect was twice as strong. The two regions had two different COVID-19 epidemics. They had different epidemic waves (see Fig 1a) peaking at different values of incidence: 100 daily cases per 100,000 inhabitants in France, 90 in NY. They also had different public health responses, as France maintained stricter and longer-lasting restrictions on movement than New York [41–43]. The vaccination campaigns were also different, with NY starting earlier, and France catching up by June 2021. This reflected also in a different number of tests administered, peaking at around 1,300 daily tests per 100,000 inhabitants in France, 1,500 in NY. France and NY have also different weather profiles: the latter has a humid subtropical-to-continental climate with abundant precipitation throughout the year, averaging 1,000-to-1,100 mm [44,45]. Storms and tropical systems deliver frequent heavy downpours, so the state typically experiences more days with rainfall exceeding 25 mm than most parts of France. In contrast, mainland France receives about 700-to-900 mm of cumulative precipitation per year on average [46], with a north-south gradient. These climatic differences imply that inhabitants of New York may be accustomed to more frequent and heavier rainfall, whereas many French regions - particularly in the south - experience fewer heavy rain events. This is consistent with previous studies on other types of extreme weather (snow, ice), which show that populations living in areas where specific weather events are more common tend to be less affected by their occurrence [47,48]. These differences may help explain why the effect of rainfall on reducing the number of performed tests, albeit being present in both regions, was stronger in France: for example, heavy rainfall such as 50 mm in 24 hours was associated with almost a 20% reduction in daily tests in France, and 9% in New York State, compared to no rain. It should be noted, however, that both values, especially if compounded over multiple days, can substantially impact detection. Such disruptions directly impact public health response directly, through measures that depend on testing such as including self-isolation and quarantine, and indirectly, by impairing a correct assessment of epidemic circulation [14,17]. This marked regional difference suggests that public health authorities in France should exercise increased vigilance during periods of heavy rainfall, as such events can impair surveillance quality and lead to an underestimation of the observed epidemic risk. Beyond public health, such weather-driven reduction in tests can delay individual treatment, increasing the likelihood of negative health outcomes [49].

We found that a larger average yearly number of rainy days was associated with a smaller reduction in tests due to rainfall (see Fig 2c and 2d). This suggests that communities have adapted to keeping normal activities unperturbed when rain is common. This association was slightly stronger in New York State than in France, which is compatible with the fact this kind of adaptation is stronger in regions more accustomed to rain. We found, however, that this relationship disappeared when relating the effect on tests to the number of days of heavy rain. Here, the relationship was not significant (France) or even negative (NY), hinting that, while communities may adapt easily to moderate rainfall, they will have a hard time adapting spontaneously to extreme weather. This factor must be integrated into infectious disease preparedness in the era of extreme weather, as spontaneous community adaptation is unlikely to suffice and proactive, top-down interventions are required.

We found evidence that rainfall may reduce testing more in poorer communities, suggesting that weather-driven disruptions to epidemic monitoring are amplified in economically vulnerable populations [19,20]. This aligns with studies from low- and middle-income settings showing that rainfall disproportionately disrupts healthcare access in poorer or remote communities [15,50]. But this also fits well with two more general considerations. First, the burden of COVID-19, and infectious disease epidemics in general, is unevenly distributed across socioeconomic strata [41,51]. Second, the effect of extreme weather is felt most in poorer areas [52]. When these two emergencies occur simultaneously, populations that are already vulnerable may face additional strain, further hindering access to care [53]. The compounding of poverty, climate change, and infectious disease vulnerability highlighted by our findings underscores the urgent need for coordinated mitigation strategies to offset these intersecting risks [2].

Our study has limitations. First, its observational design does not formally allow us to establish a causal link between rainfall and testing activity. Unknown confounders are, nonetheless, unlikely, as any such factor would need to influence rainfall itself. However, to strengthen public health guidance and inform preparedness, future work should identify the mechanisms mediating the impact of rainfall on testing, whether driven by behavioral responses to forecasts, delays in travel, or reduced transport availability. Addressing these questions will probably require individual-level studies that record personal mobility and testing decisions, moving beyond the ecological scope of our present analysis. In addition, the evidence of the association between income and the effect of rainfall is, in our study, limited to France. Our analysis does suggest this may be true in New York State, too, but the statistical power is not sufficient to make a definitive determination. Also, our type of analysis again does not allow drawing causal relationships, and our data do not allow us to control for confounders that possibly exist and need to be determined in the effort to better mitigate the compounded vulnerabilities of the poorer communities. Finally, our study is limited to two geographic settings. While the methodological framework is generalizable to other geographic areas, our results do depend on local conditions and thus describe region-specific associations. Future work should examine whether these findings replicate in other at-risk geographical areas where rainfall represents a significant environmental factor. Despite quantitative differences and variations in statistical power, the results were, however, qualitatively consistent across two contrasting contexts. This concordance supports the potential generalizability of our findings, which warrants further investigation in other regions.

Our study provides evidence of rainfall disrupting the surveillance of respiratory epidemics by reducing diagnostic testing, especially during heavy or infrequent rain and in socioeconomically disadvantaged areas. Recognizing this vulnerability is essential to avoid delays in detection and misjudgment of epidemic trends. Integrating weather-driven risk into preparedness and implementing proactive, top-down measures can strengthen epidemic resilience under a changing climate.

Conclusion

This study provides evidence that rainfall is associated with reductions in diagnostic testing for respiratory infections, using COVID-19 as a case study in two distinct geographic settings, France and New York State. Heavy rainfall was linked to decreases in the number of COVID-19 tests performed. This effect was stronger in France than in New York State, varied geographically, and was amplified in socioeconomically disadvantaged areas. We also found that regions accustomed to frequent light rainfall exhibited smaller variations in testing, while heavy or infrequent rainfall was associated with larger disruptions, suggesting limited spontaneous adaptation to extreme events. Together, these results indicate that rainfall can introduce systematic, weather-driven distortions in epidemic surveillance, leading to delays in diagnosis and treatment, biased detection of epidemic trends, and consequently public health response. As climate change is expected to increase the frequency and intensity of extreme precipitation, explicitly accounting for weather-related disruptions in surveillance systems will be essential to ensure timely and reliable monitoring of respiratory epidemics and equitable access to care.

Supporting information