Hay Fever Symptom Burden From Grass Pollen: Implications for Public Health Warning Thresholds

Abstract

This study quantifies the lagged association between grass pollen exposure and the onset of allergic rhinitis (AR) symptoms, including differences between asthmatic and non‐asthmatic subgroups. Data included 114,834 self‐reported AR symptoms collected from 25,194 uniquely identified users of a smartphone application. Reports were gathered prospectively for seven consecutive grass pollen seasons (2017–2023) in Melbourne. We adjusted for meteorological, air quality, and other time‐varying covariates. A quasi‐Poisson regression model combined with a distributed lag non‐linear model (DLNM) was used, testing 0–7 days lagged effects. An association between grass exposure and daily AR incidence of any severity was observed on “low” exposure days (19 grains/m³), when exposure was modeled cumulatively across a 3‐day lag period (RR 1.31, 95% CI 1.16–1.47). Higher grass concentrations were associated with increased risk at lag‐1, with risk rising from RR 1.08 (95% CI 0.99–1.17) at 19 grains/m³ to RR 1.19 (95% CI 1.09–1.29) at 150 grains/m³. By lag‐2, risk returned to a non‐significant baseline. Asthmatics, users reporting medication use, and users reporting more severe symptoms exhibited elevated risk at later lags. The magnitude and duration of elevated risk varied across stratified models by age and sex. We provide new population‐level evidence that grass pollen triggers allergic rhinitis symptoms predominately on the day of exposure, even at concentrations traditionally classified as “low”, with symptom risk persisting longer in asthmatics and other sensitive groups. This study highlights the value of refining communication and early‐warning systems to better support individuals who are sensitive to even low levels of exposure.

Plain Language Summary

Hay fever is a common allergic disease triggered by exposure to grass pollen. Biologically, symptoms are expected to be short lived following exposure. However, population‐based epidemiological studies have reported mixed findings, with some indicating that symptoms may remain elevated, or even peak, in the days following exposure. We examined how grass pollen exposure affects the risk and timing of hay fever symptoms across 7 days following exposure. We also looked at whether risk and timing differ for more vulnerable groups such as people with asthma. The study analyzed 114,834 symptom reports submitted by over 25,000 users of a smartphone app in Melbourne, Australia. Symptom risk was mainly elevated on the day of exposure and, to a lesser extent, the following day. Even grass pollen concentrations traditionally classified as “low” were associated with a measurable increase in symptom risk. For most users, this increased risk was short lived and returned to baseline within two to three days. Some groups were more affected. People with asthma and those reporting greater symptom burden exhibited more persistent risk, with patterns also differing by demographic. These findings support improving pollen warnings and public health messages to better reflect symptom timing and individual susceptibility.

Key Points

• Grass pollen exposure showed the strongest association with same‐day hay fever symptoms, with weaker next‐day effects

• Grass pollen concentrations traditionally classified as “low” were still associated with a measurable increase in symptom risk

• Both the timing and magnitude of the association between grass pollen exposure and hay fever symptoms were heterogeneous across subgroups

1. Introduction

Grass pollen is a major aeroallergen responsible for acute respiratory exacerbations (Davies, 2014). In south‐eastern Australia, Melbourne's diurnal cycle for grass pollen is complex and varies across a season (Mampage et al., 2025). High pollen days precede surges in emergency presentations for asthma (Erbas et al., 2018; Ito et al., 2015; Silver et al., 2018) and acute increases in allergic rhinitis (AR) symptoms (Silver et al., 2020). Internationally, smartphone‐based applications have been used to forecast and manage allergy (Bédard et al., 2020; Sousa‐Pinto et al., 2022; Tripodi et al., 2020). Previous work in Australia using the Melbourne and Canberra Pollen apps have shown that community‐reported AR symptoms track pollen exposure and can support pollen and symptom forecasting (Silver et al., 2020); the present study extends that work by quantifying short‐lag exposure‐response patterns and subgroup‐specific AR symptom risk in Melbourne before severe clinical endpoints.

Across Australia, thresholds for pollen exposure have traditionally guided risk classification and warning systems (Ong et al., 1995). These thresholds, still used nationally (AirHealth Pty Ltd, 2025; AirRater, 2025; QUT Allergy Research Group, 2025), are defined by fixed daily pollen concentrations. Ong et al. (1995) defined these thresholds arbitrarily, they did not explicitly model multi‐day effects or differences in susceptibility among vulnerable subpopulations such as asthmatics.

Current pollen thresholds used to classify exposure implicitly assume that risk is both immediate and uniform across the population. Individual susceptibility to grass pollen is highly variable and shaped by a combination of host, environmental and exposure‐related factors. Host factors such as age (Nitschke et al., 2022), sex, socioeconomic status, and medication use can influence risk (Guarnieri & Balmes, 2014). Among allergic individuals, susceptibility to pollen exposure differs with variation in Th2‐skewed inflammatory responses. Beyond individual variables, there are many environmental variables that influence AR symptoms. Meteorological conditions including temperature, rainfall and wind affect pollen movement in terms of release, allergenicity and dispersal (Schäppi et al., 1998). These meteorological conditions are also associated with AR risk (Li et al., 2025). Air pollutants including PM_2.5, NO₂, and O₃ may prime airway inflammation lowering the threshold for, and increasing the intensity of, allergen‐induced exacerbation (Guarnieri & Balmes, 2014). A population‐level investigation examining how pollen effects influence the extent to which existing threshold classifications convey health risk is therefore critical for vulnerable individuals.

Among asthmatic individuals, the frequent co‐occurrence of AR raises additional concern. AR and allergic asthma are distinct but interconnected manifestations of atopic airway disease. They share common IgE‐mediated inflammatory pathways (Bergeron & Hamid, 2005), however differ in anatomical focus and symptom expression (Leynaert et al., 2000). In asthmatic individuals, exacerbations of AR may act as an early indicator of heightened airway reactivity and precede more severe respiratory events (Bousquet et al., 2001). Allergic disease presents heterogeneously, highlighting the need to characterize risk across vulnerable populations.

Several studies have examined the lagged impact of elevated pollen exposure on respiratory health (Carlsen et al., 2022; Johnston et al., 2009; Jones et al., 2021; Osborne et al., 2017). There is an apparent disconnect between the immediate type I hypersensitivity response expected within 30 min of allergen exposure and the lagged presentation of symptoms observed in epidemiological studies. While type I hypersensitivity reactions do have a late phase, which could persist for 1–2 days after exposure (Galli et al., 2008), such late‐phase responses do not explain the lagged effects observed 2–7 days later. The temporal dynamics of pollen‐induced respiratory exacerbations remain inadequately explored.

Estimates of the duration of heightened symptom susceptibility and the shape of the lagged dose‐response relationship vary, with some studies reporting up to 7‐day lagged effects (Liu et al., 2023) while others observe single‐day lagged effects (Jones et al., 2021). One study found hospital admission for asthmatics peaked 2–5 days after grass pollen exposure (Osborne et al., 2017). Inconsistency across studies likely arises from differences in pollen composition and allergenicity (Jones et al., 2021), the susceptibility of the underlying population (Sedghy et al., 2018), intra‐seasonal timing (de Weger et al., 2011), environmental co‐exposures (Sedghy et al., 2018) and variation in outcome and exposure data quality. Since pollen exposure appears to induce delayed effects, which may manifest heterogeneously, thresholds based solely on same‐day counts may systematically underestimate risk for vulnerable subpopulations, such as asthmatics. Melbourne Pollen App data provides a unique opportunity to explicitly capture any delayed symptom burden below the severe clinical endpoints evaluated in many previous studies.

To address these knowledge gaps, our study will directly address the following research questions: How does an increase in daily grass pollen concentration affect the occurrence and timing of self‐reported respiratory symptoms among Melbourne Pollen App users reporting within 30 km of the monitoring site? Does the association between grass pollen exposure and respiratory symptoms differ for users who report a doctor‐diagnosed asthma condition or symptoms of higher severity? Do current grass pollen thresholds used by early warning systems accurately convey risk to all individuals?

2. Method

2.1. Allergic Rhinitis Symptom Data

Crowd‐sourced smartphone allergic rhinitis (AR) reports were collected via the Melbourne Pollen App for users in Melbourne during the 2017–2023 grass pollen seasons (October to December), as described by Silver et al. (2020). Users could access pollen and climate information. In exchange, users were invited to participate by submitting reports of hay fever symptoms. App submissions outside of the October‐December pollen season were excluded. Submissions outside of a 30 km radius of the pollen trap site were excluded based on GPS coordinate data to mitigate misclassification bias. This radius was informed by the landscape of Melbourne: The city lies on flat coastal plains with ocean to the south, forestry to the east, and highlands to the north and west, roughly 35 km from the CBD. The selected inclusion radius is further supported by Katelaris et al. (2004) and Pashley et al. (2009) finding that a single sampler provides a reasonable estimate of concentrations within a 30 km radius.

Participants were identified by a device Universally Unique Identifier (UUID) and could rate their AR symptoms on a five‐point scale, from no symptoms “1” to severe symptoms “5.”

• [Severe Symptoms “5”] My symptoms were obvious and intolerable

• [Moderate Symptoms “4”] My symptoms were obvious, inconvenient but still tolerable

• [Slight Symptoms “3”] My symptoms were obvious but tolerable

• [Mild Symptoms “2”] My symptoms were slight and a nuisance

• [No Symptoms “1”] No symptoms.

Users aged 18 years or older could also optionally answer seven additional questions to describe themselves and their relevant experience with AR and asthma.

2.2. Grass Pollen Exposure

Airborne pollen sampling records from October 2017 to December 2023 were collected at 09:00 hr local time using a Burkard 7‐day volumetric sampler (Burkard Manufacturing Co. Ltd., Rickmansworth, Hertfordshire, UK). For this analysis, airborne grass pollen concentration (grains/m³) was used as the primary exposure. Non‐grass pollen was also reported by subtracting the grass pollen count from the total pollen count. The pollen trap site was located at the University of Melbourne 1.7 km North of central Melbourne at coordinates −37.7971 and 144.9648. Daily concentrations represent the 24‐hr sampling window from 09:00 hr the corresponding day to 09:00 hr the following day. This window covers 15 hr (62.5%) of that calendar day and 9 hr (37.5%) of the following day (00:00–09:00). Where grass pollen has been described within this manuscript in terms of clinical, categorical levels, they were done so as originally described by Ong et al. (1995): “low” (0–19 grains/m³), “moderate” (20–49 grains/m³), “high” (50–99 grains/m³), and “extreme” (≥100 grains/m³).

2.3. Meteorological and Environmental Covariates

Meteorological and air quality data were sourced from either the Australian Bureau of Meteorology (BOM) or Environmental Protection Authority Victoria (EPA). Total daily precipitation (mm), daily minimum temperature (°C), daily maximum temperature (°C), and solar radiation (MJ/m²), were obtained from the BOM (Bureau of Meteorology, 2025). They were collected from Melbourne Airport (Site 086282). Hourly environmental exposure data was provided by the EPA and accessed using the open‐source DataVic platform (Victorian Department of Premier and Cabinet, 2012). Collected at the Alphington air quality station (−37.7784, 145.0306), hourly concentrations were available for particulate matter of 10 microns or less (PM₁₀, μg/m³), ambient nitrogen dioxide (NO₂, ppb), carbon monoxide (CO, ppm), ozone (O₃, ppb), sulfur dioxide (SO₂, ppb) concentrations, measures for vector wind speed (VWS, ms⁻¹) and direction (VWD, deg). Hourly concentrations for particulate matter of 2.5 microns or less (PM_2.5, μg/m³) were collected from the Melbourne CBD station (−37.8074, 144.97). These sites were selected pragmatically–Alphington site provided most covariate data but was missing a large period of PM_2.5 measurements in 2021. Hence, for PM_2.5, Melbourne CBD site was chosen. Data collection methods and temporal trends have been described in Figure S1 in Supporting Information S1 with smoothed trends fitted using the geom_smooth () function in ggplot2.

2.4. Statistical Analysis

Time‐series analyses were conducted using daily aggregate counts of symptom reports as the unit of analysis. Each row corresponded to a single day of the grass pollen season, indexed by date, and included the daily count of symptom submissions as the primary outcome variable. Because the outcome was the daily count of symptom submissions, estimated effects should be interpreted as relative changes in report counts within the app‐using population rather than per‐person risk.

Grass and non‐grass pollen were assigned as a single daily concentration for all submissions on a given calendar day. For air pollutant covariates (PM_2.5, O₃, NO₂), hourly concentrations were first assigned to each participant submission, then, for each day in time series, a submission‐weighted daily mean was computed by averaging the exposure of all submissions that day. Meteorological variables (solar radiation, temperature, wind speed, winds direction) were assigned as a daily mean for all participants, and rainfall was included as daily total (mm). By aligning exposures to submission times, estimates reflect contemporaneous conditions rather than calendar‐day averages.

A distributed lag nonlinear model (DLNM) was used, as first described by Gasparrini et al. (2010), to express the association between airborne grass pollen concentration (grains/m³) and submission reports of AR symptoms. The lag period considered ranged from the same day (lag 0) to 7 days later (lag 7). The final model was specified with a lag period from lag 0 to lag 3 and fitted using a quasi‐Poisson family with GCV‐selected smoothing parameters. This was to ensure final estimates and standard errors were interpretable and appropriate for handling overdispersion. The DLNM cross‐basis used natural cubic splines to describe the shape of the relationship across two dimensions: the level of pollen exposure and the time lag after exposure. Natural cubic splines fixed with 3 degrees of freedom were specified for both the lag and pollen dimension of the cross‐basis. This cross‐basis structure was validated via a grid search sensitivity analysis using Negative Binomial GAMs and ranked by REML scores. A validation for the selected smoothing structure, including the evaluation of all candidate models, is provided in Text S1 and Table S1 in Supporting Information S1.

Covariates were identified a priori from previous literature based on their potential relationships with grass pollen exposure, AR symptoms, or reporting behaviors. A smooth function of calendar time, an automated penalized spline with an upper limit for k‐index of 60, was included and evaluated using the mgcv k‐index diagnostic (Wood & Wood, 2015) to address serial correlation. Additional variables were specified to capture residual temporal biases including day of the week, a factor for each year, a factor for the major bushfires in 2019 (specified as 21 November to 31 December), and number of users from those that submitted on a given day that had also submitted in the previous 7‐day to account for short‐term, active app usage (Model summary can be viewed in Table S2 in Supporting Information S1).

The final model can be expressed as:

$$\log \left(E\left(Y_t\right)\right)={\beta }_0+\text{cb}\left({\text{grass}}_t,\text{lag}=3\right)+{\beta }_1\mathit{·}{\text{rainfall}}_t+{\beta }_2\mathit{·}{\text{temperature}}_t+{\beta }_3\mathit{·}{\text{solar}}_t+{\beta }_4\mathit{·}{\text{pollen}}_{\text{other},t}+{\beta }_5\mathit{·}{\text{PM}}_{2.5t}+{\beta }_6\mathit{·}{\text{NO}}_{2_t}+{\beta }_7\mathit{·}{\text{bushfire}}_t+{\beta }_8\mathit{·}{\text{year}}_t+{\beta }_9\mathit{·}{\text{dow}}_t+{\beta }_{10}\mathit{·}{\text{submissions}}_{7,t}+\left[f_1\left({\text{VWD}}_t\right)\mathit{·}I\left({\text{available}}_t=1\right)\right]+s\left(t,k=60\right)$$

Where $Y_t$ represents the daily count of app submission on day $t$; $\text{cb}\left({\text{grass}}_t,\text{lag}=3\right)$ represents the DLNM crossbasis function modeling the effect of grass exposure up‐to a maximum 3‐day lag; ${\beta }_1$ to ${\beta }_{10}$ are the regression coefficients for the included covariates rainfall, minimum temperature, solar radiation, other pollen, PM_2.5, NO₂, the 2019 bushfires (21 November 2019–31 December 2019), year of study, day‐of‐week, and number of device Universally Unique Identifier (UUIDs) in the previous 7 days (submissions_7); $f_1\left(\right)$ is a smooth functions of vector wind direction, included conditional on data availability since data was missing for the 2 years (2017 and 2018) of the study; and $s\left(t\right)$ is a natural cubic spline accounting for temporal trends.

Relative risk estimates were generated over the range of observed grass pollen concentrations, with the common reference value specified as the 5th percentile of pollen exposure (1 grain/m³). This was chosen since days with 0 grains/m³ may not reflect baseline conditions if biased by heavy periods of rain (Figure S2 in Supporting Information S1). We computed both the individual lag‐specific and cumulative RRs across the 3‐day lag window.

Autocorrelation and multicollinearity diagnostics, including ACF and PACF plots, Ljung‐Box tests, variance inflation factors (Table S3 in Supporting Information S1), and concurvity, were used to confirm residual independence and guide final covariate inclusion.

Where grass pollen exposure thresholds used in Australia were specifically investigated, exposure‐response relationships were modelled continuously, then, RR and CIs were estimated at the upper bound of each exposure category. Because these categories (low, moderate, high, and extreme) are used in public warning systems (AirHealth Pty Ltd, 2025; AirRater, 2025; QUT Allergy Research Group, 2025), the upper bound of each category (19, 49, 99, and 150 grains/m³) was selected as a conservative estimate to ensure that the reported risks cover all days assigned to that category.

2.5. Subgroup and Sensitivity Analyses

Several subgroup analyses were conducted. We stratified by reported symptom severity and asthma status. We also assessed effect modification by age and sex through stratified analyses. Participants reporting use of hay fever medication at the time of submitting a symptom report reflect a complex demographic since medication use affects severity of symptoms but is also taken by a those that suffer more severe symptoms–stratified analyses were also conducted to investigate this demographic. Results were presented as relative risks (RRs) with 95% confidence intervals, consistent with the primary analyses.

To compare the shape of the lagged‐risk profiles between self‐reported asthmatic and non‐asthmatic app users, stratification was used. The data set was divided based on asthma status. For each group, the primary analysis was repeated. To isolate and compare temporal dynamics of risk, without capturing behavioral differences in reporting or baseline risk, the three‐dimensional exposure‐lag‐response surfaces were normalized. For each discrete pollen concentration level, the relative risk across the 0–3 days lag period was scaled such that the maximum relative risk at any lag was set to 100%, and all other lags were expressed as a percentage of this maximum. Normalization removed the effect of absolute risk magnitude thereby mitigating behavioral biases in symptom reporting.

Sensitivity tests were performed to improve the robustness of findings. A sensitivity analysis was conducted by including additional covariates. These covariates were maximum temperature and ozone. We conducted additional analyses by expanding the spatial inclusion radius from 30 to 50 km around the pollen trap site to test the robustness of associations under increasing exposure uncertainty, considering that pollen concentrations measured at a single trap become less representative as distance increases.

2.6. Software

All analyses and visualisations were conducted using R version 4.4.2 (R Core Team, 2024). For DLNM, the dlnm 2.4.10 package in R was used (Gasparrini, 2011); for fitting the model, the mgcv 1.9–1 R package was used (Wood & Wood, 2015).

2.7. Ethics

Data collection and analysis was approved by the Human Research Ethics Committee of the University of Melbourne and Melbourne Health (1647764.1 and QA2015148). Ethics for analysis in this manuscript granted by The University of Queensland (2025/HE001610).

3. Results

3.1. Descriptive Statistics for Population Demographics

The demographic characteristics of all app users of the Melbourne Pollen App are shown in Table 1. After filtering to meet the inclusion criteria, 114,834 symptom reports were by submitted from 25,194 unique devices across 644 days. A slightly higher proportion of users identified by device ID were female (53%) than male (47%), however representation by submission count reflected a distribution more consistent with the Greater Melbourne population. Male app users generally submitted slightly more symptom reports relative to their representation (49% of reports were male). Looking at vulnerable groups including asthmatics (37% of reports were male) and asthmatics submitting severe symptom reports (32% of reports were male), females tend to report more relative to their representation. Participants identified as with asthma (34%) were more prevalent in the sample compared to the broader Melbourne population. This is also the case for hay fever (65%). The number of submissions made by these demographics was greater relative to their representativeness (36% and 71%, respectively). Participants aged 31–50 years represented roughly half of the app users, with their contribution also forming 51% of all submissions. Notably, a greater proportion of the population reported use of medication to manage hay fever symptoms. Medication use was higher amongst asthmatics, and even higher amongst asthmatics reporting severe symptoms.

Table 1.

Demographic Distribution of App Users and Their Submissions

Demographic of app user (N, %)	All submissions (any severity)	Asthmatic submissions (any severity)	Severe asthmatic submissions (severity = 5)	2021 ABS Census Greater Melbourne (Australian Bureau of Statistics, 2021)
N = 25,194 a	n = 114,834 b	n = 37,707 b	n = 3,915 b	N = 4,917,750
Sex
Female (11,649, 53%)	53,772 (51%)	23,642 (63%)	2,645 (68%)	2,497,033 (51%)
Male (10,150, 47%)	51,163 (49%)	14,064 (37%)	1,270 (32%)	2,420,718 (49%)
NA (4,689)	9,899	1	0	–
Age
<18 (1,731, 8%)	4,081 (4%)	–	–	1,055,963 (21%)
18–25 (2,807, 13%)	19,119 (8%)	2,428 (7%)	426 (11%)	512,831 (10%)
26–30 (3,302, 14%)	12,144 (11%)	3,761 (10%)	485 (12%)	389,770 (8%)
31–40 (5,427, 25%)	31,735 (29%)	10,169 (27%)	1,287 (33%)	794,359 (16%)
41–50 (4,833, 22%)	23,647 (22%)	7,873 (21%)	815 (21%)	653,545 (13%)
51–60 (2,286, 10%)	13,452 (12%)	5,340 (14%)	461 (12%)	573,032 (12%)
61+ (1,758, 8%)	14,659 (14%)	8,070 (21%)	440 (11%)	938,235 (19%)
NA (3,050)	5,997	66	1	–
Asthma c
No (14,491, 66%)	67,213 (64%)	–	–	4,531,403 (92%)
Yes (7,454, 34%)	37,707 (36%)	37,707 (100%)	3,915 (100%)	386,347 (8%)
NA (3,249)	9,914	0	0	–
Hay fever c
No (7,722, 35%)	30,897 (30%)	5,566 (15%)	426 (11%)	–
Yes (14,460, 65%)	74,011 (71%)	32,139 (85%)	3,489 (89%)	–
NA (3,012)	9,926	2	0	–
Medication d
No (11,085, 44%)	43,796 (38%)	11,220 (30%)	489 (12%)	–
Yes (16,628, 66%)	71,038 (62%)	26,487 (70%)	3,426 (88%)	–

Note. Groups are not mutually exclusive.

^a Number of unique participants identified by device (% of sample).

^b Number of submissions (% of all submissions).

^c [demographic] indicates a self‐reported doctor's diagnosis.

^d Antihistamines, decongestant, corticosteroids, immunotherapy and alternative/other.

The spatial distribution of symptom reports demonstrated higher reporting density toward Melbourne's central areas, with the greatest density of reports occurring within Inner Melbourne (Figure 1). Higher report densities also extended eastward from the inner city into the eastern suburbs. Outside of Inner Melbourne, primarily southerly, there were also pockets with greater reporting density.

Figure 1.

Spatial distribution and intensity of all allergic rhinitis (AR) symptom reports submitted between October and December, 2017–2023, in Melbourne, Australia. Individual reports are illustrated using red semi‐transparent points. Overlaid is a two‐dimensional kernel density estimate surface, estimating reporting intensity (reports per km²). Australian Geography Standard (ASGS) 2021 Mesh Block boundaries were used to delineate the inner Melbourne study area. All spatial layers were transformed to Web Mercator projection (EPSG:3857). The map was defined by a 30 km radius centered on the pollen monitoring site—reports submitted outside of this radius are not shown. The map displays reporting density and does not represent population‐adjusted risk. Basemap: © CARTO, © OpenMapTiles, © OpenStreetMap contributors. Refer to Figure S3 in Supporting Information S1 for a duplicate map without Inner Melbourne.

3.2. Preliminary Results

A 7‐day lag was first tested based on observations of higher lag periods in existing literature. The exposure‐response association was essentially flat by lag‐3, indicating no additional delayed effect at longer lags (Figure S4 in Supporting Information S1). Including lags beyond lag‐3 would bias cumulative risk estimates with non‐informative lag periods, whereby the model forces estimate of risk for days beyond reasonable effect following exposure. The primary analysis was correctly specified to the lag‐0 to lag‐3 interval.

3.3. Primary Analysis

Grass pollen exposure was associated with a non‐linear increase in relative risk (RR) for symptom reporting (Figure 2). For symptoms rated as slight‐severe, RR increased more steeply at low grass pollen concentrations and began slowing consistently at around 30 grains/m³, continuing to slow toward a flat response as grass pollen concentration approached the highest day of exposure (319 grain/m³–not shown in figure). Surface plots in Figure 2 were presented with a maximum typical threshold to be observed during a pollen season in Melbourne (150 grains/m³) due to data sparsity and wide confidence intervals.

Figure 2.

Three‐dimensional surface plots illustrating Relative Risk (RR) for hay fever symptom reporting across a 3‐day lag at increasing ambient grass pollen concentrations, stratified by the severity of reported hay fever symptom. The RR for each stratum was calculated against a common low‐exposure reference point (the 5th percentile of grass pollen concentration). RR estimates show the multiplicative increase in risk for each stratum relative to their own unique baseline risk. In each of the six plots, the ambient grass pollen concentration in grains/m³ is displayed on the z‐axis, the RR is displayed on the y‐axis, and the lag from lag 0 (no lag) to lag 3 (3 days after exposure) is displayed on the x‐axis. Data presented is for the pollen seasons Oct‐Dec from 2017 to 2023 in Melbourne, Australia. Individual plots are labeled with the appropriate ordinal category name, followed by the corresponding numerical classifier in square brackets.

RR peaked on the day of exposure and subsided rapidly (Figure 1 & Table 2). Visually, the surface plots (Figure 2) form a convex, upwards profile along the pollen axis and a decay slope along the lag axis, demonstrating an immediate acute increase in AR symptom risk, which subsides rapidly. Table 2 shows the association was greatest at lag 0, where statistical significance was consistently reached, indicating AR reports were heightened on the day of exposure. The effect decreases over the subsequent 3 days, generally reaching null‐effect by lag‐2 and remaining thereafter (Table 2).

Table 2.

Table of RR and 95% CI Extracted From Continuous Data Modeled Using a Quasi‐Poisson Regression With DLNM in Figure 2

Symptom severity	Grass (grains/m3)	Relative risk (95% confidence interval), relative to stratum baseline risk
		Lag‐0	Lag‐1	Lag‐2	Lag‐3	Cumulative
Any Symptoms [2–5]	19	1.31 (1.16–1.47)	1.08 (0.99–1.17)	0.98 (0.91–1.06)	0.99 (0.90–1.09)	1.37 (1.09–1.73)
	49	1.52 (1.34–1.72)	1.12 (1.04–1.21)	0.97 (0.90–1.04)	0.98 (0.89–1.07)	1.61 (1.28–2.03)
	99	1.72 (1.50–1.98)	1.16 (1.07–1.26)	0.96 (0.89–1.03)	0.96 (0.86–1.06)	1.83 (1.42–2.35)
	150	1.80 (1.54–2.10)	1.19 (1.09–1.29)	0.96 (0.88–1.04)	0.94 (0.84–1.06)	1.92 (1.46–2.54)
Severe [5]	19	1.55 (1.28–1.87)	1.12 (0.99–1.27)	0.95 (0.85–1.07)	0.95 (0.82–1.09)	1.56 (1.08–2.27)
	49	2.10 (1.72–2.58)	1.19 (1.06–1.34)	0.90 (0.80–1.01)	0.91 (0.80–1.05)	2.06 (1.42–2.99)
	99	2.78 (2.45–3.45)	1.26 (1.12–1.42)	0.86 (0.77–0.96)	0.88 (0.77–1.02)	2.65 (1.81–3.87)
	150	3.16 (2.49–4.00)	1.31 (1.15–1.48)	0.85 (0.76–0.95)	0.86 (0.74–1.01)	3.02 (1.99–4.59)
Moderate [4]	19	1.52 (1.31–1.77)	1.12 (1.01–1.24)	0.97 (0.88–1.07)	0.99 (0.88–1.11)	1.62 (1.20–2.18)
	49	1.89 (1.61–2.22)	1.16 (1.05–1.29)	0.93 (0.85–1.02)	0.96 (0.96–1.08)	1.97 (1.46–2.65)
	99	2.25 (1.88–2.68)	1.21 (1.09–1.34)	0.90 (0.82–0.99)	0.93 (0.93–1.05)	2.28 (1.66–3.13)
	150	2.41 (1.98–2.94)	1.25 (1.13–1.40)	0.90 (0.82–1.00)	0.90 (0.90–1.04)	2.47 (1.74–3.05)
Slight [3]	19	1.36 (1.20–1.53)	1.09 (1.00–1.19)	1.00 (0.92–1.08)	1.04 (0.94–1.15)	1.54 (1.22–1.95)
	49	1.57 (1.38–1.78)	1.13 (1.04–1.23)	0.99 (0.91–1.06)	1.02 (0.92–1.12)	1.76 (1.39–2.24)
	99	1.71 (1.48–1.98)	1.17 (1.08–1.27)	0.97 (0.90–1.05)	0.98 (0.88–1.09)	1.90 (1.47–2.46)
	150	1.72 (1.46–2.02)	1.20 (1.10–1.32)	0.99 (0.91–1.08)	0.96 (0.85–1.09)	1.98 (1.49–2.62)
Mild [2]	19	1.14 (1.02–1.28)	1.04 (0.96–1.13)	0.99 (0.92–1.07)	0.98 (0.89–1.09)	1.16 (0.92–1.45)
	49	1.17 (1.03–1.32)	1.07 (0.99–1.16)	1.01 (0.94–1.09)	0.98 (0.89–1.09)	1.24 (0.99–1.56)
	99	1.13 (0.99–1.31)	1.09 (1.00–1.18)	1.04 (0.97–1.12)	1.01 (0.90–1.12)	1.30 (1.01–1.66)
	150	1.08 (0.92–1.26)	1.08 (0.99–1.18)	1.07 (0.98–1.16)	1.04 (0.92–1.17)	1.28 (0.98–1.68)
No Symptoms [1]	19	0.99 (0.89–1.11)	0.97 (0.90–1.05)	0.98 (0.90–1.05)	1.00 (0.90–1.11)	0.94 (0.75–1.19)
	49	0.98 (0.87–1.10)	0.99 (0.92–1.07)	1.00 (0.93–1.08)	1.01 (0.91–1.12)	0.98 (0.78–1.23)
	99	0.94 (0.82–1.09)	1.00 (0.92–1.09)	1.03 (0.96–1.11)	1.02 (0.92–1.14)	1.00 (0.78–1.28)
	150	0.90 (0.76–1.06)	0.98 (0.89–1.08)	1.03 (0.94–1.12)	1.02 (0.90–1.17)	0.93 (0.70–1.23)

Note. Data shown for whole sample and stratified‐severity subgroups. Grass concentrations were defined using upper limits of grass pollen thresholds used for risk classification defined by Ong et al. (1995).

RR increased as the severity of symptoms reported increased. Higher RRs were consistently observed for more severe AR reports at equivalent grass pollen concentrations, relative to each stratum's respective baseline risk. Participants reporting a severity score of 1 (no symptoms) exhibited no significant association across any lags, consistent with expectations for the negative control group.

3.4. Comparing the Temporal Dynamics of Exposure‐Response Between Asthmatics and Non‐Asthmatics

Asthmatic participants exhibited a greater increase in RR for severe symptom reporting, relative to their baseline risk, compared with non‐asthmatics at “low” and “moderate” grass concentrations (Figure 3). At low grass pollen concentrations, asthmatics that reported severe AR symptoms exhibited a 94% increase in risk (RR 1.94, 95% CI 1.30–2.89). This magnitude was lower but still significant in non‐asthmatics reporting severe symptoms, who exhibited a 57% increase in risk (RR 1.57, 95% CI 1.28 to 1.91). Slight to moderate symptom reports generally exhibited RR of a higher magnitude in non‐asthmatic subgroups, with estimates closer in magnitude to their severe symptom subgroups. However, 95% confidence intervals overlap and widen as severity increases.

Figure 3.

Cumulative risk estimates over the 3‐day lag period, stratified by asthma status and symptom severity. RR estimates show the multiplicative increase in risk for each subgroup relative to their own unique baseline risk. A complete table of results can be found in Supporting Information S1. Data was modeled continuously, and point estimates were extracted at the corresponding grass pollen concentrations shown on the x‐axis. Data table with estimates and 95% CI are shown in Table S4 in Supporting Information S1.

After normalizing for the magnitude of risk by scaling each exposure‐lag surface so that the highest RR within each stratum equals 100%, at a population‐level asthmatics appear to exhibit heightened persistence of RR for severe hay fever symptoms compared with non‐asthmatics (Figure 4). At a “low” grass pollen concentration of 19 grains/m³, asthmatics reached their peak RR at lag 0 and decayed to 72% of that peak at lag‐3. In contrast, non‐asthmatics decayed more substantially to 60% of its own peak value over the same period. This indicates that while both asthmatics and non‐asthmatics exhibit a rapid onset of risk, the risk profile for asthmatics is more sustained, and thus potentially more burdensome.

Figure 4.

Normalized exposure‐lag‐response surfaces for asthmatics (blue) and non‐asthmatics (green). Surface plots are individually scaled across each pollen concentration increment to a maximum of 100% for each plot respectively, to isolate the temporal dynamics of risk for comparison. Percentages should be interpreted as relative values and not as absolute magnitudes.

3.5. Subgroup Analyses and “Low” Grass Pollen Concentrations

Subgroup analyses for severe symptom reporting at “low” grass pollen concentrations demonstrated clear heterogeneity in the magnitude and duration of heightened risk across subgroups (Figure 5). Across all subgroups, non‐asthmatics typically exhibited a lag‐0 RR close to the cumulative RR estimate, indicating later lags contributed minimally. By contrast, asthmatics exhibited a higher cumulative RR relative to their lag‐0 RR, with later lag estimates showing a clear downward, but non‐zero, trend. Though confidence intervals overlap and lower magnitude lag days did not reach statistical significance, there is a consistent pattern indicating later lags appear to contribute the higher cumulative risk. Across both asthmatics and non‐asthmatics, risk was highest amongst app users aged under 40 years for both same‐day risk estimates and cumulative lag estimates. Users reporting medication use demonstrated RR of greater magnitudes than users who did not. This was consistent for both individual lag estimates and cumulative lag estimates.

Figure 5.

Subgroup analysis for risk of symptoms reported as severe at low grass pollen exposure. RR estimates show the multiplicative increase in risk for each subgroup relative to their own unique baseline risk. Cumulative RR, and lagged RR from lag 0 (no lag) to lag 3 (3 days after exposure) are uniquely identified by color and shape. Data presented is for the pollen seasons Oct‐Dec from 2017 to 2023 in Melbourne, Australia. A complete table with RR and 95% CI is shown in Table S5 in Supporting Information S1.

3.6. Associations of Covariates With AR Symptom Reporting

Environmental, behavioral, and temporal covariates were associated with AR symptom reporting. Higher minimum temperature (β = 2.8e−02, p = 1.4e−05), solar radiation (β = 2.8e−02, p = <2e−16), and non‐grass pollen concentrations (β = 1.0e−03, p = 0.0018) were significantly associated with AR symptom reporting. PM_2.5 (β = 1.31e−02, p = 0.037) and NO₂ (β = 6.4e−03, p = 0.406) exhibited positive associations, but only PM_2.5 reached statistical significance. The covariate representing repeat reporting within the preceding 7 days demonstrated a positive association (β = 4.76e−04, p = 2.67e−12). The bushfire event in 2019 (β = −5.312e−01, p = 0.031) and rainfall (β = −1.37e−02, p = 0.00063) were each negatively associated with AR symptom reporting. Time of day and wind direction demonstrated significant non‐linear associations (both p < 2e−16). Both of these variables exhibited high concurvity, and therefore the main effects of these variables cannot be uniquely identified. They were retained to control for confounding and reduce residual serial correlation. Alternative specifications including ozone and maximum temperature did not materially change the estimated association between grass pollen and AR symptoms.

4. Discussion

The effect of grass pollen on allergic respiratory symptom reporting, among users of the Melbourne Pollen App, was greatest on the day of exposure, and demonstrated comparatively smaller lagged associations. Risk declined the day following exposure and reached null effect the day thereafter (Figure 2). This temporal pattern generally aligns with the biological mechanisms of a type I hypersensitivity reaction where the early‐phase response is triggered within 30 min of exposure and the late phase typically peaks 6–9 hr after exposure, subsiding within 1–2 days (Galli et al., 2008).

We found that currently used thresholds for classifying pollen risk exposure as “low” (Ong et al., 1995) may not accurately reflect symptom burden across the entire app‐using population, particularly for those individuals reporting severe symptoms or asthmatic. We observed a non‐linear dose‐response relationship whereby risk rose steeply at low‐to‐moderate concentrations of grass pollen. Even historically “low” levels pose a population‐level burden for asthmatics (cumulative RR 1.94, 95% CI 1.3–2.89 for asthmatics reporting severe symptoms). As the severity of symptom reports submitted increased, so did the magnitude of RR estimates at equivalent grass pollen concentrations, potentially reflecting greater susceptibility, variation in reporting tendencies among users experiencing more severe symptoms, or exposure to grass pollen of greater allergenicity (Buters et al., 2015).

Although medication use reflects active disease management, symptom reports from asthmatic participants co‐reporting medication use were observed to exhibit heighted risk during periods of elevated pollen exposure. Of the 3,915 severe symptom reports from asthmatics across the seven pollen seasons, 88% of reports co‐reported having taken medication (antihistamines, decongestants, corticosteroids, immunotherapy and alternative/other) to manage their hay fever symptoms. In Australia, when diagnosed with asthma, action plans are typically prescribed to enable asthmatics to manage their disease. As of 2022, 32% of asthmatics had a written action plan and 34% of asthmatics take asthma related medication daily (Australian Bureau of Statistics, 2022). Asthmatic app users reporting severe symptoms and medication use demonstrated an 86% increase in RR from their baseline at the “low” grass pollen classification. In contrast, asthmatic app users reporting severe symptoms and no medication use exhibited no significant increase in AR symptom risk. Although it is unknown whether medication was taken prophylactically, after symptom onset, or both, the heightened RR among medicated participants indicates that these individuals likely represent a more sensitive subgroup that the “low” pollen threshold may not adequately classify.

To interpret temporal observations, it is essential to consider the alignment of pollen release, monitoring practices, and symptom reporting. Melbourne's diurnal cycle for grass pollen concentration is complex and varies depending on meteorological conditions. Observations from previous grass pollen seasons in Melbourne suggest that ambient levels begin to rise early in the morning from around 09:00 hr, increase through midday, and peak in the early evening (approximately 18:00 hr) (Mampage et al., 2025). Local disturbances such as wind gusts can resuspend pollen locally, and long‐range pollen transport could also modify concentration. Pollen counts were collected daily using a volumetric sampler, with the 24‐hr count assigned to a given day reflecting exposure from 09:00 hr on that day to 09:00 hr the following day. This sampling period aligns well with the window of biologically relevant exposure of Melbourne's diurnal cycle, and consistent with observed diurnal participant reporting trends (Figure S5 in Supporting Information S1), thereby reducing temporal measurement bias and strengthening the study. Because DLNM smooths effects across lag days, any redistribution of effects attributable to temporal misalignment across lags is likely to be small and unlikely to influence the detection of delayed associations.

Successive days of exposure might extend elevated symptom risk beyond the timeframe we observed in this study, which would explain why longer 2–5‐day lags were observed to have a higher risk for adult asthma hospitalization previously (Osborne et al., 2017). Given our results indicated that AR risk persisted into the day following exposure, and in some demographic's potentially days thereafter, it can be inferred that during periods of sustained high grass pollen exposure, new exposure may coincide with residual lagged effects from the preceding days, thereby compounding overall risk. Exploratory analysis revealed crude signals that support this hypothesis (Table S6 in Supporting Information S1). However, while our study can provide insight for the delayed impact of individual daily exposures, it does not explicitly model cumulative or interactive effects of successive high exposures. Consequently, although overlapping lag effects suggest extended and potentially heightened risk, the model cannot definitively quantify these putative effects.

The priming effect provides a plausible theoretical framework for successive exposure driving longer lagged responses. Repeated intranasal allergen challenge has been shown to demonstrate a stronger immediate response to subsequent allergen exposure (Ciprandi et al., 1998; Orban et al., 2021). During the early phase of a type I hypersensitivity reaction, mast cells release histamine within minutes of allergen exposure. Histamine mediates acute AR before being rapidly degraded (Vitte et al., 2022). During the late phase, Th2 cytokine expression increases and eosinophils accumulate in the nasal mucosa (Orban et al., 2021). Together, these processes are thought to create a state of heightened airway responsiveness that lower the threshold for and increase the intensity of AR symptoms. At the population level, this provides a biologically plausible explanation for extended lag effects during periods of sustained allergen exposure.

A comparable study by Jones et al. (2021) using crowd‐sourced AirRater data in Tasmania identified similar same‐day, and lag 1 associations between grass pollen and AR symptoms using a personalized case time series (CTS) approach. CTS treats days without reported events as zero counts, although non‐use of mobile health applications may reflect under‐reporting rather than true absence of symptoms. Despite this limitation, personal forecasting is a potential answer for management of allergic respiratory disease (Kmenta et al., 2014), and our study supports this given the heterogeneity in association observed across our sample of Melbourne Pollen App users of different demographics. While Jones et al. (2021) primarily focused on population‐averaged responses across pollen taxa, and interaction with PM_2.5, our study identified consistent population‐level temporal associations with the Melbourne Pollen App userbase. Together, the consistency between individual‐level CTS findings and population‐level time‐series estimates strengthens confidence in the validity and generalizability of the observed temporal grass pollen and AR symptom associations.

The sensitivity analysis testing the spatial inclusion radius reinforced the need for accurate exposure classification amongst smaller samples. While the risk estimates for the primary analysis and asthma‐stratified analysis remained stable when the inclusion radius was increased from 30 to 50 km, the estimates for the asthmatic app users were attenuated across some smaller subgroups. This difference is likely a consequence of the interplay between data density and exposure misclassification. The primary analysis, with higher reporting density, provided a robust signal that was resilient to the noise introduced by spatial smoothing at 50 km. In contrast, the sparser asthmatic stratums exhibited a more fragile signal. Though remaining statistically significant, effect estimates for these more fragile signals were biased toward the null. This finding reinforces the specified 30 km radius for our analysis, as it provides a more reliable exposure classification and estimate for risk.

A strength of this study was the use of smartphone‐based symptom reporting, which enabled timely measurement of outcomes within a large sample of app users with rich covariate data. While interaction with the app may have been biased toward periods of more symptoms, the timely convenience of smartphone reporting nonetheless may reduce recall bias and provided large‐scale data beyond hospital‐based endpoints. Emergency department presentation or hospitalization often used as an outcome for severe disease cannot accurately reflect the experience of all individuals, particularly for asthmatics. Such outcomes are influenced by the individual's capacity to manage their symptoms, access to healthcare, health‐seeking behavior, and physician decision‐making, beyond simply clinical severity alone. Other individuals experiencing severe symptoms may manage them at home, for example, using prescribed rescue medication, and thus remain invisible to hospital‐based outcomes. Other tools for measuring outcomes including health‐diaries are also popular but may be more prone to recall bias. Importantly, app‐based symptom reporting provided population level insight into allergic disease symptom burden that may not be captured through clinical endpoints or symptom diaries alone.

Self‐reported symptom data collected via mobile apps may not perfectly reflect true symptom prevalence due to inherent selection and measurement biases. Individuals with a history of allergic disease may more proactively interact with the app leading to increased reporting from health‐conscious individuals. Demographic selection bias also appears evident and likely stems from older adults being less willing to engage with mobile phone‐based apps (Fox & Connolly, 2018). App users aged 31–50 years account for 47% of participants but comprise only 29% of the Greater Melbourne population and 38% of adult asthma prevalence. Conversely, older adults (aged 61 years and older) represent 8% of the app users, compared with 19% of Greater Melbourne and 25% of adult asthma prevalence (Australian Bureau of Statistics, 2021). Overrepresentation of users aged 31–50 years and underrepresentation of adults older than 61 years therefore cannot be explained by asthma prevalence alone. Beyond demographic selection bias, measurement bias may emerge as reporting frequency could partly reflect digital habits and mobile phone usage, which might differ by age or sex, rather than symptom burden. Consequently, the observed associations between grass pollen and AR symptoms, particularly within demographic subgroups, may have been influenced by differential participant selection and symptom reporting.

There are an additional two limitations inherent to the app design that may introduce reporting bias. On days where pollen was forecasted as “high” (≥50 grains/m³) or “severe” (≥100 grains/m³), early morning symptom reports (07:00–09:00 hr) may be influenced by push notifications (Figure S5 in Supporting Information S1). Push notification warnings were sent to app users when the pollen count was forecasted as high or severe, which introduces reporter bias. Since some users of the app rely upon these notifications to manage their symptoms, withholding potentially lifesaving information was not an option. These push notifications may bias lagged associations between the number of symptom reports and grass pollen concentration away from the null hypothesis. However, the presence of comparable associations at lower grass pollen concentrations (<50 grains/m³) (Table 2), where push notification effects are minimal, indicate that reporting bias does not materially undermine the interpretation of the observed associations. Second, when participants log into the app, they can view the pollen forecast prior to submitting their symptom reports, which may influence reporting through expectation effects. As with push notifications, this is an inherent design of the application which cannot be ethically eliminated. Participants viewing the pollen count prior to reporting could lead to inflated symptom reporting on days with elevated forecasted pollen levels. The consistency of association across lower pollen concentrations suggests that this effect is unlikely to explain the observed relationships.

There are limitations to the study that should also be acknowledged. Participants could only report once per day, precluding analysis of intra‐daily symptom variation. This study did not capture micro‐scale variability in grass pollen exposure limiting spatial inference. Tracking symptom reports over time using device identifiers to approximate individual participants has limitations. Participants may change devices or use multiple devices across pollen seasons, creating uncertainty in identifying continuous longitudinal participation. In Australia, consumers typically possess phones for 3.17 years (Islam et al., 2020). Although our model adjusted for many covariates and confounding variables, residual confounding from unmeasured factors such as allergen cross‐reactivity, sub‐pollen particles (Mampage et al., 2025), or behavioral biases cannot be excluded. An important consideration pertains to the comparability of effect estimates between different subgroups. As the outcome reflects the daily count of symptom reports, estimand effects represent relative changes in the number of reports within each subgroup, not per‐person risk. The shapes of the exposure‐response curves are directly comparable across subgroups. The magnitudes are somewhat comparable as population‐level responses from Melbourne Pollen App users, however, should be interpreted with caution as between‐group differences may reflect both biological susceptibility and reporting behavior, including app engagement or reporting habits or variation in baseline risk. To adjust for variation in active “at‐risk” app users, a covariate was included that sums the daily number of reports from individuals who also submitted data within the preceding 7 days. It was not viable to accurately compute a denominator for active app use by group as app use reflected ongoing recruitment, loss, and variable reporting consistencies across years (Figure S6 in Supporting Information S1). A computed denominator risks introducing a differential bias whereby reporting pattern differences between groups artificially inflate or deflate effect estimates.

This study provides population‐level evidence that grass pollen exposure is associated with allergic rhinitis symptom reporting on the day of exposure, and that measurable symptom risk occurs even at pollen concentrations traditionally classified as “low.” Although symptom risk declined rapidly after exposure for most users, asthmatic participants and other clinically sensitive subgroups exhibited greater cumulative risk and retained statistically significant associations at later lags. These results indicate that commonly used pollen classifications based on fixed daily thresholds may not fully capture short‐term symptom burden across heterogenous populations. Refining how pollen information is communicated, particularly to individuals who experience symptoms at lower exposure levels or over longer periods, may improve the relevance and usefulness of pollen forecasts and early‐warning systems.

Conflict of Interest

The authors declare no conflicts of interest relevant to this study.

Supporting information

Supporting Information S1

Availability Statement

The data that support the findings of this study contain potentially identifiable geolocation and health‐report information and are therefore not publicly available. De‐identified data and analytic code may be made available from the corresponding author on reasonable request, subject to ethics and data‐sharing approvals.