The impact of air pollution alert services on respiratory diseases: generalized additive modeling study in South Korea

This study investigates whether ${\text{P}}{{\text{M}}_{2.5}}$ alert messages affected the number of hospital visits and admissions for respiratory sicknesses in South Korea, from January 2015 to October 2019, using GAM. We answer the following research questions: (a) Do the two types of ${\text{P}}{{\text{M}}_{2.5}}$ alert message system—WEA and AIT—reduce the number of new patients with respiratory sicknesses? (b) Do the effects of alarms differ by the type of respiratory sickness (COPD, RTI, asthma, and pneumonia)? We hypothesize that alerts discourage people from outdoor activities so as to avoid exposure to PM, consequently reducing the number of hospital visits and admissions for respiratory sicknesses.

The study site is South Korea (figure 1). ${\text{P}}{{\text{M}}_{2.5}}$ in South Korea deserves greater attention, since the level there is the highest among OECD countries and ranked 26th among all countries in the world in 2019 (Visual 2020). In 2019, the annual average ${\text{P}}{{\text{M}}_{2.5}}$ concentration in South Korea was 24.78 $\mu {\text{g}}\ {{\text{m}}^{ - 3}}$, which is more than twice the WHO recommendation of 10 $\mu {\text{g}}\ {{\text{m}}^{ - 3}}$. The reason for such unfavorable conditions is the high population density and the high proportion of urbanized area; 82% of the population lives in urban areas (United Nations 2018, Kim et al 2019b). As the population is concentrated in urban areas, so as the sources of air pollutants, such as construction activities, traffic congestion, industrial and power generation plants (Cho and Choi 2014, Murray et al 2020).

Zoom In Zoom Out Reset image size

The Korean government enacted the Special Act on the Reduction and Management of Fine Dust in 2018 and implemented a comprehensive plan for PM management. Under the act, the detrimental effects of ${\text{P}}{{\text{M}}_{10}}$ and ${\text{P}}{{\text{M}}_{2.5}}$ were recognized, and their levels began to be reported and forecast through a public air pollution alert system. The predicted ${\text{P}}{{\text{M}}_{2.5}}$ level is divided into four grades: good (from 0 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ to 15 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$), normal (from 16 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ to 35 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$), bad (from 36 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ to 75 $\mu {\text{g}}\,{{\text{m}}^{ - 3}}$), and very bad (over 76 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$). People can access real-time information through online website and smartphone applications. An alert is issued as a watch or a warning when the measured concentration exceeds 75 $\mu {\text{g}}\ {{\text{m}}^{ - 3}}$and 150 $\mu {\text{g}}\ {{\text{m}}^{ - 3}}$, respectively, for more than 2 h. In February 2019, another alert category, emergency reduction measure, was created; it is issued when the current day's measured ${\text{P}}{{\text{M}}_{2.5}}$ concentration and the next day's predicted ${\text{P}}{{\text{M}}_{2.5}}$ concentration both exceed 50 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ (Ministry of Environment 2019) (table 1).

Table 1.  ${\text{ P}}{{\text{M}}_{2.5}}$ alert standards by concentration level.

Level	Method	Criteria
Emergency reduction measure	WEA	(a) The average concentration on the day exceeds 50 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ and the average concentration for 24 h on the next day is expected to exceed 50 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ or (b) The alert of watch or warning was issued, and the average concentration for 24 h on the next day is expected to exceed 50 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ or (c) The concentration is predicted to exceed the average concentration of 75 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ for 24 h the next day
Watch	AIT	The average concentration per hour is more than 75 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$ and lasts 2 h
Warning	AIT and WEA	The average concentration per hour is more than 150 $\mu {\text{g}}\;{{\text{m}}^{ - 3}}$and lasts 2 h

In Korea, both of the two aforementioned mobile-based warning systems are deployed: WEA and AIT (figure 2). They are operated by different entities, hence the differences in the transmission method, in the necessity for a subscription, and in the levels of the alerts. WEA messaging is organized by the Ministry of the Interior and Safety and sent by local governments. WEA transmits the alert to all people who have mobile phones (Ministry of the Interior and Safety 2015). AIT messaging is managed by the Korea Environment Corporation (K-eco) and local governments, and sends alerts via SMS only to those who subscribe. The subscribers to the service can customize the information that they wish to receive, including the area, the type of alerts, and the current or predicted level of PM. During the study period, from January 2015 to October 2019, both types of the alert—WEA and AIT were issued simultaneously for 22 days. There are 23 and 290 days that only a single type of alert was issued—WEA and AIT respectively (table 2).

Zoom In Zoom Out Reset image size

The spatial analytical unit is the primary local jurisdiction, city and ward, called sigungu in South Korea, and the temporal unit is one day from January 2015 to October 2019. The final sample size is 318 591.

The first dataset is the daily number of visits and admissions to the hospital mainly due to respiratory sicknesses (Korean Standard Classification of Diseases (KCD) codes from J00 to J99), provided by the Health Insurance Review and Assessment Service from January 2015 to October 2019 (Statistics Korea 2010). We categorized the total respiratory sicknesses (J00–J99) into four subsets: COPD (J40–J44), RTI (J00–J06, J22), asthma (J45–J46), and pneumonia (J17–J18) on the basis of the literature (Dominici et al 2006, Dastoorpoor et al 2019, Mason et al 2019). The second dataset is the record of alert message issuances via WEA and AIT systems. The record for WEA was provided by the National Disaster Safety Portal and Public Data Portal. The data consist of the alert date, the area code, the content of the alert message, and the alert level (emergency reduction measure or warning). The record for AIT was obtained from the website called Air Korea. This data includes the alert zone, the alert level (watch or warning), and the date and time of issuance.

We used the air quality record collected by the Korea Environment Corporation. These data consist of the average levels of ${\text{P}}{{\text{M}}_{2.5}}$ per hour with ${\text{P}}{{\text{M}}_{10}}$, sulfur dioxide (${\text{S}}{{\text{O}}_2}$), carbon monoxide (CO), ozone (${{\text{O}}_3}$), and nitrogen dioxide (${\text{N}}{{\text{O}}_2}$). Air pollution monitoring networks are unevenly distributed across administrative districts, and the number of those varied over time (figure 1), from 504 stations in 2015 to 676 stations in 2019 (National Institution of Environment Research 2016, 2019). We also utilized weather records observed by the Automatic Weather System at about 510 stations in South Korea. These data include the average temperature, the average wind speed, and the maximum precipitation in a day. Additionally, to control for the effect of the number of hospitals, we used hospital statistic data provided by the National Statistical Office.

The GAM is an extended version of a generalized linear model (GLM). Like GLM, it allows a non-normal distribution of dependent variables (Wood 2008, Stram 2014). The advantage of GAM is its capacity to deal with nonlinear relationships between independent and dependent variables, using smooth functions (James et al 2013). In this study, we fitted non-linear relationship with two smoothers: thin plate regression spline and cubic regression spline. Thin plate regression spline captures automatically optimized smoothing parameters by the generalized cross-validation criterion without prior knowledge of knot placement (Larsen 2015, Wood 2003). Cubic regression spline draws a cubic spline by evenly distributing a specified number of knots using covariate values (Wood 2017). We used Akaike information criterion (AIC) and un-biased risk estimator (UBRE) to select the model with different smoothers. We used the GAM model with negative binomial distribution and log link function to address the nature of the overdispersed count data (figure 3), in order to answer our research question (Lee et al 2012, Feng et al 2016, Ardiles et al 2018). The final model is given as follow:

$$\begin{align*} & \text{In}[E({\text{Respiratory}}\_{\text{Diseas}}{{\text{e}}_{it}})] \nonumber\\[3pt] & \quad = {\beta _0} + {\beta _1}{\text{Bot}}{{\text{h}}_{it}} + {\beta _2}{\text{Bot}}{{\text{h}}_{i.t - 2}} + {\beta _3}{\text{WE}}{{\text{A}}_{it}} \nonumber\\[3pt] & \quad\quad + {\beta _4}{\text{WE}}{{\text{A}}_{i.t - 2}} + {\beta _5}{\text{AI}}{{\text{T}}_{it}} + {\beta _6}{\text{AI}}{{\text{T}}_{i.t - 2}} \nonumber\\ & \quad\quad + \sum\limits_{k = 1}^3 {{f_k}} \left( {{\text{P}}{{\text{M}}_{2.5}}_{i \cdot \left( {t - 2\left( {k - 1} \right)} \right)}} \right) + \sum\limits_{k = 1}^2 {{f_{k + 3}}} \left( {{\text{PL}}{{\text{T}}_{kit}}} \right) \nonumber\\ & \quad \quad + \sum\limits_{k = 1}^3 {{f_{k + 5}}} \left( {{\text{MT}}{{\text{R}}_{kit}}} \right) + {f_9}\left( {{\text{Hospita}}{{\text{l}}_{it}}} \right) + {f_{10}}\left( {{\text{Tim}}{{\text{e}}_t}} \right) \nonumber\\ & \quad \quad + \sum\limits_{k = 1}^6 {{\beta _{k + 6}}} {\text{DO}}{{\text{W}}_k} + \sum\limits_{k = 1}^3 {{\beta _{k + 12}}} {\text{Seaso}}{{\text{n}}_k} + {\beta _{16}}{\text{SG}}{{\text{G}}_i}. \end{align*}$$

Zoom In Zoom Out Reset image size

The dependent variables are the daily number of new patients with respiratory sickness counted at location i and time t. We analyze the total respiratory disease (Total), as well as COPD (COPD), RTI (RTI), asthma (Asthma), and pneumonia (PNA), separately.

The main question variables are the issuance of alerts by the two types of alert systems: WEA and AIT. Both are specified as dummy variables; we assigned the value one on a day of alert issuance. Both indicates the days with both of the alerts are issued, WEA and AIT indicate the days with only WEA or AIT systems were active respectively. The reference group is the days with none of those alerts was issued. We also include the time lag of the days with those alerts in our model, because they could remain influential for more than a day. As the first lag has a correlation with the subject day (0.582), we employed a second lag for the issuance of alerts (${\text{Bot}}{{\text{h}}_{t - 2}}$, ${\text{WE}}{{\text{A}}_{t - 2}}$, and ${\text{AI}}{{\text{T}}_{t - 2}}$).

We considered the concentration of the air pollutants ${\text{P}}{{\text{M}}_{2.5}}$ (${\text{P}}{{\text{M}}_{2.5}}$; $\mu {\text{g}}\ {{\text{m}}^{ - 3}}$), ozone (${\text{PL}}{{\text{T}}_1};$ 0.01 ppm), and sulfur dioxide (${\text{PL}}{{\text{T}}_2}$; ppm), which have been known to have a negative relationship with respiratory diseases. We employ second and fourth lag for ${\text{P}}{{\text{M}}_{2.5}}$ (${\text{P}}{{\text{M}}_{2.5\left( {t - 2} \right)}}$ and ${\text{P}}{{\text{M}}_{2.5\left( {t - 4} \right)}}$). We excluded ${\text{P}}{{\text{M}}_{10}}$, carbon monoxide, and nitrogen dioxide due to high correlations with ${\text{P}}{{\text{M}}_{2.5}}$ (0.805, 0.569, and 0.452, respectively). Our model includes various meteorological factors (MTR), the average temperature (${\text{MT}}{{\text{R}}_1}$; Celsius), the average wind speed (${\text{MT}}{{\text{R}}_2}$; m s⁻¹), and total precipitation (${\text{MT}}{{\text{R}}_3}$; mm) in a day. Strong air current tends to disperse the concentration of air pollutants, which could offset the harmful effect (Yang et al 2020).

For the time indicator, we specified the cumulative number of days (Time). We also considered a seasonal trend (${\text{Seaso}}{{\text{n}}_1}$ to ${\text{Seaso}}{{\text{n}}_3}$; the reference is fall). The boundary layer height, wind direction, and the main cause of air pollutant depend on the season (Miao et al 2015, Kim et al 2017). With the complex effect of those factors, the average concentration of ${\text{P}}{{\text{M}}_{2.5}}$ also varies. ${\text{P}}{{\text{M}}_{2.5}}$ concentration is relatively low in summer and fall, and high in winter and spring (Park 2021). The study site saw twice as much ${\text{P}}{{\text{M}}_{2.5}}$ in winter compared with summer (Visual 2020). We also specified the variable of the day of the week (${\text{DO}}{{\text{W}}_1}$ to ${\text{DO}}{{\text{W}}_6};$ the reference is Friday), because the number of hospital visits differs by the day. To control for the locational characteristics of each spatial unit, we introduced the fixed effects (SGG). We also specified the number of hospitals (Hospital) on a quarterly basis within the administrative area to scale the patients per hospital. We present the distribution of key variables in figure 4 and descriptive statistics in appendix A.

Zoom In Zoom Out Reset image size

We show the results of GAMs in table 3. The results suggest that the model (3) of cubic regression spline smoother with 20 knots has the best fit among the three models. It has the highest deviance explained and the lowest AIC and UBRE score in models for all respiratory diseases. The model (1) of the thin plate regression spline has the worst fit, which has the lowest AIC and UBRE score in the four respiratory disease models excluding pneumonia. Therefore, we selected the model (3) of cubic regression spline smoother with 20 knots as the final model.

We present the analytical results in table 4. We found that the alert system significantly reduces hospital visits and admissions and that WEA has a greater impact than AIT.

Table 4. Analytical results of the generalized additive model assessing the effect of WEA and AIT alert issuance on all individual respiratory sicknesses and on the total.

Variable	Total	COPD	RTI	Asthma	Pneumonia
Both	−0.086*** (0.024)	−0.132*** (0.031)	−0.100*** (0.026)	−0.058** (0.029)	−0.080** (0.032)
${\text{Bot}}{{\text{h}}_{t - 2}}$	−0.039 (0.025)	−0.025 (0.031)	−0.074*** (0.026)	0.036 (0.029)	−0.059* (0.032)
WEA	−0.125*** (0.012)	−0.130*** (0.015)	−0.155*** (0.012)	−0.084*** (0.014)	−0.100*** (0.015)
${\text{WE}}{{\text{A}}_{t - 2}}$	−0.053*** (0.012)	−0.048*** (0.015)	−0.075*** (0.013)	0.003 (0.015)	−0.091*** (0.016)
AIT	−0.025*** (0.010)	−0.020* (0.012)	−0.019* (0.010)	−0.038*** (0.012)	−0.031** (0.012)
${\text{AI}}{{\text{T}}_{t - 2}}$	0.020** (0.010)	0.022* (0.012)	0.020** (0.010)	0.016 (0.011)	−0.017 (0.012)
$f\left( {{\text{P}}{{\text{M}}_{2.5}}} \right)$	8.883*** (10.515)	7.446*** (8.888)	8.594*** (10.190)	8.159*** (9.695)	7.092*** (8.484)
$f\left( {{\text{P}}{{\text{M}}_{2.5\left( {t - 2} \right)}}} \right)$	8.239*** (9.789)	6.217*** (7.489)	8.317*** (9.877)	6.500*** (7.810)	5.710*** (6.913)
$f({\text{P}}{{\text{M}}_{2.5\left( {t - 4} \right)}})$	7.655*** (9.126)	7.862*** (9.357)	7.692*** (9.168)	8.075*** (9.598)	7.366*** (8.793)
$f({\text{PL}}{{\text{T}}_1}):$O3	5.201** (6.4932)	4.806*** (6.037)	5.104** (6.381)	7.032*** (8.615)	10.623*** (12.610)
$f\left( {{\text{PL}}{{\text{T}}_2}} \right):$SO2	5.885*** (7.110)	7.179*** (8.582)	6.473*** (7.780)	3.659* (4.594)	5.313*** (6.456)
$f\left( {{\text{MT}}{{\text{R}}_1}} \right):$ temperature	15.167*** (17.101)	13.268*** (15.428)	14.958** (16.931)	12.954*** (15.108)	13.513*** (15.658)
$f\left( {{\text{MT}}{{\text{R}}_2}} \right):$wind speed	4.445*** (5.398)	4.800*** (5.800)	4.316*** (5.254)	5.124*** (6.175)	6.366*** (7.606)
$f\left( {{\text{MT}}{{\text{R}}_3}} \right):$ precipitation	10.736*** (12.464)	9.072*** (10.685)	10.446*** (12.160)	8.461*** (10.015)	7.515*** (8.966)
$f\left( {{\text{Time}}} \right)$	18.993*** (19.000)	18.987*** (19.000)	18.993*** (19.000)	18.989*** (19.000)	18.996*** (19.000)
${\text{Seaso}}{{\text{n}}_1}:$spring	0.004 (0.005)	−0.091*** (0.006)	0.021*** (0.005)	0.014** (0.006)	−0.023*** (0.006)
${\text{Seaso}}{{\text{n}}_2}:$summer	−0.010*** (0.004)	0.011** (0.005)	0.019*** (0.004)	0.035*** (0.004)	0.061*** (0.005)
${\text{Seaso}}{{\text{n}}_3}:$winter	0.166*** (0.004)	0.140*** (0.005)	0.187*** (0.004)	0.177*** (0.005)	0.374*** (0.005)
${\text{DO}}{{\text{W}}_1}:$Monday	0.259*** (0.003)	0.265*** (0.003)	0.273*** (0.003)	0.243*** (0.003)	0.270*** (0.003)
${\text{DO}}{{\text{W}}_2}:$Tuesday	−0.051*** (0.003)	−0.036*** (0.003)	−0.047*** (0.003)	−0.053*** (0.003)	−0.093*** (0.003)
${\text{DO}}{{\text{W}}_3}:$Wednesday	−0.102*** (0.003)	−0.083*** (0.003)	−0.105*** (0.003)	−0.096*** (0.003)	−0.101*** (0.003)
${\text{DO}}{{\text{W}}_4}:$Thursday	−0.063*** (0.003)	−0.049*** (0.003)	−0.060*** (0.003)	−0.074*** (0.003)	−0.053*** (0.003)
${\text{DO}}{{\text{W}}_5}:$Saturday	−0.295*** (0.003)	−0.411*** (0.003)	−0.308*** (0.003)	−0.332*** (0.003)	−0.304*** (0.003)
${\text{DO}}{{\text{W}}_6}:$Sunday	−2.144*** (0.003)	−2.358*** (0.004)	−2.044*** (0.003)	−2.312*** (0.004)	−1.581*** (0.004)
$f\left( {{\text{Hospital}}} \right)$	17.540*** (18.570)	18.582*** (18.946)	17.926*** (18.739)	18.504*** (18.922)	18.957*** (18.999)
Constant	6.293*** (0.019)	4.173*** (0.024)	5.240*** (0.021)	3.649*** (0.023)	2.122*** (0.024)
Observations	318 591	318 591	318 591	318 591	318 591
Adjusted R2	0.876	0.815	0.859	0.827	0.826
Deviance explained	88.5%	81.5%	86.1%	83.4%	84.3%
AIC	4369 309	2984 802	3930 906	2628 102	2299 649
UBRE	2185 538	1493 210	1966 320	1314 870	1150 639

Note: parametric coefficients: coefficient and standard errors in parentheses.Smooth terms (f(variable)): effective degrees of freedom (edf) and reference degrees of freedom (Ref.df) in parentheses.PLT: air pollutant, MTR: meteorological factor, DOW: day of the week.***p < 0.01, **p < 0.05, *p < 0.1.

First, days with the issuance of both alert systems have 8.6% lower number of new patients with respiratory diseases than days without (Coeff. = −0.086, p = 0.00). The alert systems were effective in reducing hospital visits and admissions for all individual respiratory sicknesses—COPD (Coeff. = −0.132, p = 0.00), RTI (Coeff. = −0.100, p = 0.00), pneumonia (Coeff. = −0.080, p = 0.05), and asthma (Coeff. = −0.058, p = 0.00). The effect last more than a day for only two kinds of respiratory sicknesses—RTI (Coeff. = −0.074, p = 0.00) and pneumonia (Coeff. = −0.059, p = 0.10).

Second, on the day of alert issuance by WEA, total hospital visits and admissions by respiratory sickness are reduced by 12.5% (Coeff. = −0.125, p = 0.00), controlling for the number of hospitals, the concentration of ${\text{P}}{{\text{M}}_{2.5}}$ and other air pollutants, weather, time, and AIT effects, considering a 1% significance level. This result implies that WEA may have altered the behavior of people with current and potential respiratory problems: they might refrain from outdoor activities in response to the alert. Such reduced chance of exposure to PM would prevent outbreaks of new symptoms or aggravations of existing illnesses. The effect of the alarm lasts two consecutive days, in that people seem to continue to be cautious in doing outdoor activities for at least two days after they get the WEA message. The two-day lag variable of WEA issuance is associated with a decrease in the number of new respiratory patients by 5.3% (Coeff. = −0.053, p = 0.00).

The WEA was effective in reducing hospital visits and admissions for all individual respiratory sicknesses examined in this study. The WEA has the highest effect on RTI, 15.5% reduction (Coeff. = −0.155, p = 0.00), followed by COPD (Coeff. = −0.130, p = 0.00), pneumonia (Coeff. = −0.100, p = 0.00), and asthma (Coeff. = −0.084, p = 0.00). As may be easily imagined, taking daily preventive measures inspired by alert messages would be the most beneficial in avoiding seasonal infectious diseases rather than chronic diseases (Biggerstaff et al 2016). However, with lower magnitude, the alert messages also worked to reduce the number of hospital visits and admission for chronic symptoms, such as COPD. Those people with chronic symptoms might be more circumspect so as not to worsen their condition (Vlaeyen and Linton 2000, Hill 2019), thus they would have been sensitive in complying with the guidelines provided by the alert. For three individual sicknesses, the duration of the alarm exceeded a day. A WEA alert issued two days prior has a reduction effect of a lesser magnitude, reducing pneumonia by 9.1% (Coeff. = −0.091, p = 0.00), followed by RTI (Coeff. = −0.075, p = 0.00), and COPD (Coeff. = −0.048, p = 0.00).

Third, on the day of AIT alert issuance, the total hospital visits and admissions for respiratory sicknesses declined by 2.5% (Coeff. = −0.025, p = 0.00), controlling for the aforementioned factors. Compared with the case of WEA, the current effect was 10 percentage points lower, suggesting a lower effectiveness for this type of alert message system. The two-day duration of the effect that was observed for the WEA system was not observed here. After two days, the effect of AIT disappeared. The two-day lag variable of AIT issuance is associated with an increase in the number of new respiratory patients by 2.0% (Coeff. = 0.020, p = 0.05). The AIT effects were smaller for all of four individual respiratory sicknesses. Unlike the WEA, the AIT has the highest reduction effect for one of the chronic symptoms, asthma by 3.8% (Coeff. = −0.038, p = 0.00), followed by pneumonia (Coeff. = −0.031, p = 0.10), COPD (Coeff. = −0.020, p = 0.10), and RTI (Coeff. = −0.019, p = 0.10).

By comparing the effects, we revealed that alert messaging through WEA is more effective than AIT. This difference can be attributed to the target audiences. Whereas WEA is for everyone with a cell phone connection, AIT is only for subscribers who are willing to receive more information regarding air quality, such as both predicted and current concentrations of ${\text{P}}{{\text{M}}_{2.5}}$. The request-based alert system has a clear limitation. According to the results of a survey conducted in 2017 in a city in the Seoul Metropolitan Area called Suwon City, 9% of the participants answered that they did not know about the AIT service, and among the rest who knew about the service, 42% had not signed up for it (Suwon City 2017). Even the number of AIT subscribers has decreased from 37 594 in 2018 to 16 785 in 2019 (Visual 2020).

Additionally, WEA and AIT are issued at different ${\text{P}}{{\text{M}}_{2.5}}$ concentration level. WEA is issued only at the warnings and emergency reduction measures are active, while AIT is issued even at lower levels of such criteria and thus issued more frequently. In 2019, WEA was sent 18 days while AIT was sent 33 days in Seoul (Seoul Metropolitan Government 2020). Therefore, it might be reasonable to guess that people might feel alert fatigue and be less responsive to AIT (Baseman et al 2013). One extant study emphasized that people do not feel the necessity of additional information because they already receive sufficient PM-related information from multiple channels, such as mobile phone applications, newspapers, and outdoor electronic boards (Min 2019). Moreover, we speculated that the way of WEA delivery is more effective than that of the AIT. WEA drew people's attention as it comes as pop-up message with vibration and alarm sounds, while AIT comes in general SNS text message style, often could be confused with other messages.

While the AIT system seems redundant, interestingly, it complements the WEA system to contribute to overall protective measures. WEA has the largest reduction effect on the number of new patients with seasonal infectious diseases, while AIT was effective for chronic diseases. By sending simple and direct alerts to a broader range of people, the former would help people to reduce the chance of short-term exposure to PM in general. The latter, however, might be mostly of interest to people with underlying medical conditions, particularly those related to the respiratory system, who would subscribe to receive additional information on PM and follow detailed guidelines to care for themselves.

In order to visualize the effects of the two alert systems and to speculate about the counterfactual condition, we stratified the sample into three groups and plotted the relationship between the number of hospital admissions and visits for respiratory sicknesses and ${\text{P}}{{\text{M}}_{2.5}}$ concentration in figure 5 for (a) the entire study period of 1761 days, (b) the 1761 days when either type of alert was not issued, and (c) the 291 days when either type of alert was issued. Because there was at least one region where the ${\text{P}}{{\text{M}}_{2.5}}$ alert was not issued, the number of days of the first two conditions are the same.

Zoom In Zoom Out Reset image size

First, for the entire study period, short-term exposure to ${\text{P}}{{\text{M}}_{2.5}}$ did not have a noticeable effect on the number of new patients with all individual respiratory sicknesses (figure 5(a)). When we control the alert issuance, a higher ${\text{P}}{{\text{M}}_{2.5}}$ concentration does not necessarily lead to an immediate increment of those patients. Second, for days without either of the alerts (figure 5(b)) the number of patients increased according to the concentration of ${\text{P}}{{\text{M}}_{2.5}}$, as suggested in the previous studies (Neuberger et al 2004). Part of the reason might be the preventive protocols that individuals follow, such as refraining from outdoor activities in face of the high risk of PM. In the condition of high ${\text{P}}{{\text{M}}_{2.5}}$ concentration when alert is issued, the alert system may have guided them. The negative relationship becomes apparent in that, on the days of higher average PM level, more jurisdictions issued alerts to warn people, consequently the number of new patients gets even lowered compared to the days of low PM levels (figure 5(c)). This is obvious evidence of the effectiveness of the warning.

We found that time and season were the other factors that influence the number of new patients. The number of respiratory disease outbreaks changes by season. In winter, the number of hospital visits and admissions for respiratory sicknesses is higher by 16.6%, compared with that of the autumn, the reference group (Coeff. = 0.166, p = 0.00). On the other hand, the number of such new patients declined in the summer by 1.0% (Coeff. = −0.010, p = 0.00). Since hospitals have different operating hours for each day of the week, the number of respiratory patients varies by the day of the week; 25.9% (Coeff. = 0.259, p = 0.00) more people go to the hospital for the aforementioned sicknesses on Monday, while 10.2% (Coeff. = −0.102, p = 0.00) and 6.3% (Coeff. = −0.063, p = 0.00) fewer people do so on Wednesday and Thursday, compared with Friday.

We investigated whether the added alert category—emergency reduction measures—changed the effects of WEA alert, by dividing our study period into the two; before and after February 2019 (table 5). After the day, the reduction effect of WEA was 14.9% lower than before. While the WEA issuance before February 2019 is associated with a decrease in the number of new respiratory patients by 12.3% at a 1% significance level (Coeff. = −0.123, p = 0.00), after the day, it has an increase effect by 2.6% but does not have statistical significance (Coeff. = 0.026, p = 0.13). We reasoned that this is not a sufficient period to reveal the effect of WEA change in additional categories. After February 2019, WEA issued only the 13 times. Therefore, it should be evaluated after a more sufficient period of time to identify the effect of the improvement.

Table 5. Comparing the WEA effect by adding the alert category.

Variable	Entire study period	Before February 2019	After February 2019
WEA	−0.125 *** (0.012)	−0.123 *** (0.020)	0.026 (0.017)
${\text{WE}}{{\text{A}}_{t - 2}}$	−0.053 *** (0.012)	−0.032 (0.020)	0.030 * (0.018)
Constant	6.293 *** (0.019)	6.467 *** (0.022)	6.302 *** (0.155)
Observations	318 591	254 356	64235
Adjusted R2	0.876	0.875	0.902
Deviance explained	88.5%	88.4%	89.4%
AIC	4369 309	3513 482	850 370.5
UBRE	2185 538	1757 540	425 768.7

^*** p < 0.01 ^* p < 0.1.