2.1 Study setting
This study was conducted in two cities of Australia – Sydney and Perth. Sydney is located on Australia’s east coast and has a temperate climate with warm summers and cool winters. It is the largest city of Australia with an estimated population of 5.5 million in 2020, living in an area of 12,368 square kilometres [25]. Perth, about 4,000 kilometers away from Sydney, is located on Australia’s west coast and has a Mediterranean climate with hot and dry summers, cool and wet winters. It is the fourth largest city in Australia with an estimated population of 2.1 million in 2020, living in a metropolitan area of 6,418 square kilometres [25]. These cities were selected to evaluate the respiratory disease hospitalization costs in different demographic and climatic contexts in Australia [26]. Three time periods were used in this study: 2010s, 2030s and 2050s. The baseline period of July 2010-June 2016 was defined as 2010s; the future corresponding periods 2030-2036 and 2050-5056 were defined as 2030s and 2050s, respectively.
2.2 Data sources
2.2.1 Hospitalisation cost data
Respiratory disease hospitalization costs: Daily respiratory disease (ICD-10-AM: J00-J99) hospital admissions and associated hospitalization costs from public hospitals over the study period from July 2010 to June 2016 in Sydney and Perth were obtained from the Australian National Independent Hospital Pricing Authority (IHPA). The daily respiratory disease hospitalization costs were aggregated by the date of admission for statistical analysis.
2.2.2 Meteorological data
Meteorological data: Daily minimum (Tmin) and maximum (Tmax) temperatures were obtained from the Australian Bureau of Meteorology. Daily mean (Tmean) temperature was calculated by the average of Tmin and Tmax. Tmin, Tmean and Tmax are the averages calculated from 17 weather observation stations in Sydney, and 13 weather observation stations in Perth (Figure S1). Daily mean temperatures for the 2030s and 2050s were projected based on three Commonwealth Scientific and Industrial Research Organisation (CSIRO)- defined Representative Concentration Pathways (RCPs). Depending on the RCP emission scenarios, the temperature in Sydney is projected to increase 0.9-1.2 ℃ for the period 2030s, relative to the reference period 1986-2005; and by 1.0-2.0 ℃ for the period 2050s (Table S1) [27]. In Perth, the temperature is projected to increase 0.8-1.0 ℃ by the 2030s, and by 0.9-1.8 ℃ for the period 2050s, relative to the reference period 1986-2005 [27].
In addition, population data for the baseline period 2010s and projection periods 2030s and 2050s were obtained from the Australian Bureau of Statistics (ABS) for the two cities [25]. For the projection periods we adopted the ABS ‘medium’ levels of fertility, life expectancy, net overseas migration and interstate flows for the two cities. Daily populations were estimated using the linear interpolation method [28].
2.3 Statistical analyses
The methods for statistical analyses had been described in our previous studies on assessing the effects of temperature on healthcare costs in Adelaide, Perth and Sydney [29–31]. Two-stage data analyses were performed to explore the association between temperature and hospitalization costs and predict future hospitalization costs. In the first stage, the relationships between daily mean temperature and daily respiratory disease hospitalization costs were estimated using a generalized linear time series regression with a distributed lag non-linear model (DLNM) [32]. In the second stage, the future temperature-attributable respiratory disease hospitalization costs were estimated based on the baseline associations, projected temperatures, and populations for 2030s and 2050s.
2.3.1 First stage
To assess the shape of the exposure-lag-response relationship, a DLNM model with Gamma distribution was fitted simultaneously to estimate the possible non-linear relationship and lagged effects of temperature on hospitalization costs [32–34]. The model controlled for long-term trends, seasonality, weekday variations and public holidays. The exposure-response curves were modelled using a natural cubic spline for temperature with three internal knots placed at the 10th, 75th, and 90th percentiles of temperature distributions, and the lag-response curves with a natural cubic spline with an intercept and three internal knots placed at equally spaced values in the log scale [35, 36]. To control for location-specific seasonality and long-term trends, a B-spline with 8 degrees of freedom (df) per year for time [bs(time,8df per year \(\times\) 6 years)] in Sydney and 4 df per year for time [bs(time, 4df per year \(\times\) 6 years)] in Perth were included in our models [34]. To control for weekday variations, the day of week (dow) was also included in the model. Public holidays (pubhol) were controlled by using a binary variable. The temperature effects were calculated relative to the optimum temperature (i.e. the temperature at which the minimum relative risk for hospitalization costs occurred), which was obtained from the cumulative exposure-response curve for respiratory disease hospitalization costs, as per the methods of Gasparini et al. [33, 37] The location-specific models were described as follows:
$$\text{I}\text{n} \text{S}\text{y}\text{d}\text{n}\text{e}\text{y}: \text{log}\left[\text{E}\left({Y}_{t}\right)\right]={\alpha } +cb\left({Temp}_{t}\right)+bs\left(time,8df per year\times 6 years\right)+{dow}_{t} + {pubhol}_{t}$$
$$\text{In Perth}:\text{ log}\left[\text{E}\left({Y}_{t}\right)\right]={\alpha } +cb\left({Temp}_{t}\right)+bs\left(time,4df per year\times 6 years\right)+{dow}_{t} + {pubhol}_{t}$$
Where,
Y t is respiratory disease hospitalization costs on day t;
α is the intercept;
cb(Temp t ) is the cross-basis natural cubic spline function for daily mean temperature with both response and lag dimension applied from the DLNM;
bs(time, 8 and 4 df per year \(\times\)6 years) is the B-spline with degrees of freedom per year multiplied by the 6 year study period, adjusted for seasonality and long-term trend in Sydney and Perth [38],
time is in days;
dow is the day of week on day t; and
pubhol is a binary variable representing public holidays on day t.
The optimum temperature for respiratory disease hospitalization costs was identified via an overall cumulative exposure-response curve in the first stage, and the outcome below or above the optimum temperature was assigned as the temperature effect due to exposures [32, 34]. The daily mean temperature was chosen as the index providing the best fit in the analysis of respiratory disease hospitalization costs. In order to completely capture the temperature-related hospitalization costs, a maximum of 14 lag days was used in the model, and the value showing the best fit for the respiratory diseases in this study. We tested these modelling choices in sensitivity analysis, which was conducted by changing temperature metrics from daily mean to minimum and maximum temperatures, df for time from 4 df to 8 df per year, and the maximum lag days of 14 days to 7-21 days, to compare and best capture the effects of temperature on hospitalisation costs. Residual analysis and autocorrelation tests were conducted to evaluate the goodness of model fit and autocorrelation. To ensure the hospitalization costs were comparable across the different years, consumer price index (CPI) data were obtained from the Australian Bureau of Statistics [39], and the daily hospitalization costs were adjusted for inflation and standardized to the second quarter of 2016 in Australian dollars (AUD).
2.3.2 Second stage
After the baseline exposure-response relationships between daily temperatures and hospitalization costs were obtained from the first stage analysis, future respiratory disease hospitalization costs were estimated based on baseline model and projected temperature changes in 2030s and 2050s. The future temperature effects on hospitalization costs were estimated under three temperature scenarios (RCP2.6, RCP4.5, and RCP8.5 emission scenarios) for the periods 2030s and 2050s. Hospitalization costs and fractions attributable to temperature with reference to the optimum temperature of minimum relative risk were calculated to show temperature attributable costs (AC) and attributable fractions (AF). The AC and AF were calculated using a method of Gasparrini and Leone [35]. The AC and AF are defined as:
$${AC}_{x,t}={AF}_{x,t}\times {C}_{t}$$
$${AF}_{x,t}=1-\text{e}\text{x}\text{p}(-{\sum }_{l=0}^{L}{\beta }_{{x}_{t-l}, l})$$
Where,
x is the daily mean temperature exposure on day t;
C t is the daily hospitalisation costs on day t.
\({\beta }_{{x}_{t-l}, l}\) is the natural logarithm of RR given exposure on day t–l after l days have elapsed.
The effects of projected daily mean temperatures above the current observed range were estimated using Monte Carlo simulation (n = 1000) [35, 40]. Estimated hospitalisation costs were also adjusted for future population estimates in the 2030s and 2050s as per ABS projections [25].
All statistical analyses were performed using R software (version 4.1.0) with the packages “dlnm”, “tsModel” and function “attrdl” [35, 41]. The statistical significance level of 0.05 was adopted for the analyses.