The selected provinces are representative for the northwest, northeast, and Red River Delta region of Vietnam. The northwest is a mountainous region with a population of 3.5 million people. Its climate is characterized by cold, dry, sunny winters. Summers in the region are hot due to a high frequency of hot and dry days and westerly winds (IPSNOR, 2009). Northeast Vietnam has a population of 13 million accounting for 15% of Vietnam’s population. The climate of this region is strongly influenced by the north-eastern monsoon (Mui, 2006). The summers are hot, humid and coincide with the rainy season, but in contrast to the northwest, dry conditions are rare because of less frequent westerly winds. Red River Delta has a population of 17 million, of which a half live in the capital city, Hanoi. Summers are also hot and rainy. The mean annual temperature in the northeast and Red River Delta regions is 23°C. All three regions have distinctive seasons, including: Spring (February-April), Summer (May-September), Autumn (September-November), Winter (November-February).
Daily hospital admission data were obtained from 7 provincial hospitals comprising (Bac Giang (BG), Tuyen Quang (TQ), Phu Tho (PT), Dien Bien (DB), Lai Chau (LC), Ha Noi (HN), Quang Ninh (QN) – 1 hospital in each province) located across northern Vietnam (Supplemental Figure 1) for the period between January 2005 and December 2015. Hospital admission data included daily counts for all causes (excluding external causes), cardiovascular diseases (I00-99; excluding: acute rheumatic fever, I00-02, and chronic rheumatic heart diseases, I05-09), and respiratory diseases (J00-99; excluding: lung diseases due to external agents, J60-70), and certain infectious and parasitic diseases (ICD10-Code: A00-B99; excluding: infections with predominantly sexual mode of transmission, A50-64, HIV, B20-24, Helminthiases, B65-83, sequelae of infectious and parasitic diseases, B90-94). Disease classification was based on primary and discharge diagnosis. We also extracted background information (e.g. age, sex, and address) of individual patients and their dates of admission and discharge from hospital records. The time-span of daily hospitalization data varied from 2 to 11 years as the electronic data management system became operational in the respective provincial hospital. Patients not from those 7 selected provinces and hospital records (<1%) with missing data were excluded from the analysis. Daily meteorological data (ambient temperature, humidity and precipitation) were obtained from the provincial hydro-meteorological stations or from the closest airport weather stations. The precipitation data for province Phu Tho was not available, therefore humidity only was adjusted for in the respective model.
Data analyses involved three steps. First, we performed descriptive statistics for exposure (temperature) and outcome (hospitalization) variables (i.e. hospitalisation for all causes, infectious diseases, respiratory and cardiovascular diseases). Second, we examined the province-specific temperature-hospitalisation relationship for each province using mean temperature centered at 240C for each province. Third, we performed a random-effect meta-analysis to estimate the pooled effect sizes of temperature on all-cause, infectious, cardiovascular, and respiratory admissions. Poisson Generalized Linear Model (GLM) and Distributed Lag Model (DLM) were used to examine the province-specific association between temperature and daily average temperature and risk of hospitalization. The province-specific Poisson regression time-series model used was:
where Yt is the observed daily count of hospital admissions on day t; α is the intercept; Tt is the daily average temperature on day t and l is the lag days; H is the daily average humidity; R is the daily cumulative rainfall. s is the Flexible spline function with 7 knots per year; DOW is the day of the week.
The linear distributed lag model of temperature for lags up to 6 days (a week) was used to examine the delayed effect of temperature on hospitalizations. A flexible spline function with 7 knots per year was used to control for long-term trends and seasonal patterns in hospitalizations (Bhaskaran et al. 2013). The model was adjusted for humidity, rainfall, and day of the week to control for potential confounding effects.
A random-effect meta-analysis was applied to calculate within-province and between-province variation and generate pooled effect size (relative risk, RR). The pooled effect sizes, which were calculated for cause-specific hospitalizations, comprising all-cause, infectious diseases (ID), cardiovascular disease (CVD), and respiratory diseases (RD), were computed by lag day (from 0 to 6 days). I-squared (coefficient of inconsistency) statistics was used to determine the heterogeneity between studies (Higgins et al. 2003).
All the data analyses were conducted using the “glm”, “db metan” packages of Stata 14.0 (Stata Corporation, College Station, TX, USA).