Association of ambient ozone pollution with respiratory disease among children: comparison among three distribution fittings of daily hospitalization in generalized additive model analysis

  Objective  To compare differences among the results of utilizing Poisson distribution, quasi-Poisson distribution, and negative binomial distribution in the generalized additive model (GAM) analysis on the association of ambient air ozone (O₃) with respiratory diseases in children for providing references to researches on the relationship between air pollutants and diseases.

  Methods  The data on 117 502 children with respiratory diseases hospitalized in Zhengzhou Children′s Hospital, Henan province during 2016 through 2019 were extracted from the FUTang Updating Medical Records (FUTURE) database; daily data of meteorological monitoring and atmospheric pollution in Zhengzhou city during the same period were also collected. Poisson distribution, quasi-Poisson distribution, and negative binomial distribution of generalized additive model were used to analyze the relationship between daily ambient air O₃ concentration and number of child hospitalization due to respiratory diseases.

  Results  The results of Kolmogorov-Smirnov goodness of fit test revealed that the distribution of daily hospitalization of children with respiratory diseases was consistent with negative binomial distribution (D = 0.055, P = 0.079), but not with Poisson distribution (D = 0.203, P < 0.001). The results of the GAM analysis with Poisson distribution, quasi-Poisson distribution, and negative binomial distribution showed that a 10 μg/m³ increase in ambient O₃ was significantly related to an increment in the number of child hospitalization due to respiratory diseases averagely at lag day 0 – lag day 3, with the relative risks (RRs) (95% confidence interval, 95% CI) of 1.0039 (1.0015 – 1.0064), 1.0041 (1.0001 – 1.0081), and 1.0041 (1.0000 – 1.0081), respectively.

  Conclusion  The study results suggest that negative binomial distribution should be adopted first when conducting a GAM analysis involving an overdispersed dependent variable for reducing false positive error.