Page 78 - 2022-31-中国全科医学

http://www.chinagp.net E-mail:zgqkyx@chinagp.net.cn ·3915·

due/non-hospitalization，and there are few articles analyzing the factors influencing the number of hospitalizations using the
countdata. Objective To fit an optimal model suitable for the annual number of hospitalizations of residents，and to analyze
its influencing factors. Methods Data came from the health service survey database of the project "innovating payment system
toimprove health benefits" carried out from July to August 2019. A total of 27 196 residents from four mountainous counties
（Yanchi，Haiyuan，Pengyang and Xiji） in southern Ningxia were selected to attend a questionnaire survey by use of multistage
stratified random sampling，and 22 427（82.46%） of them with complete and clear key information who returned responsive
questionnaires were included for analysis. Four models（Possion regression，negative binomial regression，zero-inflated
Poisson regression and zero-inflated negative binomial regression） were established with annual number of hospitalizations as
the dependent variable，and demographic factors（sex，age，marital status，education level and occupation），household
characteristics（household size，annual household income per capita，prevalence of poverty-stricken household/household
living on minimum subsistence allowances），social feature（distance to the nearest township health center） as independent
variables，then the model with the best fitting performance was selected to explore the factors associated with the annual number
of hospitalizations. Results The percentage of residents with zero annual hospitalizations was 88.30%（19 802/22 427）.
The negative binomial regression model was examined to have better fitting performance than the Possion regression model by
discrete choice experiments（O=87.665，P<0.01）. The statistic value of Vuong's test was greater than 1.96，indicating that
the data were zero-inflated. The zero-inflated negative binomial regression model had the smallest goodness of fit（18 331.87）
measured by the Akaike information criterion. And the analysis using this model revealed that sex，education level，occupation，
household size，annual household income per capita，an prevalence of poverty-stricken household/household living on minimum
subsistence allowances were associated with annual number of hospitalizations（P<0.05）. And age，marital status，education
level and occupation were associated with the zero-inflation in annual number of hospitalizations（P<0.05）. Conclusion The
zero-inflated negative binomial regression model fitted the data of annual hospitalizations of the participants best. Being female，
unemployed，or poverty-stricken household/household living on minimum subsistence allowances was associated with higher
possibility of seeking hospitalization care，while higher education level，or a larger household size was associated with lower
possibility of seeking hospitalization care. A higher potential demand for hospitalization care was found in 40-59-year-olds and
the married，which could be partially lowered by improving education level and reducing physical labor intensity level.
【Key words】 Health service utilization；Hospitalization；Zero-inflation model；Regression analysis；Root cause
analysis

在实际研究中，经常会在计数资料中遇到零膨胀现 同时为合理规划宁夏卫生资源配置、制定或调整相关政
象，即观测值为“0”的比例远大于其他取值比例。如 策向山区倾斜等提供数据支持和理论依据。
机动车 - 电动车碰撞事故发生频次、森林火灾发生次数 1 资料与方法
等。该类资料中“0”值过多且相同“0”值能够表达不 1.1 资料来源 本研究资料来源于 2019 年 7—8 月开
同含义，常会导致数据呈现过度离散，表现出较大变异， 展的“创新支付制度，提高卫生效益”项目卫生服务调
从而超出相同条件下 Possion 回归和负二项回归等传统 查数据库 ［4］ 。该项调查采用多阶段分层整群随机抽样
计数模型能够预测的范围 ［1-3］ 。为了正确展开参数估计 的方法确定调查对象，具体抽样方法为：在宁夏南部山
和统计推断，国内外学者相继提出零膨胀及其推广模型， 区 7 个县中随机抽取 4 个样本县（盐池县、海原县、彭
理论和实践的交互印证使之逐渐成为统计学的研究热点 阳县、西吉县）；再以经济发展水平好、中、差作为分
之一。但在居民住院情况的研究领域中，大多成果仍集 组依据，将各县辖区内的所有行政村划分为 3 个层次，
中在对住院率、未住院率等的分析，其个案数据结果常 每层按照 40% 的比例，采用随机数字表法抽取样本村；
局限在“是”或“否”，对于次数背后的含义和科研价 随后根据所在村庄的户主花名册进行系统抽样，每个村
值不能充分挖掘。故而本文将住院行为以次数分级，充 庄抽取 20~33 个家庭户作为样本户；将户内所有常住（居
分把握大量观测值为“0”的数据特征，拟构建 Possion 住时间≥ 6 个月）成员列为调查对象，开展后续入户调
回归、负二项回归、零膨胀 Possion 回归和零膨胀负二 查。原始数据库共有 27 196 份问卷信息，本研究去除
项回归模型，并对其进行对比分析，进而深入剖析居民 关键变量缺失或不明确的问卷后，纳入可供分析的问卷
住院次数的影响因素，从而为符合零膨胀相关特征数据 22 427 份（82.46%）。
的拟合提供方法学上的实证研究，有效弥补住院频次分 1.2 研究方法 本研究选取的因变量为居民过去 1 年
析的缺乏，以此丰富卫生服务利用研究的方向和形式， 内住院次数。考虑到患者住院情况是多因素作用的结果，