Score tests for zero-inflation and overdispersion in two-level count data
Score tests for zero-inflation and overdispersion in two-level count data
复制标题
DOI：
10.1016/j.csda.2012.11.006
复制
发表时间：
2013-05-01
影响因子：
1.8
通讯作者：
Jung, Byoung Cheol
中科院分区：
文献类型：
--
作者：
Lim, Hwa Kyung;Song, Juwon;Jung, Byoung Cheol
在泊松回归模型中，观测值要么被聚类，要么由计数的重复测量值表示，观测到的零计数的数量有时大于泊松分布的预期频率，即使在对多余的零进行建模后，过度分散也可能仍然存在。零膨胀负二项（ZINB）混合回归模型，建议分析这些数据。以前的研究提出了分数统计分别测试零通货膨胀和过度分散相关计数数据。在这里，我们还处理同时得分测试零通货膨胀和过度分散的两个水平计数数据使用ZINB混合回归模型。分数检验建议（1）零膨胀存在过度分散，（2）过度分散存在零膨胀，（3）零膨胀和过度分散同时进行。通过模拟研究评价了分数检验统计量的水平和功效。模拟结果表明，由于随机效应的变化，得分检验统计量可能偶尔低估或高估名义显著性水平。本研究提出一种参数化自助法来克服这个问题。Bootstrap检验的模拟结果表明，分数检验保持了名义水平，并提供了良好的功效。(C)2012 Elsevier B.V.保留所有权利。
In a Poisson regression model in which observations are either clustered or represented by repeated measurements of counts, the number of observed zero counts is sometimes greater than the expected frequency by the Poisson distribution and overdispersion may remain even after modeling excess zeros. The zero-inflated negative binomial (ZINB) mixed regression model is suggested to analyze such data. Previous studies have proposed score statistics for testing zero-inflation and overdispersion separately in correlated count data. Here, we also deal with simultaneous score tests for zero-inflation and overdispersion in two-level count data by using the ZINB mixed regression model. Score tests are suggested for (1) zero-inflation in the presence of overdispersion, (2) overdispersion in the presence of zero-inflation, and (3) zero-inflation and overdispersion simultaneously. The level and power of score test statistics are evaluated by a simulation study. The simulation results indicate that score test statistics may occasionally underestimate or overestimate the nominal significance level due to variation in random effects. This study proposes a parametric bootstrap method to overcome this problem. The simulation results of the bootstrap test indicate that score tests hold the nominal level and provide good power. (C) 2012 Elsevier B.V. All rights reserved.