Grups de Recerca d'Amèrica i Africa Llatines, Unitat de Bioestadística, Universitat Autònoma de Barcelona, Barcelona, Spain.
Am J Ind Med. 2009 Oct;52(10):811-6. doi: 10.1002/ajim.20739.
Sickness absence (SA) is a commonly used outcome in occupational health cohort studies. Without the use of statistical techniques that take into account that SA is a recurrent event, the probability of obtaining biased estimates of the effects related to SA is very high. The objective of this article is to examine the application of marginal models, comparing them to count-based models, when the outcome of interest is SA.
By re-sampling the data of a reference study, 1,000 samples of 1,200 individuals were generated. In each of these samples, the coefficients of two factors were estimated by fitting various models: Poisson, Negative Binomial, standard Cox model for a first occurrence, Andersen-Gill and Prentice-Williams-Peterson.
In general, differences among the models are observed in the estimates of variances and coefficients, as well as in their distribution. Specifically, the Poisson model estimates the greatest effect for both coefficients (IRR = 1.17 and IRR = 1.60), and the Prentice-Williams-Peterson the least effect (HR = 1.01 and HR = 1.26).
Whenever possible, the instantaneous form of analysis should be used for occurrences of a recurrent event. Collection of study data should be organized in order to permit recording of the most complete information possible, particularly regarding event occurrences. This should allow the presence of within-individual heterogeneity and/or occurrence dependency to be studied, and would further permit the most appropriate model to be chosen. When there is occurrence dependence, the choice of a model using the specific baseline hazard seems to be appropriate.
缺勤（SA）是职业健康队列研究中常用的结果。如果不使用考虑到 SA 是一种复发性事件的统计技术，则获得与 SA 相关的效果的有偏估计的概率非常高。本文的目的是检验边缘模型的应用，将其与基于计数的模型进行比较，当感兴趣的结果是 SA 时。
通过重新抽样参考研究的数据，生成了 1000 个 1200 人的样本。在这些样本中的每一个中，通过拟合各种模型（泊松、负二项、首次发生的标准 Cox 模型、Andersen-Gill 和 Prentice-Williams-Peterson）来估计两个因素的系数。
通常，在方差和系数的估计值以及它们的分布方面，模型之间存在差异。具体而言，泊松模型对两个系数（IRR=1.17 和 IRR=1.60）的估计效果最大，而 Prentice-Williams-Peterson 模型的效果最小（HR=1.01 和 HR=1.26）。
只要有可能，就应该使用分析的瞬时形式来分析复发性事件的发生。应该按照记录尽可能完整信息的方式组织研究数据的收集，特别是关于事件发生的信息。这应该允许研究个体内异质性和/或发生依赖性的存在，并进一步允许选择最合适的模型。当存在发生依赖性时，选择使用特定基线风险的模型似乎是合适的。
