Proportional Hazards Regression in Epidemiologic Follow-up Studies An Intuitive Consideration of Primary Time Scale
Proportional Hazards Regression in Epidemiologic Follow-up Studies An Intuitive Consideration of Primary Time Scale
复制标题
DOI：
10.1097/ede.0b013e318253e418
复制
发表时间：
2012-07-01
期刊：
影响因子：
5.4
通讯作者：
Cullings, Harry M.
中科院分区：
文献类型：
--
作者：
Cologne, John;Hsu, Wan-Ling;Cullings, Harry M.
在心脏病或癌症等慢性疾病的流行病学队列研究中，年龄的混淆可能会对研究中的风险因素的估计影响产生偏差。在这类研究中使用COX比例风险回归模型，通常建议将按时间顺序的年龄作为主要时间尺度进行非参数处理。然而，涉及生物标志物基线测量或其他因素的研究经常使用自测量以来的随访时间作为主要时间尺度，没有明确的理由。年龄的影响是通过将入职年龄建模为参数协变量进行调整的。参数调整提出了模型充分性的问题，因为它假设年龄和疾病之间存在已知的函数关系，而使用年龄作为主要时间尺度则不是。我们用图形说明了这一点，并直观地说明了为什么以随访时间为主要时间尺度的年龄调整参数方法对特定年龄的发病率提供了较差的近似值。对年龄进行适当的参数调整可能需要广泛的建模，这是浪费的，因为使用年龄作为主要时间尺度很简单。此外，基于研究开始的任意时间的随访时间的潜在危险可能在风险方面没有固有的意义。考虑到可能存在有偏见的风险估计，当年龄混淆是一个令人担忧的问题时，应将年龄视为与流行病学随访数据进行比例风险回归的首选时间尺度。
In epidemiologic cohort studies of chronic diseases, such as heart disease or cancer, confounding by age can bias the estimated effects of risk factors under study. With Cox proportional-hazards regression modeling in such studies, it would generally be recommended that chronological age be handled nonparametrically as the primary time scale. However, studies involving baseline measurements of biomarkers or other factors frequently use follow-up time since measurement as the primary time scale, with no explicit justification. The effects of age are adjusted for by modeling age at entry as a parametric covariate. Parametric adjustment raises the question of model adequacy, in that it assumes a known functional relationship between age and disease, whereas using age as the primary time scale does not. We illustrate this graphically and show intuitively why the parametric approach to age adjustment using follow-up time as the primary time scale provides a poor approximation to age-specific incidence. Adequate parametric adjustment for age could require extensive modeling, which is wasteful, given the simplicity of using age as the primary time scale. Furthermore, the underlying hazard with follow-up time based on arbitrary timing of study initiation may have no inherent meaning in terms of risk. Given the potential for biased risk estimates, age should be considered as the preferred time scale for proportional-hazards regression with epidemiologic follow-up data when confounding by age is a concern.