Radke Brian R
Economics Unit, Economics and Competitiveness Division, Alberta Agriculture, Food and Rural Development, 303 7000 113 Street, Edmonton, Alta., Canada T6H 5T6.
Prev Vet Med. 2003 Jun 26;59(4):241-56. doi: 10.1016/s0167-5877(03)00103-x.
Interval-censoring occurs in survival analysis when the time until an event of interest is not known precisely (and instead, only is known to fall into a particular interval). Such censoring commonly is produced when periodic assessments (usually clinical or laboratory examinations) are used to assess if the event has occurred. My objectives were to raise awareness about interval-censoring including its existence, the potential ramifications of ignoring its existence, the different types of interval-censored data, and the analytical methods to analyze such data (including availability in standard statistical software). Asynchronous interval-censored survival analysis was demonstrated by parametric evaluation of risk factors for the time to first detected shedding of Salmonella muenster (identified by repeated periodic fecal cultures) for a herd of dairy cows. These results were compared with those from survival analyses which ignored or approximated the interval-censoring. Ignoring or approximating the asynchronous interval-censoring in the survival analysis generally resulted in the risk factors' regression coefficients having the same signs and a decrease (often >50%) in their absolute size. All the standard errors from the three methods of approximating the interval-censoring were <40% of their interval-censored counterparts. The conclusions drawn from the asynchronous interval-censored analysis versus those from the approximations varied dramatically. (The general conclusion from the approximations was that none of the risk factors for this example warranted further consideration.) That ignoring or approximating the left- and interval-censored nature of the dependent variable resulted in biased results was consistent with the literature. In the currently available asynchronous interval-censored models, the inclusion of time-dependent covariates that vary continuously is awkward. Statistical models for the semi-parametric estimation of asynchronous interval-censored survival analysis are not generally available in standard statistical software.
在生存分析中，当直到感兴趣事件发生的时间不能精确得知（相反，仅知道其落入特定区间）时，就会出现区间删失。当使用定期评估（通常是临床或实验室检查）来评估事件是否发生时，通常会产生这种删失。我的目的是提高对区间删失的认识，包括其存在情况、忽视其存在的潜在后果、不同类型的区间删失数据以及分析此类数据的分析方法（包括标准统计软件中的可用性）。通过对一群奶牛首次检测到明斯特沙门氏菌脱落时间（通过重复定期粪便培养确定）的风险因素进行参数评估，展示了异步区间删失生存分析。将这些结果与忽略或近似区间删失的生存分析结果进行比较。在生存分析中忽略或近似异步区间删失通常会导致风险因素的回归系数具有相同的符号，并且其绝对值大小会减小（通常>50%）。三种近似区间删失方法的所有标准误差均小于其区间删失对应值的40%。异步区间删失分析得出的结论与近似分析得出的结论有很大差异。（近似分析的一般结论是，该示例的风险因素均无需进一步考虑。）忽略或近似因变量的左删失和区间删失性质会导致有偏差的结果，这与文献一致。在当前可用的异步区间删失模型中，纳入连续变化的时间依存协变量很棘手。标准统计软件中通常没有用于异步区间删失生存分析半参数估计的统计模型。
