Mortality selection and sample selection: a comment on Beckett.
Mortality selection and sample selection: a comment on Beckett.
复制标题
死亡率选择和样本选择：对贝克特的评论。
DOI：
--
复制
发表时间：
2001
影响因子：
5
通讯作者：
A. Noymer
中科院分区：
文献类型：
--
作者：
A. Noymer
死亡率选择和样本选择：贝克特的评论 * 加利福尼亚大学伯克利分校健康与社会行为杂志2001年，2001年12月42日（9月）：326-327在一篇有趣的文章中，Megan Beckett（2000）研究了晚年健康不平等的重要问题。许多研究表明，社会经济阶层之间的健康差异在年龄较大时缩小。Beckett使用国家健康和营养调查（NHANES）的面板数据表明，死亡率选择不能解释趋同的健康不平等。本文重新审视了贝克特对选择问题的研究方法，认为贝克特的研究方法虽然富有创造性，但可以有多种解释。正如贝克特所说，把导致老年人健康不平等趋同的现象看作是死亡率选择的“人工制品”。在年轻时，社会经济地位（SES）较高的人比社会经济地位较低的人有较低水平的健康问题。在老年人中，这两个群体的健康问题发生率接近均等。如果发病率的模式反映在死亡率上，那么在年龄较大时，较低SES队列（发病率和死亡率较高）与其初始规模相比将小于较高SES队列（发病率和死亡率较低）。自Vaupel、Manton和Stallard（1979）以及Keyfitz和Littman（1979）的开创性工作以来，许多人口统计学家都假设人口中存在不同的“虚弱”率，这决定了个体在任何年龄段与基线死亡风险的偏差。根据虚弱假说，那些在年轻时死亡的人往往有很高的虚弱程度，这使得幸存者的分布更加健壮。如果满足非随机死亡风险的条件，那么老年低SES队列将比“地址通信：Andrew Noymer，加州大学伯克利分校社会学和人口学系，2232 Pied-mont Avenue，Berkeley，CA 94720;电子邮件：andrew@demog.berkeley.edu”更可靠。我就知道了这种生命历程中命运的逆转被称为“队列反转”（Hobcraft，Menken，and普雷斯顿，1982）。另一方面，低死亡率、高社会经济地位队列的脆弱性分布变化更小，并且将经历更少的队列倒置。低社会经济地位队列的更大队列倒置可能足以克服低社会经济地位的莫尔劣势。然而，这个问题必须谨慎分析，因为理解死亡率选择效应的整个框架都建立在反事实的基础上。如果我们认为趋同是死亡选择的结果，我们就意味着一种内在的差异持续到老年，如果不是因为选择效应，我们就会观察到这种差异。另一方面，如果我们认为趋同是内在的，或者是医疗保险的结果（Beckett 2000），我们认为即使没有选择，我们也会看到趋同。在这两种情况下，都有一个令人不安的动词“会”。实际上，我们只能观察到确实发生的生命率，而不是那些在满足某些条件时会发生的生命率。样本选择的一般问题在社会科学中经常遇到（参见《社会科学》）。Stolzenberg and Relles 1997; Winship and Mare 1992），正如Beckett（2000）所指出的，差异死亡率只是更普遍问题的一个特例。虽然我们不能简单地“控制”（即，条件）选择偏差就像我们对待混淆变量一样，统计技术确实存在，试图抵消偏差。然而，死亡率选择是样本选择的一个非常特殊的情况，如果所讨论的因变量本身与健康有关，则更是如此。由于队列倒置，由于莫尔而进行的样本选择具有超出面板随访中非随机缺失数据的因果关系。这就是贝克特的方法的问题所在。考虑贝克特使用的统计技术来建立没有莫尔- 326的假设
Mortality Selection and Sample Selection: A Comment on Beckett* ANDREW NOYMER University of CaIifornia—BerkeIey Journal of Health and Social Behavior 2001, Vol 42 (Septmber): 326-327 In an interesting article, Megan Beckett (2000) examines the important question of converging health inequalities in later life. Many studies have shown that the differences in health across socioeconomic strata narrow at older ages. Using panel data from the National Health and Nutrition Examination Survey (NHANES), Beckett shows that the converging health inequality cannot be accounted for by mortality selection. The pre- sent comment reconsiders Beckett’s approach to the selection problem, which, while creative, is open to multiple interpretations. Consider the phenomenon that would cause converging health inequalities at later ages to be an “artifact,” as Beckett puts it, of mortality selection. At younger ages, persons with high- er socioeconomic status (SES) have lower lev- els of health problems than those with lower SES. At older ages, the prevalence of health problems in the two groups is closer to parity. If patterns in morbidity are mirrored in mortal- ity, then at older ages a lower SES cohort (higher morbidity and mortality) will be small- er compared to its starting size than a higher SES cohort (lower morbidity and mortality). Since the seminal work of Vaupel, Manton, and Stallard (1979) and Keyfitz and Littman (1979), many demographers have assumed that there are different rates of “frailty” within a population, which determine an individual’s deviation, at any age, from some baseline mor- tality risk. According to the frailty hypothesis, those who die at young ages tend to have high frailty, which skews the distribution of sur- vivors to be more robust. If this condition of nonrandom mortality risks is met, then the aged low SES cohort will be more robust than ‘Address correspondence to: Andrew Noymer, Departments of Sociology and Demography, University of Califomia-—Berkeley, 2232 Pied- mont Avenue, Berkeley, CA 94720; email: andrew@demog.berkeley.edu. when it started out. This reversal of fortune over the life course is called “cohort inversion” (Hobcraft, Menken, and Preston 1982). On the other hand, the low mortality, high SES cohort will have a much less-changed frailty distribu- tion, and will experience less cohort inversion. The greater cohort inversion of the low SES cohort could be enough to overcome the mor- tality disadvantage of being low SES. This problem must be analyzed cautiously, however, as the entire framework for understanding mortality selection effects rests on a counter- factual foundation. If we hold that convergence is a result of mortality selection, we imply that an intrinsic differential persists into older ages and that we would observe it were it not for the selection efiect. On the other hand, if we hold that the convergence is either intrinsic or the result of, for example, access to Medicare (Beckett 2000), we posit that even without selection we would see convergence. In both cases, there is the troubling verb “would.” In reality, we can only observe vital rates that do occur, not those that would occur if some con- dition is met. The general problem of sample selection is encountered frequently in the social sciences (cf. Stolzenberg and Relles 1997; Winship and Mare 1992), and as Beckett (2000) notes, dif- ferential mortality is just a special case of the more general problem. Although we cannot simply “control for” (i.e., condition on) selec- tion bias the same way we would a confound- ing variable, statistical techniques do exist that try to counteract the bias. However, mortality selection is a very special case of sample selec- tion, all the more so if the dependent variable in question is itself health-related. Because of cohort inversion, sample selection due to mor- tality has causal implications beyond nonran- dom missing data in panel followups. This is what makes Beckett’s approach problematic. Consider the statistical technique used by Beckett to set up the hypothetical of no mor- 326