ADVANCES IN STRATEGIES FOR MINIMIZING AND ADJUSTING FOR SURVEY NONRESPONSE
ADVANCES IN STRATEGIES FOR MINIMIZING AND ADJUSTING FOR SURVEY NONRESPONSE
复制标题
DOI：
10.1093/oxfordjournals.epirev.a036176
复制
发表时间：
1995-01-01
影响因子：
5.5
通讯作者：
GROVES, RM
中科院分区：
文献类型：
--
作者：
KESSLER, RC;LITTLE, RJA;GROVES, RM
自20世纪50年代以来，调查回复率的下降（1,2）使调查研究人员意识到研究无反应偏差的影响，制定程序以最小化无反应的程度，以及调整无反应的调查估计的重要性（3-5）。本报告回顾了这些领域的最新发展。讨论了面对面家庭调查中的“单位”无反应问题（当一个抽样个体根本没有被调查时）。有关单元无响应的更广泛讨论，请参见Groves(2)。关于如何在数据缺失时进行统计分析的更一般的讨论，请参见Little和Rubin (4)， Madow等人(5)和Rubin(6)。单位无响应是一个问题，至少有两个原因：首先，在一个预先指定的病例数量固定的样本中，单位响应的减少直接转化为样本大小的减少，这降低了调查估计的精度。其次，当受访者和非受访者在调查措施方面存在系统差异时，单位不回应可能导致偏见。即使回复率很高，如果非回答者与回答者在罕见结果方面存在显著差异，这种偏差也可能很重要。例如，即使心理健康调查的回复率为80%，如果被调查者中精神分裂症的患病率为1%，非被调查者中为6%，那么精神分裂症的总样本患病率将被低估2倍。可以采用两种方法来减少非反应偏差。一种是使用减少无响应率的数据收集策略。另一种方法是收集所有或一部分非受访者的信息，并将这些信息纳入样本估计以减少偏差。结合
The decrease in survey response rates since the 1950s (1, 2) has sensitized survey researchers to the importance of studying the effects of nonresponse bias, of developing procedures to minimize the magnitude of nonresponse, and of adjusting survey estimates for nonresponse (3-5). This presentation reviews recent developments in these areas. The discussion addresses the problem of" unit" nonresponse (when a sampled individual is not surveyed at all) in face-to-face household surveys. For a broader discussion of unit nonresponse, see Groves (2). For a more general discussion of how to perform statistical analysis when data are missing, see Little and Rubin (4), Madow et al.(5), and Rubin (6). Unit nonresponse is a problem for at least two reasons: First, in a sample with a fixed number of predesignated cases, a reduction in unit response translates directly into a reduction in sample size, which reduces the precision of survey estimates. Second, unit nonresponse can lead to bias when respondents and nonrespondents differ systematically with respect to survey measures. Even if the response rate is high, this bias can be important if nonrespondents differ markedly from respondents in terms of rare outcomes. For example, even if the response rate in a mental health survey were 80 percent, the total sample prevalence of schizophrenia would be underestimated by a factor of 2 if the prevalence was 1 percent among survey respondents and 6 percent among nonrespondents. Two approaches to the reduction of nonresponse bias can be adopted. One is to use data collection strategies that reduce the nonresponse rate. The other is to collect information on all or a subset of the nonrespondents and incorporate this information into the sample estimates to reduce bias. A combination