Mäkelä Pia
STAKES, Alcohol and Drug Research Group, P.O. Box 220, 00531 Helsinki, Finland.
J Stud Alcohol. 2003 Jul;64(4):589-96. doi: 10.15288/jsa.2003.64.589.
This study describes the characteristics of nonrespondents and assesses the impact of unit nonresponse on estimates of central alcohol consumption variables by examining the impact of corrective weighting.
The data came from a Finnish general population random sample of 1,932 respondents (987 women) (response rate: 78.1%). The survey was carried out in the year 2000 using face-to-face interviews. The impact of unit nonresponse was assessed by comparing results using (1) no weighting; (2) poststratified weights adjusted for age, gender and region; (3) weights obtained from a statistical model predicting response propensity; and (4) weights from the model, adjusted to match the population distribution for age, gender and region. Extensive auxiliary information used to predict response propensity came from administrative registers.
Compared with respondents, both male and female nonrespondents had fewer children, lived in urban areas and lived in southern Finland. Male nonrespondents were also older; female nonrespondents more often had only a basic education and were less often in the second-highest income quartile. The change in alcohol variables resulting from the adjustment for nonresponse was small, however, and the difference between the different weighting schemes was even smaller.
If nonrespondents' drinking differs considerably from that of respondents, this difference cannot be captured even by using extensive auxiliary information and an elaborate model predicting propensity of nonresponse.
本研究描述了无应答者的特征,并通过检验校正加权的影响,评估单位无应答对酒精消费中心变量估计值的影响。
数据来自芬兰普通人群的1932名应答者(987名女性)随机样本(应答率:78.1%)。该调查于2000年采用面对面访谈进行。通过比较以下结果评估单位无应答的影响:(1)不加权;(2)按年龄、性别和地区调整的事后分层权重;(3)从预测应答倾向的统计模型获得的权重;(4)根据模型得到的权重,并调整以匹配年龄、性别和地区的总体分布。用于预测应答倾向的大量辅助信息来自行政登记册。
与应答者相比,男性和女性无应答者的子女都较少,居住在城市地区且居住在芬兰南部。男性无应答者年龄也较大;女性无应答者通常只有基础教育,且较少处于第二高收入四分位数。然而,因无应答调整导致的酒精变量变化很小,不同加权方案之间的差异甚至更小。
如果无应答者的饮酒情况与应答者有很大差异,即使使用大量辅助信息和精心构建的预测无应答倾向的模型,也无法捕捉到这种差异。