Stata 在 svy 回归后确定有影响的观察结果 [英] Stata identify influential observations post svy regression

问题描述

使用Stata svy命令时，如:

When using the Stata svy command, such as:

svy: logistic graduate age female i.math i.english

应该完成各种后续步骤.例如，寻找重要的异常值或高杠杆点.如果没有 'svy' 元素，以下命令将起作用:

there are various follow-up steps that should be completed. For example, looking for significant outliers or high leverage points. Without the 'svy' element the following commands would work:

predict p
predict stdres, rstand
scatter stdres p, mlabel(snum) ylab(-4(2) 16) yline(0)

但是，当使用 svy 前言运行逻辑回归时，它只会产生以下错误:

However, when the logistic regression was run with the svy preface it simply produces the following error:

svy 估计后不允许选项 rstandard

option rstandard not allowed after svy estimation

太好了.什么是允许的?人们如何看待重要的异常值或高杠杆点?

Great. What is allowed? How does someone look at significant outlier or high leverage points?

推荐答案

@NickCox 的评论是正确的——在将诊断扩展到复杂的调查设置方面没有做太多工作.原因之一是从技术上讲，调查推断是非参数的:推断的对象不是变量之间的某种理想化关系，而是人口普查回归，所有异常值"都在整个人口可能拥有.不可能受到异常值的严重影响；只有估计方程，标准误差是稳健的"无论如何(即使用三明治公式而不是 Hessian.)

@NickCox is right in his comment -- there is not much work done in extending diagnostics to complex survey settings. One of the reasons is that technically speaking, survey inference is nonparametric: the object of inference is not some idealized relation between variables, but the census regression, with all the "outliers" that the full population might have. There is no likelihood that will be badly affected by outliers; there are just estimating equations, and the standard errors are "robust" anyway (i.e. use the sandwich formula rather than the Hessian.)

目前的工作主要由 Rick Valliant 完成(R 包 svydiags:https://cran.r-project.org/web/packages/svydiags/，他的学生李建柱的论文:https://drum.lib.umd.edu/bitstream/handle/1903/7598/umi-umd-4863.pdf?sequence=1&isAllowed=y;有一些我无法立即找到的论文发表的后续论文.)

The work that is out there has mostly been done by Rick Valliant (R package svydiags: https://cran.r-project.org/web/packages/svydiags/, dissertation by his student Jianzhu Li: https://drum.lib.umd.edu/bitstream/handle/1903/7598/umi-umd-4863.pdf?sequence=1&isAllowed=y; there were some follow up papers published out of that dissertation I could not find right away.)

(这感觉更像是针对 CrossValidated/stats 而不是 SO/Stata 的讨论.)

(This all feels more like discussion for CrossValidated/stats rather than SO/Stata.)

这篇关于Stata 在 svy 回归后确定有影响的观察结果的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！