t*****8 发帖数: 157 | 1 大家generate propensity score 的时候, 是把所有相关的varibales都放到logistic
model里面, 还是只是把significant 的variables?
谢谢大家了。 |
w*******9 发帖数: 1433 | 2 两者都不是,关于这个model selection有很多文章,但是没有一个统一的定论。最后
的结论是1:包括的covariates越少越好。2如果你选择了某个看起来promising的model
,结果某个variable match的不好,就再把那个variable放进去。3要多考虑functions
of covariates like logx, x^2。4评价model的最终标准是match的怎样,一个t-test
是不够的,KStest也要考虑,而且you also need to match any functional forms of
the covariates。所以这是一个back and forth很烦人的过程。
logistic
【在 t*****8 的大作中提到】 : 大家generate propensity score 的时候, 是把所有相关的varibales都放到logistic : model里面, 还是只是把significant 的variables? : 谢谢大家了。
|
t*****8 发帖数: 157 | |
c*****1 发帖数: 2460 | 4 Use Bayesian logistic regression instead. |
y*****w 发帖数: 1350 | 5 1. Use those covariates that are related to both the grouping variable and
the outcome variable in the study. For example, if your study is to analyze
whether a drug treatment is highly associated with mortality, and the
demographics and baseline characteristics are not balanced between the
treatment groups, then you would need to (1) run univariate logistic
regressions to identify those covairates that are highly associated with
drug treatment, (2) run univariate Cox regression models to identify those
covariates that are highly associated with time to first mortality. Choose
those that show a high level of association in both cases as covariates to
get the PS.
2. Use either a matching method by applying calipers of width 0.2 of the
standard deviation of the logit of the PS or a 5->1 digit greedy matching. |