可以test L1 L2 penalty (LASSO)

or use each of the explainary variable to conduct univariate regression with the dependent variable to detect the relationship between each variable and the 'Y', then make your decision.

试试

谢谢！您给我的命令是二元logit的逐步回归，请问mlogit的逐步回归命令也是一样的吗？我照着您给的命令把logit换成mlogit做了，结果所有的变量都显示：
note: o.gender dropped because of estimability
note: o.gage dropped because of estimability
note: o.marriage dropped because of estimability
note: o.account dropped because of estimability
（以下省略若干个变量）
begin with full model
p < 0.1000            for all terms in model
一个变量都没有被剔除，这是不是意味着不用删掉任何一个变量？可是这样模型就显得好大呀。

谢谢！lasso太高深一点了吧，一时半会学不会耶。

请 help stepwise (似乎不支援 mlogit)!