In fact, even if a tool variable passes the test, it is not necessarily a good tool variable.

Teacher Xu Yiqing from Stanford University recently published an article reprinting the conclusions of dozens of papers from several top journals, and found that instrumental variables were abused to some extent.

A good tool variable must first work logically and then pass the test. And I don't know what the test the subject said is. A single tool variable needs to pay attention to two conditions: externality and correlation, and multiple tool variables need to pay attention to over-identification.

So there may be two reasons for the problem of "not obvious after adding tool variables":

Without adding instrumental variables, the influence of other factors on the explained variables is mistaken for the significant role of the core explanatory variables. In this case, it is necessary to logically judge which factors will affect the core explanatory variables and whether they can be controlled as much as possible. If not, we need to use other methods for causal identification, such as instrumental variable method and tendency matching method. If it is panel data, we should also consider the replacement method.

Adding tool variables becomes inconspicuous. In two cases, 1) When the tool variables are properly selected, the causal relationship between the core explanatory variables and the explained variables is correctly identified, which may not be significant, and the tool boundary blocks the "back door" of unobservable factors; 2) When the tool variables are not selected properly, adding the wrong tool boundary may distort the causal relationship. Refer to teacher Xu Yiqing's article mentioned earlier.