交互项中主效应不显著, 交互项显著可怕吗?

凡是搞计量经济的，都关注这个号了

邮箱：econometrics666@sina.cn

所有计量经济圈方法论丛的code程序, 宏微观数据库和各种软件都放在社群里.欢迎到计量经济圈社群交流访问.

关于Interaction terms in regression，咱们推送的文章相当多了，里面的材料大多能够解决日常遇到的问题。不过如果咱们不细致地去看、去了解，就始终解决不了那个疑问，结果就是一而再再而三地请求他人相助。

1.交互项与分组回归的区别是什么? 异质性分析

2.计量回归中的交互项到底什么鬼? 捎一本书给你

3.计量经济学中"交互项"相关的5个问题和回应

4.面板数据中去中心化的交互项回归什么情况

5.内生变量的交互项如何寻工具变量, 交互项共线咋办

6.省份/行业固定效应与年份固定效应的交乘项固定效应

7.U型和倒U型关系及其调节效应, 把非线性关系推进一点

8.中介和调节效应自助法检验，针对非正态截面数据

9.控制、调节和中介变量，系说

10.具有调节变量的中介效应程序和数据, 独家解读相关结果

11.具有调节变量的中介效应分析, moderated mediation

12.中介和调节效应操作指南, 经典书籍和PPT珍藏版

计量经济圈社群日常

计量经济圈社群是一个非常专业的学术组织，当然大多数内容都只能在社群内部进行交互流通。

看了计量经济圈社群的优质解答后，下面再读一读相关materials，一定会对你有较大帮助的。计量经济圈公众号里的很多内容可以直接通过菜单栏搜索到，在慢慢覆盖绝大部分应用微观实证领域的方法。

下面这个短文虽然是英文，但是非常易于理解。可以让你得到如下答案：1.连续自变量去中心化进行交乘易于解释，2.交互项显著但主项不显著也没什么大不了，3. 怎么画交互项的交互图(margins, marginsplot)。

若看得不痛快，建议用PC端浏览

对中间那些解释不感兴趣的话，直接看上面这三小段结论

完整版本的PDF已经上传到社群，当然也可以直接打印公众号里的文章。

再对这个问题进行延伸一下，即遇到如下四种情况作何解释：

主项和交互项都显著，

交互项不显著但至少一个主项显著，

交互项显著但主项效应更大，

交互项显著且比主项效应更大。

1.The outcomes of a two-factor analysis are quite complex. However, you can think of four basic results: 1. Nothing at all is significant. After cussing furiously, you should think about whether or not you’ve designed a study with sufficient power. Is it worth pursuing your question with a more powerful design? 

2. The interaction is not significant, but at least one main effect is significant. Here’s where the plot of means is important. Does it look like an interaction is present? If so, then you might (again) consider that your study did not have sufficient power. If so, then think about ways to increase power. If, however, your graph reveals lines that are roughly parallel, you might want to consider that the interaction is of no statistical or practical significance, which should lead you to focus on any main effects that are significant. In your focus on the marginal means, you are essentially reverting to a one-way ANOVA. 

3. The interaction is significant, but is dominated by the main effects. In general, I think that when an interaction is significant, one has to be a bit careful in interpreting the main effects. Nonetheless, K&W illustrate situations in which one may wish to consider the main effects as important—even in the presence of a significant interaction. I think that their advice is particularly salient when the interaction may disappear with an appropriate transformation of the data. 

4. The interaction is significant, and it dominates the main effects. I think that it will more often than not be the case that the significant interaction makes the main effects difficult to interpret in isolation. That is, the interaction so qualifies the interpretation of the main effect that you’re better off focusing on the interpretation of the interaction and ignoring any interpretation of the main effects.

下面这些短链接文章属于合集，可以收藏起来阅读，不然以后都找不到了。