Difference in differences analysis with fixed effects

[John Snow]

Hi everybody,

I'm currently trying to conduct a special diff-in-diff analysis with stata. My research is about sovereign credit ratings and their impact on cross-border M&A volumes. One of the goals of my study is to test if the impact of sovereign credit ratings on M&A changed as a result of the crisis. So I dont want to test if the volume or the ratings changed because of the crisis, but the impact of the ratings on the M&A volume. I think that with a common diff-in-diff analysis this cannot be tested, right? So Im really struggeling with this and hope to find some answers here.

Thanks in advance for you help!

[Phil Bromiley]

I am not sure I understand exactly what you are trying to do. Most of the diff-in-diff analyses I know of have some entities subject to a shock and others not. But all countries would have had the crisis.

You can run an xtreg with M&A as dv and rating as the iv and then interact rating with a dummy indicating before or after the crisis. Thus, you'll have rating i.crisis#rating on the rhs. [Look up factor variables in the documentation.] Then the parameter on the interaction is indicating the change from the pre-crisis effect.

[John Snow]

So what i actually want to test is if the coefficient (the effect) of the rating changed due to the crisis. I e.g. found that before the crisis the coefficient is 1 and afterwards it is -5 in my treatment group. The question now is if this change from positive to negative was due to the crisis or due to something else.

[Carlo Lazzaro]

John:
I would follow Phil's sound advice along with posting what you typed and what Stata gave you back, so that those interested in your thread can get the whole story.

[Phil Bromiley]

Again, I would normally not use treatment group terminology or diff-in-diff if all entities are subject to the treatment/event. With the data, you can estimate whether the parameter has changed and test whether the change is statistically significant.

Your question "if this change ...was due...", is impossible to answer. You can rule out specific alternative explanations by including them in the model. You can "test" for omitted variables (but I'm skeptical of such tests). But there is no general way to rule out all other possible explanations.

As Carlo suggests, some data (using dataex), what you typed and what Stata gave back would get you better help from this list.