[Linghui Han]

The policy at the national level changes industry market concentration level at the industry level and then ultimately affects GDP growth rate at the provincial level.

[Linghui Han]

Not sure my last post was posted successfully. It reads:

The policy at the national level leads to changes in market concentration level at the industry level and ultimately affects GDP growth rate at the provincial level.

[Clyde Schechter]

So #60 clears up a lot. But now I am seeing this differently, and not sure I have it right. You have a policy intervention that causes changes in market concentration. You then believe that market concentration leads to change in GDP growth, and want to test/estimate that causal effect. This sounds to me more like an instrumental variables model, where the policy is an instrument for change in market concentration. (And if it is, somebody else will have to help you from here, as I do not use instrumental variables regression and know very little about it.) Or am I off base here? If I'm off base, where am I going wrong?

The other thing I'm not sure I get is the focus on the biggest industries in a province to define the groups. Is that just because the policy's effects on market concentration vary by industry and you are trying to focus on what is most prominent in any given location? Or is there something else: do you expect the relationship between market concentration and GDP growth to also vary according to industry size? Or yet something else?

Sorry this is so protracted, but I don't work in this field, and there are probably many things that you know about relationships among these constructs that I am unaware of. So thank you, too, for your patience with me.

[Linghui Han]

The problem using the policy change as an instrument is that the market concentration level in industries is invariant across provinces. If use it as an instrument, I have to aggregate it up to the provincial level which make it undesirable.

On using the biggest industries in a province to define the groups, I am indeed trying to focus on what is most prominent in any given province since the policy first affects market concentration and then on GDP growth.

Are my points clear?

[Linghui Han]

The problem using the policy change as an instrument is that the market concentration level in industries is invariant across provinces. If use it as an instrument, I have to aggregate it up to the provincial level which make it undesirable.

On using the biggest industries in a province to define the groups, I am indeed trying to focus on what is most prominent in any given province since the policy first affects market concentration in industries and then on GDP growth in provinces.

Are my points clear?

[Linghui Han]

If a province has a industry who contributes more than 50% of its GDP and the market concentration level in this industry rises after the policy intervention, the GDP growth rate in this province drops.

[Clyde Schechter]

I see. Very clear. Well, leaving the instrumental variables approach aside, I do not see a DIDID here, because I only see an interaction between your group variable and a pre-post intervention time indicator. I don't see any third variable modifying those effects. So it looks like just plain DID. That's how it looks to me. I have to say, however, that I am a bit uncomfortable in this advice because the content is out of my field. I'm concerned that there may be other aspects of this that an economist would spot that I am missing and that might lead to a better approach. There are many economists active on the Forum. But I don't know if any of them are following this thread, especially now that it has become so long. Think about reposting this issue in a new thread to see if you can get some advice from within your discipline. I'd be happy to learn that they agree with me, but I'd prefer not to steer you in the wrong direction if they don't.

[Linghui Han]

To me, it looks like two steps of DID. The first step of DID is the differential market concentration change between industries before and after the policy intervention. The second step of DID is the differential GDP growth rate change between different province groups before and after the policy intervention. Does this sound correct to you?

[Clyde Schechter]

Well, sort of, but not quite. It seems to me that what you have here is a pair of simultaneous regressions that you are proposing to estimate separately. But that will give you wrong answers in the second step because it will not properly account for the covariation between the policy change and the market concentration change, or, for that matter, any direct correlation between the policy and GDP growth.

[Linghui Han]

I see. Thank you very much! That's very helpful!

[Linghui Han]

I think I have one more question: When to use DIDID? Would you mind giving a simple example? Thank you very much!

[Clyde Schechter]

So suppose that some of your provinces governments are headed by a man, and others by a woman. Then you might wonder if the effect of the policy differs by the sex of the provincial head. That would be a DIDID model, and the regression would contain a three way interaction provincial head sex # treatment group # pre_post treatment.

[Linghui Han]

So at the provincial level it has to be distinguishable between control and treatment group.

[Clyde Schechter]

No, it doesn't have to be. Any attribute that, at any level, distinguishes subsets of the data could serve. It could be at the industry level, or the firm level, really at any level.

[Linghui Han]

I am still wondering to create a DIDID setting: I classify the industries into three categories (market concentration rising/declining/constant, variable name -indgroup-), three groups of provinces as classified in #60 (variable name -provgroup), and policy intervention dummy (variable name -time-), and run the following code:

xtreg gdpgrowthrate i.provgroup##i.indgroup##i.time control_vars, fe vce(cluster id)

Does this sound reasonable?