Instrumental Variable

[adele chapman]

Hi all, I have been conducting a stata research project for my dissertation for university and was looking for advice on panel IV strategies.

I am exploring the relationship between environmental opinions (independent) on green energy uptake (dependent). I am using control varibles such as age, sex, income, and education. Lastly, I am grouping these results by constituency. My IV estimation will be living in a rural area which is also binary.

I am having trouble identifying the best command to write as my dependent variable is binary, so I cannot do xtivreg, and i have 4 different environmental opinions so i may need to repeat the iv proccess against all 4 variables.

I am also unsure if i need to include all my control variables or constituencies in the command or if i should just focus on the relationships of my main variables.

If anyone has a solution or any advice please let me know!

[George Ford]

I suspect you need to account for the correlations of the outcomes.

cmp?

Or maybe a control function with sureg or reg3?

[Hemanshu Kumar]

Just out of curiosity, why would living in a rural area be a good instrument for environmental opinions? In other words, why would it make sense to think that living in a rural area only impacts green energy uptake via environmental opinions, and no other channel (conditional on age, sex, income and education, of course)?

[Roman Mostazir]

Generalised structural equation model (-gsem-) comes very handy, in such situtations, in terms of fitting the models with various variables and testing indirect paths.

[Jeff Wooldridge]

I agree with Hemanshu that this seems a shaky IV strategy. But I'd back up even further and ask about the question you're trying to answer. Will you be surprised to find that people who think the environment is important will tend to use green energy? Maybe you can find information on local policies and see if they affect green energy uptake. Currently, I'd say your analysis is descriptive, and using rural as an IV for opinions is stretching. What causal effect are you trying to recover?

If you decide to pursue this, you can use cmp where the endogenous explanatory variable follows an ordered probit and the outcome is a probit. But I think a flexible regression where you include dummy variables for each different opinion level (with the lowest as the base) is more honest about what can be learned.

[adele chapman]

Hi thank you for all your responses,

I'll start with Hemanshu, I am basing this off literature that argues those who live in rural areas care more about maintaining the environment around them compared to those in urban areas.
Since, yesterday I have also used another instrumental variable of an individual voting for the green party, which argues a similar thing. My hope was that these IV would not relate to green energy tariff uptake alone as having strong environmental opinions is required.

Yes Roman, GSEM sounds like a viable option as I have conducted a multilevel mixed effects logistic regression using the meprobit command for my main analysis.

I will agree with you Jeff that the IV is shaky, I am unfortunately limited for variables as I am using the UK understanding society survey dataset. I tried using a two stage residual inclusion method and it seems my IV strategy is too weak to argue that there is no endogeneity. I will try the CMP strategy that you and George recommended.

My wariness with probit was that this removed a lot of the constituencies for the analysis due to collinearity or 'predicts failure perfectly', would using 'cluster(constituecies)' at the end of the probit command work for cmp?