regression panel data: OLS / GEE

[Timea De Wispelaere]

hello!

I have a panel dataset on which I would like to perform a regression. my dependent variable is the change in carbon intensity during the year, and my dependent variables are a number of firm characteristics, for 230 firms.

Companynumber year firmsize profitability leverage capitalintensity CAPEX KZindex pctcarbonintensitychange
2.442e+08 2013 18.7566 .0005925 .0765077 .3196759 .1901186 -.0483869 -.0484771
2.442e+08 2014 18.84879 .005932 .1401661 .2698615 .2208322 .1099032 -.0484771
2.442e+08 2015 18.7756 .0086093 .0432784 .2847554 .2438501 -.2025951 -.0253539
2.442e+08 2016 18.67199 .0370381 .03602 .3608946 .264327 -.4171754 -.0003668
2.442e+08 2017 18.74567 .0932389 .0619554 .4206807 .3075858 -.3962344 -.0053627
2.442e+08 2018 18.74328 .0089255 .0107424 .4338437 .2702132 -.4307564 -.0175255

To probe causality, I want to understand the effect of the firm characteristics of previous year on the percentage change in carbon intensity of the current year. If I perform an OLS on this data:

regress pctcarbonintensitychange L.firmsize L.profitability L.leverage L.capitalintensity  L.CAPEX L.KZindex

my R squared value is extremely low (0,02). This is striking because doing the same analysis with one wave data (regressing the firm characteristics of 2012 on the percentage change of carbon intensity between 2013 and 2018) I have an R squared value of 0,4664. I am now doubting whether an OLS regression is suitable for the goal of my regression.

However, I know performing a GEE could be the answer. This gives me also viable output. However, my knowledge on GEE is not profound and I would prefer using OLS in order to be sure to interpret the output correctly.

What is your view on this? do you think I could use an OLS regression? and if this is the case, what should be changed to my code in order for my regression to be more informative?

Kind regards,
Timea De Wispelaere

[Phil Bromiley]

You didn't get a quick answer. You'll increase your chances of useful answer by following the FAQ on asking questions-provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

The first thing I would be tempted to suggest is that you probably want to use level of carbon not percentage change. You're using levels of everything else and it's a little confusing to use levels versus percentage change. It is also extremely likely that part of the reason for your very low explained variance is the percentage change dependent variable.

With panel data like this, you should be looking at the timeseries/cross-sectional procedures in Stata. Look at the PDF documentation, read the introductory material, and read xtreg carefully. Xtreg is probably the most obvious way to attack this data. While you can use GEE, it only gives you access to random effects that assume the unobserved variables you are controlling for our uncorrelated with the included variables while you may want to consider a fixed effects estimator.

[Timea De Wispelaere]

thank you very much!

[Timea De Wispelaere]

I have taken a look at the files you recommended. I see why xtreg is appropriate in my case. However, in my course econometrics I learned that autocorrelation should be tested and remediated for. Therefore, if I for example test for autocorrelation with the Preusch Godfrey LM test, and this test points to an autocorrelation, isn't a GLS the better one? or does xtreg already takes this autocorrelation into account?

Kind regards,
Timea De Wispelaere

[Carlo Lazzaro]

Timea:
if you detect heteroskedasticity and/or autocorrelation in your dataset, just go -xtreg- with -cluster- (or -robust-) standard error.

[Timea De Wispelaere]

Thank you very much Phil and Carlo!