short term and long term analysis

[Nour Ben Ouhida]

Hello everyone,

I want to test the short-term and long-term impact of a certain independent variable on a dependent variable (I have panel data). How is it possible to distinguish long-term analysis from short-term analysis using stata?
I looked around for answers and it's said that panel cointegration methods are useful for the long term and that the error correction model can be used for the short term. Though, I didn't find the right leads on how to apply them. I would appreciate it if anyone could guide me.
Thank you in advance.

[Jared Greathouse]

Please give your example data using the dataex command. For us to provide meaningful feedback, you must provide your example data using the dataex command, the real data from an easily importable source (i.e., Github), or the equivalent of a toy example.
Otherwise, anything we say is simply a waste of time. Note, that I'm not trying to be mean in saying this, I'm saying this because if we can't see your dataset as you do with a minimal worked example, anything we suggest is just guesswork. The reason that I'm emphasizing this is because questions like this one likely have a relatively simple fix, but even simple fixes can be wildly overcomplicated without a minimal worked example of a dataset and code that you've tried to accomplish your task.

So please, provide us with your example data that encapsulates the problem.


I'll reply later with an example of my own, but seeing yours would be best.

[Nour Ben Ouhida]

Thank you Jared for your valuable feedback
Actually I haven't tested it on stata because I simply didn't know how. I'm working on the impact of carbon disclosure on the firm's financial performance in the short-term and long-term. After checking the literature, I chose ROA as a dependent variable for the short term effect and Tobin's Q for the long term effect. I assumed that the relationship that exists between them is linear and constructed the next equations:

(SG stands for sales growth)
after conducting several tests, I found out that the fixed effect model is the most appropriate for my analysis. However, I didn't know what I should do to distinguish the short-term from the long-term analysis

[River Huang]

Please "ssc install xtdcce2" (check out https://www.stata-journal.com/articl...article=st0536 and https://journals.sagepub.com/doi/abs...6867X211045560) and help.

[Jared Greathouse]

So wait, talk to me about how many firms we're considering here and what the specific design is. Instrumental variables? Difference-in-differences? How many treated units are there?

[Nour Ben Ouhida]

I have 248 firms and 11 years (2005 to 2015). I'm going to test the impact of carbon disclosure on carbon performance and financial performance so I have 3 dependent variables which are ROA, Tobin's Q and carbon intensity (a proxy for carbon performance). As for control variables I'm going to include leverage, firm's size and sales growth in the analysis since they can influence the dependent variables.

[Jared Greathouse]

Wait, 248 firms, how many ever disclose their carbon levels?

You gotta compare the ones that did disclose, to firms that didn't disclose.

[Nour Ben Ouhida]

I'm sorry I forgot to mention that I'm working on 5 countries, the BRICS to be exact so it's 50 firms per country on average. My study is only concerned with firms that disclose so I removed firms that didn't have disclosed carbon information (the initial sample was 964, after removing these firms it was rounded to 248). The article I'm basing my thesis on assessed the impact of disclosure in the short and long term however, they didn't explain the methodology they used,it was only mentioned that ROA should be used for the short term effect since it's an accounting measure and use Tobin's Q for the long term because it mirrors investors' perceptions of the firm's future growth and long-term profitability, which is why I'm stuck now because my econometric and stata knowledge is basic. I only know that I should regress my models using the fixed effect regression
I attached the article if you want to take a look at it

[Jared Greathouse]

This isn't making sense. How can you evaluate the effect of a policy or intervention when you're only including firms that were treated? Why wouldn't you compare the 248 firms to firms that didn't have the intervention?

To truly know the causal effect of a drug, you have to compare the outcomes of those who took the drug to those who didn't take the drug. You can't just give the drug to 50 people and not have a comparison group of people. So before we get into Tobin's Q or any statistical questions, let's focus on your research design. My advice to you is to include those other untreated units.

[Nour Ben Ouhida]

that's true it makes more sense, thank you for the insight
so I need to divide my database into two sub-groups, firms that disclose and those that do not then for the analysis I should go for the difference-and-difference approach since it deals with comparing a group to another control group?

[Jared Greathouse]

If this were my problem I would use synthetic controls, but with 248 units I would need to do some subgroup analysis because averaging those causal effects in event time would be an absolute monster. So if this were really my paper, I actually would use synthetic controls since it's the higher version of difference-in-differences.

In fact, you could try with the sdid command which is sort of a synthesis between synthetic controls and difference-in-differences (ssc inst sdid), but I would equally advise you to read the paper describing the estimator in the American Economic Review

[Nour Ben Ouhida]

I'll check it out and do my own research to understand better how to proceed
thank you Jared for your help and guidance. I really appreciate it