Add time and industry effects to panel data

[Mislav Sagovac]

I suppose this is a straightforward question but I want to be 100% sure: How to add time and industry effects to panel data?

I have panel data of firms for period 2000-2014. I thing it would be appropriate to control for time and industry effects (I plan to use fixed effects estimation).
Should I add time dummy for every year ot time trend? Also should I add dummy for every industry? But industry is not changing over time so this should not be a good explanatory variable for fixed effect?

[Clyde Schechter]

Whether to use a separate indicator (dummy variable) for each year (except one reference year, of course) or use a linear time trend depends on what you think is happening over time. If you expect a consistent linear trend in the outcome variable over time (after adjusting for everything else), then a single "continuous" time trend variable is appropriate. If you think of each year as representing a new, more or less independent "shock" to the outcome variable, then separate indicators are most appropriate. And, in principle, there are other possibilities as well, depending on what kind of trajectory you expect your outcome to follow over time. Without knowing the content of your study, it is impossible to say. (And even if I knew the content of your study, I probably couldn't advise you because I am not knowledgeable in finance or econometrics.)

With industry being constant over time within firm, it cannot be used as a covariate in the fixed effects model. Even if you put it in, Stata will quickly drop it due to its collinearity with the firm fixed effects. The other way of thinking about this is that by using a fixed effects model, you are already adjusting for industry--you are just not getting a separate estimate of the industry effects.

[Mislav Sagovac]

Clyde,

Thank you very much for your answer. That was very helpful.

Is it meaningful to include time dummies in fixed effects model? I have panel of firms and would like to estimate sensitivity between cash flows and investment for different regimes. I wanted to control for some secular factors that influence investment through time.

[Clyde Schechter]

I wanted to control for some secular factors that influence investment through time.

Generally speaking, yes, it is meaningful to include time dummies in this circumstance. It's just that if there is something more systematic about secular factors that influence investment through time, then there are better ways to do it than time indicators. But if the secular factors just bounce around haphazardly from year to year and don't display trends, then time indicators is as good away as any. I have no expertise in finance or econometrics, so you will have to be the judge of whether that is the case in your project.

[Mislav Sagovac]

Thank you again for your answer.

[Carlo Lazzaro]

Mislav:
as an aside to Clyde's helpful insights, as far as time dummies creation is concerned, you should rely on -fvvarlist- and its capabilities.

[Phil Bromiley]

You might also consider including the mean of the dependent variable for the industry (excluding the firm of interest) as a control. This allows for industry-specific time effects which are more likely than all industries being influenced the same by time. It also uses far fewer degrees of freedom.

With firm fixed effects, you can see the not having an industry effect two ways. First, strictly statistically, industry will not estimate because it doesn't vary within firms (assuming no firm changes industry). Second, since firm is nested within industry, having the firm takes care of industry.

[Sanaullah Farooq]

Hi, This is also a straightforward question related to above question. How do I create dummy for Time and Industry effects in my pooled data?

[Carlo Lazzaro]

Sanaullah:
welcome to the list.
For the future, please consider starting a brand new thread.
That said, please take a look at -help fvvarlist-.

[Paul Mao]

Whether to use a separate indicator (dummy variable) for each year (except one reference year, of course) or use a linear time trend depends on what you think is happening over time. If you expect a consistent linear trend in the outcome variable over time (after adjusting for everything else), then a single "continuous" time trend variable is appropriate. If you think of each year as representing a new, more or less independent "shock" to the outcome variable, then separate indicators are most appropriate. And, in principle, there are other possibilities as well, depending on what kind of trajectory you expect your outcome to follow over time. Without knowing the content of your study, it is impossible to say. (And even if I knew the content of your study, I probably couldn't advise you because I am not knowledgeable in finance or econometrics.)

With industry being constant over time within firm, it cannot be used as a covariate in the fixed effects model. Even if you put it in, Stata will quickly drop it due to its collinearity with the firm fixed effects. The other way of thinking about this is that by using a fixed effects model, you are already adjusting for industry--you are just not getting a separate estimate of the industry effects.

@Clyde

Would it make sense to apply the same logic to firm fixed effects?

I am trying to understand what would be the difference between:

reghdfe y x1 controls, absorb(firm) cluster(firm) ==> control for firm time invariant unobserved factors

reghdfe y x1 controls, absorb(firm#c.year) cluster(firm) ==> control for trends of these time invariant unobserved factors?

Would there be any instances justifying the use of firm#c.year rather than the dummy firm fixed effect ?

Thank you.

[Clyde Schechter]

I cannot really make any sense of absorbing firm#c.year. If you actually try both of the commands you show, you will see that the second produces a -nocons- version of the first. I'm not sure why. And I don't know what absorbing firm#c.year would even mean.

Anyway, if you want to incorporate a time trend into your model (as opposed to time indicators), just add c.year to the list of regressors. If you want to include a firm#c.year interaction, that, too, should be added to the list of regressors. Save -absorb()- for discrete variables and their interactions.