time fixed effects

[Madhur Bandranaike]

Hi everyone, I am using a panel data for 40 U.S. states over period 1999-2010.Observations are aggregated at state level. Hausman test suggest use of fixed effects regress. Does stata command "xtreg y x1, fe" takes care of time fixed effects in it or we need to include indicator variable i.year for time fixed effects. My panel setting is "xtset state year" and all data is calculated at state level. Thanks for your answers. Madhur France

[FernandoRios]

Hi Madhur,
You DO need to include the time indicator to control for time fixed effects.
xtreg y x1 i.year,fe
HTH

[Roberto Liebscher]

Just as an addition: Using the fe option with the xtreg command adds fixed effects of your panel variable to your model. So in your case the fixed effects model estimates the change of y for a one unit increase in x1 controlling for unobserved state level heterogeneity. What your eventually left with is the time-series variation of y because cross-sectional variation is capture by your fixed effects. If you think part of your time series variation in y is to be explained by overall time trends or other time series (for instance seasonal) patterns, you should add time-series dummies as suggested by Fernando.

[Madhur Bandranaike]

Thanks a lot for your explanations. Problem is when I include time fixed effects, results on main variables turn insignificant, while most of time dummies are highly significant. For further insight I have attached results with and without time fixed effects and a snapshot to show data structure. From explanation of Roberto, it seems that I don't need to include time fixed effects in my model given my sample structure.

[Stephen Jenkins]

Problem is when ...

"Insignificance" may be the right answer or, rather, summarise all that you can say given the data you have. Maybe it is "time" that is much more important than whatever "x" is for your state-level outcome variable. What does "theory" suggest about the specification of your model -- e.g. why so many lags of "x"? Are there not other characteristics of states that are relevant? [If they are constant, however, the state fixed-effect will wipe those out. If "x" doesn't vary much, then I suspect there's not much variation to identify the parameters associated with "x" anyway.] Note well Roberto's remarks about what is left to identify your model (year to variation) when you use state fixed effects.

[Roberto Liebscher]

Building on Stephen's post it might be a good idea to look at descriptive statistics about the standard deviation of your x-variables. This could give you some sense of what is going on:

xtsum x*

This summary table shows how x varies between states (cross-sectional variation) and within states (time-series variation). Like Stephen noted, if time-series variation in your x variables is low relative to cross-sectional variation

then [...] there's not much variation to identify the parameters associated with "x" anyway

.

[Madhur Bandranaike]

Thanks to Roberto and Stephen for response. Referring to Roberto's suggestion to chek standard deviation, this is indeed the case for one of my main variable of interest that time-series variation is low relative to cross-sectional variation, while opposite is true for 2nd explanatory variables. Remaining Xs represent controls featuring state and industry trends.
Once again thanks for very useful comments.