I have panel data for bond and I am trying to estimate the following model:
depvar x1 x2 x3 x4, fe
Where x1 and x2 are firm specific variables while x3 and x4 are time-variant variables. So in the t observation x3 and x4 are the same for all the unit observed.
Differnet software give me the estimation but I'm not sure to make the right process.
It is correcto to use Fixed Eff. or it is better to switch to RE?
Moreover, I have the same problem in a similar situation where depvar is binomial? Can I use a logistic panel regregression?
Thanks for your help and your time.
Hi,
Yes you can use cross-sectional invariant variables in panel data, but be cautious if it is correlated with the constant term. You need to do a Hausmantest to decide whether FE or RE is appropriate.
Kind regards
No, it can not be used, Although the panel data is a combination of data time-series and cross-section, but the method of panel data is only devoted to the use of research data time-series data, not cross-section. Cross-section data only help explain the characteristic example: will obviously vary between regions (geography, population, demographics, socio-economic, commodities). This method can use fixed effect (FE) with Least Square Dummy Variable
Another question like Hausman test to decide whether FE or RE and logistic methods have been answered by S.C.Thushara and Daniel Melerrieks.
Dear Alberto, .
in addition to previous comments:
It depends on your situation. If you can use RE as none of your explanatory variables are correlated with the error term there is no problem. If you are afraid of potential endogeneity (induced by the fixed effect/firm effect), you usually use FE. If the time-invariant variable is not of particular interest you can also go with FE. The coefficient will not be estimated but it is controlled for the time-invariant variables. If you are interested in the effect of the invariant variables you can use the Hausman-Taylor estimator depending on the properties of your other variables.
A further test to decide if FE is appropriate is a test proposed by Mundlak (1978). In case you use STATA:
http://blog.stata.com/2015/10/29/fixed-effects-or-random-effects-the-mundlak-approach/
Best regards,
Flo
Mundlak, Y. 1978: On the pooling of time series and cross section data. Econometrica 46:69-85.
Dear Alberto,
In my opinion, Florian Schuberth's insights above are correct. I would like to stress that RE has great risk of bias if individual effects are correlated with any of the regressors in your study. The choice between RE and FE is never arbitrary and it must be taken on the grounds of a test of consistency of RE estimators (e.g. Hausman). If RE would indeed yield bias results (which is very often the case), for a short panel, I would also recommend Hausman and Taylor's procedure, but if you have a long panel, there might be good case for dynamic methods, such as Arellano and Bond's. Both of these procedures will correct potential bias.
Best,
Ricardo
Your choice of approach will depend on whether you are a novice using panel data or whether you have some econometric expertise.
For a novice, I recommend you begin by running pooled OLS, Fixed Effects and Random Effects. All these methods have their own strengths and weaknesses, so we allow the data to guide us on the best model by applying the appropriate specification test: F-Test, Breush-Pagan Test and Hausman Test. These tests will decide for us the most appropriate model to use.
If you have some econometric knowldge, then you should be aware that socio-economic data is susceptible to endogeneity bias arising from measurement errors in data, reverse causality etc. To deal with this in your data, you will have to use either an Instrumental Variable estimator (such as 2SLS) or the generalised method of moments (gmm) estimator. GMM is preferred because recent advances suggest that it yields more a more precise estimator and has better finite sample properties. For GMM, you apply the one step or two step GMM devised by Arrelano and Boover (1995) and Blundel and Bond (1998).
c'est possible, mais c'est le modele à effet aleatoire qui sera conseillé.
Florian Schuberth I know that time-invariant variables will be "absorbed" by the fixed-effects approach. x3 and x4 are time-variant variables in Alberto Burchi ’s model ，but most of answer dissuss time-invariant variables
Another question askedby Arnav Kumar, Can Panel Data analysis be done if Independent variables vary over time within a cross section but are invariant (same) between cross sections?, is similar with Alberto Burchi ’s . Vince Daly's answer says，to avoid multicolinearity, you must then not also include time-period fixed effects. I wonder why it cannot include time-period fixed effecets?
Thanks for your help and your time.
If you don't sample out of the population, FE is always the right choice since it will control for the firm-specific effects. RE is likely to suffer from omitted variable bias if you don't have a well-specified model.
Using a combination of firm-specific and time-variant global variables is also not a problem.
- Badges
- Science topic