Wave dummies in FE

[Rohan Kundra]

Hello,

I am currently estimating panel models using an unbalanced panel dataset for 65 years and above. It consists of 15 European countries observed across multiple waves (roughly every 2-3 years). There are in total 7 waves from 2004 to 2020. I have used robust standard errors that are clustered at the individual level. I have controlled for wave effects (i.wave) to capture time specific effects.

 xtpoisson doctor_visits c.age i.gender i.education i.income i.livingalone i.morbidity i.depressed i.smoking i.wave, fe irr vce(robust)

The dependent variable is a count variable measured over the previous 12 months. The main variables morbidity level and depression are significant, but age is not. I ran the regression without i.wave and voila, age became significant and positive (1.020 increase in doctor visits with one year increase in age, c.p).

I also ran a RE model, but with i.wave and i.country both and age is here significant and positive.

My question being, is this expected in such cases? Should I then keep the wave effects in the FE model? Hausman Test prefers FE. How should I interpret and present this finding?

Thank you in advance.

[Andrew Musau]

Time effects are indeed essential in this empirical context. They capture unobserved period-specific shocks that could influence all units simultaneously, such as technological progress or institutional changes over the sample period. Controlling for these ensures that the estimated coefficients reflect within-period variation rather than broader temporal trends. The choice of fixed effects (FE) is also appropriate and does not require further justification via a Hausman test. In this setting, FE estimation provides consistent results under standard assumptions and is the natural specification for addressing unobserved heterogeneity across units.

[Felix Bittmann]

Maybe you can get away with c.year? As long as the general trends are linear, this might be sufficient.

[Carlo Lazzaro]

Rohan:
as the previous helpful comments imply, is not the estimator that gives you back the highest number of statistically significant coefficients the one you should look up to.

[Rohan Kundra]

Hi @Andrew Musau @Carlo Lazzaro,
Thank you for your replies. I was only wondering why the age variable loses its significance after controlling for waves and thus wanted to make sure that there wasn't any mistake from my side while estimating the model. Looks like the effects of age cannot be distinguished from the time effects in FE.

Hi @Felix Bittmann
Thank you for the suggestion. However, the waves are irregularly spaced. Although I did try, there's not much difference. The variables are all stable.

[George Ford]

The waves will be directly correlated (almost perfectly, but not exactly) with age. The wave tracks people over time.

[George Ford]

A thought. I suspect if you want a suitable coef on age, then use a CRE model, which you'll have to do by hand since xtpoisson does not include that option. Might just set age as the first wave.

[Carlo Lazzaro]

Rohan:
please note that, in a FE model, including -i.waves- in the right-hand side of the regression equation is almost mandatory.
Therefore, I would sacrifice -age-, that in all likelihood is almost perfectly correlated with -i.wave-, as George Ford suspects.

[Rohan Kundra]

Thank you George Ford & @Carlo Lazzaro

I will proceed with i.wave as recommended.

CRE models are not part of my syllabus, so I will stick with Poisson.