Should age quintiles be used as a continuous variable when some individuals have several years of panel data including age?

[DavidHarmsen]

Should age quintiles be used as a continuous variable when some individuals have several years of panel data including age?

The age quintile for some specific individuals could change if I'm correct when the specific individuals become a few years older

[Nick Cox]

This question is of the form "Given X, should I do Y?" with most of X not stated. So the answer is mostly of the form "it depends".

But what's wrong with using age as a predictor? Age quintiles just throw away information. Even if you expect things to happen at certain ages, you should categorise in terms of those ages.

[DavidHarmsen]

That's a very good point Nick! Therefore, I'm now going to use a 'normal' age regression without age quintiles and an age regression with age quintiles to see and compare the effects.

It's about the effect of age of individuals on stockreturns, so some individuals could hold a portfolio for 5 years in a row.
Using i.c. quintiles instead of i.quintiles seems like a seldom used method.

I remember that a teacher of my school said that we should use groups of ages to see if old individuals perform better than very young or young individuals. I do not expect sudden changes at certain ages therefore I use quintiles.

[Clyde Schechter]

Modeling age effects can be complicated. Relationships between modeled outcomes and age can take on all sorts of forms. A handful are simply linear. But increasing or declining effects that are exponential, or that approach a maximum asymptote, or U- (and inverted U-) relationships are all seen in various phenomena. Before you settle on a particular specification of age, be it as a single continuous variable, or some set of age groups*, or some curvilinear form, or a linear spline, you would be well advised to explore your data graphically to see what is reasonable for your problem.

To amplify Nick's comment a bit, discrete age groups are, as he says, sometimes appropriate when the dividing points correspond, at least approximately, to real jump-points in the phenomenon under study. That's a fairly rare circumstance to begin with, and that the jump-points would correspond to quintles of the sample age distribution would be an extraordinary coincidence.

Finally, I don't know what "i.c. quintiles " means: it isn't any kind of legal Stata syntax. If you mean using c.quintiles in your regression, that means that your regression will be based on a number that ranges over {1, 2, 3, 4, 5} and that you are modeling the outcome as a linear function of that number. That is very unlikely to be a good way to specify the outcome-age relationship; you can almost certainly do better than that. Again, explore graphically first to see what makes sense.

[DavidHarmsen]

Very helpful advice Clyde. I'll first make a graph without age groups to see if there are real jumps. If there is only 1 big jump which I expect based on my current results. Then I'll make 3 age groups. 1 group with ages before the jump, 1 group with ages in the jump and 1 group with ages after the jump and maybe 2 extra age groups, 1 with the youngest ages and 1 with the oldest.