If I would like to create a regression model including time-invariant explanatory variables (e.g. gender) as well as time-variant variables (e.g. income, health) for a longitudinal investigation of panel data - which technique would you recommend?
Is there any technique, that allows for the integration of both sorts of explanatory variables?
I've read about fixed effects regression, first difference and random effects, but have not found a promising answer, yet...
It seems to me, given that you have varying panel lengths per respondent, that you would want to use the actual score at each period as the outcome, and not a change score. The change score would be more appropriate if you had two time periods for every person and thus could use a difference-in-differences approach. Here, I can imagine that change scores will get messy given the unbalanced panels. As for models, if this is a Likert-type scale, then you should consider using an ordinal logit or probit regression. In Stata, you'd use xtologit or xtoprobit.
Hi Anna - I think this paper may help you:
http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=9452105&fulltextType=RA&fileId=S2049847014000077
alternatively it can be accessed from my ResearchGate page https://www.researchgate.net/publication/233756428
In sum, I think Random effects is what you want, and the above paper shows how you can avoid some of the biases often associated with RE models whilst still modelling t-invariant predictors (basically by separating effects into within- and between-individual effects). It's focus is Political Science data (so repeated measures of countries, rather than people) but the argument is the same for panel data too.
Hope that helps!
Andy
Article Explaining Fixed Effects: Random Effects Modeling of Time-Se...
Frederik asks the right questions. It's hard to provide counsel without knowing what it is you are measuring, what is the data type for your dependent variable, who are your units, how many time periods do you have, what are you interested in knowing. Some researchers want to only control for the clustering and thus do not care about the random components of the model (they focus on the fixed effects), whereas others care about the random components and don't care much about the fixed effects...
I think we'll be able to provide better recommendations when we know what it is you need.
I have just started thinking about the details, so by now I assume I want "too much", respectively I am interested in too much.
I wanted to explore my options...
Outcome variable: life satisfaction (11-point, end-defined scale)
However, I have not yet decided, whether I will use the absolute values, the deviation between two points of time (life satisfaction at t2 - life satisfaction at t1) or the mean deviation between more than two points of time, which would result in a continuous variable.
Number of persons: approx. 950.
Oberservations: from 2 to max. 4 for each person.
I want to figure out the determinants (t-variant and -invariant) as well as their interplay.
It seems to me, given that you have varying panel lengths per respondent, that you would want to use the actual score at each period as the outcome, and not a change score. The change score would be more appropriate if you had two time periods for every person and thus could use a difference-in-differences approach. Here, I can imagine that change scores will get messy given the unbalanced panels. As for models, if this is a Likert-type scale, then you should consider using an ordinal logit or probit regression. In Stata, you'd use xtologit or xtoprobit.
You may find these resources generally useful
http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#multipleresp
I am suggesting random effects as you are interested in time-invariant effects.
This gives a gentle stage-by-stage example
http://www.bristol.ac.uk/media-library/sites/cmm/migrated/documents/repeated-measures.pdf
And if you want Stata example of concepts and application, then try the online free Lemma course - Module 15
http://www.bristol.ac.uk/cmm/learning/online-course/
Please i need this article titled: Stock Market Seasonality: International Evidence
- Badges
- Science topic