Finding appropriate model for panel data analysis

[Markus Liesche]

Hi all,


am trying to analyze the effect on inheritance receipt on private pension savings of households.

I've got serval cross sectional survey data sets from different years containing data of around 2000 households each.

I've managed to to append the different survey data sets which results in an unbalanced panel data set because the repeated cross section surveys suffer from respondent attrition. My data format is long (see datex screenshot)

The dataset allows me to observe the same 2047 households over a period from 2006 to 2010.

I created a dummy variable indicating one if a household received an inheritance as well as variables for their respective pension savings and demographic control variables.


Running an OLS regression for single years delivers significant results for the effect of inheritance receipt on pension savings (see screenshot).

Since I have access to data of the same households over different years is it recommendable to run a panel regression using fixed effects to control for unobserved heterogeneity between the different households?

Since I'm interested in effects of a time variant dummy variable (inheritance received 0 /1 ) and time in-variant predictors (gender, education-level, region of residence) as well is a hybrid model recommendable?

I'm struggling to figure out which model would be appropriate to analyze the the assumed impact of inheritance on private pension savings .



I would be grateful for any advice.

[Clyde Schechter]

`Why is inher_gift time-invariant? Nobody received an inheritance during the time under observation? Really? Hard to believe in a data set of this size.

This is a complicated scenario. Any variable can have both within-household and cross (between)-household effects. And those effects can be quite different. I would expect that the within and between effects of inheritance on household savings would, in fact, be quite different. The between-household effect would reflect the fact that if your forebears were wealthy enough to leave you an inheritance, you grew up in a more prosperous environment, are likely to yourself be more prosperous, and therefore more likely to have done some saving. On the other hand, regardless of your prior circumstances, having a chunk of wealth drop into your lap from the sky certainly facilities saving more money. So I think that either a fixed-effects model (which only estimates within-effects) or a between-effects model (which only estimates between-effects, duh!) or a random effects model (which measures a combined effect on the assumption that they are equal) completely misses the boat here. There is a hybrid model that can give separate estimates of between and within effects. -xthybrid- by Francisco Perales, available from SSC, does this.

That said, if it is really true that inher_gift is time-invariant, then your data are simply uninformative about the within-household effect, so at best you can only get a between-households estimate. To me that sounds like it really falls short of the mark.

As an aside, your heart was clearly in the right place showing a screenshot of -dataex- output. But, actually, that defeats the entire purpose of -dataex- and gives us the worst of both worlds. The whole purpose of -dataex- is so that those who want to help you can copy what you post, paste it into the Stata do-file editor, and run it to create a faithful replica of your data, with appropriate metadata. By putting it in a screenshot, that purpose is defeated. And since -dataex- output is difficult for human eyes to read, it also fails to provide a "see it at a glance" understanding of the data. -dataex- output should always by pasted directly into the Forum editor. It even begins and ends with lines telling you where to start and end copying. This wasn't critical in your case, because it wasn't necessary to work directly with your data to answer your question. But, in general, it would be a problem.

[Markus Liesche]

Dear Clyde,

thanks for your fast response. You're absolutely right, the dummy variable "inher_gift" which indicates whether a person received an inheritance or gift in a specific yearis time-variant.

Since you mentioned a hybrid model would be the most suitable model to give separate estimates of between and within effects I assume this will help me to analyze the effect on inheritance receipt on private pension savings of households.

[Clyde Schechter]

Yes, I think that the hybrid model fit by -xthybrid- would be suitable for your goal. Be sure to read -help xthybrid- so you understand fully how the model itself works and learn the syntax and options.

By the way, the hybrid model is a variation on the Mundlak correlated random effects model. I prefer the hybrid approach because I think the results are easier to understand and explain.