Descriptive analysis with panel data

[lisa bäcker]

Hello everyone,

I need to work with panel data.
At the moment I make the descriptive analysis and I am not sure if I can use the "normal" commands

tabstat AGE INCOME SEX DEPOSIT, statistics( mean sd min p25 p50 p75 p90 max)
pwcorr DEPOSIT AGE INCOME SEX , star(.1)

It would be great to get a response!
Thank you.
Kind regards,
Lisa

[Nick Cox]

Much depends on what you want precisely. tabstat certainly permits by: or by() to allow panelwise summaries.

For another example, pooling panels in a correlation might make sense scientifically, but the panel and dependence structure will usually make nonsense of P-values.

Conversely, pooling panels with very different magnitudes will often produce pure garbage.

It's hard to generalise. Here's one arbitrary example.

webuse grunfeld, clear

* ssc inst sepscatter assumed 

sepscatter invest mvalue, sep(company) ysc(log) xsc(log) ///
legend(pos(3) col(1)) xla(50 100 200 500 1000 2000 5000) ///
yla(1000 500 200 100 50 20 10 5 2, ang(h))

The pooled pattern seems to make some sense to me, but the structure of between- and within- panel variation is essential to interpretation. pwcorr will not adjust P-values (you show interest in the horrible practice of starring, so N.B.) and as said the P-values can't be trusted but the correlations are still summary measures.

[lisa bäcker]

But following your example. It is allowed to us

pwcorr invest mvalue kstock

And I can still interpret it normally? Or should I just leave it at all?

[Nick Cox]

If "normally" means take the P-values literally, or even seriously, then absolutely not, as I flagged.

Correlations obey the Cauchy-Schwarz inequality regardless of dependence structure, so magnitudes mean just about they usually mean, with the caveat of the example that "overall correlation" could just be an artefact of aggregation.

[lisa bäcker]

ok, now I understand it.
But is it possibly to solve the problem if I just use the command pwcorr for just one period? (so I don't have different points in time any more)
And how can I solve the problem that I do need the descriptive statistics of my panel data for the analysis ( mean sd min p25 p50 p75 p90 max)?
thanks in advance
lisa

[Nick Cox]

Sorry, I don't understand what you are asking about tabstat that I haven't explained. You may need to show an example.

A cross-section is a cross-section. What it means depends on the panel design. There is no problem of time dependence for a single time. There might be other kinds of dependence, e.g. spatial.

[lisa bäcker]

ok, sorry I am so stressed, I read over your explanation for tabstat.
I will try it!
Thank you very much, nick!
This is such a big help for sure!