Select cases for plot

[Jeremiah Jaggers]

I am doing a mixed model for longitudinal data with 13 time points. There are over 90,000 observations, so the line graph his nearly impossible to interpret. Is it possible to select a set number of cases so that it is easier to read?

[Nick Cox]

If I understand this correctly, you want a random subset of panels (cases = observations in my experience) to reduce a spaghetti plot to something with more evident structure.

Here is some technique.

webuse grunfeld, clear

xtset

xtline invest, overlay ysc(log)  

set seed 31459

gen select = cond(year == 1935, runiform() < 0.5, 0)

bysort company (select) : replace select = select[_N]

xtline invest if select, overlay ysc(log)

There are 10 companies: I asked for a 50% sample for one year and then extended the selection to all years for the same panels. My code produced a plot with just 5 companies. With different random numbers a different number would not have been surprising.

The code ignores details specific to the dataset, such as getting better axis labels -- except that ysc(log) is used as natural for that data: it may be useless or even inapplicable for your data; I mention it only because it might be helpful.

Your numbers will be different and you are likely to need to want to use a much smaller probability of selection.

A refinement on the idea is that say 40 panels displayed as a 2 x 2 display with say 10 in each might work well enough to give flavour.