Just wondering if anyone has any inventive (but relatively simple) ideas about how to approach comparing panel data from two different assessment methods that collect the same variables. Basically, I have data from people who reported about events on each day for 30 days (daily diary) and then at the end of the 30 days, they went through each day of the last 30 days and tried to recall these events for each day. I've already used ICCs and straightforward agreement to figure out absolute match between these two methods on a given day, but I'd like to know whether people how far people are off (in days) in reporting these events. They also report specific characteristics of each event (e.g., whether it was in the evening or not, whether they were with someone when it happened; both categorical), so I'd like to figure out, if both reports don't match, are people reporting basically the same kinds of events, but on the wrong day? If so, how many days are they off? I typically use Stata, but any thoughts would definitely be helpful. Also, the data basically looks like this (but can be converted to wide from, if needed):

enter image description here

  • $\begingroup$ Please help me understand the form of your data. What are the "categorical variables"? What do the column headers mean? Which indicates the same event? Which indicates it was remembered? $\endgroup$ – EngrStudent Jan 29 '15 at 19:39
  • $\begingroup$ Sorry, "id" is the participant's unique identification number, "day" is the day that is being assessed, "eventdd" is whether an event was reported on the daily diary, "eventrc" is whether an event was reported via recall (so if eventdd==1 & eventrc==1 would indicate that an event was reported via both assessment methods, so the two agree), "eventddch1" and "eventddch2" are characteristics of the event reported via daily diary, and "eventrcch1" and "eventrcch2" were characteristics of events reported via recall. Does that make sense? $\endgroup$ – twray Jan 30 '15 at 5:00
  • $\begingroup$ Check my related answer on panel data analysis. It contains references to both R-specific resources and a general comprehensive book. $\endgroup$ – Aleksandr Blekh Jan 30 '15 at 20:16
  • $\begingroup$ Can you give me an estimate of the number of variables (columns)? $\endgroup$ – EngrStudent Feb 1 '15 at 13:32

How I would approach it (yes, this is a disclaimer).

Let's say that for a single user, for a single event, there are 10 features of the event. Let's go further and say that there are 5 measures of the features. That means you have 10 columns and 5 rows for this subset.

enter image description here

But what goes "in the box"? It looks like you have binomial data. I am guessing that you put a "1" if a feature is indicated in one measure and a "0" if it is not.

So what questions can I ask of the data?

  • are all measures created equal?
  • are all features created equal?

Thoughts that drive approaches:

  • If all measures were identical, then all rows would be identical. If the measures were all junk, they might as well be random.
  • I wonder if two measures are just variations on the same measure. I wonder if two features are just variations on the same feature. Can they be combined? (think PCA)

(INCOMPLETE: I'm still thinking here, and will get back to it when I get a chance)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.