Hi,
I have a data set composed of five categorical variables which indicates by the "state" variable. the presence or absence of individuals in different rooms. This data is collected over one week or more.
variables:
day, hour_of_day,state,room,individuals
ex:
'Sunday', 12, 'Present', 'Bedroom', 'John Doe '
....
Since the 24 different levels of the hour_of_day may not properly reflect different categories of the day, It would be better to merge some levels and find intervals of day which better categorize the user activity.
My first idea is to use contingency tables or Poisson regression with different categories and compare their goodness of fit using the Likelihood Ratio Test (G2) or the Pearson chi-squared X2 statistics. But I don't know if this is correct?
I would also appreciate any answer about existing functions in R!
Mohammad
Why reduce the dimension of the data?
Basically, you have, for each individual, one time series containing 168 (= 7 x 24) data points. Do a time series analysis of this. If you want, you can include day-of-week as a factor, and apply some cyclical element for the daily cycle (e.g. a Fourier series).
I don't want to reduce the dimension but reduce hours into interval of day.
For the time series, I don't have any numerical observation, I just have presence or absence. Moreover a time series per individual would not reflect the marginal relationship between other variables (day, hour_of_day,state,room),
I assume that this is all videotaped, so that it is possible to determine that 'John Doe' entered the bedroom at 3:29 pm and left at 6:42 pm. We know that most people are very set on a 24 H clock.
Try making variables like "Time from Midnight to First Kitchen" or "Time from Bedroom more than 4 hours to first dining room". Variables of this nature may help avoid some of the problems.
If you have the data minute by minute, then you can try different levels of categorization. You already have by hour, so you can make by 2H, 4H, and 6H. If you have the data by minute then you can have more levels. This approach solves the problem of the observer selecting levels that mismatch the pattern being observed. If I get home from work at 5:30 and the observer sticks to a hour schedule, then there is half an hour of critical activity that is missed by the observer. This first half hour may be very important to what happens at 6:00.
I worry that "John Doe" wants to go for a ride to the beach. So at 3:59 he goes to the bedroom, grabs his sunglasses, smooths his hair, and leaves. It is now 4:02. Does this count as John Doe in Bedroom just like if it was night and he was going to sleep? That would seem to be a problem.
This looks like a nested structure, but it is crossed - time could be nested within person but it could also be nested within room. Do you have an overarching identifier that links room to person, or are they separated?
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification