Hi all,
I'm looking at accelerometer data. I used two accelerometers at once for a backup, but now I need to correlate the two datasets provided. I'm only looking at one axis on each, but one set of data has ~5000 more data points than the other. This may sound like a lot, but the datasets on average hold about 215,000 points of data. During data collection, I paired the two so they began collecting data and ended at the same time, but still one observed more movements than the other. Is there anyway around this to compare the two to say that both pick up about the same type of data than the other? A correlation test in R isn't working since there are two different amounts of data in each column. I do, however, have the mean and SD of each set.
Thanks so much in advanced.
I assume the values are recorded from a process running over some time, so that subsequent readings are not completely independent from each other.
If you have a time stamp, you can pair by time (take the smaller datra set and for aech value find the one closest [in time] to that value).
If you don't have a time stamp but you can assume that the two series were essentially measured in parallel and over the identical total time span with almost constant time-steps, I would delete ~5000 evenly spaced values from the larger series (these are seen as "extra recordings" available only in B but not in A). Example: say you have exactly 5000 viewer values in series A than in B (which has 215000 values). If you remove every (215000/5000) = 43 th value vom B (#1, #44, #87, ... #214958), you get a series with remaining 21000 values that should match those in series A.
This should work well even if the precise time-points of the recordings in the two series are not identical - as long as the changes in the signal over time are mostly small between subsequent recordings relative to the changes over the entire time span.
Lacey Valois , if your time series are simply recorded at two different frequencies and are thus not synchronised, you can for example use the Hayashi-Yoshida approach to compute a covariance, without removing any observation:
Hayashi, T., & Yoshida, N. (2005). On covariance estimation of non-synchronously observed diffusion processes. Bernoulli, 11(2), 359-379.
In addition, if you have some doubts on the accuracy of the times of the recording, you can increase the time step when computing the correlation (taking for example into account one observation in two). By doing so, you should see an increase of the correlation (known as Epps effect).
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics