I want to see if there is an increase in my dependent variable 'Y' within a session for a group A (10 participants).
My research question: IS there a significant increase in 'Y' within-session for group A?
For every second, dependent variable 'Y' and sensor noise 'X' were sampled. 'Y' will be eliminated if the 'X' = YES. The session time varies for each participant (between 20 to 50 seconds). There are 10 participants. Every participant have around 20 sessions.
I have attached the data format that was collected for every session for 10 participants (as image)
My questions are:
Q1: What statistical method should I use to show if the variable 'Y' significantly increased or decreased or no pattern observed within session for the group? I am aware of basic methods such as ANOVA.
Q2: I was thinking about the following method: (1) treat each second as repeated measures and perform ANOVA with time as IV, and Y as DV for all 10 participants? In this case, the varying session time becomes a problem. How to handle this?
Q3: is there any time-series statistical method that I can use to analyse the change in 'Y' within a session for the sessions numbers 1, 10 and 20 (individually) for all 10 participants?
Thank you
Instead of ANOVA you can use linear mixed effects, as this is robust to unbalanced numbers of data points. I'm not sure about the time series issue, as I haven't dealt with that before, but I guess whatever methods are used for dealing with temporal autocorrelations in OLS regression may also be applicable for linear mixed effects.
Dear Srilekha,
you should try making a second dummy time variable (copy of your time), then you can include time as fixed (continuous) and random effect. The participants and sessions could be modelled as a fixed effect. If this doesn't work (estimation problems), try to model the sessions as random effect. The following paper might help on how to do it in SAS (especially page 14):
http://www2.sas.com/proceedings/sugi29/188-29.pdf
The comparison between session 1,10 and 20 could then be collected using linear contrasts (if you model them as fixed effects).
Another issue for you might be the first few measurements of Y (also depending on the variation in Y). If the 'Base'-measurement differs a lot between the participants, you might want to consider including it as a fixed effect into your model to account for this difference.
I hope this helps.
Stephen, Haresh, and Ann Christina give good advice on making strong and complex models of your data. I’ll just add another possibility in case you’re looking for the virtue of simplicity (e.g., compelling graphs). You might make each participant session the unit of analysis by summarizing each with a single number. One possibility is to calculate the correlation (r) between time and “Y” treating the “X” = “yes” as missing data. That would give you a single number for each participant session that represents the fit between time and a linear increase/decrease in “Y.” Alternatively, using the same (time,Y) pairs, you could calculate the slope (m) of the best ft line. This would represent the average consistent rate of change in Y with time for each participant session. By turning each participant session into a single number, as the unit of analysis, you eliminate the problem of each participant having different length sessions and different missing values for time. Choosing which single number you use depends on the hypothesis you’re testing. Once you have a single number, the analyses become straight-forward. For example, if you would like to know if time matters within a session, conduct a one sample t-test comparing r or m with 0 for the sample (or just for a participant). If you would like to know if the r or m changes over sessions, correlate session with the r or m. Hope this helps some Srilekha. ~ Kevin
As your variable is time dependent, never go for correlation analysis.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis