Dear all,
I'm comparing accelerometer-derived vs self-reported measures of sedentary behaviour and I've collected 30 studies which provide the information I need.
Basically, for every study I get a measure of sedentary time from accelerometers and a self-reported tool. These all come from the same sample.
So after analysing the data collected I found that the accelerometer data are non-normally distributed whereas the self-reported data are.
The difference of the means is normally distributed but there is an extreme outlier.
I understand if I was analysing each individual study I could use either a paired t-test or the Wilcoxon test BUT what should I use after pooling the data of 30 studies???
I do not need to work out if accelerometer data are correlated to self-reported data.
I need to know if the amount of time obtained from accelerometers is significantly different from that estimated by self-report methods.
How should I proceed?
Thanks in advance,
Luca
P.M. I do not need to carry out a meta-analysis
A meta-analysis is probably the most appropriate analysis procedure. That coupled with using a raw mean difference for your effect size. This assumes that your 30 studies are all of the studies conducted on this research question or that they represent a random sample of all studies. The meta-analysis procedure also assumes normality for all of the data coming from your 30 studies. It sounds like you are violating the latter assumption but I don't see a better alternative than a meta-analysis. In the calculation of the overall effect size with a meta-analysis, a greater weighting is given to the studies with a greater precision (i.e., those with larger sample sizes). That is logical.
Hi Luca
You might consider a methods comparison analysis - accelerometer versus self-report, with 30 data points. This involves creating a Bland-Altman plot, testing for zero bias, establishing the limits of agreement and potentially, Deming or Passing–Bablok regression.
I totally agree with Adrian, you must create a Bland-Altman plot.
Hi guys, thanks a lot for your help, Adrian and Kabir, I had actually done a Bland-Altman plot...I've just done it again 'cause yesterday night I happened to find another study relevant to my thesis project.
I've plot the results of 31 studies, calculated the bias and lower and upper LoA (+/- 2*SD) but I'm unsure how to interpret it because I've never done one before and at Master level it wasn't even mentioned in the data analysis module...
So I've attached the plot here for some clarification.
My main doubt is if I have to use either the line of identity or the bias line to interpret the results because, if I consider how scattered the various results are and by taking into account the line of identity, up to 500 min/day self-report methods do underestimate the time spent in sedentary behaviour.
Between 500 and 700 min/day there's an equal distribution above and below the line of identity with 2 outliers on each side.
I could conclude that up to 500 min/day self-report methods are not ideal to estimate time spent in sedentary behaviour. However, between 500 and 700 min/day they could be used to estimate sedentary time.
On the other hand, if I consider the line of bias as a reference, it appears to be an equal amount of points above and below of it and I'm still unsure on how to interpret this situation...
On a tutorial on youtube a guy does a linear regression analysis to check for proportional bias but he doesn't have a pool of data...just a set of data (I did it just to see what I would get but I don't think I can use it with a pool of data).
Could somebody shed some light please?
Thank you,
Luca
- Badges
- Science topic
- Similar topics
- Medicine
- Public Health