I am starting to write a protocol of a systematic review whose research question involves the relation between two different outcomes. I am struggling to find out how to pool these data and present them in such a way that allows me to conclude whether these outcomes have some kind of relation or not - i.e whether one varies as a function of the other.
I presume there is one input and their are two outcomes from each subject which are rather independent and you want to analyse if any correlation exists between two outcome (which can have totally different measuring units). If I am correct, assume x1,x2... are the input and y1,y2,y3.. and z1,z2,z3 are two sets of corresponding outcomes. Test whether X-Y have any correlation or not (start with simple linear to polynomial and check each time with hypothesis test if the coefficient term is zero or not ). If coefficients are statistically non-zero, correlation is confirmed. Repeat the same with X-Z data. If both of the correlations are valid statistically, Y-Z must be correlated. The correlation equation can be derived from the two earlier equation formulated.
Dear Biswanath,
thank you for your answer. Actually, I have two inputs, and each one has its own outcome (a clinical outcome: pain; a laboratory outcome: some gait variables). I want to test if pain reduction is associated with "biomechanical improvement". The main question is that I want to do this with pooled data from randomized controlled trials (therefore, I would extract means, standard deviations, standard errors, etc..).
Don't know if this is possible, indeed..
Perhaps one approach might be "Seemingly unrelated regression", which performs regressions on different datasets with different labels (own outcomes) and takes into account the correlation of residuals and, which gives an additional option to perform feature (variable) selection and/or elimination of colinear features. You can do that e.g. in RapidMiner ("SUR"-operator).
Does anyone have experience with multivariate meta-analysis? Is this method appropriate for my goal?
Giovanni:
I am not sure if Network Meta-analysis is applicable for your project.
It is worth looking, at least to get an idea how to combine or compare the two sources of info.
A numbers of publications is available, the latest is:
A simple regression model for network meta-analysis. OA Epidemiology 2013 Jul 22;1(1):7. http://www.oapublishinglondon.com/oa-epidemiology
http://www.oapublishinglondon.com/images/article/pdf/1374838008.pdf
Radina,
thanks for your kind reply. I became familiar with network meta-analysis some days ago when I was studying some complex meta-analysis methods. As far as I know, the network meta-analysis can be applied to compare treatments, and, more specifically, to pool data of studies with more than two arms.
My goal is a little bit different. I am looking for a method that allows me to determine in what extent the improvement in pain and function is related to biomechanical outcomes. I am not trying to predict, I am trying to analyze multiple outcomes at the same time - i.e if the changes on my independent variables (biomechanical variables) has significant effects on the dependent variables.
cheers
Giovanni:
this chapter reminded me of your question. It is worth reading
http://www.stat.cmu.edu/~hseltman/309/Book/chapter14.pdf
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics