Hello,
For meta-analysis it is advised to correct for reliability of the measurements. For this, the raw correlation are transformed to reflect unattenuated correlation coefficients using the following formula: corrected r = raw r / √(𝛼 - dependend variable) * √(𝛼 - independend variable).
What is the preferred method if 𝛼 is missing and also not reported in the original source for the instrument or follow up test of psychometric properties?
What can one do if the independend variable has been measured and reported as a composite score of an instrument that only reports alpha for sub-scales?
Is there a way to handle the situation when only alpha for sub-scales is known?
Thank you in advance.
Hello Saskia,
Different opinions exist. Do note, that if you invoke the adjustment for unreliability in the estimated correlation of two measures, you are behaving as if you have access to true scores (error-free observations). Since that is virtually never the case in real life, the correlation estimates will not match with what people are likely to observe. Instead, they'll be biased upwards.
If you're describing relationships among constructs, then it may make sense to use the adjustment. If you're describing what sorts of observed/manifest relationships exist among pairs of measures, you're likely better off abstaining from adjustment, in my opinion. Real data are noisy; to behave as if that's not the case is being a bit unrealistic.
Good luck with your work.
Hello Mr Morse,
thank you very much for your input.
I do describe relationships among psychometric constructs. Additionally, for my work I am following the guidelines for meta analysis of Borenstein (2009), so I feel confident in my choice to attenuate.
I certainly appreciate your answer, tho. It will help me with my discussion probably.
- Badges
- Science topic
- Similar topics
- Correction