Rule of thumb reliability of scales for using manifest variables in path analysis?

Question

I am using path analysis to test hypotheses in my study. I have been told by a few people in passing that I can use manifest variables rather than latent variables as long as my reliabilities are high. The Cronbach's alpha reliabilities of my constructs are 0.81, 0.88, 0.87, 0.90 and 0.92.
I know in general, above 0.7 is considered acceptable reliability for scales, but I don't know if that's different for path analysis.

Coffman and MacCallum (2005) write that low reliabilities bias the parameter estimates, and that those parameters will in general be lower than if reliabilities were perfect. However, they didn't provide a cut-off... I.e. what level of reliability is generally deemed acceptable before necessitating the use of latent variables? Should I assume that 0.7 is the cutoff? And if so, is there a reference I can cite for this that specifically relates to path analysis?

Thank you in advance.

EDIT: In my analyses I have used manifest variables. What I really want to know is whether I am justified in using manifest variables, or whether it's necessary for me to use latent variables.

Answer 1

Just as null hypothesis significance tests ("sig. or non-sig.?") have been widely criticized for answering a question that interests few people, the use of a cutoff in assessing reliability is not generally very helpful. It's better to treat reliability as something that's measured on a scale. And note that, yes, for some contexts 0.7 is highly satisfactory whereas for others (e.g., educational testing) even 0.9 might not be.

Also important is the way reliability is established. I take it from your description that you've measured it in a single way (perhaps via Cronbach's alpha) when using multiple methods (test-retest? interrater? etc.) might shed additional light on the situation.

Further, reliability can sometimes be at odds with validity. A measure that is largely unidimensional will have a high Cronbach's alpha but may capture a given construct less completely than a measure that has a lower alpha but wider coverage. So the dimensionality and construct validity of your measures is worth considering carefully.