As much as I agree with Thomas, and though theoretically possible, it's advisable to have a minimum of 3 items per construct (latent variable).
2- or 1-item construct is fundamentally weak, in particular with covariance-based SEM such that provided by AMOS, LISREL. An option is to use PLS-based SEM, such as smartPLS, warpPLS. Even these, you'll need to have a strong theoretical backings as why only 2 or 1 indicator (item) is used.
You may want to check these refrences:
a. why PLS-SEM -> PLS-SEM indeed a silver bullet by Hair.
b. A single-item construct -> Multivariate analysis, a book by Hair et al. 2010. (Find out in factor analysis/composite scale section.
As much as I agree with Thomas, and though theoretically possible, it's advisable to have a minimum of 3 items per construct (latent variable).
2- or 1-item construct is fundamentally weak, in particular with covariance-based SEM such that provided by AMOS, LISREL. An option is to use PLS-based SEM, such as smartPLS, warpPLS. Even these, you'll need to have a strong theoretical backings as why only 2 or 1 indicator (item) is used.
You may want to check these refrences:
a. why PLS-SEM -> PLS-SEM indeed a silver bullet by Hair.
b. A single-item construct -> Multivariate analysis, a book by Hair et al. 2010. (Find out in factor analysis/composite scale section.
Inmaculada, you can, but you should also be aware that measurement error in a mediator--more likely to be a problem with a single variable--biases parameter estimates predicting the outcome. So any steps you can take to maximize the reliability of the mediator are important.