As in conventional multiple regression analysis with observed (manifest) predictor variables, you could run two separate latent regression models:

- one model with only X as predictor of Y

- a second model in which you have both X and Z as predictors of Y

For each model, you should get an overall R-squared value for Y as part of the completely standardized solution. You can then take the difference between the two R-squared values to see how much Z adds to R-squared above and beyond X. You could then say something along the lines of "X accounts for xx% of the variability in Y, and Z adds another xx% when X is also in the model." In case you have missing data, you would have to make sure that you address the missing scores appropriately (e.g., by using full information maximum likelihood) so that the sample sizes do not differ between the two models.

Since predictor variables (here X and Z) are typically at least to some extent correlated, you cannot usually make these sorts of R-squared computations based on the standardized path coefficients. The standardized path coefficients are partial regression coefficients that account for the partial redundancy (covariance, correlation) between X and Z. But you can (often, not always) judge from the standardized path coefficients which predictor is "stronger" (the one with the larger absolute value of the path coefficient).

Christian Geiser thanks a lot for very insightful comment. Could you elaborate a bit more on "but you can (often, not always) judge from.." - when can me make this type of inference and when this type of judgement should be avoided?

With high collinearity among predictor variables, the standardized regression coefficients may become difficult to interpret. See, for example:

Article Suppressor Variables in Path Models

Ok, I got it. Thanks again!