Is it possible to use SEM for forecasts?I like the new JMP-Pro 15 Structural Equation Model (SEM) very much, because you can easily model very complex processes here as well - even if the calculation time is very long.
Besides these advantages in the analysis, I would also like to use the created model for the prediction. Can someone give me a hint how to do that? I found no Prediction Profiler on the platform.
Thanks in advance!
Re: Kann man SEM auch für Vorhersagen nutzen?Hi @markus,
So glad to hear you like the new SEM platform! Getting a forecast depends on the model you're fitting (and there are all kinds of models one can fit in SEM!). If you have a regression model (X --> Y), then using the Parameter Estimates is the key. The regression coefficient represents the change in the outcome, Y, for every unit change in the predictor, X. Assuming you have an estimate of the intercept of Y, you can assemble a standard regression equation, such that Y_hat = y_intercept + reg_coef * X, where you can plug-in different X values to see how your predicted Y (Y_hat) changes.
Another case could be a model where W --> X --> Y (called a mediation model, here X "mediates" the relation between W and Y). If you're interested in the effect W has on Y, then you would multiply the two regression coefficients to obtain the "indirect" effect of W to Y and then use that for your prediction equation.
Now, the good news is that we're planning to add the profiler in JMP Pro 16 so this will be simplified in the future! The other good news is that estimation in 15.1 will be much faster than 15.0, so be sure to upgrade as soon as 15.1 is available (sometime in Feb/2020).
Re: Can you use SEM for predictions?Thank you Laura!
The model consists mainly of 3 parts which are connected to each other:
1) Some Xs, where 2 are covariants, indirectly predict a series of Ys through a latent factor.
2) Two of these Ys become Xs for the next stage and predict a Y along with some other Xs
3) This Y becomes a new X for level 3 and predicts indirectly, via another latent factor, that a number of final Ys will come together with some other X's
How do I handle these estimates?
For clarity, I add a structure diagram.
