Hello John-Kåre Vederhus ,

yes the rsqure is very low...sorry, and I would share your interpretation. Rsq has to issues.

First if your goal is purely predictive, you want to explain as much as possible (this is what I call "modeling from right to left"

Second, if your goal is causal (i.e., you want to estimate the effect of one or several variable than rsqure is not unimportant but not the primary target. It is not unimportant as a low rsqure signals there may be many further excluded causes of your DV and it is likely that these show some kind of biasing effect (endogeneity).

Hence, if you have the second goal perhaps you can identify one or better more instrumental variable that help to identify the effect of your indepent variables and shield it from endogeneity despite low rsqure and the other influences.

Best,

Holger

Thanks, Holger Steinmetz ! I believe prediction is what I'm looking for, it's a latent path analysis. I'm examining the influence of exposure to parental alcohol problems on adulthood QoL. Then direct and simple path analysis has R2 = 0.02, which is unexpected but still an OK result (some studies show resilience; there is no necessary link between adverse childhood experiences and adult Health/QoL).

However, I also wanted to examine the hypothesis that parental alc exposure is Associated With perceiving childhood as difficult (through a difficult childhood questionaire - DCQ). I then modeled DCQ as a mediator. The R2 of the DCQ construct was 26%. Can I interpret this as an moderate or considerable influence? See pic. The indirect influence from parental alc exposure to QoL was quite similar to the more simpler model, though, and I want to do a more complex mediation With other potential pathways between DCQ and QoL, i.e., mechanism by which adverse childhood experiences influence adulthood QoL (that model is not uploaded here).

Best

John-Kåre

Hello,

I don't think you strive for prediction as you use many revealing causal concepts (mediator, path analysis, influence). With a prediction model, you would optimally incorporate many variables, optimally in a nonlinear model (e.g., machine learning) and test the hell out of it.

You try to model a typical mediation model. One essential problem (much related to my point about endogeneity) is that a two variable model (or three) is not very solid as you will probably dozens of unobserved confounders. In addition, you have latent variables with many indicators that ring my bell (I doubt that the measurement models are valid). I am a bit sensitive today, maybe, as I struggled two day with a misfitting factor model that had 5 factors with 3 indicators each. Even this did not work.

My recommendation would, to first strive for a fitting measurement model and then try to incorporate either theoretically based candidates for confounders or--better--search for an instrument. When you have a fitting full mediation model (i.e. no direct effect) this at least provides some evidence that the mediator-outcome-link is not confounded but that depends on a) the sample size and b) the effect of the exposure (the power for testing the downstream link between mediator and outcome has a rather low power).

As a summary, don't worry bout the rsquare but about endogeneity.

Below is something to read

HTH

Holger

Antonakis, J., Bendahan, S., Jacquart, P., & Lalive, R. (2010). On making causal claims: A review and recommendations. The Leadership Quarterly, 21, 1086-1120. doi:10.1016/j.leaqua.2010.10.010

Duncan, G. J., Magnuson, K. A., & Ludwig, J. (2004). The endogeneity problem in developmental studies. Research in human development, 1(1-2), 59-80. doi:10.1080/15427609.2004.9683330

Baiocchi, M., Cheng, J., & Small, D. S. (2014). Instrumental variable methods for causal inference. Statistics in Medicine, 33(13), 2297-2340.

Maydeu-Olivares, A., Shi, D., & Rosseel, Y. (2019). Instrumental variables two-stage least squares (2SLS) vs. maximum likelihood structural equation modeling of causal effects in linear regression models. Structural Equation Modeling: A Multidisciplinary Journal, 26(6), 876-892.

OK, thanks Holger Steinmetz !

Actually, my measurement model worked fine, but I note your concern about causality.

best

John-Kåre

Factorial loads seem to refer to multicollinearity

Antonio Serpa Barrientos