Effect size typically refers to a difference in means for 2 groups, divided by the variable's SD.
Perhaps your covariate x is a dichotomous variable representing 2 groups. In this case you can compute the estimated mean of s for each group and divide by the square root of the estimated variance of s. Alternatively, you can compute the estimated mean of the observed outcome for the 2 x groups and divide by the outcome SD.
See also the standard literature on effect size, including Cohen's writings.
H Priess posted on Thursday, June 04, 2009 - 3:00 pm
As noted above, effect size often refers to the difference in two groups' means, divided by SD. In the case of latent growth models with binary predictors (e.g., gender), could one use the standardized coefficient "StdY" as a measure of effect size (provided predictor was coded 0,1 so that the coefficient was equal to the group difference), or is this not an equivalent calculation?
So the DV is a growth factor like "i" centered at a certain time point, where StdY makes this DV have SD=1 so the slope on gender is the effect size because it is the mean difference wrt gender in i divided by its SD? Makes sense.
Hello, I want to know effect sizes in a multigroup LGM (twins versus singletons) with a binary predictor (gender, coded 0,1), but unfortunately Bengt's last post (4.45pm) is too cryptic for me. The influence of gender on the intercept is significant in twins and not in singletons. This could be a power issue, given the smaller sample size for the singleton sample. Could you please clarify how I can use the standardized beta coefficient to check if the effect sizes for sex differences are indeed greater in the twin than in the singleton sample?
Hi, What if it's just one group and an effect sizes is needed for the effect of a continuous variable on intercept and slope? For instance, I'm in the midst of preparing a Montecarlo model for a simple LGCM with time invariant and time varying covariates. Now, I will be assessing the effects of various continuous predictors on the level and trend factors. I need to select values for my slopes and those values should correspond to some measure of effect. In this case my dependent variable is caloric intake (continuous variable) measured at three time points. One of my predictors is a measure of anxiety (also continuous). If I want to assume a medium effect based on the Cohen's criterion, how would I determine the start value for my slope? I'd appreciate any guidance you could give.
Perhaps you can think in terms of how many standard deviations the i and s change as a function of one standard deviation change in the covariate. For instance, a small effect might be that s changes 0.25 SD units.
With growth, there is a choice in what the DV should be - for instance, should it be s or should it be y_t at the last time point? The y_t DV has a difference variance than s, and therefore the effect on y_t is different than the effect on s. The effect on y_t is perhaps more a more tangible concept.
I'm running an LGM for a treatment group effect on 5 times.
Is it possible to test the effect of the group (0 or 1), the effect of another covariate and their interaction in the same model ? And is it possible to have the effect sizes of the 2 mains effect and the effect size of the interaction ?
Sarah Lowe posted on Tuesday, March 11, 2014 - 7:40 am
I am working on a revision, and the Editor would like an indicator of effect size for cross-lagged paths. My co-authors and I have already provided standardized (stdYX) coefficients and 95% CIs, but it seems like the Editor wants something more.
His suggestion was to 1) run the model with and without each cross-lagged path, 2) calculate the change in R2 for the DV, and 3) convert to an F statistic.
At first, this sounded OK to me, but then I ran the models and noted that some of the R2s were actually *larger* with the (non-significant) cross-lagged path excluded (e.g., .57 with the path, .53 without the path), meaning that the addition of the path was associated with a -R2.
This has me questioning whether this approach to estimating effect size makes sense. I would appreciate your thoughts on this matter.
It doesn't sound like the editor's suggestion as you describe it makes sense because when you leave out significant cross-lagged paths the model becomes mis-specified and its estimates are of unknown value. I think standardized CIs make more sense.
Sarah Lowe posted on Tuesday, March 11, 2014 - 12:39 pm
Reviewers for a paper want me to calculate effect sizes for the standardized estimates that I got from doing an SEM model with an endogenous factor that has 6 categorical indicator variables. How would I do this? Is it possible? Thanks!
I don't know how effect size comes into play here. That has to do with differences between groups. Perhaps you want to describe this on SEMNET.
Anna Lee posted on Thursday, March 29, 2018 - 2:28 pm
Hi Drs. Muthen, I have a cross-lagged model with logistic, poisson, and OLS paths. My variables both appear to be continuous, but one is discrete with multiple categories. I'd like to calculate an effect size for all three paths, but this question specifically refers to the OLS paths. Can you advise me on how to calculate an effect size for these paths? I'm wondering if I should use the STDYX function to calculate the linear regression of y on x (pg. 641 of Mplus User's Guide) - and if so, what I should report for the effect size (i.e. beta values? r-squared?
Dear Dr. Muthen, I have a two-wave bivariate cross lagged analysis including also some control variables. The reviewer asked me to calculate to effect sizes. Is there any option to calculate them from the Beta estimates of the results? Thank you very much in advance
They want more standard measures than the betas. They would like to know the effect that the first wave variable (considered as the cause) has on the second wave variable (considered the effect) having controlled the stability (auto-regressive effects) and the control variables. I was thinking to report R square values including the cross-lagged relationships and without them, in order to demonstrate the variance explained when the cause is included in the model? Would it make sense? Thank you again!