Видео

Technisch kann man mit N = 80 eine Mediation rechnen, aber die Power ist halt recht klein, so dass man nur einen relativ großen Effekt mit diesem Design finden kann.

Q1 If normal results and bootstrap results differ, I would trust the bootstrap results. Q2 No. The interaction tests whether the (conditional) effect is significantly different for different values of the moderator. That some conditional effects are significant and some are not does not indicate that they are significantly different from each other.

​@@RegorzStatistik I appreciate your quick response. I understood the answer to the first question. I'd like to ask the second question again. If the bootstrap results for the interaction term are non-significant, can we still consider the conditional effects meaningful if they are significant at the mean and mean+1SD of the moderator variable? I'm wondering if the significance of the bootstrap results is a prerequisite for interpreting conditional effects.

@@sojeongyim5344 If the conditional effects were significant at M and +1SD then for those values of the moderator you have significant results. If you had (in advance) a hypothesis: There is an effect for moderator values at 1 SD then (and probably only then) that result would be "meaningful". In the general case, I would not look at (or interpret) conditional effects in the case of a non significant interaction - because in that case you have not shown that the effect differs for different values of the moderator. The conditional effects are a means to investigate HOW the moderation works. But if there is no moderation then this question is not relevant.

@@RegorzStatistik Thank you for the detailed explanation-it really helped me understand!

If only the X and Y relationsship is nonlinear (and not the X-M and M-Y relationsships) you could include X² as a covariate for the prediction of Y, thus using polynomial modeling for the direct effect.

@@RegorzStatistik thank you for the insight!

I have read that there is something like nonparametric LPA that does not assume normality. But I don't know if there is an R implementation for that.

Das lässt sich (wie leider meistens bei R Fehlermeldungen) aus dem Fehlercode alleine für mich nicht erschließen. One possible explanation is here: stackoverflow.com/questions/63775190/error-in-if-is-matrixmodel-ncolmodel-3-stopmodel-argument-must-b

Strange. I have rerun my code and I don't get an error message.

@@RegorzStatistik OK. Thanks for checking and for this informative video.

The assumption is met if the variables you include in your final model (transformed or not) have a linear relationsship.

Yes, I would rebuild exactly the model I run with PROCESS to test the assumptions.

Entscheidend ist, ob die Null im Konfidenzintervall liegt oder nicht. Es ist in Ihrem Fall durchaus möglich (und bei der relativ kleinen Untergrenze des KI auch wahrscheinlich), dass es auch vereinzelte Ergebnisse unter Null gibt - entscheidend ist nur, dass dies weniger als 2,5% der Ergebnisse sind, dass die Ergebnisse also außerhalb des Konfidenzintervalls liegen. (2,5 statt 5,0, weil es ja an beiden Seiten des Konfidenzintervalls einen Bereich außerhalb gibt).

@@RegorzStatistik Vielen Dank! :)

Auch eine Moderationsanalyse ist eine multiple Regression - wenn das Konfidenzintervall für die Interaktion die Null nicht mit einschließt, dann liegt eine Moderation vor. Wobei ich für eine Moderationsanalyse Model 1 von Hayes PROCESS Macro für R nutzen und dort als Zusatzparameter Bootstrapping anfordern würde. Dann bekommt man am Ende des Outputs für die Regressionsergebnisse auch noch Bootstrap-Resultate und kann so prüfen, ob das Bootstrap-KI die Null einschließt oder nicht.

Yes. If you use the multicategorical option for the IV, then PROCESS constructs two dummy variables for the three conditions. And those dummy variables are tested as ordinary binary predictors in a multiple regression, with a t-statistic.

Yes, TOST has the same assumptions as the parametric tests that form its basis. I haven't come across that problem in praxis, yet. One possible solution could be: Using the CI-approach (90% CI), but using bootstrapping for the CI of the group differences, since bootstrapping does not assume normality and is fairly robust against heteroscedasticity - e.g., running a linear regression with bootstrapping, and as the sole predictor using the grouping variable: ruclips.net/video/PgVnhPTIOgQ/видео.html

Leider nein.

I think, according to the definitions by Muller, it is not a moderated mediation.

You could look at Schoemann and Jorgensen (2021), full reference information see the video description. "Traditional methods of assessing model fit do not take this dependence among observed variables into account and, thus, provide incorrect estimates of model fit. To assess model fit for latent interaction model we recommend first fitting a model without the latent interaction variable, or any product-indicators, and assessing model fit on this “main effects” model." (section 1.3, last paragraph)

@RegorzStatistik I am grateful for your guidance... Thank you, brother...

I think you could use the extracted residuals as well for a homoscedasticity test (scatterplot predicted values and residuals, as in normal regression). However, for level 2 it is not trivial to get the predicted values.

Unfortunately, no, I don't have any recommendations for that.

That depends on the variable. In PROCESS there is an option for multicategorical data for the IV and the MOD, not for MED and DV, if you have a single Likert item that you treat as ordinal (a Likert scale made up of a couple of Likert items can be seen as on an interval scale, so that is not problematic).

@@RegorzStatistik Thank you for your response. The research variables I am using are childhood trauma (X), malevolent creativity behavior (Y), aggressive behavior (M), and resilience (W). All instruments use a Likert scale. There is a question from another statistician about whether the PROCESS model, which uses a moderation-mediation (multiple regression) approach, requires interval data for analysis, while the data generated from the Likert scale is ordinal and cannot be further processed.

@@bagusadi1277 PROCESS is based on regressions and has the same requirements as multiple regression.

@@RegorzStatistik Thank you for your attention and response. I truly appreciate it. I am still quite new to research, especially at the master’s level, where more complex analytical techniques are involved. I have also read various journals and other literature regarding differing opinions on analyzing ordinal versus interval data. It seems that some statisticians are against using regression or multiple regression approaches to analyze ordinal data obtained from Likert scales, even when assumption tests (independence, linearity, homoscedasticity, multicollinearity, and normality) have been conducted. I would also like to hear your opinion regarding other studies I have encountered, which suggest an additional process of transforming ordinal data into interval data using approaches like MSI or Z-scores to enable analysis using Hayes’ PROCESS model, which is based on regression and multiple regression. What are your thoughts on this approach? To be honest, the differences and debates around data transformation and the analysis of ordinal versus interval data using Hayes’ PROCESS model have left me quite confused. Thank you so much for taking the time to share your thoughts and insights.

@@bagusadi1277 You can find my thoughts about this topic here: ruclips.net/video/rFvSQIQ0Yis/видео.html In general, I would look at your are of research and the published papers there. How do they treat Likert items or Likert scales?

Ja, eine Moderation ist im Prinzip eine ganz normale Regression mit denselben Annahmen. (Allerdings ist es relativ häufig so, dass bei einer Moderation eine gewisse Multikollinearität vorliegt - das lässt sich aber nicht grundsätzlich ändern und ist im Modell begründet.)

Nicht wirklich, das habe ich primär aus Vorlesungen an der Uni gelernt und weniger aus Büchern.

Dafür würde ich zusätzlich Model 4 rechnen.

Das kann ich nach dieser Schilderung alleine leider nicht beurteilen.

I don't think there is an index of moderated mediation for PROCESS model 59.

Das weiß ich nicht. Haben Sie schon probiert, nach der Tabellennummer nur eine Zeilenschaltung und keine Absatzschaltung einzufügen? Vielleicht erkennt Word das dann als zusammenhängend? (Habe ich aber noch nicht ausprobiert)

@@RegorzStatistik vielen Dank! Das hat tatsächlich funktioniert :)

Dieser konkrete Fall ist m.E. nicht geregelt, aber es ist durchaus vorgesehen, als Mitautoren auch Leute mit "Spezialized Roles" aufzuführen, 9.10 APA 7. Beispiele dort sind Herausgeber (Ed.) Und bei einem Film: (Writer) und (Director) Insofern könnte man dann m.E. auch die Illustratorin mit bei den Autoren angeben mit der Rolle in Klammern danach.

I think in PROCESS this would be the case for a moderated mediation, e.g., model 7 (with one of the IVs as IV and the other IV as moderator).

Currently, there is only nonparametric survival (Kaplan Meier). But they plan to add other models to the survival module as well: jasp-stats.org/2024/01/09/non-parametric-survival-analysis-in-jasp/

I think by including an additional parameter in the ICC function: icc(null_model, ci = T) For more information about the possible parameters of the icc function: ?icc

Ja, es müssen genau die Modelle nachgebaut werden, die auch im PROCESS-Output erscheinen.

I would report the results. It is not that uncommon that you don't get the results you expected.

@@RegorzStatistik thank you, do we need to run model 4 before doing model 59

It provides results that are robust to a violation of the homoscedasticity assumption.

I would report the regression coefficient and the bootstrapped confidence interval.

I don't know about that. If I have heteroscedaticity I simply use other standard errors (HC3, HC4) or bootstrapping.

If BOTH is significant, then both interactions taken together explain significantly additional variance of the DV (beyond the other predictors in the model). Whereas the test for an individual interaction shows whether that interaction explains significantly additional variance of the DV beyond the other predictors including the other interaction. It is basically the same situation we know from multiple regression: You can have a significant correlation between a predictor and the criterion variable but at the same time not have a significant regression weight for that predictor in the context of a multiple regression. If one of the interactions is not significant, then that does not imply that its effect is zero; for that reason the effect size for BOTH will be larger thant the effect size for the other, significant interaction, in most cases.

​@@RegorzStatistik thanks for your reply. If one interaction is not significant, and the other interaction is significant but both has a value greater than either square change for either oof the interactions can we talk about synergies between the moderators? I was trying to see when two moderators exist together (both are different HR practices) whether their moderation effect will produce a higher moderating effect or not

@@giuc100 If the R² for BOTH is smaller than the sum of the two individual R² values of the interactions, then there is some amount of overlap between the two interaction effects.

I am not sure that I understand your question. You can run a PLS SEM with formative constructs as shown in the last part of the video.

I think there are many videos about how to import a csv file. I guess you will find them my searching in RUclips with this search string: r import csv

I haven't come across this situation, yet, so I don't know. Here are my guesses: ad 1: For something like that I would prefer using LCGA (latent class growth analysis or GMM (growth mixture modeling) instead of LPA. ad 2: In this case I think I would combine those measures into one value and include that in the LPA.

@@RegorzStatistik Actually, I plan to identify latent groups based on three variables measured longitudinally. However, these three measures were not assessed at the same time points. For example, x was measured at ages 11 and 14, y was measured at 103 months, 9 years, and 11 years, and z was measured at ages 8, 10, 12.5, 13.5, and 17.5. As you can see, the intervals within each variable and between variables are not equal. Do you think I can still use GMM in this situation?

​@@ahmadvalikhani6290 In that case I don't think GMM will work (but I have no experience with a data situation like this). One problem for a LPA could be that z (with 5 timepoints) will dominate the analysis. If you run a LPA I think I would run at least two models: one with the full set of variables and one with only 2 or 3 ages for y to counter that threat (but that is only my intuition, not based on any literature).

@@RegorzStatistik Thanks. I will try to use cross-sectional data. Apart from that, I have kept encountering this error, but I could not find the reason or solution for it: The 'variances'/'covariances' arguments were ignored in favor of the 'models' argument. Warning in (function (data, modelName = NULL, nboot = 999, level = 0.05, : some model(s) could not be fitted! Warning in (function (data, modelName = NULL, nboot = 999, level = 0.05, : only 1-component model could be fitted. No LRT is performed! Warning: Mclust could not estimate model 2 with 5 classes. Warning: Mclust could not estimate model 6 with 2 classes. Warning: Mclust could not estimate model 6 with 3 classes. Warning: Mclust could not estimate model 6 with 4 classes. Warning: Mclust could not estimate model 6 with 5 classes. Warning: One or more analyses resulted in warnings! Examine these analyses carefully: model_2_class_5, model_3_class_5, model_6_class_2, model_6_class_3, model_6_class_4, model_6_class_5

@@ahmadvalikhani6290 I guess some models could not be estimated, e.g. convergence issues. Nothing more I can say based on this.