Hi Mike. Thanks for the video. The technique you mentioned reminds me of the stepwise function found in SPSS. Not surprisingly the two major contributing variables to the fit are the two most statistically significant items as measured by probability. Again - neat video. Thanks Chris

Succint, straighforward. Great video. Thanks!

@mikecrowson2462

Жыл бұрын

You are very welcome!

Hi Mike, Great video to start me on the quantile regression! So do you suggest starting with my full model and then take out predictors one by one and look at the MAEs and see which of them contribute the most to the prediction? I have the post-anxiety scores as my response variable. But then utilizing mixed methods ANOVA I found only the pre-anxiety scores as the significant predictor. None of age, SES, Gender, Background education, and the use of our intervention to reduce anxiety were not significant predictors. As my response variable is skewed, my supervisor suggested doing the quantile regression, because any transformation did not have much effect on the normality distribution of the response variable.

@mikecrowson2462

3 жыл бұрын

Hi Mitra, I'm glad you my video helpful. Let me comment on different portions of your posting. First, when you say 'mixed methods' Anova I'm assuming you are referring to a repeated measures anova and you entered age, SES, gender and education as covariates. In that case, the covariates are essentially testing whether the change from pre- to post is moderated by those variables. The tests of the covariates in the repeated measures ANOVA is essentially like create a difference score on your repeated measure (Pre - Post) and then regressing that difference score onto those variables. In that regression model, predictors that are significant are effectively predicting CHANGE from pre-to-post-test and not scores at post. If you model your regression with Post-test scores being the outcome variable and you include pre-test scores, age, education, etc. as predictors, that is perfectly ok. Logically, one would expect that the pretest scores would serve as the dominant predictor in the model (and potentially wash out at least some of the other effects) since the regression slope captures the autocorrelation in the repeated measures (in a nutshell, one would expect that the same individuals responding to the same measure at two different testing occasions would have correlated responses). If other predictors are significant in the regression after controlling for pretest scores, then you have evidence of their effects on posttest scores after controlling for any pretest differences. That may or may not be something of particular interest to you. Now regarding the issue of skew on your response variable. OLS regression does not assume normality of the response variable but rather of the conditional response variable (with the latter reflected in the residuals derived from the model). So if you are studying skew to check your assumptions for the model, then you should be checking the residuals from your regression model. That's not to say that skew on the DV is unrelated to skew on the residuals. However, it is possible for residuals to be non-normally distributed due to other factors (e.g., omitted non-linearities, etc.). If you have non-normality in your residuals you can be thankful for the Central Limit Theorem for the fact that any potential biasing effect of that non-normality on your standard errors decreases as your sample size grows larger. A transformation of the DV is one possible approach that might reduce the problem of skew in your residuals; however, the downside is that you cannot interpret the regression slopes based on the original metric for the DV. It's ok to do it. But if the scaling of the DV has some substantive meaning behind it, it can be problematic interpretation-wise. Next (gee I feel like I'm writing a book here, but I have regression diagnostics on the brain!) you should always check your data for outliers and influential cases (see kzread.info/dash/bejne/p6COr7msqdfeeKw.html) for possible cases that may be swaying your results. Sometimes one or a few cases might have a pronounced influence on the regression parameters and exert bias in their estimation. Finally, quantile regression (such as median regression which I talk about in the video) is one approach to not having to deal so much with concerns about residual outliers or heteroskedastic variance, so it is a viable approach to parameter estimation. So to your question about MAE. In retrospect, I was thinking at the time about a possible effect size index that might be useful in reporting. And what you described is what I meant. However, an easier way to report on relative effects of predictors is to standardize those predictors and run your regression with those variables included. In effect, you will have a partially standardized solution that you can use to rank order your predictors in terms of their unique contributions to the model. Sorry for the long book here. You raised several issues that led the stat prof in me to want to make sure you had some pertinent concepts addressed :) Good luck with your research!

Hello Mike, Thanks for the video. Can you take a video about the piecewise regression. Is it possible to be run by SPSS? Thanks again.

Thank you so much, I have a question , how can I calculate the quantiles for a specific p, using Rankit-Cleveland method. Pleaase helpp

My version of SPSS doesn't have the option to do this, which ver is yours?

@mikecrowson2462

3 жыл бұрын

Hi there. I believe when I put this video together I was on version 26 (although now I'm up to 27).