Видео

Hi. Thank you for the feedback. The additional line is mainly down to determining the size per group, The first part of the additional line of n/(df of # measurements/timepoints) is clear cut, but then when I tried to align with G*Power, I found that I was regularly 2 off, i.e., # groups, for the data used. The adjustment was purely to align with G*Power. I'd imagine this adjustment is due to rounding in the sample size calculations as there is a lot happening behind the scenes. You could argue that the "+ # groups" is not needed, but then there will be a minor difference between R and G*Power in output. I just found the adjustment a minor "trick" that appeared to align both packages. Hope this helps.

A few options out there, but I generally use the eta_sq() function in the sjstats package.

I'd probably just plot the 95% CI of the means, as opposed the estimated marginal means, if all you're interested in is a visual. A video on interaction plots that I have may help here ... ruclips.net/video/IFrL4f86nIM/видео.html.

The output that complements the effect size are the 95% CI limits, which are different to the standard error values. The CI is the mean +- margin of error. When the margin of error is the standard error divided by sqrt(n).

I would generally work out the AUC for each individual sample.

@@LaceyMathsStatsConsultancy so If we will do for each single sample then by using all values we will perform statistical test then?

@user-pm8gz3jr1w Yes ... you set up a for loop to calculate each AUC for each sample, then create a dataframe mapping each sample to corresponding AUC.

@@LaceyMathsStatsConsultancy thank you sir for your assistance

Sir how I will create a Loop for this analysis can You please tell me. Like I will fix something in Loop take one by one sample sets for calculations of AUC?

Not really. This is for the user to ensure that if there are multiple explanatory continuous variables in the model, that these variables are not autocorrelated, e.g., using variance inflation factor. Also, it's important that if the purpose is model fitting, that the model residuals are checked for normality and randomness.

@@LaceyMathsStatsConsultancy Thanks! Sorry I actually meant temporal autocorrelation. I find it difficult to get my head around that concept or how to properly implement it in repeated measures models!

@@designedfore I'd imagine in this case, it would be a matter of checking the autocorrelation of the residuals from the model. So if the model residuals are not approximately normally distributed and random, then there is a potential issue with the model. I can look to do a video on this, if you feel it may be of help?

@@designedfore - some interesting info on temporal autocorrelation on www.flutterbys.com.au/stats/tut/tut8.3a.html

Hi Jess, That's a package separate to RStudio which can be downloaded from: www.psychologie.hhu.de/arbeitsgruppen/allgemeine-psychologie-und-arbeitspsychologie/gpower.

Yes, you are correct. It is the sample size outputted from the power calculation divided by the number of levels to the within subjects factor less one + number of levels to the between subjects factor. So, if using the notation from the script: It's the ceiling to [n/(nm - 1) + ng]. Is that ok?

@@LaceyMathsStatsConsultancy Awesome! thank you very very much!!

eff_size outputs partial eta-squared. There are some nice points, with references to support, on eta-squared and partial eta-squared, and when to use on: stats.stackexchange.com/questions/15958/how-to-interpret-and-report-eta-squared-partial-eta-squared-in-statistically.

@@LaceyMathsStatsConsultancy ​ thanks! great post, indeed! I've just opened a post on CV asking about that. My case is very similar to yours (youtube won't let me paste my CV link here). I have a mod1 <- lmer(Y ~ GROUP * YEAR + (1|PARTICIPANTS) and then I'm doing: emm1 <- emmeans(mod1,~YEAR|GROUP) and emmeans(mod1,~GROUP * YEAR) . Each group has 2 levels (Group 1 and Group 2) and Year has 2 levels (year 1 and year 2) too. Then I'm calculating as you've said in the video: eff_size(emm1, sigma = sigma(mod1), edf = df.residual(mod1))

so are the partial n² suitable in this case?

@Larissa Cury Partial eta-squared is what I'd use in the scenario you've outlined, clearly outlining it's classification in the Methods section of a report.

@@LaceyMathsStatsConsultancy thank you! what do you mean by cleary outlining its classification?

The last few lines of script give an example of how to report the effect size statistics with corresponding p-values, em means and CIs. All that would be left is to write the results to Excel and then you are good to go.

That warning is just informing the user that there are interaction effects. Typically, the interaction effects are your primary outcome, with the main effects being your secondary outcome. The warning is not an error but just a form of signposting in case the user is not aware of interaction effect(s) being present. The effect sizes from the main effect are still accurate.

@@LaceyMathsStatsConsultancy so when I'm performing emmeans(model,~ B|C) %>% pairs(adjust="Tukey"), am I accouting for that? I don't get the warning (being the model: Y ~ B * C + (1|ID))

@@LaceyMathsStatsConsultancy thank you!

That's great to read. Thanks for the feedback!

@@LaceyMathsStatsConsultancy You are every welcome

No problem ... a link to the script is included in the description ⬆.

Thanks for the feedback. Within the argument to stat_boxplot, you can include width=0.25, but this upsets the position of the whiskers, however, including position_dodge(width=0.75) will fix the alignment, e.g., stat_boxplot(geom = "errorbar",width=0.25, position = position_dodge(width = 0.75)) Hope this helps.

Hope the script from the following link is a help: drive.google.com/drive/folders/1qxJh5U5hzqznJG-vXlW7XvZ4uDk-S0g5?usp=sharing

No problem. Script can be copied from the following link: drive.google.com/drive/folders/1wvHXV-_ZPFTjTCy5FP2SgGGLBni1i_mD?usp=sharing

www.psychologie.hhu.de/arbeitsgruppen/allgemeine-psychologie-und-arbeitspsychologie/gpower

Thank you for the feedback. Delighted you found the video helpful.

Thank you...a follow-up video is on the to-do list.

Hi Dear professor Lacey! I'm moving to Cork in 2022 and I'll apply to Data science course(Higher Diploma) in CIT. Will you be teaching in that course next year? Hugs from Brasil! Hope to meet you soon.

No problem ... thanks for the feedback!

Great...thank you for the feedback!

@@LaceyMathsStatsConsultancy with 400 plus variables how do i recode all missing values to 999 for every variable at once

@@nyashapatiencemandeya1128 This syntax should do the trick: RECODE ALL (SYSMIS=999). EXECUTE.