Tutorials for statistical software - as long as necessary and as short as possible. R, SPSS and even Excel are covered. Branching out to other helpful and related topics for getting data analysis done.
Background: Dr. Björn Walther (PhD in Economics), doing simliar content on my German RU-vid channel. ru-vid.com_BjoernWalther
Hello, thanks for the video! Just a question, how does this differ from G*Power's one-way ANOVA? Or in other words, what makes this calculation a Kruskal-Wallis test per se?
Hey there, it is the same calculation as for the one-way ANOVA, since the goal is the same and power differences are often negligible between the two, except in rare cases. Some researchers will add 15% to the sample size to be safe. Cheers, Björn.
@@statorialsOh okay, thanks for the clarification! Do you by any chance know where I can find the actual calculations G*Power uses to ultimately reach these sample sizes? I have been running power analysis in Stata, Python, & R using the same inputs but I get different Ns!
I would have pointed to the GPower manual, but that is not revealing the formulas used. Cohen (1988) is always a good start to search, but I could not see anyhting either while taking a quick glance.
i have 3 groups and gender variable for between subject design. for anova results only the group variable has significant main effect while the interaction and gender show no significant effect. how would i perform the Post-hoc test in R?
Hey there, it sounds like you have no repeated measures when you have only between subject factors, hence, I would recommend going with the two-factorial ANOVA and not a mixed ANOVA. df %>% anova_test(dv ~ iv1*iv2) Post-hoc testing depends on the interaction effect. If you have none, you will use something like that: df %>% pairwise_t_test(dv ~ iv1, p.adjust.method = "bonferroni", detailed = TRUE) with iv1 being your group variable. Cheers, Björn.
Hi I wonder, i cannot tick the box for comparing means for Factor Interactions. I could only compare the means for between subject only, or wtihin subject only. Is there anything i could do? I'm using SPSS 27, btw.
Hi there, if you have specified a between-subjects factor, SPSS automatically uses it within products in EM Means. The only thing I can think of, is that you specified a custom model under the Model button that consits only of the respective "main" effects. Maybe check on that. Cheers, Björn.
@@statorials Yes, actually, under EM Means Button, there's no Compare Simple Main Effect's Box to tick, like what's shown in your video. Only Compare Main Effect's Box. I still don't know how i could specified it under the Model Button. Will there be any problems if I try to manually use independent t test to determine the mean difference for each measurement as a post-hoc analysis? without Bonferroni correction. Thank you very much
Hmm, I suppose SPSS 27 could have EM Means not included (yet). You can of course do the t-tests yourself. You can use ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-PW4OlfCLrM8.html or ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-INg9oHbmsJY.html as orientation. You only have to take care when doing independent or paired t-tests when you observe between-subjects effects and within-subjects effects or even when doing the interaction effect. Cheers, Björn.
Hallo Björn, vielen Dank für deine hilfreichen Videos! 🙏 Kann man für die Post-Hoc-Tests auch eine Effektstärke (Cohen's d) berechnen und wenn ja, wie? Über eine (zeitnahe) Antwort wäre ich sehr dankbar. Die Abgabe der MA rückt immer näher. 😅
Hallo Lisa, danke für das Lob! Ja, das ist möglich. Auf dem deutschen Kanal ist das erste von zwei Videos schon online: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-f6JRhnCmePc.html Die Effektstärke für die Haupteffekte kommt in genau einer Woche. Viele Grüße, Björn.
@@statorials Danke für deine schnelle Antwort - das hilft mir so sehr! 🙏 Tatsächlich hätte ich noch eine Frage: Bei mir kam ein signifikanter Interaktionseffekt raus und die Post-Hoc-Tests haben gezeigt, dass sich die Gruppen zum Prä- und zum Post-Messzeitpunkt signifikant unterscheiden. Ist es sehr problematisch, dass sie sich auch zum Prä-Messzeitpunkt unterscheiden? Denn das würde man ja eigentlich nicht erwarten. Die Werte der EG steigen vom Prä- zum Post-Messzeitpunkt auch signifikant an. Die der KG nicht. Das ist also erwartungsgemäß. Ich danke dir jetzt schon vielmals!
Hallo Lisa, gerne! Zu deiner Frage: Wie stark ist der Gruppen-Unterschied in der Prä-Messung mit Cohen's d denn? Wenn es ein noch kleiner Unterschied ist, wäre das sicher verkraftbar und könnte auch problemlos als vertretbar eingeordnet werden. Die Tatsache, dass sich nur die EG steigert, unterstützt entsprechend die Hypothese, dass die Intervention hilfreich war. Viele Grüße, Björn.
@@la_lisamarie Hallo Lisa, zwar ist der Effekt schon tendenziell mittel, dennoch würde man den hier relativieren können, wenn man gleichzeitig die Steigerung nur bei der EG hat. Ist die EG oder KG höher in der Prä-Messung höher? Also überholt die EG die KG oder wird der Unterschied zwischen EG und KG größer, wenn die EG auch bei der Prä-Messung höher lag? Viele Grüße, Björn.
Hallo Björn, vielen Dank für deine ganzen hilfreichen Videos zur mixed ANOVA in R! 🙏 Ich muss im Rahmen meiner Masterarbeit das partielle Eta-Quadrat berechnen. Ist das hier auch möglich? Ebenso muss ich für die Post-Hoc-Tests der mixed ANOVA eine Effektstärke (Cohen's d) berechnen. Könntest du mir hier weiterhelfen?
Hi, Thank you for providing the tutorial. I have a couple of questions. I want to calculate the number of participants required for the following study. The study is longitudinal in nature and has two groups. Data is collected at two time points. The experiment has three conditions and a control condition. I am planning to compare the two groups across time and the different experimental conditions within each group and across time. How do I calculate the number of participants required in a group for a large effect size? Also, while selecting, should I select between factors, within or interaction? Thank you
Hi there, I would recommend doing a mixed ANOVA approach that covers between and within effects: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-5AflJkhaln0.html The estimated marginal means (posthoc-testing) you mentioned, cannot be directly considered in the sample size calculation though, which is not a deal breaker. Cheers, Björn.
Hi, thanks for this great tutorial, with 16 groups and 8 measurements I obtained total sample size of 272, this means 17 per group. Does it mean that per each measurement I need 17 replicates or 17/8, which is 2.125, which would make 3 replicates. This concerns a pot experiment with 4 variables (e.g., presence/lack of plants, presence/lack of contaminant 1, presence/lack of contaminant 2, presence/lack of contaminant 3). Thank you for help!
Hi Adam, glad you like it. The way I read you post, it appears to me that you have 8 measurements per 272 individuals that will be split into 16 groups. Do you have an additional between-subject factor with the pots, meaning your subjects will be put into 16 groups and within each group you have another experiment? What is the distinction between the groups, if not? Cheers, Björn.
@@statorials Hi Björn, thanks for the reply. So the experiment is divided into 4 variables: presence of plants (yes or no), presence of contaminant (yes or no), type of soil (type 1 or 2), aeration (yes or no), which gives 16 combinations. There would be 8 samplings every week. We would like to used repeated measures ANOVA afterwards and we're interested if 3 replicates per each group would be enough. All the groups would be affected in the same way by changing temperature. Thanks. Adam
Hi! I was wondering why you call it a mixed ANOVA if you are doing a repeated measures ANOVA? Even though linear mixed-effects models can analyse the same data as RM-ANOVA, these are actually 2 different types of models.
Hey there, the term "mixed ANOVA" is used because it includes both within-subjects (repeated measures) and between-subjects factors. in the traditional ANOVA model world. While the term "mixed" in mixed ANOVA may be misleading in regard to linear mixed models it is, as you pointed out, a different class of models. Depending on your research field, the terms may vary a bit, like split-plot ANOVA or something with mixed, like mixed-design ANOVA or mixed-model ANOVA. Cheers, Björn.
@@statorials i have 3 levels for my within subjects factor and two show a low p value for the levenes test. However, my sample sizes for the between groups only have a difference of 11 participants and the standard deviations for each group do not vary much. Is there another rule of thumb I can use to evaluate this assumption ?
I don't understand why for rank biserial you calculate it with spearman and point biserial with pearson. rank biserial is not run with rank_biserial of effectsize?
Hey there, I dug into that a bit and Spearman is an approximation (at best). I'll swap out the video in the oncoming weeks when I had time to record a new video regarding that. One can use the effect size calculation of the two-sample Wilcoxon test, meaning Glass rank biserial correlation coefficient or Cliff's delta. For your question, you can "correlate" an ordinal an a nominal variable using Chi² test of independence and use the effect size Cohen's ω to quantify the association. Cheers, Björn.
Thank you Dr. Björn! Please can you correct me if my understanding is wrong? If we have paired data ( before treatment , after treatment) first we need to check normality assumption ( for residuals NOT for each group separately ) using Shapiro-wilk then if p_value > 0.05 , the data is normally distributed, so we can use paired T-test if p_value <0.05 , the data is not normally distributed , so we can use Mann-Whitney Wilcoxon test. thanks again.
Hi there, you are almost correct in your proceedings. 1) I would be careful though when it comes to the Shapiro-Wilk test and its sensitivity in larger samples, rejecting the Null Hypothesis because of negligible differences. Consider a Q-Q-plot for that task. 2) If that looks not approximately normally distributed, go with the Wilcoxon paired test aka. Wilcoxon-signed rank test (ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-9sX62Qagrs0.html), NOT the Mann-Whitney-Wilcoxon test. Cheers, Björn.
Hi, thanks for your video. When I tried to runa the post hoc test I have an error. "Error in `map()`: ℹ In index: 1. ℹ With name: V1. Caused by error in `complete.cases()`: ! not all arguments have the same length Run `rlang::last_trace()` to see where the error occurred." I guess it is because I remove some extreme outliers and data is unbalanced. Any way to follow with the comparison (Sum of Squares IV perhaps) DO you know how to do it?
Thanks for the tutorial. If using the "determine" option and the effect size specification as in Cohen (1988), should the variance explained by effect and the error variance entered be the squared values or not ("absolute values")? I find it a bit tricky to find values from comparable studies. For the variance explained by effect, could I enter the change observed in the main outcome variable in a comparable study, while using the error between measurements of the (same) main outcome variable from another study?
Hello Sara, my G*Power does not show error variance under the Determine option for the mixed ANOVA. Which version of G*power are you using? You could also use the sum of squares, if provided by the studies, to calculate the Eta-squared and then transform it to f. See here, how to do it: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-YPO0hBKCUQc.html . By the way, it should be sressed that the effect siize for the mixed ANOVA is based on the interaction effect. Cheers, Björn.
@@statorials Thanks for the reply, Björn. I have the same version as you. The error variance will appear under the Determine option if choosing "as in Cohen (1988) - recommended" under Options.
Thanks for the feedback! You're welcome! I have not read about that anywhere yet. Sure, it would follow the same logic as lowering the threshold with large samples. If you can dig up a source, you could surely justify that, although a lower lower alpha for a smaller sample is more favourable. ;-) Cheers, Björn.
Great video! But how do I interpret the pairwise comparison charts (the triangle shaped ones)? I've been searching and I'm yet to find out what do they mean.
Thanks for he positive feedback. The triangle is just a different way of presenting the information of the pairwise comparisons for the Kruskal-Wallis test. Depending on how you defined your Alpha in the K-W-Test, the lines represent, if the comparisons of the groups from start of the line compared to the end of the line is "significant", aka. the p-value low enough that the effect can be be observed due to chance. So, a blue line means that the groups that are connected through the line, are different in regard to mean ranks. A red line means they are not. If you have more groups the form will have more lines - one for each pairwise comparison. Cheers, Björn.
