Відео

glad you liked it!

oh, yeah, I see what you mean. in this case I did not look at MCAR etc. yet, but I'll put it on the list for the future videos. until then have a look at this, and similar, articles: cran.r-project.org/web/packages/finalfit/vignettes/missing.html

Glad you enjoyed it! Thanks for watching!

You are very welcome! Thanks for watching!

@@yuzaR-Data-Science I have a question about the video around the 8:10 mark. A line is drawn at misclassification_cost = 213, but why doesn't it go slightly lower? There are two more points below the red line where misclassification_cost can be confirmed. Could this be due to a default setting for tol_metric or something similar? Let me know if you'd like any refinements!

deep question! I think that balancing sensitivity and specificity is also important. so that, whe the sum of sens-spec is literally maximazed (goes to those 2 points), either sensitivity or specificity will be smaller or bigger than the counterpart, while they should be very similar. again, floating cutpoint problem exists and the cutpoint could be a bit below or a bit above the "one optimal cutpoint". I never saw the issue about floating optimal cutoff raised in any paper. so, knowing this you'd have an advantage in inference ;)

@@yuzaR-Data-Science Thank you for your reply! In the code for cutpointr, the specified metric = misclassification_cost should take priority. This metric is calculated based on the number of False Positives (FP) and False Negatives (FN), so it seems unlikely that the logic would be influenced by considerations of the balance between sensitivity and specificity. Do you mean to suggest that the balance between sensitivity and specificity might be influencing the algorithm in a way that is independent of misclassification_cost? I would greatly appreciate your advice.

@@yuzaR-Data-Science After examining the data in detail, the results for the points where you drew the green line were as follows: misclassification_cost = 212 (fp = 59, fn = 153) misclassification_cost = 212 (fp = 60, fn = 152) misclassification_cost = 213 (fp = 60, fn = 153) misclassification_cost = 213 (fp = 68, fn = 145) misclassification_cost = 213 (fp = 69, fn = 144) Additionally, the optimal_cutpoint was the average of the predicted_glm values for the two points where misclassification_cost = 212. Given this, it seems logical that the result of res |> summary() should report misclassification_cost = 212 instead of 213. Wouldn't you agree?

sure. any groups you have in you dataset ;)

I only provide code for the members. There is another button near "subscribe", it's name is "join". So, in this way you would support me, but you don't need to do that at all. Just stop the video and type up the code. it's even good for learning. if you decide to support me, you can have access to any code from any video I made.

bro, I tried to do tidymodels videos already, only got a few, because they are more for predictions, not for inference. and honestly, tidymodels did not provide enough interpretability while always made a way to the result in a special own not-completely intuitive way. one day I will definitely do videos on tidymodels, but they are less useful for inference at the moment and there are many other useful and interesting topics to cover. hope you stick along for some time until I make tidymodels content. cheers mate

by the way, did the end work? :) I don't see it working in channels analytics, but I definetely create some suspense at the end to keep people binch-learning ;) thanks for noticing!

@@yuzaR-Data-Science Thank you brother. Will surely wait. But I totally agree that you cover the things that you are doing they are really more important for R users. Thank you again mate.

@@yuzaR-Data-Science Well, I would honestly tell, maximum people are just watching videos and go by because they don't like statistics and are just watching it for a project or assignment. Then there will be a small proportion who like statistics and but not mainly R. Then there will be even smaller of both statistics and R lovers and they wouod have time. So what you are doing is good enough. Just see the percentage of people who will be effected by it. And surely. For those who don't know what to see next but want another topic. It gets the click through for sure.

thanks mate!

oh man, thanks a lot for such a generous feedback! statistics is too wide of a field to have one single ressourse. There is classic stats, which I do here. There is stochastic and some folks follow it religiusly. And there is maths side of stats, which can be very beginner-unfriendly. So, what to do? If you are beginner, read many books about classic stats for non-statisticians. I am obviously biased, because I work in animal science, but my favorite is: "Statistics for veterinary and animal science". It's practical, not too wordy, but very clear and simple. Books for biologists of epidemiologist are also good for beginners. If you are advanced, go to ISLR book, it's free and online. This one is a good link between classic stats, math behind it and intro to ML and AI. If you are into stochastic and Bayesian framework - Statistical Rethinking. Main thing though: start analyzing data accepting that you will be wrong. My goal is to be less wrong every time I do something. If I don't know or doubt my analysis, it sucks in the beginning, but, as Richard Feynman said, it will show your gaps in knowledge and where you should invest more time. Ask 3 statisticians the same question and you'll often get 3 different answers. So, the answer to your question is - many ressources instead of one and don't try to be right, try to be less wrong :)

@@yuzaR-Data-Science Thank you, such a detailed and informative answer. Thank you for taking the time to write it. Yes indeed I wanted more of it related to health sciences. Will surely check all of these.

glad I could help. cheers

sure, I just renewed the link for "emmeans" post. Just go to community (members space), search for "emmeans" and klick on link. this link will be active for 3 days. there is both code and explanations for the code. let me know where you was able to get it. cheers

I already have a video an imputation of missing values. Check out this one: ua-cam.com/video/Akb401i32Oc/v-deo.html

ROC curves are primarily used to evaluate the performance of binary classification models (like logistic regression), not linear regression or survival analysis.

@yuzaR-Data-Science Alright,thank you.

You are welcome 🙏

Glad it's useful! Thank for your nice feedback and for watching! Hope the videos are not too long and not too boring. Let me know if you'd prefer shorter less dense ones

Glad you think so! Thanks for cool feedback! :)

Perfect! Glad it's useful and and I was lucky it was a good timing :)

thanks for nice feedback! :) yes, you are right, the video would have become to long. what do you think if I would have made several separate videos, like one for multiple cutoffs, one for bootstrapping etc.? I don't to make videos for the sake of making videos ... but try to have a dense coherent high info story. but there is a trade off for everything :) so, what would you prefer?

@@yuzaR-Data-Science Several separate videos, sounds good! More videos on about model performance will always be useful. In general, the topics of model diagnostics, model performance, and variable importance are three topics I need to delve into the most right now. Your videos are full of enthusiasm and knowledge, so any material on these topics will be useful.

Have you seen my review on the performance package? That video is probably exactly what you want. It’s general though. Any kind of model has unique assumptions set. I have also several videos on few different models ready to watch. Hope they can help

Hope you enjoyed it! Thanks for watching 🙏

thanks mate! I hope you'll enjoy the second video on quantile regression too

@yuzaR-Data-Science please send the link of your 2nd video

here we go: ua-cam.com/video/4nJD2tpZFDs/v-deo.html

first of all - sure, it can, secondly, you don't need to because bootstrapping will work the non-normality out ;) that's the whole purpose of the method

thanks :) you probably only subscribed. I send the code to people, who joined. It's a different button, where people support me monthly with a small pay. but you don't need to do that at all, you just can pause the video and write down the code, it's even better for a learning purposes to tipp out the code yourself. kind regards

Not at all, you can use same visualization functions for both, mixed and not mixed models

sure, vif works for both numeric and categorical variables. how it's calculated - I don't know exactly, just superficial formula like 1/(1 - summary(model)$r.squared) - but I treat it like a car: I don't know how engine works, but I know how to drive. so, if your vif is below 5 or in some cases below 10, you can accept the results. when vif is above 10 you'll find some multicollinear variables (both numeric and categorical)

thanks you very much for the feedback! will do survival analysis with R similar to this one! I have two very old not very good and not R, but a bit theoretical videos on survival analysis on this channel. I don't think they are helpful, but you want, you could check them out.

good question, I have a video on performance package. check this out and you'll see the demo with answers your question :)

Good question mate! The role of tidymodels in med science is small but steadily growing. I also wanted to make more videos on tidymodels, but they don’t provide much inference. They predict but often don’t explain why and how. With images and videos in medicine they will gain more insight influence though. So it’s definitely useful to learn and I’ll cover them in the future videos too.

Great suggestion! It’s already on my list! Please, stay tuned

Confounding topic is definitely on my list! So, it’ll be produced in the near future! Until then you might find glmulti package 📦 and my video on it useful. It’s better than backwards variable selection. Just be careful not to use too many predictors. First, sort out only the ones you want to really analyze. Then use univariable and sort all out with p value below 0.2. Then use glmulti. Also make sure vif is below 5 in all predictors

@yuzaR-Data-Science Can you help me share your video title?

@@dinhluongnguyen3610 sure, here it is: ua-cam.com/video/Im293ClFen4/v-deo.html

@@yuzaR-Data-Science Thank you so much. A great, if you give pdf file.

sure, here is the pdf for donwload for the next 3 days: we.tl/t-gIzPga6bjP . If you see this message after 3 days, let me know and I'll upload again

Of coarse, when you join my channel, I send you the pdf with code and explanations (transcripts) of any video. But, please, don't feel like you have to join! You just can pause the video and type up the code, it's free and not much of a code. Please, only join if you want to support my work and you'll get the benefit of getting the transcripts. Kind regards, Yury

Thanks so much 🙏 I also struggled in the beginning of learning R, so I try to teach the way I would have loved to be teached

Thank you very much Muhammad for such a nice feedback! Sure, in the beginning we'll all had difficulties to switch to R. I came from Matlab and NCSS to R. And also needed to box myself through the error messages. The good news is - the error messages are finite. The are only a few (20-50) error messages, you quickly learn how to deal with. After it error message will become a help. Levels are easy, you can determine the order yourself: library(dplyr) library(forcats) # install the packages, if they don't load df <- data.frame( category = c("B", "A", "C") ) # Reorder levels df <- df %>% mutate(category = fct_relevel(category, "A", "C", "B")) # Print the reordered factor levels levels(df$category)

@@yuzaR-Data-Science Thank you very much. Will also be looking forward to more video.

thanks

Glad it helped! I don't know about the methods switch after update. But I know that the author of the package is responsive and you can ask him via github or twitter or any other way. He'll give much better answer.

@yuzaR-Data-Science i will try that, thank you

@@thomaswiggersmller8983 👍

no, you can sort them out via p-values, e.g. <0.2, but the variable importance should only be asked from the multivariable models. folks in the medicine have little idea about stats, thus, take their methods with a grain of salt and consult a statistician ;)

@@yuzaR-Data-Science thanks for replying! Just to clarify would you put all of the available predictors in a multivariate model and then based on p-values <0.2 adjust the model accordingly? (like backwards selection process?)

hey, unfortunately the robustbase is not supported yet. but you can make a request on the github of the author. most other models work well

Thank you! ❣

@@FabianMueller-n6g you are welcome! :)

Glad you enjoyed it, Elias! I am working on roc curve and optimal cutpoint video right now. Hope it will deliver the things you are interested in. Stay tuned. Kind regards from holidays in Australia