Here's a tip to get more value out of plots. Try to guess what you expect to see on the plot before you make it. Write these predictions down if it helps keep you honest. Then look at the plots and compare to your predictions. I find this exercise helps draw my attention toward the things that I didn't expect based on theory or intuition.
I once invented something I called "shared parameter models", but a few months later I learned that I had independently come up with mixed effects models which by various names have been in the literature for quite a while.
My chair would lose his mind over this. He'd say, "You always start with theory to guide your model development and if theory doesn't suggest it, you don't model it." He had no use for looking to see what the data has to say. I took many a beating in his seminars over that point.
You need both. If there's a model the visuals suggest that doesn't make sense theoretically, don't model it. What I'm complaining about is when people use neither--they just test one model after another, then visualize it (maybe).
Don't worry about being too technical, if i don't get it the 1st time i will get it on the 2nd or 3rd watch. Yeah, i probably don't speak for everyone...
@@QuantPsych I have some suggestions if you do choose to dig into the topic while staying within your focus of linear models. Most raw data is non-negative or non-numerical and asymptotic normality cannot always be relied upon. So I would recommend exploring the topic with generalized linear models with Gamma and Poisson conditional likelihoods as they illustrate modelling cases such as duration or counts respectively. Multinomial logistic regression is a good choice for (non-hierarchical) categorical data.
Visualization is definitely underrated.. Thank you for your research papers, as we need them to emphasize the proper rules and methods.
Here's a tip to get more value out of plots. Try to guess what you expect to see on the plot before you make it. Write these predictions down if it helps keep you honest. Then look at the plots and compare to your predictions. I find this exercise helps draw my attention toward the things that I didn't expect based on theory or intuition.
I once invented something I called "shared parameter models", but a few months later I learned that I had independently come up with mixed effects models which by various names have been in the literature for quite a while.
You’re right in time! I’ve just started doing that part of my diploma!
Thank you for the previous videos on hierarchical data, they saved me!!
My chair would lose his mind over this. He'd say, "You always start with theory to guide your model development and if theory doesn't suggest it, you don't model it." He had no use for looking to see what the data has to say. I took many a beating in his seminars over that point.
Theory-driven model development is okay, but it tends to work best when you are interested in what Dustin calls a "confirmatory approach".
You need both. If there's a model the visuals suggest that doesn't make sense theoretically, don't model it.
What I'm complaining about is when people use neither--they just test one model after another, then visualize it (maybe).
Thank you so much for this series Dustin! Could you clarify if this works in a Bayesian framework? Just curious. Thanks!
These visualization techniques do not depend on whether you take a Frequentist or Bayesian interpretation of probability.
Yes it does!
Don't worry about being too technical, if i don't get it the 1st time i will get it on the 2nd or 3rd watch. Yeah, i probably don't speak for everyone...
Good to hear! I do think I need to have more videos on at least some of the more technical topics
I can only agree with emphasizing the use of plots in model development. Plots are not perfect, but they can make a big difference.
Yep!
That was illuminating ... Thank you !
Gracias.
when quant psych makes a video, i stop and i click
me too...despite the fact that I have to go teach class in 15 minutes and this is a 25 minute vid.
At 5:02 in the third row is a mistake: 3-3.2=-.2 instead of .2.
I'm curious what can be learned from these residual plots when the likelihood of the data is non-normal.
That's one area I haven't branched into yet. But if the data are severely non-normal, they'll probably fail.
@@QuantPsych I have some suggestions if you do choose to dig into the topic while staying within your focus of linear models. Most raw data is non-negative or non-numerical and asymptotic normality cannot always be relied upon. So I would recommend exploring the topic with generalized linear models with Gamma and Poisson conditional likelihoods as they illustrate modelling cases such as duration or counts respectively. Multinomial logistic regression is a good choice for (non-hierarchical) categorical data.
lol :D Ridiculously useful and fun lol XD
Dustin is the goofiest statistician I've yet to come across (which is a compliment).