Fleeting/Random Thoughts

Dason

Ambassador to the humans
#1
Just figured we could do with a fleeting/random thoughts thread. If you've got something on your mind but don't think it deserves an entire thread this is the place to post!

Please don't use this thread to post homework questions - that's essentially what all of the statistics board is for.

As for me: I just wish I knew exactly what was going on with my summer funding. I've been told that I'm guaranteed summer funding but I don't know if it'll be by doing consulting for the Ag department or if I'll be getting an early start on the RA I'm starting in the fall.
 
Last edited:

CowboyBear

Super Moderator
#2
As for me: I just wish I knew exactly what was going on with my summer funding. I've been told that I'm guaranteed summer funding but I don't know if it'll be by doing consulting for the Ag department or if I'll be getting an early start on the RA I'm starting in the fall.
Nice idea for a thread. Which of the two would you rather be doing?

My fleeting thoughts:
1. What assumptions do we make about the distribution of residuals in a negative binomial regression? (surely they're not integers, so not negative binomially distributed themselves?)
2. Why are peer reviewers in psych more interested in pushing their pet theories than, I don't know, actually making useful comments about data analysis?
3. There's a conference coming up in a nice snowy place, and I'd like to go, but I can't quite decide which topic to try and present about.

Also, happy easter all!
 

spunky

Doesn't actually exist
#3
1. What assumptions do we make about the distribution of residuals in a negative binomial regression? (surely they're not integers, so not negative binomially distributed themselves?)
i dunno whether you've already had looked into this book or not, but i believe joseph hilbe wrote something specifically on negative binomial regression, right? i would like to think he either provides some guidance about how the residuals should look like or at least a few references... he's always such a pleasure to read!! (i've never done a negative binomial regression myself as i suffer from poissonitis and prefer to model my count data as poisson RVs whenever possible, lolz)

2. Why are peer reviewers in psych more interested in pushing their pet theories than, I don't know, actually making useful comments about data analysis?
oh! i like TOTALLY hear you there... i think we're both in very similar areas, right? (i'm in educational measurement/psychometrics) and i do get this feeling that people in our areas just love to jump over the data analysis and engage in bizarre philosophising while people like us kind of tap on their shoulders and say "well, yes but the data says... excuse me, sir/madam, the DATA... THE DATAAAA!" [/QUOTE]

3. There's a conference coming up in a nice snowy place, and I'd like to go, but I can't quite decide which topic to try and present about.
from what i read on your posts, you seem to be very porficient in structural equation modeling. what about something on latent grow modeling? i'm not sure what's hot in stats for psychology nowadays but in the field of educational measurement, oh boy, if you're into hierarchical linear models/growth curve modeling then you're part of the IN crowd... and if you do both at the same time they worship you...

Also, happy easter all!
happy easter to you too!! see? since you taught me the "/quote trick" i cant stop using it nowwww..

feelting thought of mine:
there's this new stats fad in the U.S. called value added models which people (read politicians) are using to measure how effective teachers are. it's HIGHLY controversial... to the point that a teacher in LA committed suicide after being ranked as "ineffective". i guess it's because, if the people behind this thing have their way, results from these models could be used to fire people, grant promotions, accommodate funding for schools, etc... now if the Obama administration through its "Race to the Top" program endorses these models, i can only wonder what a republican administration would do with them (these models were developed by economists so you get the idea).

there's been talk here in british columbia, canada to bring the models up north and since i am working very, very hard in becoming an expert in both hierarchical and latent growth models i didnt know until the summer holidays started that i was a de-facto member of the team who's brining them up here... which makes me wonder... does one say yes? no? maybe? there's lots of pluses to this (the funding $$$ is so good i could probably quit my stats consulting job and only do research, i'd pretty much come out of this experience with a dissertation already written, etc.) but now i read about people committing suicide and stuff and i'm like... "darn... what do responsible/ethical researchers do in these cases?"
 

Dason

Ambassador to the humans
#4
Nice idea for a thread. Which of the two would you rather be doing?
I'd probably prefer to start on the RA. It'd be nice to get some work done without having to do classwork as well. I've also had a bad crop of clients lately so consulting was getting a little grating...

1. What assumptions do we make about the distribution of residuals in a negative binomial regression? (surely they're not integers, so not negative binomially distributed themselves?)
Note that I've never actually performed a negative binomial regression but my understanding is that you're just talking about a generalized linear model using a negative binomial family and some sort of link.

In most of the cases with generalized linear models we can create standardized residuals. We typically don't place assumptions on the residuals themselves (as opposed to the case in general linear models). We're just putting an assumption on the distribution of the response variable conditioned on it's expected value. We also put assumptions on the form we expect the expectation function to take. So really the distribution we expect the residuals to take in a negative binomial regression is just a shifted negative binomial distribution at each x so that the mean is 0.

What is done sometimes is to standardize the residuals. One way to do that is to say that we have a guess as to what the distribution at each x value should look like - so if we plug the observed value into the cdf of our expected distribution then things should turn out looking approximately uniform. This isn't as clear cut in the case of discrete outcomes (and it doesn't really work for bernoulli outcomes) as it is for continuous variables because the inverse cdf transform isn't exactly uniformly distributed but it does still create some sort of standardized residual that should be approximately uniformly distributed.

I'm just spitting out what I remember from our section on model checking in my advanced methods course and I don't have access to my notes right now so I could have spewed out some stuff thats wrong. But it seems to make sense to me.
 
Last edited:

CowboyBear

Super Moderator
#5
i dunno whether you've already had looked into this book or not, but i believe joseph hilbe wrote something specifically on negative binomial regression, right? i would like to think he either provides some guidance about how the residuals should look like or at least a few references... he's always such a pleasure to read!! (i've never done a negative binomial regression myself as i suffer from poissonitis and prefer to model my count data as poisson RVs whenever possible, lolz)
I actually picked up his book the other day! Interesting that you've found him a pleasure to read in the past - I haven't read any of his other stuff, but I've found the negbin book rather hard going :( He seems to assume a fairly sturdy background in mathematical stats and calculus, which is a pity because negative binomial regression is a solution (or set of solutions) to a really practical applied research problem (overdispersed count data). The book could potentially be useful to anyone from dentists to criminologists and lots of other researchers not particularly expert in stats, but he seems to have decided to pitch it at a level that cuts out a lot of the potential market. I might need to give it another go though!

oh! i like TOTALLY hear you there... i think we're both in very similar areas, right? (i'm in educational measurement/psychometrics) and i do get this feeling that people in our areas just love to jump over the data analysis and engage in bizarre philosophising while people like us kind of tap on their shoulders and say "well, yes but the data says... excuse me, sir/madam, the DATA... THE DATAAAA!"
Lol, that sounds very familar! My background's pretty mixed actually - fair bit of clinical training, but research has been a mix of health psych, psychometrics and a bit of criminology. And now heading more towards the environmental psych side! I wonder sometimes whether journals should make a policy that every article needs to be reviewed by at least one individual with expertise in the research methods being used (e.g. a quantitative psychologist, a statistician, or what have you).

from what i read on your posts, you seem to be very porficient in structural equation modeling. what about something on latent grow modeling? i'm not sure what's hot in stats for psychology nowadays but in the field of educational measurement, oh boy, if you're into hierarchical linear models/growth curve modeling then you're part of the IN crowd... and if you do both at the same time they worship you...
Heh, I probably don't know enough about latent growth modelling to bend the ear of anyone else, but HLM is definitely the cool kid around here too!

feelting thought of mine:
there's this new stats fad in the U.S. called value added models which people (read politicians) are using to measure how effective teachers are. it's HIGHLY controversial... to the point that a teacher in LA committed suicide after being ranked as "ineffective". i guess it's because, if the people behind this thing have their way, results from these models could be used to fire people, grant promotions, accommodate funding for schools, etc... now if the Obama administration through its "Race to the Top" program endorses these models, i can only wonder what a republican administration would do with them (these models were developed by economists so you get the idea).

there's been talk here in british columbia, canada to bring the models up north and since i am working very, very hard in becoming an expert in both hierarchical and latent growth models i didnt know until the summer holidays started that i was a de-facto member of the team who's brining them up here... which makes me wonder... does one say yes? no? maybe? there's lots of pluses to this (the funding $$$ is so good i could probably quit my stats consulting job and only do research, i'd pretty much come out of this experience with a dissertation already written, etc.) but now i read about people committing suicide and stuff and i'm like... "darn... what do responsible/ethical researchers do in these cases?"
Wow, that's interesting - I hadn't heard of these value added models. I wonder if one avenue could be to get involved in the work, but to start considering and writing about the ethical dimensions of this type of modelling? E.g. asking questions like, do these models provide sufficient certainty to justify making decisions that have major consequences for people's careers or lives? Could be quite an interesting area to become an expert in!
 

CowboyBear

Super Moderator
#6
Dason said:
Note that I've never actually performed a negative binomial regression but my understanding is that you're just talking about a generalized linear model using a negative binomial family and some sort of link.
Yep, I was thinking of the case of a log link (which seems to be usual).

In most of the cases with generalized linear models we can create standardized residuals. We typically don't place assumptions on the residuals themselves (as opposed to the case in general linear models). We're just putting an assumption on the distribution of the response variable conditioned on it's expected value.
Ah, that makes a lot of sense. Thanks!
 

Dason

Ambassador to the humans
#9
I had a couple of ideas for StatisticsPedia articles that I was going to write and when I finally got some time to sit down and write them today I had no motivation. Sadness.

I'm not sure if I mentioned this before but I got married about a month ago. I bring it up because today we finally got around to getting her name legally changed. But that's about all I feel like I got accomplished today.
 

Dason

Ambassador to the humans
#11
Well I wanted to edit my first post because I feel like I need to elaborate on a few things and why things need to be done a certain way. I also wanted to make a few articles on some simple things that we get lots of questions on (like figuring out probabilities using z-tables and such). Maybe one on generalized linear models through a concrete example.
 

Dragan

Super Moderator
#12
I had a couple of ideas for StatisticsPedia articles that I was going to write and when I finally got some time to sit down and write them today I had no motivation. Sadness.
Well, don't be sad. Publishing research takes time and a lot of patience.

I don't usually do this, but to generate some motivation here's an article I recently published in Journal of Probability and Statistics.

http://www.hindawi.com/journals/jps/2011/497463/

You can download the pdf. And, I would also point out that if you look at Appendix A, BGM helped me derive the values of delta 4 and delta 5 in a previous post here on Talk Stats. In short, Talk Stats helps.

I hope this provides some motivation.
 

Link

Ninja say what!?!
#13
I had a couple of ideas for StatisticsPedia articles that I was going to write and when I finally got some time to sit down and write them today I had no motivation. Sadness.
LOL. I think that definitely describes me. I'm just hoping that some people start the articles so that I can bum off their momentum.

I'm not sure if I mentioned this before but I got married about a month ago. I bring it up because today we finally got around to getting her name legally changed. But that's about all I feel like I got accomplished today.
Wow. Congratulations man! On a related note, I'm getting ready to propose to the love of my life. Wish me luck!
 

Dason

Ambassador to the humans
#14
Wow. Congratulations man! On a related note, I'm getting ready to propose to the love of my life. Wish me luck!
!!! Best of luck. She'll say yes with probability 1. Just make sure you don't propose in a set of measure zero. Seriously though that's awesome.
 

CowboyBear

Super Moderator
#16
I'm not sure if I mentioned this before but I got married about a month ago. I bring it up because today we finally got around to getting her name legally changed. But that's about all I feel like I got accomplished today.
Whoa, I totally missed this bit!! Congrats =)

And good luck Link - I'm sure you'll get a yes! Romance is in the air at talkstats...
 

Lazar

Phineas Packard
#17
Congrats all!

Anywho in relation to my random thought, we were lucky to have Steve West at our institute this week (one of the Authors on the Cohen Cohen Akin and West Stats bible). His visit got me thinking, do we put to much emphasis on bias reduction at the cost of considering efficiency (se inflation) or power? Two examples discussed in our lab (the second from the presentation by Prof West):
1. We have been pushing latent aggregation methods for multi-level models which are excellent at reducing bias but do considerable increase standard errors.
2. The use of latent variables for interactions appears to reduce bias but often kills statistical power.
 

spunky

Doesn't actually exist
#19
Anywho in relation to my random thought, we were lucky to have Steve West at our institute this week (one of the Authors on the Cohen Cohen Akin and West Stats bible).
well, isnt that exciting? i think i'm gonna be buried with my Cohen, Cohen, Akin & West... it's all torn and full of annotations all around the margins but heck, it's always been good to me. i think i'd have asked him to autograph him for me, since Dr. Cohen's no longer with us... :(

His visit got me thinking, do we put to much emphasis on bias reduction at the cost of considering efficiency (se inflation) or power? Two examples discussed in our lab (the second from the presentation by Prof West):
1. We have been pushing latent aggregation methods for multi-level models which are excellent at reducing bias but do considerable increase standard errors.
2. The use of latent variables for interactions appears to reduce bias but often kills statistical power.
meh, i've always had a little bit of a love-hate relationship with power analysis (or power in general, perhaps because say 90% of my undergrad education in stats was strictly fisherian) but to me, i've always thought that a long as you control for type-1 error, you should be ok... if there REALLY is something going on in your data, you should be able to catch it. and i think the whole process of "catching" what's going on with your data should be an exercise on research design and the logic substantiating the theory where you're coming from, rather than using the structure of hypothesis testing to get some assurance of significance... although, i guess for most people it's always easier to rely on a few clicks here and there to get sample sizes rather than doing the job of sitting down and planning a well-designed experiment...
 

Dason

Ambassador to the humans
#20
But isn't planning a usable sample size part of sitting down and planning a well-designed experiment? I think it's emphasized too much as the only thing that matters because there are definitely aspects of a study designed that don't get considered nearly as much as they should but sample sizes still plays an important part don't you think?