• October 31, 2014
October 31, 2014, 5:19:12 AM *
Welcome, Guest. Please login or register.

Login with your Chronicle username and password
News: Talk about how to cope with chronic illness, disability, and other health issues in the academic workplace.
 
Pages: 1 [2]
  Print  
Author Topic: Statistics suggestions  (Read 7580 times)
nomenclature
New member
*
Posts: 13


« Reply #15 on: March 10, 2012, 12:59:40 AM »

Let me check I have this right: you were planning on doing a multiple regression, but then it was suggested to you that it isn't appropriate.  You investigated further and found divided opinions, with a single other method proposed as an alternative.  If that's right, then I think:

1) It's probably not one of those things that Thou Shalt Not Do, like violating the fundamental assumptions of the regression.  You will be able to find enough support both in your department and in the literature to make the case that your model is only as imperfect as statistical models generally are, not far far worse.

2)  You've had a specific alternative method suggested.  Make a model or two using that model and your OLS regression, and compare.  How do the underlying statistics in the two models differ, and how do your results compare?

This is what I needed to hear. It is really one of those things where there are divided opinions on whether it is appropriate to use multiple regression. Actually, there were 4 other alternative methods suggested. I have been so wrapped up in finding the RIGHT way to do it. I think you are saying that as long as I can provide sufficient support for my model, I won't wander too far off the grid on what's acceptable.

Logged
nomenclature
New member
*
Posts: 13


« Reply #16 on: March 10, 2012, 1:08:43 AM »


Do you have a dichotomous outcome variable or clustered data or something? I'm surprised. If it's something like analyzing clustered data without using MLM/HLM, maybe you should go with the opinions of people in your field. Some areas just gloss over some statistical considerations while other areas are not wiling to bend the rules. I really don't understand why the stats office you consulted with is not allowed to help you figure out whether multiple regression is right given your data. Something is weird.

PM me if you want to share but don't want to broadcast the specific issue on the fora for some reason. I can't be your stats consultant, but you have piqued my interest and maybe I can google up something for you.

Laurel_knx: I sent you a PM, fellow googler :)
Logged
totoro
Friendly Neighborhood Troll and
Distinguished Senior Member
*****
Posts: 4,962


« Reply #17 on: March 10, 2012, 9:04:45 AM »

I have found that different disciplines all have their culture and conventions on what is acceptable statistical practice and what the most common methods they use are. So if you are in psychology there is little point asking an economist what to do and vice versa, for example.
Logged
laurel_knx
Senior member
****
Posts: 631


« Reply #18 on: March 11, 2012, 11:30:55 AM »



Do you have a dichotomous outcome variable or clustered data or something? I'm surprised. If it's something like analyzing clustered data without using MLM/HLM, maybe you should go with the opinions of people in your field. Some areas just gloss over some statistical considerations while other areas are not wiling to bend the rules. I really don't understand why the stats office you consulted with is not allowed to help you figure out whether multiple regression is right given your data. Something is weird.

PM me if you want to share but don't want to broadcast the specific issue on the fora for some reason. I can't be your stats consultant, but you have piqued my interest and maybe I can google up something for you.
Laurel_knx: I sent you a PM, fellow googler :)

I responded and called shenanigans on some of your "advisors".
Logged
oldfullprof
Ridiculous
Distinguished Senior Member
*****
Posts: 11,359

Representation is not reproduction!


« Reply #19 on: March 11, 2012, 3:00:01 PM »

Add a stat person to your committee, even if you don't want to.  You don't want to look ridiculous later.
I like "even if you don't want to" because you are right. I don't want to but it sounds like I have to .....

We had a masters kid who decided to use his own brand of factor analysis on his thesis.  I was the outside member and his two mentors were only qualitative.  I made him take it out.  It was crazy.  He had plenty of stuff besides the factor analysis.
Logged

Taste o' the Sixties
juillet
Senior member
****
Posts: 347


« Reply #20 on: March 15, 2012, 12:23:29 PM »

I have found that different disciplines all have their culture and conventions on what is acceptable statistical practice and what the most common methods they use are. So if you are in psychology there is little point asking an economist what to do and vice versa, for example.

Hmm, depends on what it is, I would think.  I think at a complex level there's a lot of argument both between and within fields about appropriate methods, but at the more basic and intermediate methods social scientists share a lot of the same practices.  I think an economist or a quantitative sociologist can help an intermediate psychologist with their stats and vice versa, and I do social science stats consulting had have consulted across a wide variety of fields at my university.
Logged
kron3007
Distinguished Senior Member
*****
Posts: 1,121


« Reply #21 on: March 16, 2012, 9:08:46 AM »

Why dont you look at some similar papers in your field and see what approach they used?  This way you will see which of the suggested options are common in your field.  If anyone questions your approach you can defend your position based on the precedent that has been set in your field and provide references to support your decision.  Obviously you will need to make adjustments based on your data, but this would at least point you in the best (or at least a) direction.    

You should remember that when you submit your paper, people in your field will be reviewing it.  As such, it dosnt really matter what a statistician decides is the best approach (if this distinction can even be made), it matters what people in your field use.  

Just as an example, during my MSc defense one of my examiners basically implied that there were flaws in my stats that basically made all of my analyses flawed.  While he had a point, I was able to defend my position by stating that this is the common practice in my field (with references) and that based on his assertion my entire discipline is in error and not valid.  Sometimes (especially in biology) the reality just does not fit into statisticians little boxes and there is no mathematically perfect approach.
« Last Edit: March 16, 2012, 9:14:57 AM by kron3007 » Logged
msparticularity
Distinguished Senior Member
*****
Posts: 18,502

Assistant Professor cum bricoleur


« Reply #22 on: March 16, 2012, 4:11:49 PM »

Let me check I have this right: you were planning on doing a multiple regression, but then it was suggested to you that it isn't appropriate.  You investigated further and found divided opinions, with a single other method proposed as an alternative.  If that's right, then I think:

1) It's probably not one of those things that Thou Shalt Not Do, like violating the fundamental assumptions of the regression.  You will be able to find enough support both in your department and in the literature to make the case that your model is only as imperfect as statistical models generally are, not far far worse.

2)  You've had a specific alternative method suggested.  Make a model or two using that model and your OLS regression, and compare.  How do the underlying statistics in the two models differ, and how do your results compare?

This is what I needed to hear. It is really one of those things where there are divided opinions on whether it is appropriate to use multiple regression. Actually, there were 4 other alternative methods suggested. I have been so wrapped up in finding the RIGHT way to do it. I think you are saying that as long as I can provide sufficient support for my model, I won't wander too far off the grid on what's acceptable.



Combine this line of thought with what kron3007 says: you need to provide your support in the form of recent work in your field that is reasonably similar to your own. Keep in mind, too, that being able to discuss the limitations of your approach and the range of confounds and challenges it invites will go a long way toward establishing your competence to select appropriate models. (Keep in mind that the diss is more than just a diss: it's a demonstration of your overall knowledge and skills, rather than just a narrow exercise to show you could get through this particular round of data collection and analysis.)

And yeah, there are lots and lots of Thou Shalt Nots in data analysis, and a remarkable number of them exist only in the minds of specific individuals.
« Last Edit: March 16, 2012, 4:15:30 PM by msparticularity » Logged

"Once admit that the sole verifiable or fruitful object of knowledge is the particular set of changes that generate the object of study...and no intelligible question can be asked about what, by assumption, lies outside." John Dewey

"Be particular." Jill Conner Browne
elie_s_dad2
New member
*
Posts: 27


« Reply #23 on: March 19, 2012, 8:11:33 AM »

Statisticians seem to have a bad rep among scientists based on a couple of the side comments on this thread.  Probably not surprisingly, statisticians face many frustrations while deal with scientists (social or otherwise) as well.

Obviously, I don't know anyone on the thread personally, but from my experience:

-It's not that easy to give advice to scientists who may not know enough technical language to properly frame their question, much less do the analysis themselves and interpret the results.

-As someone mentioned statistical models are all flawed.  As the statistician, I'm usually the one more likely to mention the flaws of a model because the field matter expert is inherently interested in the positive result that proves their idea.  The purpose of going back to that painful 'abstract theory' is that doing so can help you understand the flaws of your model and how to mitigate them.

-It would be frustrating for anyone to be so misunderstood by a non-technical audience that they are later told that they said something contradictory or incorrect.

-After studies, scientists tend to view 'the statistics part' as displaying a secret code that proves they were right all along.  Statisticians who tell scientists their results are ambiguous are 'bad' and statisticians (or quantitative colleagues) who confirm the scientist was right all along are 'good'.

Just my 2 cents American.
Logged
nomenclature
New member
*
Posts: 13


« Reply #24 on: March 19, 2012, 10:02:52 PM »

Quote

Combine this line of thought with what kron3007 says: you need to provide your support in the form of recent work in your field that is reasonably similar to your own. Keep in mind, too, that being able to discuss the limitations of your approach and the range of confounds and challenges it invites will go a long way toward establishing your competence to select appropriate models. (Keep in mind that the diss is more than just a diss: it's a demonstration of your overall knowledge and skills, rather than just a narrow exercise to show you could get through this particular round of data collection and analysis.)

And yeah, there are lots and lots of Thou Shalt Nots in data analysis, and a remarkable number of them exist only in the minds of specific individuals.

It is
Logged
nomenclature
New member
*
Posts: 13


« Reply #25 on: March 19, 2012, 10:14:43 PM »

Oops ... my bad for the posting above.

I was going to say that this has been a wonderful learning experience for me. I have learned that I need to let go of my previous notion that there is one right way to analyze my data. Thanks everyone.
Logged
Pages: 1 [2]
  Print  
 
Jump to:  

Powered by MySQL Powered by PHP Powered by SMF 1.1.9 | SMF © 2006-2008, Simple Machines LLC Valid XHTML 1.0! Valid CSS!
  • 1255 Twenty-Third St., N.W.
  • Washington, D.C. 20037
subscribe today

Get the insight you need for success in academe.