Hi everyone,
This isn't a STATA per se question, but more generally a statistical one (followed by a short STATA Q.) I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? If not, are there any other appropriate methods for testing associations betwen categorical outcomes and categorical exposures (both with more than 2 categories)? Also, does one specify the 'i.' in front of BMI (as described above) in ordical logistic regression in STATA to specify the different categories or is that only for the exposure variables? Thanks! /Amal. Amal Khanolkar, PhD candidate, Centre for Health Equity Studies (CHESS), Karolinska Institutet, 106 91 Stockholm. Ph# +46(0)8 162584/+46(0)73 0899409 www.chess.su.se * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
Also look at mlogit for straight categorical outcomes without ordering.
Do not put "i." in front of BMI. Stata will take care of the ordering. P On 11/11/2010 10:23 AM, Amal Khanolkar wrote: > Hi everyone, > > This isn't a STATA per se question, but more generally a statistical one (followed by a short STATA Q.) > > I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? If not, are there any other appropriate methods for testing associations betwen categorical outcomes and categorical exposures (both with more than 2 categories)? > > Also, does one specify the 'i.' in front of BMI (as described above) in ordical logistic regression in STATA to specify the different categories or is that only for the exposure variables? > > Thanks! > > /Amal. > > > Amal Khanolkar, PhD candidate, > Centre for Health Equity Studies (CHESS), > Karolinska Institutet, > 106 91 Stockholm. > > Ph# +46(0)8 162584/+46(0)73 0899409 > www.chess.su.se > * > * For searches and help try: > * http://www.stata.com/help.cgi?search > * http://www.stata.com/support/statalist/faq > * http://www.ats.ucla.edu/stat/stata/ >  E. Paul Wileyto, Ph.D. Assistant Professor of Biostatistics Tobacco Use Research Center School of Medicine, U. of Pennsylvania 3535 Market Street, Suite 4100 Philadelphia, PA 191043309 2157467147 Fax: 2157467140 [hidden email] * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
In reply to this post by Amal Khanolkar
Yes, but that strikes me as just throwing away information.
Nick [hidden email] Amal Khanolkar I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
Hi Nick,
Because STATA would create equal intervals between the 3 categories of BMI?? Would mlogit be a better choice then? /Amal. Amal Khanolkar, PhD candidate, Centre for Health Equity Studies (CHESS), Karolinska Institutet, 106 91 Stockholm. Ph# +46(0)8 162584/+46(0)73 0899409 www.chess.su.se ________________________________________ From: [hidden email] [[hidden email]] On Behalf Of Nick Cox [[hidden email]] Sent: 11 November 2010 16:28 To: '[hidden email]' Subject: st: RE: Ordinal logistic regression Yes, but that strikes me as just throwing away information. Nick [hidden email] Amal Khanolkar I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
Not at all my meaning; BMI always contains much more information than any crude categorical reduction of it.
N.B. http://www.stata.com/support/faqs/res/statalist.html#spell The fact that you have, it seems, categorical predictors is not itself a reason for categorising the response. Nick [hidden email] Amal Khanolkar Because STATA would create equal intervals between the 3 categories of BMI?? Would mlogit be a better choice then? Nick Cox Yes, but that strikes me as just throwing away information. Amal Khanolkar I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
In reply to this post by Nick Cox
I usually feel the same way about reducing information, but in some
cases the clinicallyrelevant categories are of greater interest than the continuum. Mary Ellen On 11/11/2010 9:28 AM, Nick Cox wrote: > Yes, but that strikes me as just throwing away information. > > Nick > [hidden email] > > Amal Khanolkar > > I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? > > > * > * For searches and help try: > * http://www.stata.com/help.cgi?search > * http://www.stata.com/support/statalist/faq > * http://www.ats.ucla.edu/stat/stata/ > >  Mary Ellen MackesyAmiti, Ph.D. Research Assistant Professor Community Outreach Intervention Projects (COIP) School of Public Health m/c 923 Division of Epidemiology and Biostatistics University of Illinois at Chicago ph. 3123554892 fax: 3129961450 * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
I sympathise with the idea, but that is a different issue.
If I wanted to forecast floods, I would use river discharge as a response, make quantitative predictions, and then the very last step is to see whether discharge means that the river is above some important threshold. Degrading my data to river discharge = {low, medium, high} at the outset is neither necessary nor helpful. How does obesity differ? Nick [hidden email] Mary E. MackesyAmiti I usually feel the same way about reducing information, but in some cases the clinicallyrelevant categories are of greater interest than the continuum. On 11/11/2010 9:28 AM, Nick Cox wrote: > Yes, but that strikes me as just throwing away information. Amal Khanolkar > I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
An interesting discussion seems to be unfolding... :)
But to clear matters from my end: I've already checked associatiosn with continous BMI using linear regression. I just wanted to cross check associations with clinically important BMI cutoffs: overweight and obesity. And if mlogit ot ologit was the right choice or if something else suits this categoricalcategorical associations. /Amal. Amal Khanolkar, PhD candidate, Centre for Health Equity Studies (CHESS), Karolinska Institutet, 106 91 Stockholm. Ph# +46(0)8 162584/+46(0)73 0899409 www.chess.su.se ________________________________________ From: [hidden email] [[hidden email]] On Behalf Of Nick Cox [[hidden email]] Sent: 11 November 2010 17:20 To: '[hidden email]' Subject: RE: st: RE: Ordinal logistic regression I sympathise with the idea, but that is a different issue. If I wanted to forecast floods, I would use river discharge as a response, make quantitative predictions, and then the very last step is to see whether discharge means that the river is above some important threshold. Degrading my data to river discharge = {low, medium, high} at the outset is neither necessary nor helpful. How does obesity differ? Nick [hidden email] Mary E. MackesyAmiti I usually feel the same way about reducing information, but in some cases the clinicallyrelevant categories are of greater interest than the continuum. On 11/11/2010 9:28 AM, Nick Cox wrote: > Yes, but that strikes me as just throwing away information. Amal Khanolkar > I would like to know if BMI categorised into normal, overweight and obese could be considered as ordinal data and if so if be used as the outcome in 'ordinal logistic regression' with categorical exposures? * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
Administrator

In reply to this post by mmamiti
On Thu, Nov 11, 2010 at 4:13 PM, Mary E. MackesyAmiti <[hidden email]> wrote:
> I usually feel the same way about reducing information, but in some cases > the clinicallyrelevant categories are of greater interest than the > continuum. Its completely arbitrary though. Besides which BMI isn't a robust indicator of obesity as it doesn't work for people who are very fit and have lots of well honed muscles (their BMI often puts them in the "obese" category when they are anything but). Plenty of information on why not to categorise continuous variables at http://biostat.mc.vanderbilt.edu/twiki/bin/view/Main/CatContinuous On Thu, Nov 11, 2010 at 4:20 PM, Nick Cox <[hidden email]> wrote: > How does obesity differ? It doesn't, but clinicians seem to struggle with these concepts. Neil  "Our civilization would be pitifully immature without the intellectual revolution led by Darwin"  Motoo Kimura, The Neutral Theory of Molecular Evolution Email  [hidden email] Website  http://kimuranoip.org/ Photos  http://www.flickr.com/photos/slackline/ * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
In reply to this post by Amal Khanolkar
You can do that too, by producing your coarsened variables _after_ modelling, not _before_.
Nick [hidden email] Amal Khanolkar An interesting discussion seems to be unfolding... :) But to clear matters from my end: I've already checked associatiosn with continous BMI using linear regression. I just wanted to cross check associations with clinically important BMI cutoffs: overweight and obesity. And if mlogit ot ologit was the right choice or if something else suits this categoricalcategorical associations. * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
In reply to this post by nshephard
It may be completely arbitrary, or there may be some standards in the field that say cutpoints are at 30 and 35. However, the original poster should realize that the category indicators are not considered as codes with equal spacing and that ologit takes that into account.
Another issue is that you can use a glm to select the distribution you want rather than forcing a linear regression. There may be some information that suggests a proper model in your field of research. Tony Peter A. Lachenbruch Department of Public Health Oregon State University Corvallis, OR 97330 Phone: 5417373832 FAX: 5417374001 Original Message From: [hidden email] [mailto:[hidden email]] On Behalf Of Neil Shephard Sent: Thursday, November 11, 2010 8:29 AM To: [hidden email] Subject: Re: st: RE: Ordinal logistic regression On Thu, Nov 11, 2010 at 4:13 PM, Mary E. MackesyAmiti <[hidden email]> wrote: > I usually feel the same way about reducing information, but in some cases > the clinicallyrelevant categories are of greater interest than the > continuum. Its completely arbitrary though. Besides which BMI isn't a robust indicator of obesity as it doesn't work for people who are very fit and have lots of well honed muscles (their BMI often puts them in the "obese" category when they are anything but). Plenty of information on why not to categorise continuous variables at http://biostat.mc.vanderbilt.edu/twiki/bin/view/Main/CatContinuous On Thu, Nov 11, 2010 at 4:20 PM, Nick Cox <[hidden email]> wrote: > How does obesity differ? It doesn't, but clinicians seem to struggle with these concepts. Neil  "Our civilization would be pitifully immature without the intellectual revolution led by Darwin"  Motoo Kimura, The Neutral Theory of Molecular Evolution Email  [hidden email] Website  http://kimuranoip.org/ Photos  http://www.flickr.com/photos/slackline/ * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
In reply to this post by Amal Khanolkar
Dear all,
I tried out the links Neil suggested. As expected, dichotomizing generally lead to a loss of power. However, when it did not, this was due to outliers (in y and x). AS BMI is susceptible to occasional genuine extreme outliers, there is some sort of argument for dichotomizing. Another argument would be the provision of results in a familiar form for clinicians, even at the expense of loss of power. (There are WHO guidelines for BMI cutoffs, which are in clinical use). False outliers in BMI, due to confusing pounds & kg or inches and cm are another matter... BW Paul Seed From: [hidden email] [mailto:[hidden email]] On Behalf Of Neil Shephard Sent: Thursday, November 11, 2010 8:29 AM To: [hidden email] Subject: Re: st: RE: Ordinal logistic regression On Thu, Nov 11, 2010 at 4:13 PM, Mary E. MackesyAmiti <[hidden email]> wrote: > I usually feel the same way about reducing information, but in some cases > the clinicallyrelevant categories are of greater interest than the > continuum. Its completely arbitrary though. Besides which BMI isn't a robust indicator of obesity as it doesn't work for people who are very fit and have lots of well honed muscles (their BMI often puts them in the "obese" category when they are anything but). Plenty of information on why not to categorise continuous variables at http://biostat.mc.vanderbilt.edu/twiki/bin/view/Main/CatContinuous On Thu, Nov 11, 2010 at 4:20 PM, Nick Cox <[hidden email]> wrote: > How does obesity differ? It doesn't, but clinicians seem to struggle with these concepts. Neil   "Our civilization would be pitifully immature without the intellectual revolution led by Darwin"  Motoo Kimura, The Neutral Theory of Molecular Evolution * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
Confusion about pounds and kg or inches and cm gives rise to errors about 2 fold, which should often be obvious on careful checking.
If worried about outliers, you could always work on a transformed scale (meaning, transform the variable or use a nonidentity link function). Logarithms seem natural for a positive number that is really a ratio. Nick [hidden email] Seed, Paul I tried out the links Neil suggested. As expected, dichotomizing generally lead to a loss of power. However, when it did not, this was due to outliers (in y and x). AS BMI is susceptible to occasional genuine extreme outliers, there is some sort of argument for dichotomizing. [...] False outliers in BMI, due to confusing pounds & kg or inches and cm are another matter... * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/ 
Free forum by Nabble  Edit this page 