categorising height and bmi variables using panel data

[Jordan Driffill]

I am looking to see the effect of height, BMI and marriage on wages using the NCDS. I am sorting out each wave of data getting the relevant variables and creating dummies etc. I need to categorise the height and BMI into tall/not tall, obese/not obese. As I am taking height as time invariant, given that the first height was recorded when individuals were 23, I know I can categorise these for each wave as I go. However with BMI, I don't really understand how regression will work as it is not necessarily true if individual 1 is obese in the first period that they will be obese in the second, third... etc I realise I don't need to know the ins and outs of how the regression on my data will work, but I was just wondering if anyone could clear this up. Do I still categorise each individuals weight in each period? I am appending my data so it is in long format and hence naming my variables the same, will the categories just work like a dummy variable for each period with a 1 if obese in period one and a zero otherwise (and so on)?

Any help would be appreciated,
thanks

[Nick Cox]

I stopped at "I need to categorise the height and BMI into tall/not tall, obese/not obese": why do you think this is in any sense a good idea? There is no obvious reason to discard most of the information in those variables.

[Jordan Driffill]

Because the aim of my study is primarily to find the effect of beauty on wages and as there is no data in the NCDS on ratings of beauty, I am approximating it by grouping height and weight classes into tall/not tall and obese/not obese. After I have done this I can have an attractive group - those who are tall and not obese, and an unattractive group - those who are not tall and obese. Again, any help would be appreciated.

[Nick Cox]

The objective does not to me imply that you should throw away information. I don't know how height or BMI relates to beauty but I am pretty clear that there are no universal thresholds that could possibly justify binary classifications. Personally, I won't willingly provide code for analysis ideas that seem statistically dubious or worse. It's a hazard in any forum that unwelcome advice might seem not helpful, or even unhelpful.

[Jordan Driffill]

Your reply was unnecessarily rude in my opinion, I don't care if you can't comprehend how height and BMI relate to beauty, you are not alone in that. I am just trying to categorise individuals by height so I can complete my dissertation. I don't have time to justify myself to you. I didn't know the point of this forum was for my work to be ridiculed and undermined as a result of asking a simple (badly worded) question.

[Clyde Schechter]

Jordan,

Nick's comment was blunt, but I wouldn't characterize it as rude. He's trying to keep you from going down a bad analytic path. I said more or less the same thing in, I think, a more light-hearted tone, in a response a few moments ago to another post of yours that asked only about height.

Now, in the case of BMI, there is, of course, the standard medical classification of < 17 as underweight, 17-25 as normal, 25-30 as overweight, 30+ as obese (units of kg/m^2). As an epidemiologist, I can tell you that even within medical research this classification is of little value and analyzing the BMI data as a continuous variable almost always results in clearer and more valid findings. The one exception to that is that there is a refinement to this classification that distinguishes grades of obesity with cutpoints at 35 and 40 kg/m^2, and there is some suggestion that the benefits of surgical obesity treatments are discontinuously greater in the highest group. The medical profession also categorizes continuous variables with abandon because their entire payment system is based on discrete diagnostic categories--but that doesn't make it useful to do so when you're trying to understand how the world works in any other respect.

But you are looking at something completely different. You are looking at the effect of anthropometrics on wages. That is almost certainly going to be mediated primarily through other people's perceptions of a person's stature and body-shape. Different people will have different cutoffs for what they perceive as tall vs not-tall. And those will be applied inconsistently because in most workplace situations, nobody actually knows your measured height and weight. As for obesity, two people with the same BMI be perceived rather differently by the same observer given that there are different distributions of body fat, and also given that clothing may emphasize or obscure those. So even if there were some true discrete difference between people above and below some height cut-off, or above and below some BMI cut-off, what you will observe in terms of effects on wages will be very much blurred out from there. For that reason, again, you are much better off keeping your data in continuous form. If you were to use, say, the standard definition of obesity as beginning with BMI = 30, you are saying that a person with BMI 30.1 is radically different from a person with BMI 29.9, but is the same as somebody with a BMI of 65. That is clearly nonsensical and you can only defeat your attempt to find anthropometric-wage relationships by doing this.

Please, please, do not go down this awful road.

[William Lisowski]

Jordan - For many of us, proceeding down the path you suggest would ultimately raise problems during the review process, when these same issues raised by Nick and Clyde would be raised by statistically rigorous reviewers, and we'd be sent back to re-do our analyses in a way that meets the norms for statistical analysis.

But let me try a different expression of what I think we are advising. Many of us deal with concepts that cannot be measured directly. In the US, we don't know how "intelligent" an individual first-year college student is. But we may know their scores on the admissions tests they've taken in the application process. We don't try to use those test results to define "bright" versus "dim". We instead include those scores directly, and in our writing justify their inclusion as measures of intelligence. Similarly, our suggestion for your analysis is to use height and BMI directly in your regressions, rather than oversimplify the measures. You don't need to define "beautiful" to measure the effects of beauty.

In your initial post back in late March, you concluded by saying you would investigate the possibility of getting help from a statistician on dealing with missing values. Given the challenging quantitative analysis you are attempting for your dissertation, I recommend that, if your committee is not already fully populated, you seek someone in your field with a strong background in statistics, or someone in statistics who has worked in your field, as a member of your committee, to help guide your statistical choices, rather than leave it to what you can learn from Statalist participants.

[Clyde Schechter]

Comment directed at the entire Forum.

In principle I endorse what William says in #7. It is excellent advice. I would like to add, though, that I have given similar advice to some Forum members using the message board or ordinary e-mail and the conversations that result are eye-opening. Even at what we might call top-flight universities, professional level statistical consultation is often unavailable to students working on master's theses or doctoral dissertations, or available only if you make such a fuss that you alienate your department. In my opinion this is nothing short of scandalous! What on earth do people pay tuition for if not that? I can certainly understand business enterprises not providing training to employees. It might be short-sighted, but, then again, probably many employees claimed to have those skills when they took their jobs. But teaching and training are core missions of a university. How can they justify not making these resources available to those who generally are in greatest need of them?

I realize this post is a bit off-topic for this Forum and I apologize for raising it. But this has come up quite often, and it leaves me extremely frustrated.

[William Lisowski]

I appreciate Clyde's insight. My dissertation committee, in an earlier millennium, had three members from three different disciplines, and I had no shortage of advice from each of them (not always consistent advice, mind you, but easier to deal with than no advice).

I guess my recommendation would be that anyone proposing a quantitative dissertation should seek one committee member, or access to someone in the program, who does similarly quantitative work in the same field to ensure that the work done for the dissertation meets the expectations of others in the same field doing similarly quantitative work. Lacking that, you risk the committee after-the-fact asking an outside party to review the work that the committee are unequipped to assess. The committee member doesn't need to write your Stata code for you, but does need to be able to critically assess the quantitative literature in the field.

Jordan, please understand that all our advice is meant to be constructive, but that we all have different ways of expressing it. We wouldn't be doing this if we didn't want to help. Or if we really wanted to troll, we wouldn't be in a "real names" forum. Sometimes the best help we can give is helping the questioner reframe the question.

[Steve Samuels]

By "NCDS" I assume that you mean the UK National Child Development Survey (e.g.http://discover.ukdataservice.ac.uk/series/?sn=2000032). I'm not familiar with this study , but I did a search for "waist circumference" and "arm circumference" on that page and turned up a number of hits. If these variables, especially the waist measurement, are available, then you have a chance of creating a defensible classification of body type. I'd be surprised if someone hasn't already created such a metric, given the widely acknowledged notion of "pear shape", for example.

Like the other posters on this page, I've attended many doctoral defenses.
It's far better that you hear criticisms here first. On a whim, I googled "effects of beauty on wages". There is indeed a large literature, and I'm sure that you are well-acquainted with it. Some studies indeed look at the effects of BMI alone, but it is clear from review articles that other measures are preferred.

[Nick Cox]

Clyde, William and Steve have all expanded in detail on the kind of view that lay behind my brief comments. As their posts showed, the questions I raised will occur to many experienced researchers, including those who might examine or grade a dissertation.

Jordan: I am sorry that you were upset by my comments, but I think you are misreading my intent. Far from trying to undermine your research, I am worried on your behalf that it can't stand well on its own two feet. I have supervised and examined hundreds of dissertations at Bachelor's level and about 30 Ph.D. theses and it's a lot more satisfying to me to try to divert people from poor ideas before they write them up than to give them a low grade.

[William Lisowski]

Nick's post suggests that I may have misunderstood the level at which Jordan is working, due to cultural and/or temporal differences. 40 years ago in the U.S., "dissertation" (the term Jordan used in his first post to Statalist) was reserved for original research done as part of the requirement for a Ph.D., and it was expected to be of publishable quality. At the Master's and Bachelor's level, a "thesis" was written. Nick uses these terms differently. If Jordan's dissertation is not in support of a doctorate then I held it, and the dissertation process he is part of, to an inappropriate standard.