Measuring change in child weight status in relation to local environmental factors

Photo by Stas Kulesh

Travis’ Note: Today’s post comes from friend and colleague Megan Carter. Megan has recently submitted her PhD thesis, so please join with me in congratulating her and wishing her the best of luck in her upcoming defense!  More on Megan can be found at the bottom of this post.  You can find out how to submit your own guest post to Obesity Panacea here.

Today’s post is great thesis defense prep for me. I’ve deposited the “behemoth,” it’s now in the hands of my examiners, for all to critique.  It’s scary, but I’m sort of looking forward to it. I’ll be able to close a chapter of my life and open another. One of my research areas of interest is in how local environments may influence childhood obesity.  A portion of my thesis allowed me to explore this area, and exercise my epidemiological and statistical skills, especially since I got to work with a large, population-based cohort study. I’d like to briefly share this chunk of my thesis now that the associated papers are published.

I worked with the Québec Longitudinal Study of Children, a large cohort study (n=2120) that has followed children from 5 months of age (in 1997/98) to present (conducted and funded primarily by the Québec Government).  Measured height and weight data were collected at 4, 6, 7, 8, and 10 years of age, along with a range of behavioural, environmental, social, and academic measures.  Using this cohort to conduct my study was exciting to me because collection of multiple measured heights and weights over a significant amount time does not usually happen on such a large scale in Canada (see problems associated with parent-reported heights and weights here on page 30).

Also available were measures of the local environment, and potential confounders such as socioeconomic status (SES) – many of which were measured over time as well – another thing that does not usually happen.  The local environmental factors were intriguing to me (see table below and references at the end).  I was interested in seeing how these could influence the change in child weight status over time while also controlling for things like SES, and early life factors such as birthweight and early infant growth.

Table 1 from Carter et al., 2012 (PLOS ONE). Click to view full size.

In order to determine if children were getting heavier over time (from 4-10 y), I used a standardized body mass index measure; this uses an external population of comparison (in this case, the reference developed by the Centers for Disease Control) as the outcome. Change in BMI by itself isn’t a great measure among children because an increase in BMI is a part of normal growth and development, and “normal” changes in BMI occur differently at different ages and between the sexes – you need to account for this.

Without getting into too much of the statistical nitty gritty, I decided to take two approaches, as they provide two different but complementary ways of looking at the data, which enhances my (and hopefully others’) understanding of the problem :

  1. Random effects modeling is one of the most common methods of accounting for within person correlation among repeated outcome measures, for estimating trends (weight trajectory), and estimating explanatory variables’ potential impacts on trends. It assumes that on average, all children follow the same trajectory of weight change over time (it does allow individual children to have their own distinct trajectories but I think that’s more of a statistical discussion for another day). I’ll refer to this as the single growth curve method.
  2. Group-based trajectory modeling, on the other hand, does not assume children all follow the same trajectory, but allows for multiple mean sub trajectories of growth to be estimated in the sample.  An algorithm groups children together who have similar patterns of BMI Z-score responses over time.  Thus, each group can have its own distinct trajectory shape and prevalence in the sample.  Explanatory variables that do not occur over time, but before the trajectory period, can be included in the model to determine if they can explain why some children belong to one type of trajectory group and not another.  Explanatory variables that are measured over time can be included in the model to determine how they might impact change in weight status within a particularly trajectory group.  This method cannot estimate groups that actually exist in the population; rather these sub trajectories are like lines on a topographical map – meant to guide us toward a better understanding of the etiological underpinnings of certain phenomena. This approach is being used more and more in the obesity literature. I’ll refer to this as the multiple growth curve method.

Using the single growth curve method, children on average were estimated to follow a slight accelerating curve (trend was u-shaped) – meaning BMI Z-scores were decreasing slightly towards zero (mean of the CDC population) early on from 4-10 y, and then were increasing slightly above zero towards the latter part of the six year period.  In the multiple growth curve model, four groups were estimated (figure below), where one group was estimated to have a high, stable weight trajectory (considered obese by CDC standards).  The two groups in the middle did not change much over time, but did exhibit increases, and the last group started off at a very low BMI Z-score and experienced the sharpest rise over time.

Figure 1 from Carter et al., 2012 (PLOS ONE). Click image to view full-size.

When all was said and done, the only consistent finding that emerged using the two methods was a potential relationship between living location and weight status, although the nature of the relationship was far from clear. For example, an inverse relationship was found between living in a census agglomeration (an area with a medium population size) and BMI Z-score, as compared to living in a census metropolitan area (an area with a large population size) in the single growth curve analysis. Judging from the multiple growth curve results, this association may depend on the trajectory group (e.g. the association was seen in two of the four groups, not all groups).  The multiple growth curve method also suggested that living in a rural area may be associated with weight gain among children in the high, stable end of the BMI Z-score distribution.

What was notably consistent between the two methods was that early life factors (e.g. obesity status of the mother at 1.5 y, smoking during pregnancy) were strongly associated with weight gain in the single growth curve method, and in the multiple growth curve method, strongly predicted membership in the high, stable trajectory group.

There are a few things I’ve learned from this part of my project:

  1. There is never one true model or one best statistical method.
  2. The local environment is complex with many interacting elements.  There will always be a debate about how to best measure it.
  3. Adding longitudinal measurement adds another layer of complexity – for modeling and understanding what results mean.
  4. Early life factors were very salient risk factors for the development of early excess weight – this represents a window of opportunity for intervening. Colleagues in HALO have been investigating this very time frame.
  5. An area for research development => investigating if early life factors mediate effects of local environmental factors during pregnancy. For example, perhaps mom’s exposure to certain environmental factors may make her unborn baby more likely to be born at high birthweight and then go on to develop obesity. This area of research is beginning to develop with respect to endocrine disruptors, but there may be other relevant factors such as area SES.
  6. An appreciation for the difference between “lack of evidence of effect” and “evidence of lack of effect.” Similar to the fact that we cannot say definitively that A causes B based on the results of any one study, we cannot say that just because we did not find a relationship in our study, that none exists. There are lots of reasons for not finding an association in a study (especially an observational one) that I’ll save for a more statistically minded blog post.

About the author: Megan Carter is a PhD candidate at the University of Ottawa in the Population Health Doctoral program. Her research interest is in social epidemiology – particularly with respect to how features of the physical and social environments may influence childhood obesity development and household food insecurity. You can follow her at or on Twitter @megpophealth

Related Posts Plugin for WordPress, Blogger...
This entry was posted in Guest Post, Obesity Research, Peer Reviewed Research and tagged . Bookmark the permalink.

One Response to Measuring change in child weight status in relation to local environmental factors

  1. BISE says:

    I like your texts very much. That is why I would to use them in my article, if it is possible. I am interesting in that topic, and I need you help. Please, say YES. Thank You.