1 Elisabeth Loder, MD, MPH BMJ Editorial Team [email protected]Dear Prof. Loder, We appreciate the comments provided by the reviewers of the BMJ for our manuscript, “Trajectory of body shape in early and middle life and all-cause and cause-specific mortality: results from two prospective US cohort studies”. The comments were quite useful and we thank you for the opportunity to submit the paper for your continued consideration after revisions. The revised manuscript with “track change” of the edits made from the previous article and the specific responses to referees’ comments have been uploaded in accordance with your specifications. Thank you in advance for further consideration of our manuscript. Please feel free to contact me with any questions. On behalf of my colleagues, I hope that this revised manuscript will now be acceptable for publication in the BMJ. Sincerely yours, Mingyang Song, M.D., Sc.D. Clinical and Translational Epidemiology Unit and Division of Gastroenterology, Massachusetts General Hospital and Harvard Medical School, 55 Fruit St, Boston, MA 02114 [email protected]
24
Embed
Elisabeth Loder, MD, MPH BMJ Editorial Team Dear Prof. Loder, · 2016. 5. 5. · 1 Elisabeth Loder, MD, MPH BMJ Editorial Team [email protected] Dear Prof. Loder, We appreciate the comments
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
We have updated the manuscript and all the tables and figures in the revised version using the
new method. All the numbers referred in this response letter are from the updated results. We
have also updated the method description for rescaling of the BMI at age 50 in the
Supplementary Materials, as shown below.
Supplementary Materials
Rescaling of BMI at age 50 years
To minimize random variation, we assessed the average BMI from age 37 to
43 to represent the BMI for age 40, and the average BMI from age 47 to 53 as
the BMI for age 50. We then rescaled the BMI at age 50 into 9 categories,
consistent with the grouping of somatotypes (ranging from 1 to 9) at younger
ages. The rescaling was conducted by using a linear mixed effects model, in
which we assumed there is a linear relationship between BMI and somatotype
at each age. We then used such relationship at younger ages to derive the
somatotype at age 50 from the corresponding BMI at age 50. To allow this
relationship to vary among individuals, we included a random intercept and a
random slope in the model, as specified below:
��� = �� + ���� + �� + ���� + ���
13
where ��� denotes the BMI for individual � at age and �� denotes the
corresponding somatotype, �� and �� specify the fixed intercept and slope, �� and �� specify the random, subject-specific intercept and slope, and ��� represents the within-subject measurement error. Furthermore, it is assumed
that �~��0, ����, ���~��0, ���, and that � and ��� are mutually independent.
We used an unstructured variance-covariance matrix (i.e., without making any
particular assumption about the covariance structure) for estimation.
We modeled BMI instead of somatotype as the dependent variable because
BMI is approximately normally distributed, whereas somatotype is a discrete
variable ranging from 1 to 9. Thus, the conditional mean BMI for individual � at age can be written as:
�[���|��, ��] = �� + ���� + �� + ���� Based on the model output, we then calculated the somatotype rating at age 50
��� as ��[�� |�!�,�"�]#$!#�!�$"%�"�
&. To keep the estimated somatotypes within the
range of 1 to 9, for ��� that was larger than 9, we rounded it as 9, and for ���
that was smaller than 1, we rounded it as 1.
There were 4 and 3 time points available for the linear mixed effects modeling
in women and men, respectively, as summarized below.
Time point Women
(Nurses’ Health Study)
Men
(Health Professionals Follow-up Study)
1 BMI: age 18
Somatotype: age 20
BMI: age 21
Somatotype: age 20
2 BMI: age 30
Somatotype: age 30
BMI: age 40
Somatotype: age 40
3 BMI: age 40
Somatotype: age 40
BMI: in 1988
Somatotype: in 1988
4 BMI: in 1988
Somatotype: in 1988
The details about anthropometric assessments at each time point are described
below. It should be noted that there may be missing data for each of the time
points, but the missing data can be accommodated by the linear mixed effects
model.
In women (Nurses’ Health Study):
• Time point 1: in 1980 participants were asked to recall their body weight
at age 18. We then calculated their BMI, and linked it to their somatotype
ratings at age 20.
• Time points 2 and 3: participants were queried about their current body
weight in our biennial follow-up questionnaires. We used these data to
calculate their BMI at age 30 (the mean BMI from age 27 to 33) and 40
14
(the mean BMI from age 37 to 43), and then linked it to their somatotype
ratings at age 30 and 40.
• Time point 4: we calculated participants’ BMI in 1988 based on the
follow-up questionnaires, and then linked it to their somatotype ratings in
1988 when we also queried about their current body shape.
In men (Health Professionals Follow-up Study):
• Time point 1: in 1986 participants were asked to recall their body weight
at age 21. We then calculated their BMI, and linked it to their somatotype
ratings at age 20.
• Time point 2: participants were queried about their current body weight
in our biennial follow-up questionnaires. We used these data to calculate
their BMI at age 40 (the mean BMI from age 37 to 43), and then linked it
to their somatotype ratings at age 40. We did not include the BMI at age
30, because almost all participants were already at their 40s when
enrolled into the cohort and their BMI data at age 30 were not available.
• Time point 3: we calculated participants’ BMI in 1988 based on the
follow-up questionnaires, and then linked it to their somatotype ratings in
1988 when we also queried about their current body shape.
Point #2: Related to this, Figure 1 shows that 28% of individuals were excluded as they had
“missing somatotype data for more than two different age points”, or put more simply, fewer
than 4 somatotypes. This serious data loss strikes me as unnecessary – Figure 1 shows that the
group trajectories are essentially linear (despite the cubic fit), which means that anyone with 2
or more somatotypes could reasonably be analysed. Even requiring a minimum of 3 ought to
reduce the dropout appreciably.
We appreciate the reviewer’s thoughtful comments. However, lowering the required number of
complete data points would not reduce the dropout appreciably, because most of the excluded
participants had missing somatotype data for 5 or 6 age points. This is understandable because all
the somatotype data from age 5 to 40 were collected at the same time, and the participant who
provided somatotype data for one age was likely to answer all the somatotype questions. The
following table shows the number of the excluded participants due to missing somatotype data
according to the number of missing data points.
Number of the age points that
have missing somatotype data Women Men
3 208 305
4 86 149
5 26,315 5,186
6 7,323 8,061
Total 33,932 13,701
Therefore, the gain in sample size will be minimal (208 in women, 305 in men) if we change the
15
exclusion criterion to missing somatotype data for more than three (instead of two) age points.
Point #3: A second data exclusion is for BMI < 18.5 kg/m2. I can see that this is meant to reduce
reverse causation, but since the whole purpose is to model BMI, it seems illogical to omit
individuals whose BMI is arbitrarily low. In any case the numbers involved are tiny and will
make no difference to the results.
Yes, we excluded participants with BMI of <18.5 kg/m2 to reduce reverse causation, because
these participants likely have some underlying chronic disease that may predispose them to early
death. As the reviewer pointed out, because of the small numbers involved, our results were
robust to this exclusion. For example, the HRs (95% CI) of all-cause mortality for the heavy-
stable/increase group in women were 1.45 (1.38-1.54) in the analysis including participants with
BMI of <18.5 kg/m2, and 1.48 (1.40-1.57) after excluding these participants. Since BMI <18.5
kg/m2 technically falls outside the range of normal BMI, we think that it is preferable to exclude
them.
Point #4: Related to point 6 above, the essential linearity of the group trajectories indicates that
a simple random-slope-random-intercept model would lead to broadly the same conclusions, and
the relative size of the slope and intercept random effects in predicting mortality would quantify
the importance of mean BMI versus BMI gain.
We agree that the essential linearity of body-shape measurements with age makes the random-
slope-random-intercept model a potential choice to assess subject-specific trajectory. Also, as the
reviewer pointed out, comparing the relative size of the slope and intercept random effects in
predicting mortality would give us some sense about the importance of mean BMI versus BMI
gain. However, we believe that the approach we employed has some unique advantages. First, it
avoids the difficulty in interpretation of the results obtained from statistical modeling of different
dimensions of adiposity (e.g., mean BMI versus BMI gain in the approach suggested by the
reviewer), because these measures are not biologically independent. For example, the mean BMI
at age 50 depends largely on how BMI changes over early years, and therefore it is difficult, if
not biologically meaningful, to separate their independent effects on mortality. In contrast,
instead of comparing different adiposity measures, our approach focuses on comparison of
different groups of individuals who have distinct body-shape trajectory profiles, and addresses a
more tangible and clinically relevant question: how does mortality differ among individuals with
different trajectories of body shape.
Another advantage of classifying individuals into distinct groups is that it provides an intuitive
way to probe into the population heterogeneity in the susceptibility of body shape change across
the lifespan, which does not only yield an easily-digested message for the public but also has
great implications for future research. For example, further studies examining the relative
16
contributions of individuals’ genetics and behaviors to their trajectory profiles will provide
critical insights into tailored prevention strategies. If it is found that, for instance, the heavy-
stable/increase group has a larger genetic component, whereas the lean-marked increase group is
more behaviorally oriented, public health strategies of behavioral change should then be targeted
towards the latter group. In addition, these approaches, for example identifying risk genes, are
much more feasible for mutually exclusive phenotypes than by using multivariable analysis,
where the meaning of a variable is conditional on other variables.
A potential limitation of our approach is information loss due to discrete grouping. However, the
good discrimination of our trajectory-building model (mean posterior probability of trajectory
assignment: 0.9) and well-tracked change in BMI across trajectories indicate that the trajectories
we identified can parsimoniously summarize, without a significant loss of information, the
predominant features of lifetime body shape in the study population. Additionally, it was
reassuring that we obtained similar results after excluding participants with suboptimal trajectory
assignment.
Point #5: Is there a way to superimpose on the group trajectories of Figure 1 the mean BMI
values appearing in Table 1? It would provide some validation of the group allocation.
We tried to add to Figure 1 all the mean BMI values in Table 1. But because the Y axis
encompasses the somatotype ratings for 6 different ages and across 5 different trajectory groups,
superimposing all the BMI data does not seem to work in Excel. Instead, we have added the
mean BMI values for age 50 only to the figures, as shown on the next page.
17
18
Point #6: The hazard ratios in Tables 2 and subsequently are adjusted for a string of covariates
including lifestyle factors such as physical activity, alcohol consumption and dietary score. Are
these factors not on the causal pathway, and hence should not be adjusted for? The research
question involves the shape trajectories versus mortality, which must be due at least in part to
lifestyle differences – so why adjust for them if the interest is in the trajectories themselves?
It is a challenge to control for confounding in such studies as our current one of the exposure
across the lifespan, because some covariates can be both mediators and confounders for the
exposure-outcome relationship. Here we adjusted for several lifestyle factors because they may
influence individuals’ body-shape trajectory and are also important predictors for mortality. Not
adjusting for these factors may lead to confounding bias (i.e., the observed relationship between
body shape and mortality may be due to other lifestyle factors that are related to body shape).
However, as the reviewer pointed out, there can also be over-adjustment because individuals
with certain body shapes may be more likely to change their lifestyle that can in turn influence
their mortality risk. We tried to minimize such over-adjustment by using the cumulative average
lifestyle information collected throughout early and mid-life to age 50, because such long-term
habitual exposure is less susceptible to temporary lifestyle changes that were caused by body-
shape change at some time points during follow-up.
We also compared the crude and adjusted hazard ratios among never smokers. As shown in the
table below, the results were quite similar; indicating that neither confounding nor over-
adjustment had a substantial influence on our results, after we account for the confounding effect
by smoking.
19
Crude and adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) of all-cause and cause-
specific mortality according to trajectories of body shape from age 5 to 50 among never smokers of
women in the Nurses’ Health Study and men in the Health Professionals Follow-up Study