Balancing potential bias with identifying enough events in observational data

We are designing a retrospective study using the EHR. We intend to look at the effect of ‘weight cycling’ on developing cardiometabolic disease (type 2 diabetes, MI, etc.). Weight cycling has a few definitions, but is typically considered to be clinically significant weight loss (5% in 3 months or 10% in 6 months) followed by clinically significant gain (or vice versa). Our plan is to do a landmark analysis with a landmark period of 5 years to assess weight behavior and calculate several weight-related variables (number of cycles, other indices of weight fluctuation, max weight, etc.), after which we will assess outcomes with the goal of identifying the most important weight-related predictor. Currently, for a patient to qualify for the study we are requiring that they have a weight at least once every 18 months during the landmark period. There are a few reasons for this.
• If weights are spaced too far apart, we won’t have a sense of the individual’s fluctuation which is our main focus. It would be biased to treat all of these people as ‘weight stable’ when in truth we don’t know what their weight behaviors are.
• If we use an absolute number of weights (say 3 weights in 5 years) there are many patients who had 3 weights very close together (say a 3 day hospitalization) then did not have contact with the system again, or visited extremely infrequently. We also would not have a sense of their weight behavior and they are unlikely to add value.

However, there very likely is a bias that patients with more weights in the system are likely to be more ill. They could also be weighed more simply because they have a condition which makes them prone to unintentional weight loss. We are going to exclude certain cancers, but we can’t exclude all conditions that could potentially lead to unintentional weight loss.

We are wondering if using an 18 month requirement sounds plausible and not too biased. Our hope is 18 months is long enough to capture weights of healthy people such as those getting annual visits without being so far apart that we are not able to draw conclusions about that individual’s variability.

Another note is that the landmark period start is being defined not as the first date the patient has in the EHR, but at the beginning of a period wherein they have weights 18 months apart for 5 years (if they have that). So a patient with a weight in 1997 and then who is not seen again until 2001 where they are seen annually would have their landmark period start in 2001.

Our main question is if this sounds reasonable? Would readers feel that findings from this cohort could be applicable to their own patient populations? Does anyone have ideas for any potential pitfalls or recommendations? Thank you!

1 Like

Hi Alison

A few questions/comments from a family physician who works with an EMR:

  1. Assuming, for the sake of argument, that you could construct a granular retrospective weight “trajectory” for a large number of patients (a questionable assumption- see below), what potential spectrum of findings might the study generate? Would the results be phrased something like this?: “We found that more frequent and/or sizeable weight fluctuations were associated with an increased future risk for diabetes…” (?)
  2. Do you foresee that these types of findings would be clinically actionable in a specific way? If so, how? Patients who are most likely to have sizeable weight fluctuations are those who are actively trying to lose weight (but who have trouble keeping it off). Since we already know that patients with a higher BMI are at increased risk for diabetes and hypertension (and therefore coronary disease), documenting an association between more frequent and/or sizeable weight fluctuations and future diabetes/coronary disease would probably not surprise too many physicians. Nor would such a finding suggest a particular clinical intervention. Specifically, we wouldn’t advise patients to stop trying to lose weight in order to decrease the risk of fluctuating weight. Patients will be at risk for diabetes and other complications if they don’t lose weight, regardless of whether they fluctuate. Disentangling the contribution of high absolute BMI from the contribution of weight fluctuation with regard to future diabetes risk seems pretty difficult (if not impossible). And again, I’m not sure that the result of even a reliable disentangling would be clinically actionable.
  3. Even if you could imagine a potentially clinically actionable result from such a study, I doubt that EMR records would provide a sufficient number of weight readings for your purposes. Most primary care physicians will document a body weight in only a small fraction of their progress notes (usually because weight isn’t related to the reason(s) for the visit).

I’m not sure if any of this helps. I might be misconstruing the purpose of your study…


Thank you for the thoughtful comments-- I appreciate your insight.

To answer your questions–

  1. Our plan is to model multiple weight-related potential risk factors to see their effect on predicting a cardiometabolic outcome. The weight-related predictors we would use include maximum weight, weight cycling, overall weight change, average weight, and a few others. The goal would be to see if weight cycling is a stronger predictor than BMI. We also intend to stratify by BMI to see if there are differences in risk from the weight-related predictors amongst BMI category.

  2. The thought is that if weight cycling turns out to be a stronger predictor in our models than BMI, then we would encourage future research considering if for some patients counseling on weight stability may be more valuable than counseling on weight loss. Given how so many weight loss attempts are unsuccessful or only successful in the short-term, counseling on weight stability may be sort of a harm-reduction approach-- yes it is much better to lose weight consistently in the long term in any situation but for those without much success or stability, maybe it is worth considering attempting a weight-stable approach. I do see your points and I think they are good ones.

  3. Thank you for your thoughts on this. I think one of the reasons why EMR data are attractive for something like this is it requires quite a long period of observation which just isn’t really compatible with our resources. But thank you I appreciate your insight. The synthetic derivative we are working with does contain a large amount of longitudinal weight data and has consistent weights on about 10-20% of folks which admittedly is not a large proportion and is likely to be enriched with people who are sicker, but is still a large group of people. We also capture weights that were obtained as part of vitals checks and not necessarily entered into the visit note if that helps clarify at all or gives you any additional thoughts.

Thank you!


Thanks for the clarifications. I think I understand the rationale behind your study now.

Before embarking on this research, it might be useful to consider what type of study result your team would assume that clinicians would also find compelling. You could then present these “hypothetical” findings to a group of primary care physicians and ask for their opinions. Specifically, you would want to know whether any particular result of your study would cause clinicians to recommend that their patients stop trying to lose weight. My own gut feeling :slight_smile: is that recommending against weight loss would be a hard sell, both for physicians and patients…

Sometimes research can be useful because it might help to disentangle mechanistic explanations for clinical events. To this end, maybe your main goal is to feed into a larger research agenda (though you would want some clarity, a priori, on who would carry on the research, and how) (?) But if you are instead expecting that your observational study might be able to change clinical practice, I think it’s really important to acknowledge that the bar for such a goal is set pretty high. If you’re going to invest lots of time and money in any research, it’s important to think through all possible study outcomes and the clinical response they might prompt (or not).

1 Like