Body Mass and Risk from COVID-19 and Influenza

2020-04-06 at 10:11 pm 7 comments

Understanding the factors affecting whether someone infected with COVID-19 will become seriously ill is important for treatment of patients, for forecasting and planning, and — with factors that can be changed — for personal decisions aimed at reducing risk. Despite our current focus, influenza also remains a serious disease, so understanding its risk factors is also important.

Here, I’ll look at some of the evidence on how body mass — formalized as Body Mass Index (BMI, weight in kilograms divided by squared height in metres) — influences prognosis for respiratory diseases. Information specific to COVID-19 is still scant, but there is more data on influenza and on other respiratory infections (which includes coronaviruses other than COVID-19). Information on how BMI relates to general mortality should also be helpful.

Below, I’ll look at two relevant papers, plus a preliminary report on COVID-19. To preview my conclusions, it seems that being underweight and being seriously obese are both risk factors for serious respiratory illness. Furthermore, it seems that “underweight” should include the lower part of the “normal weight” category as defined by the WHO. Official advice in this respect seems dangerously misleading.

Body mass and respiratory illness

The most relevant study addressing how body mass affects outcomes of respiratory disease that I could find is the following:

J.-A. S. Mozer, et al (2018) Underweight, overweight, and obesity as independent risk factors for hospitalization in adults and children from influenza and other respiratory viruses, Influenza and Other Respiratory Viruses. Available from the journal site, and also here.

This study looked at several thousand patients who sought medical attention for an influenza-like illness in Mexico between April 2010 and March 2014. Of these, 3248 were adults 19 or older who were not excluded for some reason (I will ignore the younger patients, which the paper analyses separately). Among these patients, 1083 were classified as having severe disease, as indicated by their having been hospitalized.

For each patient, they recorded age, sex, and presence of a chronic medical condition (yes/no). They converted BMI to one of the categories defined by the WHO:

    underweight (BMI < 18.5)
    normal (18.5 ≤ BMI <25)
    overweight (25 ≤ BMI <30)
    obese (30 ≤ BMI <35)
    morbidly obese (BMI ≥ 35)

They used the “normal” category as the reference. They performed tests for presence of influenza virus (various strains) as well as several other respiratory viruses, including coronaviruses NL63, OC43, 229E, and (from June 2012 on) HKU1, classifying each patient as having influenza (and possibly other viruses), having non-influenza respiratory virus, or having no detected virus.

In their main analysis, the latter two groups were combined, resulting in a group of 553 patients with influenza and a group of 2695 patients with other viruses or no detected virus. Logistic regression models for whether or not the illness was severe were fit separately for these two groups, with age, sex, presence of chronic conditions, and BMI category as covariates. The results are shown in their Table 2, reproduced below:

If p is the probability of severe illness, the “odds” of severe illness are defined to be p/(1-p). The “odds ratios” shown above indicate how the modeled odds change if the covariate value is changed (changed by one year for age).

For example, if the model says that for a normal-weight influenza patient of a certain age, sex, and chronic medical condition status the odds of severe illness are 1.5 (corresponding to a probability of 0.6), the odds of severe illness for another influenza patient with the same covariates except that they are underweight would (according to the model) be 1.5×5.20=7.8 (corresponding to a probability of 0.89).

An odds ratio greater than 1 indicates that the risk of severe illness is greater for the “active” value of the covariate, odds ratios less than 1 that the risk is less. The numbers in parentheses are the confidence interval for the estimated odds ratios. A p-value of less than 0.05 is taken as an indication that the estimated odds ratio is likely to be on the correct side of 1, and such estimates are in bold (except that the overweight estimate for influenza is mistakenly bolded when it shouldn’t be). A p-value less than 0.05 also corresponds to the confidence interval for the odds ratio not containing the value 1.

Unfortunately, the paper fails to say which way around sex was coded, so we can’t tell whether it’s males or females who have a reduced risk if they have influenza (odds ratio of 0.54) and an increased risk if they have a different respiratory disease (odds ratio of 1.83). This result seems odd, but could be due to various non-obvious factors (eg, a difference in relative propensity of males and females to seek medical attention for influenza versus other illnesses).

The authors checked for interactions between the effect of BMI and other covariates, and fit additional logistic regression models to subsets of the data when interactions were significant. However, the model above, which divides the data only into influenza and non-influenza groups, captures their main conclusion, which is expressed in the first sentence of their discussion as follows:

We have demonstrated that adults that are underweight or morbidly obese are more likely to be hospitalized from an influenza-like illness, regardless of the causative agent of the illness, than normal-weight adults.

One should keep in mind that this is an observational study. The relationships seen between body mass and seriousness of illness do not necessarily represent cause and effect. For example, it could be that in Mexico smoking is especially common among underweight people, and that that, rather than their weight itself, is the reason why underweight patients had higher risk. This caveat is important if one is considering modifying BMI to reduce risk, but not if one is using BMI (without other information) to forecast outcomes in a similar population.

The data for this study isn’t publicly available, and the paper does not include full details of the analysis (eg, the output of the statistics package for all the models they fit). From just looking at the paper, though, I do not see any serious problems with the analysis that would invalidate their conclusion above.

The analysis could be improved, however. Looking at whether a possible interaction is statistically significant, and on that basis either ignoring the interaction or doing completely separate analyses for subgroups, is a rather crude approach. One could likely get better estimates (with smaller uncertainties) by building a hierarchical Bayesian model, in which many potential interactions can be modelled, while main effects can still predominate if appropriate. Similarly, grouping together all patients who don’t have influenza, ignoring the information on which specific viruses they have, throws away information that might produce useful insight if a hierarchical Bayesian approach were used, in which information regarding different viruses can be pooled to the extent that they seem to have similar effects, while still allowing any differences to be seen.

More seriously, the use of the WHO categories for BMI throws away important information, especially regarding relationships for low values of BMI. The odds ratios for serious illness for underweight patients versus normal weight patients are 5.20 for influenza and 2.88 for non-influenza illness. These odds ratios are quite substantial, and raise the question of whether patients towards the low end of the WHO “normal” weight range also have substantially higher risk of serious illness.

To help answer this question, and to shed some light on whether there is a cause-and-effect relationship of BMI to serious illness, I’ll next look at studies on how BMI relates to mortality, from all causes, not just respiratory illness.

Body mass and all-cause mortality

The following large-scale study assessed the impact of BMI on mortality from all causes, while trying to avoid problems of “reverse causation” (ie, that people may have lower weight because they’re dying, rather than be dying because they have lower weight):

The Global BMI Mortality Collaboration (2016) Body-mass index and all-cause mortality: individual-participant-data meta-analysis of 239 prospective studies in four continents, The Lancet, available from the journal site, and also here. There is an appendix of supplemental material (here or here). See also the editorial comment (here or here).

They looked at estimates from 239 earlier studies regarding the impact of BMI on mortality, and combined them (accounting for their uncertainties) to give overall estimates of how BMI relates to mortality, as well as estimates for sub-groups according to sex, age, and geographical region. They primarily looked at mortality 5 or more years after the measurement of BMI (to avoid reverse causation) amongst adults who never smoked (to avoid confounding due to smoking being a cause of lower weight) and who did not have any of various chronic diseases when the BMI was measured (again to avoid reverse causation). Here are their estimates for all regions and all ages (from 20 to 90 years), but separated by sex:

The “hazard” gives the probability of dying in a small period of time (eg, one year) in the immediate future. The modelled hazards for groups of people with differing characteristics can be compared by the “hazard ratio” for the two groups. The plot above shows hazard ratios with respect to people with BMI between 22.5 and 25.0 as the reference (ie, that group’s hazard ratio is fixed at 1). Note that the vertical scale is logarithmic (each division represents a doubling of the hazard). The vertical lines on each point give the uncertainty range; the size of the square shows how much information the estimate is based on.

This study used more BMI categories than those defined by the WHO, which were used in the first study I discussed above. Here is the finer division used in this study, and the correspondence with the larger WHO categories:

    15.0-18.5   underweight
    18.5-20.0   normal
    20.0-22.5      "
    22.5-25.0      "
    25.0-27.5   overweight
    27.5-30.0      "
    30.0-35.0   obese
    35.0-40.0   morbidly obese
    40.0-60.0      "

One conclusion we might draw from this plot is noted in the paper:

…there was a substantially higher mortality not only among those in WHO’s underweight category, but also in those with BMI 18·5 kg/m2 to <20 kg/m2, suggesting that in excessively lean adult populations underweight remains a cause for concern.

The study also looks at mortality due to various causes, including respiratory disease. Hazard ratios for death from respiratory disease are shown in the following plot:

From Figure 4

Respiratory diseases include respiratory infections such as influenza and coronavirus, but also some chronic conditions, so this is not a perfect match to our current interest. It nevertheless seems notable that risk of death from respiratory disease is elevated not only in the WHO “underweight” category (BMI<18.5) but also in the parts of the WHO “normal weight” category with BMI between 18.5 and 22.5. This strengthens the suspicion that the high odds ratios for influenza-like illness being serious found in the first study I discussed, for patients with BMI<18.5, may also apply to patients in the lower part of the WHO “normal weight” range. The worry that the results for influenza-like illness might be affected by reverse causality or confounding by smoking is also lessened by the results of this study, which was designed to avoid these problems.

Considering that COVID-19 and influenza tend to be more serious in older patients, it is of interest that the study also looks at how the mortality to BMI relationship differs by age. This plot shows how the hazard ratios vary with both BMI and age (with age 40-59 and BMI 22.5-25 as reference):

From eFigure 7 in the supplementary information

The BMI range with the smallest estimated hazard ratio increases with age. For the 80+ age group the minimum is in the BMI 25.0 to 27.5 range, the low end of the WHO “overweight” category, and for the 70-79 age group this category and the top end of the “normal weight” category (BMI 22.5 to 25.0) have very nearly the same hazard ratio. The paper finds that the minimum hazard ratio for people age 70 and above is at a BMI of 24, but this conclusion is likely highly influenced by the functional form of the curve that they fit — visually a value above 25 seems more plausible. The claim in the paper that

This analysis has shown that both overweight and obesity (all grades) were associated with increased all-cause mortality…Our results challenge recent suggestions that overweight and moderate obesity are not associated with higher mortality…

does not seem to be entirely supported by their results, with respect to people age 70 and above.

The paper unfortunately gives no breakdown by age for different causes of death. Comparing the plots above for all-cause mortality and for mortality due to respiratory disease, we can see that the BMI with minimum hazard ratio is larger for respiratory disease mortality than for all-cause mortality by roughly 2 units. It is not clear that the same shift will apply to all age groups, but it does seem likely that for people age 70 and above — those most at risk from COVID-19 and influenza — the minimum risk for mortality from respiratory disease is for a BMI in the “overweight” category, there is substantial increased risk for BMI less than 20 or more than 35, and there is moderately increased risk for BMI in the range 20-22.5 or in the range 30-35.

Of course, it is not clear that these results for influenza-like illness and respiratory disease in general will hold for a new disease such as COVID-19.

Some preliminary COVID-19 data

Some data on 775 COVID-19 patients admitted to critical care in the UK is now available here. The source is as follows:

These data derive from the ICNARC Case Mix Programme Database. The Case Mix Programme is the national clinical audit of patient outcomes from adult critical care coordinated by the Intensive Care National Audit & Research Centre (ICNARC). For more information on the representativeness and quality of these data, please contact ICNARC.

Some commentary on this preliminary release of data can be seen here. As noted in this commentary and in the report itself, the data on outcomes for patients admitted to a critical care unit is biased because not all patients have an outcome yet. (Outcomes that happen sooner will be overrepresented.) So I’ll look only at what might be inferred from the data on patients admitted to critical care, as compared to the general population. (But note that since the epidemic has not emerged uniformly over the UK, the general population is not quite the right group to compare with.) The report also has data on viral pneumonia patients (non-COVID-19) admitted from 2017 to 2019, which provides another helpful comparison.

One notable and somewhat puzzling fact is that compared to the viral pneumonia patients, the COVID-19 patients are much less likely to have very severe comorbidities. It’s not clear whether this is because this information is not being completely recorded, or because COVID-19 patients with severe comorbidities die before they can get to critical care, or because such patients are refused admission to critical care, or because (most likely, I think) this is due to COVID-19 being unusually dangerous for mostly-healthy people. The proportion of male patients is also higher (70.5%) for COVID-19 than for non-COVID-19 viral pneumonia (54.3%). These are indications that COVID-19 and non-COVID-19 viral pneumonia are not entirely similar, so some caution is warranted when trying to apply results from other diseases to COVID-19.

The report has data on BMI for the COVID-19 patients, along with age and sex matched BMI figures for the general population. Unfortunately, they use broad categories similar to the WHO categories. The BMI<18.5 category has only 9 patients, too few for meaningful results. For the others, the proportion in each category is as follows (Table 1 and Figure 5):

    BMI      COVID-19   general population   odds ratio
18.5 to <25    26.6%           26.0%          1.00 (reference)
  25 to <30    34.4%           42.0%          0.80
  30 to <40    31.0%           28.5%          1.06
     40+        6.7%            2.5%          2.62

I’ve added the odds ratios for admission to critical care for COVID-19, computed from the other numbers. Since only a small fraction of the population is admitted to critical care, the odds ratio here is the same as the “relative risk”. (I couldn’t calculate odds ratios for non-COVID-19 viral pneumonia due to the lack of figures on BMI in the general population adjusted for age and sex to match that data.)

It seems that having a BMI of 40 or more (extreme obesity) substantially increases the risk of getting a serious case of COVID-19. Less extreme obesity may lead to a small increase in risk. Being “overweight” according the WHO categorization (BMI from 25 to 30) seems to reduce the risk by a fair amount, compared to having a “normal weight” (BMI from 18.5 to 25).

We can again see here the bad effects of people using the WHO BMI categories, especially the supposedly “normal weight” category, which as I discuss above includes a range from 18.5 to 20.0 that has elevated mortality from all causes, and especially from respiratory disease, as well as the range from 20.0 to 22.5 which also has elevated respiratory disease mortality. Because of its failure to give a finer breakdown by BMI, we cannot tell from this report whether these lower weight categories have higher risk of requiring critical care for COVID-19. We also cannot tell how the risk in the upper end of the “normal weight” category (BMI from 22.5 to 25) compares to the risk in the “overweight” category (BMI 25 to 30). The “overweight” category itself could usefully be broken down into two categories, as in the study of all-cause mortality discussed above.

UPDATE: I now discuss a newer version of this data at the end of this post.

Conclusions

The studies I’ve reviewed above support one official COVID-19 assessment that is seen, for example, in this statement from the CDC (see here):

Groups at Higher Risk for Severe Illness

People with severe obesity (body mass index [BMI] of 40 or higher)

It might even be justified to extend this to BMI of 35 or greater, though I think going any lower would be dubious.

However, I’ve seen no official advice that a low BMI puts one at risk of severe illness and death. I have come across other indications along these lines, however. For example, this quote from a doctor (see here):

When you are critically ill you lose 7 per cent of your muscle mass for every day that you are on a ventilator – it becomes clear why, if you are 85 and weigh 45 kilos, we are not keen to put those people on ventilators, irrespective of Covid.

Indeed, some official advice tells you that a BMI of 18.5 is just fine. Here are excerpts from the Canadian Guidelines for Body Weight Classification in Adults:

Research studies in large groups of people have shown that the BMI can be classified into ranges associated with health risk. There are four categories of BMI ranges in the Canadian weight classification system. These are:

  • underweight (BMI less than 18.5);
  • normal weight (BMIs 18.5 to 24.9);
  • overweight (BMIs 25 to 29.9), and
  • obese (BMI 30 and over).

Q10: I am concerned that my 25-year old daughter is underweight, but according to the weight classification guidelines she is within the ‘normal weight’ category.

A: In general, Canadian adults who have a BMI within the ‘normal weight’ category have the least risk of developing weight-related health problems…

It is pretty clear that this is dangerously false statement. It’s especially dangerous at the present time, since low weight is more of a risk factor for respiratory illness than for some other diseases.

Recognizing the likely risk to people with low BMI (around 20 or less) would influence public health policy. First, it would mean devoting effort to better estimating the risk of low BMI, for COVID-19 and other diseases. A good start on this would be to stop grouping all individuals with BMI in the 18.5 to 25 range together as “normal weight”, discarding the information that is needed to investigate this issue. It would also affect where interventions are directed. In particular, in some parts of the world many people have low BMI due to food shortage. COVID-19 may be a reason to ramp up efforts to solve this problem.

At an individual level, BMI is something that can potentially be modified, as a way to reduce risk (from COVID-19 or generally). Some caution is needed here. For one thing, BMI is not a perfect measure of underlying physical condition. Perhaps there are some people for whom a BMI of 18.5 is healthy, due to unusual genetic factors, for example. Also, attempts to drastically change BMI in a short period of time could be counterproductive. Certainly, people whose BMI has been changed by such a drastic effort are not typical subjects of the studies I reviewed above, so the results of those studies may not apply to them.

Nevertheless, I think people with a BMI over 35 should endeavour not to gain weight, and people with a BMI under 20 should take care not to lose weight — both unfortunately possible as people’s routines change while avoiding COVID-19. It might also be good for people to undertake moderate efforts to move their BMI into a range from about 21 to 27 (perhaps slightly higher for people over 70).

UPDATE (2020-04-11): There is now an updated report from ICNARC on UK patients with COVID-19 in critical care (here, with an appendix). They now have information on 3883 COVID-19 patients in critical care. This new data doesn’t substantially change the conclusions I drew above from the earlier report on 775 patients, but does allow more precise statements.

Here’s a update with the new data on the odds ratios for entering critical care for COVID-19 for different BMI categories:

    BMI     critical care    general     odds
            for COVID-19    population   ratio
    <18.5       0.64%         0.64%      0.99 (0.65,1.51)
18.5 to <25    25.79%        25.37%      1.00 (reference)
  25 to <30    35.04%        42.28%      0.82
  30 to <40    31.33%        28.87%      1.07
     40+        7.20%         2.84%      2.49 (2.19,2.83)

I’ve now put in a line for BMI<18.5, since with 21 cases in this category some meaningful information can be obtained. I’ve included 95% confidence intervals for the odds ratio for this category and for the BMI 40+ category (with 235 cases); the other categories have 842 to 1144 cases and so have more precise estimates.

Note that the odds ratios in this situation are the same as the relative risks (ratios of probabilities of being in critical care) with respect to the reference category, since the probability of someone in the UK being in critical care for COVID-19 is very small.

Entry filed under: COVID-19, Science, Statistics, Statistics - Nontechnical.

Software for Flexible Bayesian Modeling – New release The Puzzling Linearity of COVID-19

7 Comments Add your own

  • 1. Joshua Pritikin  |  2020-04-07 at 8:45 am

    You talk about BMI as something that is easy to modify. However, lots of people struggle with weight control. For a science-based approach, check out https://nutritionfacts.org/how-not-to-diet/

    Reply
    • 2. Radford Neal  |  2020-04-07 at 9:09 am

      Yes, weight loss is something people struggle with. Perhaps weight gain for underweight people can also be hard. In both cases, it must have a lot to do with why their weight was unhealthy to start with. But I think avoiding extra weight gain or loss due to lifestyle changes stemming from COVID-19 could be more feasible.

      Reply
  • 3. Aaron Galloway  |  2020-04-07 at 9:15 pm

    I’m curious to what extent BMI – especially with those who are underweight or on the very low end of the “normal” range – could be thought of as observed variables of an underlying health condition that functions as a latent variable. And that this latent variable could be the key risk factor to COVID-19.

    I would appreciate any clarification on this point.

    Reply
    • 4. Radford Neal  |  2020-04-07 at 9:34 pm

      That is indeed a large concern – the “reverse causation” I mention in my post. The study of BMI and mortality tried to address this concern by excluding smokers and those with known chronic conditions (which could cause current low BMI and later death), and then looking only at deaths at least five years after the BMI measurement (if low BMI was due to an unknown health condition, they may have died before five years).

      Of course, it’s not going to be a perfect way of handling the problem.

      Reply
  • 5. nicolaschopin  |  2020-04-08 at 3:05 am

    Hi,
    you might find this paper (in French, from le Monde) interesting:
    https://www.lemonde.fr/planete/article/2020/04/07/les-personnes-obeses-sont-plus-fragilisees-par-le-virus_6035831_3244.html

    It says that preliminary French data seems to point to the same direction; e.g. there is this one hospital in Nice where 95% of COVID patient are either “obese” or “overweight”. (The paper mingles the two categories, which is a bit confusing, however.)

    Reply
    • 6. Radford Neal  |  2020-04-08 at 10:19 am

      Hi,

      The le Monde article unfortunately doesn’t give any figures on the general incidence of “overweight” or “obese” in France (ideally adjusted to match the age, sex, and geographical origin of the patients). Without this, one can’t draw any conclusions from the BMI stats for COVID-19 patients regarding how risk of serious illness relates to BMI.

      The article refers to data on 196 patients in the UK of which 32% were “overweight” (BMI 25-30) and 41% were “obese” (BMI 30+). This was perhaps an earlier version of the dataset linked in my post, which has 775 patients, with 34.4% being “overweight” and 37.7% being “obese”. But note that the fraction of “overweight” patients in this data set is LESS than in the general population (matched for age and sex). Being overweight seems to be protective, rather being a risk factor.

      Reply
  • 7. Radford Neal  |  2020-04-08 at 1:03 pm

    There is also discussion of this post at lesswrong, at

    https://www.lesswrong.com/posts/79ukbkLWwdSCYKfF4/body-mass-and-risk-from-covid-19-and-influenza

    Reply

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

Trackback this post  |  Subscribe to the comments via RSS Feed


Calendar

April 2020
M T W T F S S
 12345
6789101112
13141516171819
20212223242526
27282930  

Most Recent Posts