The lowest-mortality BMI: What is the role of nutrient intake from food?

In a previous post (), I discussed the frequently reported lowest-mortality body mass index (BMI), which is about 26. The empirical results reviewed in that post suggest that fat-free mass plays an important role in that context. Keep in mind that this "BMI=26 phenomenon" is often reported in studies of populations from developed countries, which are likely to be relatively sedentary. This is important for the point made in this post.

A lowest-mortality BMI of 26 is somehow at odds with the fact that many healthy and/or long-living populations have much lower BMIs. You can clearly see this in the distribution of BMIs among males in Kitava and Sweden shown in the graph below, from a study by Lindeberg and colleagues (). This distribution is shifted in such a way that would suggest a much lower BMI of lowest-mortality among the Kitavans, assuming a U-curve shape similar to that observed in studies of populations from developed countries ().

Another relevant example comes from the China Study II (see, e.g., ), which is based on data from 8000 adults. The average BMI in the China Study II dataset, with data from the 1980s, is approximately 21; for an average weight that is about 116 lbs. That BMI is relatively uniform across Chinese counties, including those with the lowest mortality rates. No county has an average BMI that is 26; not even close. This also supports the idea that Chinese people were, at least during that period, relatively thin.

Now take a look at the graph below, also based on the China Study II dataset, from a previous post (), relating total daily calorie intake with longevity. I should note that the relationship between total daily calorie intake and longevity depicted in this graph is not really statistically significant. Still, the highest longevity seems to be in the second tercile of total daily calorie intake.

Again, the average weight in the dataset is about 116 lbs. A conservative estimate of the number of calories needed to maintain this weight without any physical activity would be about 1740. Add about 700 calories to that, for a reasonable and healthy level of physical activity, and you get 2440 calories needed daily for weight maintenance. That is right in the middle of the second tercile, the one with the highest longevity.

What does this have to do with the lowest-mortality BMI of 26 from studies of samples from developed countries? Populations in these countries are likely to be relatively sedentary, at least on average, in which case a low BMI will be associated with a low total calorie intake. And a low total calorie intake will lead to a low intake of nutrients needed by the body to fight disease.

And don’t think you can fix this problem by consuming lots of vitamin and mineral pills. When I refer here to a higher or lower nutrient intake, I am not talking only about micronutrients, but also about macronutrients (fatty and amino acids) in amounts that are needed by your body. Moreover, important micronutrients, such as fat-soluble vitamins, cannot be properly absorbed without certain macronutrients, such as fat.

Industrial nutrient isolation for supplementation use has not been a very successful long-term strategy for health optimization (). On the other hand, this type of supplementation has indeed been found to have had modest-to-significant success in short-term interventions aimed at correcting acute health problems caused by severe nutritional deficiencies ().

So the "BMI=26 phenomenon" may be a reflection not of a direct effect of high muscularity on health, but of an indirect effect mediated by a high intake of needed nutrients among sedentary folks. This may be so even though the lowest mortality is for the combination of that BMI with a relatively small waist (), which suggests some level of muscularity, but not necessarily serious bodybuilder-level muscularity. High muscularity, of the serious bodybuilder type, is not very common; at least not enough to significantly sway results based on the analysis of large samples.

The combination of a BMI=26 with a relatively small waist is indicative of more muscle and less body fat. Having more muscle and less body fat has an advantage that is rarely discussed. It allows for a higher total calorie intake, and thus a higher nutrient intake, without an unhealthy increase in body fat. Muscle mass increases one's caloric requirement for weight maintenance, more so than body fat. Body fat also increases that caloric requirement, but it also acts like an organ, secreting a number of hormones into the bloodstream, and becoming pro-inflammatory in an unhealthy way above a certain level.

Clearly having a low body fat percentage is associated with lower incidence of degenerative diseases, but it will likely lead to a lower intake of nutrients relative to one’s needs unless other factors are present, e.g., being fairly muscular or physically active. Chronic low nutrient intake tends to get people closer to the afterlife like nothing else ().

In this sense, having a BMI=26 and being relatively sedentary (without being skinny-fat) has an effect that is similar to that of having a BMI=21 and being fairly physically active. Both would lead to consumption of more calories for weight maintenance, and thus more nutrients, as long as nutritious foods are eaten.

The lowest-mortality BMI: What is its relationship with fat-free mass?

Do overweight folks live longer? It is not uncommon to see graphs like the one below, from the Med Journal Watch blog (), suggesting that, at least as far as body mass index (BMI) is concerned (), overweight folks (25 < BMI < 30) seem to live longer. The graph shows BMI measured at a certain age, and risk of death within a certain time period (e.g., 20 years) following the measurement. The lowest-mortality BMI is about 26, which is in the overweight area of the BMI chart.



Note that relative age-adjusted mortality risk (i.e., relative to the mortality risk of people in the same age group), increases less steeply in response to weight variations as one becomes older. An older person increases the risk of dying to a lesser extent by weighing more or less than does a younger person. This seems to be particularly true for weight gain (as opposed to weight loss).

The table below is from a widely cited 2002 article by Allison and colleagues (), where they describe a study of 10,169 males aged 25-75. Almost all of the participants, ninety-eight percent, were followed up for many years after measurement; a total of 3,722 deaths were recorded.



Take a look at the two numbers circled in red. The one on the left is the lowest-mortality BMI not adjusting for fat mass or fat-free mass: a reasonably high 27.4. The one on the right is the lowest-mortality BMI adjusting for fat mass and fat-free mass: a much lower 21.6.

I know this may sound confusing, but due to possible statistical distortions this does not mean that you should try to bring your BMI to 21.6 if you want to reduce your risk of dying. What this means is that fat mass and fat-free mass matter. Moreover, all of the participants in this study were men. The authors concluded that: “…marked leanness (as opposed to thinness) has beneficial effects.”

Then we have an interesting 2003 article by Bigaard and colleagues () reporting on a study of 27,178 men and 29,875 women born in Denmark, 50 to 64 years of age. The table below summarizes deaths in this study, grouping them by BMI and waist circumference.



These are raw numbers; no complex statistics here. Circled in green is the area with samples that appear to be large enough to avoid “funny” results. Circled in red are the lowest-mortality percentages; I left out the 0.8 percentage because it is based on a very small sample.

As you can see, they refer to men and women with BMIs in the 25-29.9 range (overweight), but with waist circumferences in the lower-middle range: 90-96 cm for men and 74-82 cm for women; or approximately 35-38 inches for men and 29-32 inches for women.

Women with BMIs in the 18.5-24.9 range (normal) and the same or lower waists also died in small numbers. Underweight men and women had the highest mortality percentages.

A relatively small waist (not a wasp waist), together with a normal or high BMI, is an indication of more fat-free mass, which is retained together with some body fat. It is also an indication of less visceral body fat accumulation.

The China Study II: Does calorie restriction increase longevity?

The idea that calorie restriction extends human life comes largely from studies of other species. The most relevant of those studies have been conducted with primates, where it has been shown that primates that eat a restricted calorie diet live longer and healthier lives than those that are allowed to eat as much as they want.

There are two main problems with many of the animal studies of calorie restriction. One is that, as natural lifespan decreases, it becomes progressively easier to experimentally obtain major relative lifespan extensions. (That is, it seems much easier to double the lifespan of an organism whose natural lifespan is one day than an organism whose natural lifespan is 80 years.) The second, and main problem in my mind, is that the studies often compare obese with lean animals.

Obesity clearly reduces lifespan in humans, but that is a different claim than the one that calorie restriction increases lifespan. It has often been claimed that Asian countries and regions where calorie intake is reduced display increased lifespan. And this may well be true, but the question remains as to whether this is due to calorie restriction increasing lifespan, or because the rates of obesity are much lower in countries and regions where calorie intake is reduced.

So, what can the China Study II data tell us about the hypothesis that calorie restriction increases longevity?

As it turns out, we can conduct a preliminary test of this hypothesis based on a key assumption. Let us say we compared two populations (e.g., counties in China), based on the following ratio: number of deaths at or after age 70 divided by number deaths before age 70. Let us call this the “ratio of longevity” of a population, or RLONGEV. The assumption is that the population with the highest RLONGEV would be the population with the highest longevity of the two. The reason is that, as longevity goes up, one would expect to see a shift in death patterns, with progressively more people dying old and fewer people dying young.

The 1989 China Study II dataset has two variables that we can use to estimate RLONGEV. They are coded as M005 and M006, and refer to the mortality rates from 35 to 69 and 70 to 79 years of age, respectively. Unfortunately there is no variable for mortality after 79 years of age, which limits the scope of our results somewhat. (This does not totally invalidate the results because we are using a ratio as our measure of longevity, not the absolute number of deaths from 70 to 79 years of age.) Take a look at these two previous China Study II posts (here, and here) for other notes, most of which apply here as well. The notes are at the end of the posts.

All of the results reported here are from analyses conducted using WarpPLS. Below is a model with coefficients of association; it is a simple model, since the hypothesis that we are testing is also simple. (Click on it to enlarge. Use the "CRTL" and "+" keys to zoom in, and CRTL" and "-" to zoom out.) The arrows explore associations between variables, which are shown within ovals. The meaning of each variable is the following: TKCAL = total calorie intake per day; RLONGEV = ratio of longevity; SexM1F2 = sex, with 1 assigned to males and 2 to females.

As one would expect, being female is associated with increased longevity, but the association is just shy of being statistically significant in this dataset (beta=0.14; P=0.07). The association between total calorie intake and longevity is trivial, and statistically indistinguishable from zero (beta=-0.04; P=0.39). Moreover, even though this very weak association is overall negative (or inverse), the sign of the association here does not fully reflect the shape of the association. The shape is that of an inverted J-curve; a.k.a. U-curve. When we split the data into total calorie intake terciles we get a better picture:

The second tercile, which refers to a total daily calorie intake of 2193 to 2844 calories, is the one associated with the highest longevity. The first tercile (with the lowest range of calories) is associated with a higher longevity than the third tercile (with the highest range of calories). These results need to be viewed in context. The average weight in this dataset was about 116 lbs. A conservative estimate of the number of calories needed to maintain this weight without any physical activity would be about 1740. Add about 700 calories to that, for a reasonable and healthy level of physical activity, and you get 2440 calories needed daily for weight maintenance. That is right in the middle of the second tercile.

In simple terms, the China Study II data seems to suggest that those who eat well, but not too much, live the longest. Those who eat little have slightly lower longevity. Those who eat too much seem to have the lowest longevity, perhaps because of the negative effects of excessive body fat.

Because these trends are all very weak from a statistical standpoint, we have to take them with caution. What we can say with more confidence is that the China Study II data does not seem to support the hypothesis that calorie restriction increases longevity.

Reference

Kock, N. (2010). WarpPLS 1.0 User Manual. Laredo, Texas: ScriptWarp Systems.

Notes

- The path coefficients (indicated as beta coefficients) reflect the strength of the relationships; they are a bit like standard univariate (or Pearson) correlation coefficients, except that they take into consideration multivariate relationships (they control for competing effects on each variable). Whenever nonlinear relationships were modeled, the path coefficients were automatically corrected by the software to account for nonlinearity.

- Only two data points per county were used (for males and females). This increased the sample size of the dataset without artificially reducing variance, which is desirable since the dataset is relatively small (each county, not individual, is a separate data point is this dataset). This also allowed for the test of commonsense assumptions (e.g., the protective effects of being female), which is always a good idea in a multivariate analyses because violation of commonsense assumptions may suggest data collection or analysis error. On the other hand, it required the inclusion of a sex variable as a control variable in the analysis, which is no big deal.

- Mortality from schistosomiasis infection (MSCHIST) does not confound the results presented here. Only counties where no deaths from schistosomiasis infection were reported have been included in this analysis. The reason for this is that mortality from schistosomiasis infection can severely distort the results in the age ranges considered here. On the other hand, removal of counties with deaths from schistosomiasis infection reduced the sample size, and thus decreased the statistical power of the analysis.

The China Study II: Wheat flour, rice, and cardiovascular disease

In my last post on the China Study II, I analyzed the effect of total and HDL cholesterol on mortality from all cardiovascular diseases. The main conclusion was that total and HDL cholesterol were protective. Total and HDL cholesterol usually increase with intake of animal foods, and particularly of animal fat. The lowest mortality from all cardiovascular diseases was in the highest total cholesterol range, 172.5 to 180; and the highest mortality in the lowest total cholesterol range, 120 to 127.5. The difference was quite large; the mortality in the lowest range was approximately 3.3 times higher than in the highest.

This post focuses on the intake of two main plant foods, namely wheat flour and rice intake, and their relationships with mortality from all cardiovascular diseases. After many exploratory multivariate analyses, wheat flour and rice emerged as the plant foods with the strongest associations with mortality from all cardiovascular diseases. Moreover, wheat flour and rice have a strong and inverse relationship with each other, which suggests a “consumption divide”. Since the data is from China in the late 1980s, it is likely that consumption of wheat flour is even higher now. As you’ll see, this picture is alarming.

The main model and results

All of the results reported here are from analyses conducted using WarpPLS. Below is the model with the main results of the analyses. (Click on it to enlarge. Use the "CRTL" and "+" keys to zoom in, and CRTL" and "-" to zoom out.) The arrows explore associations between variables, which are shown within ovals. The meaning of each variable is the following: SexM1F2 = sex, with 1 assigned to males and 2 to females; MVASC = mortality from all cardiovascular diseases (ages 35-69); TKCAL = total calorie intake per day; WHTFLOUR = wheat flour intake (g/day); and RICE = and rice intake (g/day).

The variables to the left of MVASC are the main predictors of interest in the model. The one to the right is a control variable – SexM1F2. The path coefficients (indicated as beta coefficients) reflect the strength of the relationships. A negative beta means that the relationship is negative; i.e., an increase in a variable is associated with a decrease in the variable that it points to. The P values indicate the statistical significance of the relationship; a P lower than 0.05 generally means a significant relationship (95 percent or higher likelihood that the relationship is “real”).

In summary, the model above seems to be telling us that:

- As rice intake increases, wheat flour intake decreases significantly (beta=-0.84; P<0.01). This relationship would be the same if the arrow pointed in the opposite direction. It suggests that there is a sharp divide between rice-consuming and wheat flour-consuming regions.

- As wheat flour intake increases, mortality from all cardiovascular diseases increases significantly (beta=0.32; P<0.01). This is after controlling for the effects of rice and total calorie intake. That is, wheat flour seems to have some inherent properties that make it bad for one’s health, even if one doesn’t consume that many calories.

- As rice intake increases, mortality from all cardiovascular diseases decreases significantly (beta=-0.24; P<0.01). This is after controlling for the effects of wheat flour and total calorie intake. That is, this effect is not entirely due to rice being consumed in place of wheat flour. Still, as you’ll see later in this post, this relationship is nonlinear. Excessive rice intake does not seem to be very good for one’s health either.

- Increases in wheat flour and rice intake are significantly associated with increases in total calorie intake (betas=0.25, 0.33; P<0.01). This may be due to wheat flour and rice intake: (a) being themselves, in terms of their own caloric content, main contributors to the total calorie intake; or (b) causing an increase in calorie intake from other sources. The former is more likely, given the effect below.

- The effect of total calorie intake on mortality from all cardiovascular diseases is insignificant when we control for the effects of rice and wheat flour intakes (beta=0.08; P=0.35). This suggests that neither wheat flour nor rice exerts an effect on mortality from all cardiovascular diseases by increasing total calorie intake from other food sources.

- Being female is significantly associated with a reduction in mortality from all cardiovascular diseases (beta=-0.24; P=0.01). This is to be expected. In other words, men are women with a few design flaws, so to speak. (This situation reverses itself a bit after menopause.)

Wheat flour displaces rice

The graph below shows the shape of the association between wheat flour intake (WHTFLOUR) and rice intake (RICE). The values are provided in standardized format; e.g., 0 is the mean (a.k.a. average), 1 is one standard deviation above the mean, and so on. The curve is the best-fitting U curve obtained by the software. It actually has the shape of an exponential decay curve, which can be seen as a section of a U curve. This suggests that wheat flour consumption has strongly displaced rice consumption in several regions in China, and also that wherever rice consumption is high wheat flour consumption tends to be low.

As wheat flour intake goes up, so does cardiovascular disease mortality

The graphs below show the shapes of the association between wheat flour intake (WHTFLOUR) and mortality from all cardiovascular diseases (MVASC). In the first graph, the values are provided in standardized format; e.g., 0 is the mean (or average), 1 is one standard deviation above the mean, and so on. In the second graph, the values are provided in unstandardized format and organized in terciles (each of three equal intervals).

The curve in the first graph is the best-fitting U curve obtained by the software. It is a quasi-linear relationship. The higher the consumption of wheat flour in a county, the higher seems to be the mortality from all cardiovascular diseases. The second graph suggests that mortality in the third tercile, which represents a consumption of wheat flour of 501 to 751 g/day (a lot!), is 69 percent higher than mortality in the first tercile (0 to 251 g/day).

Rice seems to be protective, as long as intake is not too high

The graphs below show the shapes of the association between rice intake (RICE) and mortality from all cardiovascular diseases (MVASC). In the first graph, the values are provided in standardized format. In the second graph, the values are provided in unstandardized format and organized in terciles.

Here the relationship is more complex. The lowest mortality is clearly in the second tercile (206 to 412 g/day). There is a lot of variation in the first tercile, as suggested by the first graph with the U curve. (Remember, as rice intake goes down, wheat flour intake tends to go up.) The U curve here looks similar to the exponential decay curve shown earlier in the post, for the relationship between rice and wheat flour intake.

In fact, the shape of the association between rice intake and mortality from all cardiovascular diseases looks a bit like an “echo” of the shape of the relationship between rice and wheat flour intake. Here is what is creepy. This echo looks somewhat like the first curve (between rice and wheat flour intake), but with wheat flour intake replaced by “death” (i.e., mortality from all cardiovascular diseases).

What does this all mean?

- Wheat flour displacing rice does not look like a good thing. Wheat flour intake seems to have strongly displaced rice intake in the counties where it is heavily consumed. Generally speaking, that does not seem to have been a good thing. It looks like this is generally associated with increased mortality from all cardiovascular diseases.

- High glycemic index food consumption does not seem to be the problem here. Wheat flour and rice have very similar glycemic indices (but generally not glycemic loads; see below). Both lead to blood glucose and insulin spikes. Yet, rice consumption seems protective when it is not excessive. This is true in part (but not entirely) because it largely displaces wheat flour. Moreover, neither rice nor wheat flour consumption seems to be significantly associated with cardiovascular disease via an increase in total calorie consumption. This is a bit of a blow to the theory that high glycemic carbohydrates necessarily cause obesity, diabetes, and eventually cardiovascular disease.

- The problem with wheat flour is … hard to pinpoint, based on the results summarized here. Maybe it is the fact that it is an ultra-refined carbohydrate-rich food; less refined forms of wheat could be healthier. In fact, the glycemic loads of less refined carbohydrate-rich foods tend to be much lower than those of more refined ones. (Also, boiled brown rice has a glycemic load that is about three times lower than that of whole wheat bread; whereas the glycemic indices are about the same.) Maybe the problem is wheat flour's  gluten content. Maybe it is a combination of various factors, including these.

Reference

Kock, N. (2010). WarpPLS 1.0 User Manual. Laredo, Texas: ScriptWarp Systems.

Acknowledgment and notes

- Many thanks are due to Dr. Campbell and his collaborators for collecting and compiling the data used in this analysis. The data is from this site, created by those researchers to disseminate their work in connection with a study often referred to as the “China Study II”. It has already been analyzed by other bloggers. Notable analyses have been conducted by Ricardo at Canibais e Reis, Stan at Heretic, and Denise at Raw Food SOS.

- The path coefficients (indicated as beta coefficients) reflect the strength of the relationships; they are a bit like standard univariate (or Pearson) correlation coefficients, except that they take into consideration multivariate relationships (they control for competing effects on each variable). Whenever nonlinear relationships were modeled, the path coefficients were automatically corrected by the software to account for nonlinearity.

- The software used here identifies non-cyclical and mono-cyclical relationships such as logarithmic, exponential, and hyperbolic decay relationships. Once a relationship is identified, data values are corrected and coefficients calculated. This is not the same as log-transforming data prior to analysis, which is widely used but only works if the underlying relationship is logarithmic. Otherwise, log-transforming data may distort the relationship even more than assuming that it is linear, which is what is done by most statistical software tools.

- The R-squared values reflect the percentage of explained variance for certain variables; the higher they are, the better the model fit with the data. In complex and multi-factorial phenomena such as health-related phenomena, many would consider an R-squared of 0.20 as acceptable. Still, such an R-squared would mean that 80 percent of the variance for a particularly variable is unexplained by the data.

- The P values have been calculated using a nonparametric technique, a form of resampling called jackknifing, which does not require the assumption that the data is normally distributed to be met. This and other related techniques also tend to yield more reliable results for small samples, and samples with outliers (as long as the outliers are “good” data, and are not the result of measurement error).

- Only two data points per county were used (for males and females). This increased the sample size of the dataset without artificially reducing variance, which is desirable since the dataset is relatively small. This also allowed for the test of commonsense assumptions (e.g., the protective effects of being female), which is always a good idea in a complex analysis because violation of commonsense assumptions may suggest data collection or analysis error. On the other hand, it required the inclusion of a sex variable as a control variable in the analysis, which is no big deal.

- Since all the data was collected around the same time (late 1980s), this analysis assumes a somewhat static pattern of consumption of rice and wheat flour. In other words, let us assume that variations in consumption of a particular food do lead to variations in mortality. Still, that effect will typically take years to manifest itself. This is a major limitation of this dataset and any related analyses.

- Mortality from schistosomiasis infection (MSCHIST) does not confound the results presented here. Only counties where no deaths from schistosomiasis infection were reported have been included in this analysis. Mortality from all cardiovascular diseases (MVASC) was measured using the variable M059 ALLVASCc (ages 35-69). See this post for other notes that apply here as well.

The China Study II: Cholesterol seems to protect against cardiovascular disease

First of all, many thanks are due to Dr. Campbell and his collaborators for collecting and compiling the data used in this analysis. This data is from this site, created by those researchers to disseminate the data from a study often referred to as the “China Study II”. It has already been analyzed by other bloggers. Notable analyses have been conducted by Ricardo at Canibais e Reis, Stan at Heretic, and Denise at Raw Food SOS.

The analyses in this post differ from those other analyses in various aspects. One of them is that data for males and females were used separately for each county, instead of the totals per county. Only two data points per county were used (for males and females). This increased the sample size of the dataset without artificially reducing variance (for more details, see “Notes” at the end of the post), which is desirable since the dataset is relatively small. This also allowed for the test of commonsense assumptions (e.g., the protective effects of being female), which is always a good idea in a complex analysis because violation of commonsense assumption may suggest data collection or analysis error. On the other hand, it required the inclusion of a sex variable as a control variable in the analysis, which is no big deal.

The analysis was conducted using WarpPLS. Below is the model with the main results of the analysis. (Click on it to enlarge. Use the "CRTL" and "+" keys to zoom in, and CRTL" and "-" to zoom out.) The arrows explore associations between variables, which are shown within ovals. The meaning of each variable is the following: SexM1F2 = sex, with 1 assigned to males and 2 to females; HDLCHOL = HDL cholesterol; TOTCHOL = total cholesterol; MSCHIST = mortality from schistosomiasis infection; and MVASC = mortality from all cardiovascular diseases.

The variables to the left of MVASC are the main predictors of interest in the model – HDLCHOL and TOTCHOL. The ones to the right are control variables – SexM1F2 and MSCHIST. The path coefficients (indicated as beta coefficients) reflect the strength of the relationships. A negative beta means that the relationship is negative; i.e., an increase in a variable is associated with a decrease in the variable that it points to. The P values indicate the statistical significance of the relationship; a P lower than 0.05 generally means a significant relationship (95 percent or higher likelihood that the relationship is “real”).

In summary, this is what the model above is telling us:

- As HDL cholesterol increases, total cholesterol increases significantly (beta=0.48; P<0.01). This is to be expected, as HDL is a main component of total cholesterol, together with VLDL and LDL cholesterol.

- As total cholesterol increases, mortality from all cardiovascular diseases decreases significantly (beta=-0.25; P<0.01). This is to be expected if we assume that total cholesterol is in part an intervening variable between HDL cholesterol and mortality from all cardiovascular diseases. This assumption can be tested through a separate model (more below). Also, there is more to this story, as noted below.

- The effect of HDL cholesterol on mortality from all cardiovascular diseases is insignificant when we control for the effect of total cholesterol (beta=-0.08; P=0.26). This suggests that HDL’s protective role is subsumed by the variable total cholesterol, and also that it is possible that there is something else associated with total cholesterol that makes it protective. Otherwise the effect of total cholesterol might have been insignificant, and the effect of HDL cholesterol significant (the reverse of what we see here).

- Being female is significantly associated with a reduction in mortality from all cardiovascular diseases (beta=-0.16; P=0.01). This is to be expected. In other words, men are women with a few design flaws. (This situation reverses itself a bit after menopause.)

- Mortality from schistosomiasis infection is significantly and inversely associated with mortality from all cardiovascular diseases (beta=-0.28; P<0.01). This is probably due to those dying from schistosomiasis infection not being entered in the dataset as dying from cardiovascular diseases, and vice-versa.

Two other main components of total cholesterol, in addition to HDL cholesterol, are VLDL and LDL cholesterol. These are carried in particles, known as lipoproteins. VLDL cholesterol is usually represented as a fraction of triglycerides in cholesterol equations (e.g., the Friedewald and Iranian equations). It usually correlates inversely with HDL; that is, as HDL cholesterol increases, usually VLDL cholesterol decreases. Given this and the associations discussed above, it seems that LDL cholesterol is a good candidate for the possible “something else associated with total cholesterol that makes it protective”. But waidaminet! Is it possible that the demon particle, the LDL, serves any purpose other than giving us heart attacks?

The graph below shows the shape of the association between total cholesterol (TOTCHOL) and mortality from all cardiovascular diseases (MVASC). The values are provided in standardized format; e.g., 0 is the average, 1 is one standard deviation above the mean, and so on. The curve is the best-fitting S curve obtained by the software (an S curve is a slightly more complex curve than a U curve).

The graph below shows some of the data in unstandardized format, and organized differently. The data is grouped here in ranges of total cholesterol, which are shown on the horizontal axis. The lowest and highest ranges in the dataset are shown, to highlight the magnitude of the apparently protective effect. Here the two variables used to calculate mortality from all cardiovascular diseases (MVASC; see “Notes” at the end of this post) were added. Clearly the lowest mortality from all cardiovascular diseases is in the highest total cholesterol range, 172.5 to 180; and the highest mortality in the lowest total cholesterol range, 120 to 127.5. The difference is quite large; the mortality in the lowest range is approximately 3.3 times higher than in the highest.

The shape of the S-curve graph above suggests that there are other variables that are confounding the results a bit. Mortality from all cardiovascular diseases does seem to generally go down with increases in total cholesterol, but the smooth inflection point at the middle of the S-curve graph suggests a more complex variation pattern that may be influenced by other variables (e.g., smoking, dietary patterns, or even schistosomiasis infection; see “Notes” at the end of this post).

As mentioned before, total cholesterol is strongly influenced by HDL cholesterol, so below is the model with only HDL cholesterol (HDLCHOL) pointing at mortality from all cardiovascular diseases (MVASC), and the control variable sex (SexM1F2).

The graph above confirms the assumption that HDL’s protective role is subsumed by the variable total cholesterol. When the variable total cholesterol is removed from the model, as it was done above, the protective effect of HDL cholesterol becomes significant (beta=-0.27; P<0.01). The control variable sex (SexM1F2) was retained even in this targeted HDL effect model because of the expected confounding effect of sex; females generally tend to have higher HDL cholesterol and less cardiovascular disease than males.

Below, in the “Notes” section (after the “Reference”) are several notes, some of which are quite technical. Providing them separately hopefully has made the discussion above a bit easier to follow. The notes also point at some limitations of the analysis. This data needs to be analyzed from different angles, using multiple models, so that firmer conclusions can be reached. Still, the overall picture that seems to be emerging is at odds with previous beliefs based on the same dataset.

What could be increasing the apparently protective HDL and total cholesterol in this dataset? High consumption of animal foods, particularly foods rich in saturated fat and cholesterol, are strong candidates. Low consumption of vegetable oils rich in linoleic acid, and of foods rich in refined carbohydrates, are also good candidates. Maybe it is a combination of these.

We need more analyses!

Reference:

Kock, N. (2010). WarpPLS 1.0 User Manual. Laredo, Texas: ScriptWarp Systems.


Notes:

- The path coefficients (indicated as beta coefficients) reflect the strength of the relationships; they are a bit like standard univariate (or Pearson) correlation coefficients, except that they take into consideration multivariate relationships (they control for competing effects on each variable).

- The R-squared values reflect the percentage of explained variance for certain variables; the higher they are, the better the model fit with the data. In complex and multi-factorial phenomena such as health-related phenomena, many would consider an R-squared of 0.20 as acceptable. Still, such an R-squared would mean that 80 percent of the variance for a particularly variable is unexplained by the data.

- The P values have been calculated using a nonparametric technique, a form of resampling called jackknifing, which does not require the assumption that the data is normally distributed to be met. This and other related techniques also tend to yield more reliable results for small samples, and samples with outliers (as long as the outliers are “good” data, and are not the result of measurement error).

- Colinearity is an important consideration in models that analyze the effect of multiple predictors on one single variable. This is particularly true for multiple regression models, where there is a temptation of adding many predictors to the model to see which ones come out as the “winners”. This often backfires, as colinearity can severely distort the results. Some multiple regression techniques, such as automated stepwise regression with backward elimination, are particularly vulnerable to this problem. Colinearity is not the same as correlation, and thus is defined and measured differently. Two predictor variables may be significantly correlated and still have low colinearity. A reasonably reliable measure of colinearity is the variance inflation factor. Colinearity was tested in this model, and was found to be low.

- An effort was made here to avoid multiple data points per county (even though this was available for some variables), because this could artificially reduce the variance for each variable, and potentially bias the results. The reason for this is that multiple answers from a single county would normally be somewhat correlated; a higher degree of intra-county correlation than inter-county correlation. The resulting bias would be difficult to control for, via one or more control variables. With only two data points per county, one for males and the other for females, one can control for intra-country correlation by adding a “dummy” sex variable to the analysis, as a control variable. This was done here.

- Mortality from schistosomiasis infection (MSCHIST) is a variable that tends to affect the results in a way that makes it more difficult to make sense of them. Generally this is true for any infectious diseases that significantly affect a population under study. The problem with infection is that people with otherwise good health or habits may get the infection, and people with bad health and habits may not. Since cholesterol is used by the human body to fight disease, it may go up, giving the impression that it is going up for some other reason. Perhaps instead of controlling for its effect, as done here, it would have been better to remove from the analysis those counties with deaths from schistosomiasis infection. (See also this post, and this one.)

- Different parts of the data were collected at different times. It seems that the mortality data is for the period 1986-88, and the rest of the data is for 1989. This may have biased the results somewhat, even though the time lag is not that long, especially if there were changes in certain health trends from one period to the other. For example, major migrations from one county to another could have significantly affected the results.

- The following measures were used, from this online dataset like the other measures. P002 HDLCHOL, for HDLCHOL; P001 TOTCHOL, for TOTCHOL; and M021 SCHISTOc, for MSCHIST.

- SexM1F2 is a “dummy” variable that was coded with 1 assigned to males and 2 to females. As such, it essentially measures the “degree of femaleness” of the respondents. Being female is generally protective against cardiovascular disease, a situation that reverts itself a bit after menopause.

- MVASC is a composite measure of the two following variables, provided as component measures of mortality from all cardiovascular diseases: M058 ALLVASCb (ages 0-34), and M059 ALLVASCc (ages 35-69). A couple of obvious problems: (a) they does not include data on people older than 69; and (b) they seem to capture a lot of diseases, including some that do not seem like typical cardiovascular diseases. A factor analysis was conducted, and the loadings and cross-loadings suggested good validity. Composite reliability was also good. So essentially MVASC is measured here as a “latent variable” with two “indicators”. Why do this? The reason is that it reduces the biasing effects of incomplete data and measurement error (e.g., exclusion of folks older than 69). By the way, there is always some measurement error in any dataset.

- This note is related to measurement error in connection with the indicators for MVASC. There is something odd about the variables M058 ALLVASCb (ages 0-34), and M059 ALLVASCc (ages 35-69). According to the dataset, mortality from cardiovascular diseases for ages 0-34 is typically higher than for 35-69, for many counties. Given the good validity and reliability for MVASC as a latent variable, it is possible that the values for these two indicator variables were simply swapped by mistake.

The China Study one more time: Are raw plant foods giving people cancer?

In this previous post I analyzed some data from the China Study that included counties where there were cases of schistosomiasis infection. Following one of Denise Minger’s suggestions, I removed all those counties from the data. I was left with 29 counties, a much smaller sample size. I then ran a multivariate analysis using WarpPLS (warppls.com), like in the previous post, but this time I used an algorithm that identifies nonlinear relationships between variables.

Below is the model with the results. (Click on it to enlarge. Use the "CRTL" and "+" keys to zoom in, and CRTL" and "-" to zoom out.) As in the previous post, the arrows explore associations between variables. The variables are shown within ovals. The meaning of each variable is the following: aprotein = animal protein consumption; pprotein = plant protein consumption; cholest = total cholesterol; crcancer = colorectal cancer.

What is total cholesterol doing at the right part of the graph? It is there because I am analyzing the associations between animal protein and plant protein consumption with colorectal cancer, controlling for the possible confounding effect of total cholesterol.

I am not hypothesizing anything regarding total cholesterol, even though this variable is shown as pointing at colorectal cancer. I am just controlling for it. This is the type of thing one can do in multivariate analyzes. This is how you “control for the effect of a variable” in an analysis like this.

Since the sample is fairly small, we end up with insignificant beta coefficients that would normally be statistically significant with a larger sample. But it helps that we are using nonparametric statistics, because they are still robust in the presence of small samples, and deviations from normality. Also the nonlinear algorithm is more sensitive to relationships that do not fit a classic linear pattern. We can summarize the findings as follows:

- As animal protein consumption increases, plant protein consumption decreases significantly (beta=-0.36; P<0.01). This is to be expected and helpful in the analysis, as it differentiates somewhat animal from plant protein consumers. Those folks who got more of their protein from animal foods tended to get significantly less protein from plant foods.

- As animal protein consumption increases, colorectal cancer decreases, but not in a statistically significant way (beta=-0.31; P=0.10). The beta here is certainly high, and the likelihood that the relationship is real is 90 percent, even with such a small sample.

- As plant protein consumption increases, colorectal cancer increases significantly (beta=0.47; P<0.01). The small sample size was not enough to make this association insignificant. The reason is that the distribution pattern of the data here is very indicative of a real association, which is reflected in the low P value.

Remember, these results are not confounded by schistosomiasis infection, because we are only looking at counties where there were no cases of schistosomiasis infection. These results are not confounded by total cholesterol either, because we controlled for that possible confounding effect. Now, control variable or not, you would be correct to point out that the association between total cholesterol and colorectal cancer is high (beta=0.58; P=0.01). So let us take a look at the shape of that association:

Does this graph remind you of the one on this post; the one with several U curves? Yes. And why is that? Maybe it reflects a tendency among the folks who had low cholesterol to have more cancer because the body needs cholesterol to fight disease, and cancer is a disease. And maybe it reflects a tendency among the folks who have high total cholesterol to do so because total cholesterol (and particularly its main component, LDL cholesterol) is in part a marker of disease, and cancer is often a culmination of various metabolic disorders (e.g., the metabolic syndrome) that are nothing but one disease after another.

To believe that total cholesterol causes colorectal cancer is nonsensical because total cholesterol is generally increased by consumption of animal products, of which animal protein consumption is a proxy. (In this reduced dataset, the linear univariate correlation between animal protein consumption and total cholesterol is a significant and positive 0.36.) And animal protein consumption seems to be protective again colorectal cancer in this dataset (negative association on the model graph).

Now comes the part that I find the most ironic about this whole discussion in the blogosphere that has been going on recently about the China Study; and the answer to the question posed in the title of this post: Are raw plant foods giving people cancer? If you think that the answer is “yes”, think again. The variable that is strongly associated with colorectal cancer is plant protein consumption.

Do fruits, veggies, and other plant foods that can be consumed raw have a lot of protein?

With a few exceptions, like nuts, they do not. Most raw plant foods have trace amounts of protein, especially when compared with foods made from refined grains and seeds (e.g., wheat grains, soybean seeds). So the contribution of raw fruits and veggies in general could not have influenced much the variable plant protein consumption. To put this in perspective, the average plant protein consumption per day in this dataset was 63 g; even if they were eating 30 bananas a day, the study participants would not get half that much protein from bananas.

Refined foods made from grains and seeds are made from those plant parts that the plants absolutely do not “want” animals to eat. They are the plants’ “children” or “children’s nutritional reserves”, so to speak. This is why they are packed with nutrients, including protein and carbohydrates, but also often toxic and/or unpalatable to animals (including humans) when eaten raw.

But humans are so smart; they learned how to industrially refine grains and seeds for consumption. The resulting human-engineered products (usually engineered to sell as many units as possible, not to make you healthy) normally taste delicious, so you tend to eat a lot of them. They also tend to raise blood sugar to abnormally high levels, because industrial refining makes their high carbohydrate content easily digestible. Refined foods made from grains and seeds also tend to cause leaky gut problems, and autoimmune disorders like celiac disease. Yep, we humans are really smart.

Thanks again to Dr. Campbell and his colleagues for collecting and compiling the China Study data, and to Ms. Minger for making the data available in easily downloadable format and for doing some superb analyses herself.

The China Study again: A multivariate analysis suggesting that schistosomiasis rules!

In the comments section of Denise Minger’s post on July 16, 2010, which discusses some of the data from the China Study (as a follow up to a previous post on the same topic), Denise herself posted the data she used in her analysis. This data is from the China Study. So I decided to take a look at that data and do a couple of multivariate analyzes with it using WarpPLS (warppls.com).

First I built a model that explores relationships with the goal of testing the assumption that the consumption of animal protein causes colorectal cancer, via an intermediate effect on total cholesterol. I built the model with various hypothesized associations to explore several relationships simultaneously, including some commonsense ones. Including commonsense relationships is usually a good idea in exploratory multivariate analyses.

The model is shown on the graph below, with the results. (Click on it to enlarge. Use the "CRTL" and "+" keys to zoom in, and CRTL" and "-" to zoom out.) The arrows explore causative associations between variables. The variables are shown within ovals. The meaning of each variable is the following: aprotein = animal protein consumption; pprotein = plant protein consumption; cholest = total cholesterol; crcancer = colorectal cancer.

The path coefficients (indicated as beta coefficients) reflect the strength of the relationships; they are a bit like standard univariate (or Pearson) correlation coefficients, except that they take into consideration multivariate relationships (they control for competing effects on each variable). A negative beta means that the relationship is negative; i.e., an increase in a variable is associated with a decrease in the variable that it points to.

The P values indicate the statistical significance of the relationship; a P lower than 0.05 means a significant relationship (95 percent or higher likelihood that the relationship is real). The R-squared values reflect the percentage of explained variance for certain variables; the higher they are, the better the model fit with the data. Ignore the “(R)1i” below the variable names; it simply means that each of the variables is measured through a single indicator (or a single measure; that is, the variables are not latent variables).

I should note that the P values have been calculated using a nonparametric technique, a form of resampling called jackknifing, which does not require the assumption that the data is normally distributed to be met. This is good, because I checked the data, and it does not look like it is normally distributed. So what does the model above tell us? It tells us that:

- As animal protein consumption increases, colorectal cancer decreases, but not in a statistically significant way (beta=-0.13; P=0.11).

- As animal protein consumption increases, plant protein consumption decreases significantly (beta=-0.19; P<0.01). This is to be expected.

- As plant protein consumption increases, colorectal cancer increases significantly (beta=0.30; P=0.03). This is statistically significant because the P is lower than 0.05.

- As animal protein consumption increases, total cholesterol increases significantly (beta=0.20; P<0.01). No surprise here. And, by the way, the total cholesterol levels in this study are quite low; an overall increase in them would probably be healthy.

- As plant protein consumption increases, total cholesterol decreases significantly (beta=-0.23; P=0.02). No surprise here either, because plant protein consumption is negatively associated with animal protein consumption; and the latter tends to increase total cholesterol.

- As total cholesterol increases, colorectal cancer increases significantly (beta=0.45; P<0.01). Big surprise here!

Why the big surprise with the apparently strong relationship between total cholesterol and colorectal cancer? The reason is that it does not make sense, because animal protein consumption seems to increase total cholesterol (which we know it usually does), and yet animal protein consumption seems to decrease colorectal cancer.

When something like this happens in a multivariate analysis, it usually is due to the model not incorporating a variable that has important relationships with the other variables. In other words, the model is incomplete, hence the nonsensical results. As I said before in a previous post, relationships among variables that are implied by coefficients of association must also make sense.

Now, Denise pointed out that the missing variable here possibly is schistosomiasis infection. The dataset that she provided included that variable, even though there were some missing values (about 28 percent of the data for that variable was missing), so I added it to the model in a way that seems to make sense. The new model is shown on the graph below. In the model, schisto = schistosomiasis infection.

So what does this new, and more complete, model tell us? It tells us some of the things that the previous model told us, but a few new things, which make a lot more sense. Note that this model fits the data much better than the previous one, particularly regarding the overall effect on colorectal cancer, which is indicated by the high R-squared value for that variable (R-squared=0.73). Most notably, this new model tells us that:

- As schistosomiasis infection increases, colorectal cancer increases significantly (beta=0.83; P<0.01). This is a MUCH STRONGER relationship than the previous one between total cholesterol and colorectal cancer; even though some data on schistosomiasis infection for a few counties is missing (the relationship might have been even stronger with a complete dataset). And this strong relationship makes sense, because schistosomiasis infection is indeed associated with increased cancer rates. More information on schistosomiasis infections can be found here.

- Schistosomiasis infection has no significant relationship with these variables: animal protein consumption, plant protein consumption, or total cholesterol. This makes sense, as the infection is caused by a worm that is not normally present in plant or animal food, and the infection itself is not specifically associated with abnormalities that would lead one to expect major increases in total cholesterol.

- Animal protein consumption has no significant relationship with colorectal cancer. The beta here is very low, and negative (beta=-0.03).

- Plant protein consumption has no significant relationship with colorectal cancer. The beta for this association is positive and nontrivial (beta=0.15), but the P value is too high (P=0.20) for us to discard chance within the context of this dataset. A more targeted dataset, with data on specific plant foods (e.g., wheat-based foods), could yield different results – maybe more significant associations, maybe less significant.

Below is the plot showing the relationship between schistosomiasis infection and colorectal cancer. The values are standardized, which means that the zero on the horizontal axis is the mean of the schistosomiasis infection numbers in the dataset. The shape of the plot is the same as the one with the unstandardized data. As you can see, the data points are very close to a line, which suggests a very strong linear association.

So, in summary, this multivariate analysis vindicates pretty much everything that Denise said in her July 16, 2010 post. It even supports Denise’s warning about jumping to conclusions too early regarding the possible relationship between wheat consumption and colorectal cancer (previously highlighted by a univariate analysis). Not that those conclusions are wrong; they may well be correct.

This multivariate analysis also supports Dr. Campbell’s assertion about the quality of the China Study data. The data that I analyzed was already grouped by county, so the sample size (65 cases) was not so high as to cast doubt on P values. (Having said that, small samples create problems of their own, such as low statistical power and an increase in the likelihood of error-induced bias.) The results summarized in this post also make sense in light of past empirical research.

It is very good data; data that needs to be properly analyzed!