A very interesting article was recently published in Lancet that sought to understand which factors correlate, on a country level, with covid related outcomes. The study was observational, so it can only show correlation, not causation, but it can still give pretty strong hints as to which factors protect people from covid, and which factors increase the risk of being harmed.
The most interesting thing about the study, from my perspective, was that it sought to understand what effect lockdowns, border closures, and widespread testing have in terms of decreasing the number of covid deaths. Although correlation does not automatically imply causation, if there is a lack of correlation, then that strongly suggests a lack of causation, or at least, that any causative relationship that does exist is extremely weak. And considering the amount of money, effort, and resources that have been poured in to lockdowns this year, and that continue to be poured in to them right now, it would be pretty disappointing if lockdowns had such a minimal effect that there was no noticeable impact on mortality whatsoever. Am I right?
But I get ahead of myself. The study chose to limit itself to looking at the 50 countries with the most recorded cases of covid-19 as of the 1st of April 2020. My interpretation is that they chose the top 50 most affected countries, rather than looking at all 195 countries, due to resource constraints. Data was gathered up to the 1st of May 2020. All information gathered was in the form of publicly available facts and figures. Data gathered included information about covid, income level, gross domestic product, income disparity, longevity, BMI (Body Mass Index), smoking, population density, and a bunch of other things that the researchers thought might be interesting to look at. The authors received no outside funding and reported no conflicts of interest.
There are a few problems here that become apparent straight away. First of all, as mentioned, all the data in this study is observational, so no conclusions can be drawn about cause and effect.
Second, May was relatively early in the pandemic, and it’s now November, so we’re missing about half a year’s worth of covid data. On the other hand, the pandemic had already peaked in much of the world by May 1st, and lockdown measures had at that point been in place for months in most countries, so it should be possible to get a pretty good idea about what effect lockdown has in terms of decreasing covid deaths, even using only the data available up to May 1st.
Third, the analysis builds on publicly available data, often provided by different governments themselves, with widely varying levels of trustworthiness, and with different ways of classifying things. As an example, data from Sweden is infinitely more reliable than data from China. And while certain countries have used quite inclusive criteria when deciding whether someone has died of covid or not, other countries have been much more strict. The countries with stricter definitions will tend to have lower covid death rates than the countries with more generous definitions. This lack of homogeneity in how things are defined can make it harder to see real patterns.
Fourth, the reseachers who put this study together gathered an enormous amount of data, pretty much everything they could think of under the sun that might in some way correlate with covid statistics. That means that this study amounts to “data trawling”, in other words, going through every relationship imaginable without any a priori hypothesis in order to see which relationships end up being statistically significant. When you do this, you’re supposed to set stricter limits than you normally would for what you consider to be statistically significant results. They didn’t do this. We’re going to discuss this problem in more detail later in the article.
Before we get in to the results, I’ll just mention one more thing. The results are presented as relative risks (not absolute risks), which tends to make results look more impressive than they really are, and the statistical significance level is presented in the form of confidence intervals, not p-values (not a problem in itself, just a different way of presenting data). If you haven’t already done so, I strongly recommend you read my guide to scientific method before reading further, in order to make sure you understand all the terms used and gain maximal value from the content. Anyway, let’s look at the results.
The factors that most strongly predicted the number of people who died of covid in a country were rate of obesity, average age, and level of income disparity. Each percentage point increase in the rate of obesity resulted in a 12% increase in covid deaths. Each additional average year of age in the population increased covid deaths by 10% . On the opposite end of the spectrum, each point in the direction of greater equality on the gini-coefficient (a scale used to determine how evenly resources are distributed across a population) resulted in a 12% decrease in covid deaths. All these results were statistically significant.
Another factor that had an effect that was significant, but more weakly so, was smoking. Each percentage point increase in the number of smokers in a population was correlated with a 3% decrease in covid deaths.
Ok, let’s get to the most important thing, which the authors seem to have tried to hide, because they make so little mention of it. Lockdown and covid deaths. The authors found no correlation whatsoever between severity of lockdown and number of covid deaths. And they didn’t find any correlation between border closures and covid deaths either. And there was no correlation between mass testing and covid deaths either, for that matter. Basically, nothing that various world governments have done to combat covid seems to have had any effect whatsoever on the number of deaths.
We’re going to come back to this incredible fact in a little bit, but first we’re going to go off on a little tangent. As mentioned, the researchers didn’t correct for the fact that they were looking at a ton of different relationships, rather than just one single relationship between two variables. As I have discussed previously in my article on scientific method, the more relationships you look at, the more strictly you have to set the cut-off for statistical significance, since you will otherwise just by chance get a lot of relationships that seem significant but aren’t.
If you set a p-value of 0,05 (5% probability that a significant relationship was seen in a study even though there isn’t one in the real world), then one in twenty relationships you look at will be statistically significant just by chance. The 5% cut-off is intended to be used when looking at a single relationship, not when looking at multiple relationships. Now, in this study, the authors used confidence intervals instead of p-values, but that doesn’t change anything. A 95% confidence interval is equivalent to a p-value of 0,05, and so the same rules apply.
When you look at multiple relationships at the same time, you are supposed to correct for it. One way to correct is by using a method called the Bonferoni correction formula. This formula is very simple to understand. Say you have a p-value of 0,05 when looking at one relationship (the standard p-value in medical science). If you instead look at two relationships, you divide your p-value by two, thus getting a new p-value for significance of 0,025. If you are looking at ten relationships, you divide by ten, thus getting a new p-value of 0,005.
The authors who performed this study used a 95% confidence interval, as though they were only looking at one relationship between two variables. But they were in fact looking at a ton of variables (they never even specify how many) and a huge number of relationships, so they should have set their confidence interval much more widely.
They did have some results that they claimed were statistically significant, which I haven’t bothered to mention yet, because they’re certainly not significant after statistical correction.
For example, the authors claim a significant correlation between the Gross Domestic Product and covid deaths (relative risk 1,03, 95% confidence interval 1,00 to 1,06), and a significant correlation between the number of nurses per million population and covid deaths (relative risk 0,99, 95% confidence interval 0,99 to 1,00). But if you adjust, as they should have done, for looking at a large number of variables, then there is no way these results would still have been statistically significant. Sorry nurses.
So, what can we conclude from all this?
First of all, lockdowns do not seem to reduce the number of covid deaths in a country. Oops. Based on this data, if you want to decrease the number of covid deaths, you should encourage more people to start smoking, and possibly also start a communist revolution, to equalize wealth as far as possible.
Just kidding. As I’ve mentioned, the data is observational, so we can’t say anything about causality. What we can say from this is that lockdowns don’t seem to work – if they have any effect at all, it is too weak to be noticeable at a population level.
The other important finding from this study, from my perspective, is the strong link between obesity and risk of dying from covid. We can’t say that obesity in itself increases risk of dying – people who are obese have so many different biological systems malfunctioning at the same time that it’s impossible to say whether obesity is the cause of increased risk of death or just a marker of poor health in general.
Regardless, obesity is the strongest covid risk factor that we can do something about. And even if it isn’t the obesity itself that kills people, when we fix the obesity, we also fix the many derangements in metabolism and immune function that go along with it. So it is reasonable to think that efforts to decrease the rate of obesity in the population would decrease the number of people dying of covid. That is where we should be putting our efforts as a society right now – making people healthier so that their bodies are able to fight off covid (and cancer, and heart disease, and dementia, and all the other things that preferentially kill people with sub-optimal health).
You might also be interested in my article about whether vitamin D can be used to treat covid, or my article about whether a low fat or low carb diet is more effective for weight loss.