Controls and Confounding

The text below is a transcript of the video above, edited with AI assistance for clarity and flow. Where the two differ, the video is the authoritative version.

You’ll learn to

You have learned about confounding variables and you have learned about multivariate regression. Now it is time to bring those two pieces of knowledge together and understand how multivariate regression helps mitigate concerns about confounding variables, without being a definitive answer to them. In this video you will learn to:

  • Explain why multivariate regression helps address confounding.
  • Interpret what happens when a coefficient “goes away”: when the same variable sees its coefficient diminish or lose statistical significance as you go from a bivariate regression to a multivariate regression.
  • Recognise the limits of controlling for confounders in multivariate regression.

Recall: Confounding variables

Let’s recall the definition. A confounding variable is a variable that affects both X and Y, and that is not itself affected by X. It creates a spurious association, and it might generate the false impression of a causal relationship between the two variables. We have identified confounding as a threat to causal claims.

Multivariate regression lets us control for potential confounders. It is important to understand what we mean when we say that we control for something, and that should become clear as we go through the examples.

How controlling works

When we have two variables, X and Y, and we are suspicious about a confounder, we can include that confounder in a multivariate regression. Once the confounder is included, the coefficient on the independent variable describes the effect of the independent variable on the dependent variable holding the confounder constant.

If the coefficient of the independent variable shrinks or disappears, it is likely that confounding was present. We are going to look at two examples.

Example 1: Immigration and the Leave vote

The first example comes from the previous video: immigration and the Leave vote. When we run a bivariate regression with the Leave vote as the dependent variable and immigration as the independent variable, we see that regions with more immigrants voted Leave less.

Someone could look at that and think: maybe people who live in a place with more immigrants come into contact with those immigrants, like them, become more open-minded, and then vote against Leave because they want immigrants from Europe to come in. You could make all sorts of assertions and imagine a causal relationship between these two things if you only look at the bivariate regression.

In the bivariate regression the relationship is negative, with a slope of \(b = -0.41\), and it is statistically significant (\(p < 0.001\)). We can reject the null hypothesis that the slope is really zero. We can be confident that there is a correlation, an association, between these two variables. What we cannot be confident of is the causality. One of the reasons is that the level of education of the region may be a confounding variable here: something that could be driving both immigration and the Leave vote.

Education as a confounder

Why would education be a potential confounder in this example? More educated regions tend to attract more immigrants. Immigrants tend to go to places like London, and large cities in general are more attractive to immigrants because there are more jobs and better economic opportunities. Educated regions also voted Remain, as we have seen in previous examples.

So we have a reasonable expectation that education, measured as the percentage of people in a region with higher education, affects both the explanatory variable and the outcome: both the independent variable and the dependent variable, both X and Y.

Controlling for education

We can then control for education. We start from the bivariate regression, and when we move to a multivariate regression and add a potential confounder, we are controlling for that confounder.

Bivariate Multivariate
Immigration -0.41 0.08
Education (not included) -1.14

In the multivariate regression we are looking at the correlation between immigration and the Leave vote holding education constant. The comparison now applies only to regions with similar levels of education. Once we do that, the coefficient for immigration is near zero, and it is no longer statistically significant. The effect of immigration disappears once education is held constant, once we compare only between regions with similar levels of education.

This is an indicator that education is possibly a confounder here. When we observe this type of result, a bivariate regression where immigration is negative and statistically significant, and a multivariate regression where it is near zero and no longer statistically significant, then education is probably a confounder. We should not conclude that the causal narrative holds, the one where a place having more migrants sets off a reaction in which people become more open to immigration. There is a confounder, a key variable that is likely driving both.

Example 2: Health spending and CO2 emissions

The second example comes from our earlier discussion of confounding variables: health care spending and carbon dioxide emissions. Suppose we run a bivariate regression where the dependent variable is carbon dioxide emissions and the independent variable is health care expenditure, both per capita.

We do observe a strong positive correlation between the two variables. The slope is \(b = 0.73\) and the p-value is smaller than 0.001. So yes, countries that spend more on health pollute more. We do see this association. But does one thing cause the other? That is unlikely. So what is going on?

Controlling for GDP per capita

Let’s control for GDP per capita. We identified national wealth, or income, as a potential confounder: richer countries spend more on health care and also pollute more.

Bivariate Multivariate
Health spending 0.73 -0.04
GDP per capita (not included) 0.91

In the bivariate regression, with carbon dioxide emissions as the dependent variable, health care spending has a positive association. Once GDP per capita is in the picture, in a multivariate regression controlling for GDP per capita, the coefficient for health spending is near zero and no longer statistically significant. The relationship between health care spending and carbon dioxide emissions vanishes once GDP per capita is accounted for: once we look at the correlation between health spending and emissions while comparing only countries with similar levels of GDP per capita.

Once again, this is a strong indicator that our previous assumption about GDP per capita being a confounding variable was correct. It is empirical evidence consistent with GDP per capita being a confounder in this case. National income is driving both variables: both health care spending and carbon dioxide emissions per capita.

Does controlling for confounders prove causation?

Suppose we have a bivariate regression, we control for confounders, and we still find a statistically significant result. Does that prove causation? The answer is no. There can always be confounding variables that we cannot observe, and there can always be confounding variables that we are not aware of. On top of that, controls do not address concerns about reverse causality.

So yes, multivariate regression helps address concerns about confounding variables, but it is still not going to give us smoking-gun evidence about causality. There are other types of design and other types of estimation available to social science researchers that get closer to establishing causality. Adding controls helps you get closer to it, but it does not get you all the way there.

Recap

  • Adding a confounder to a regression can reveal spurious relationships. In both examples, including a confounder made the coefficient from the bivariate regression go away.
  • A coefficient that “goes away” is a strong suggestion that confounding was present when we were looking only at the bivariate regression.
  • Controls reduce confounding concerns, but they are not in themselves sufficient to establish a causal relationship.

Why this matters

Most social science data is observational; in a lot of cases you cannot really do experiments. Multivariate regression is a very important tool for addressing confounders, but understanding its limits is also important for interpreting research critically.