31 Adjustment: What Is Adjustment in Observational Research?

Transcript

we've talked so much in this course about how studies can go wrong and how interpretation can go wrong it's important to address the fact that there's some ways to fix these problems and one of these pro one of the ways to fix the problems with observational research in particular is called adjustment so we're going to talk about adjustment we'll d... Read More

Key Insights

• Association is not causation: The rosé analysis produced a p-value of .004 after hundreds of observations, yet the causal diagram treats the rosé-to-shortness connection as nonexistent. The example shows that a statistically notable association can arise even when the proposed exposure does not cause the outcome. Identifying the structure behind the data is therefore essential to interpretation.
• Raw percentages can mislead: Thirty percent of rosé drinkers were under five foot four inches, compared with 15 percent of non-rosé drinkers. Those figures accurately describe the observed sample, but they do not explain why the groups differ. Without examining variables connected to both drinking behavior and height, the comparison can support a false causal story.
• Confounders connect both variables: A confounder is a third factor linked to both the exposure and the outcome of interest. In the bar example, sex may influence whether someone drinks rosé and is also related to average height. Those two connections can induce a correlation between rosé and shortness without any genuine shrinkage effect.
• Randomization changes the problem: The rosé observations came from a bar rather than a randomized trial, leaving open the possibility that drinkers and non-drinkers differ in relevant ways. The fresh-fruit study likewise did not randomize people to eating or avoiding fruit. Adjustment is introduced specifically as a way to address such confounding in observational research.
• Stratification creates fairer comparisons: Instead of comparing all rosé drinkers with all non-drinkers, stratification compares exposure groups within the same sex category. This holds the proposed confounder constant within each analysis. The method is simple because sex is treated categorically, allowing the researcher to inspect women and men separately rather than combining them.
• Equal rates reveal confounding: Among women, 30 percent of rosé drinkers and 30 percent of non-rosé drinkers were short. The equality contrasts with the unstratified comparison of 30 percent versus 15 percent. Once sex is controlled through stratification, the apparent rosé association disappears, supporting the explanation that the original result was produced by group composition.
• Adjustment has many forms: Stratification is described as one of the easiest ways to adjust for a potential third factor, but statistical adjustment extends much further. More complex mathematics can handle factors that are not categorical, along with continuous factors, non-normal distributions, multiple groups, and several variables at once. These approaches share the same underlying objective.
• Measurement enables adjustment: The logic of adjustment depends on measuring the third factors that might influence the relationship of interest. Researchers then use those measurements to separate the exposure-outcome relationship from the confounders’ influence. If a confounder is unmeasured or poorly measured, the current fields warn that its influence may remain and bias the result.
• Fruit consumption needs context: The 2016 New England Journal study followed people across China after asking how much fresh fruit they ate. Those eating more fresh fruit were less likely to experience heart attacks, strokes, and other cardiovascular outcomes. Because fruit intake was not randomized, that favorable association could reflect characteristics of fruit eaters rather than fruit alone.
• Income offers one explanation: People with more money may be able to afford more fresh fruit, while people with more money also tend to live longer because they have access to more healthy things. Income can therefore be related to both the exposure and the outcome. Adjusting for income attempts to cut that alternative causal line.
• Lifestyle factors can overlap: Smoking status and salt preference are also proposed as possible confounders in the fruit study. Non-smokers may eat more fruit, while people who prefer less salt may also consume more fruit. If low salt intake is protective against cardiovascular outcomes, part of the apparent fruit benefit could instead reflect that associated behavior.
• Adjustment remains limited: Accounting for measured confounders does not guarantee that the adjusted relationship is causal. Unmeasured or poorly measured factors can continue influencing the result, and linear assumptions may miss thresholds or non-linear relationships. Adjustment tries to make observational comparisons resemble a randomized trial, but the existing material explicitly says it cannot replace evidence from a well-designed randomized trial.

Questions & Answers

Q: What is adjustment in observational research?

Adjustment is the broad mathematical process of trying to represent the pure relationship between an exposure and an outcome. It addresses confounders, which are third factors connected to both variables being studied. Researchers measure those factors and account for them so their influence is less likely to distort the exposure-outcome comparison. The method is especially relevant when participants were not randomly assigned to the exposure.

Q: Why did rosé drinking appear related to shortness?

The researcher observed that 30 percent of rosé drinkers were under five foot four inches, compared with 15 percent of non-rosé drinkers. The data came from observations in a bar, not from random assignment, so the two drinking groups could differ in other ways. Sex was proposed as a confounder because women may be more likely to drink rosé and are, on average, shorter than men. That shared factor created an apparent association even though rosé did not cause shrinkage.

Q: How does stratification adjust for a confounder?

Stratification separates the data into groups defined by a potential confounder and repeats the comparison within each group. In the rosé example, the analysis is conducted separately among women and among men. This holds sex constant while comparing rosé drinkers with non-drinkers. The method works here because it prevents differences in the groups’ sex composition from driving the apparent relationship with height.

Q: What happened after the rosé data were stratified by sex?

Among women, 30 percent of rosé drinkers were short and 30 percent of non-rosé drinkers were short. The difference seen in the overall sample therefore disappeared within the female group. This result indicates that sex, rather than rosé, could explain the original association with shortness. Stratification exposed that explanation by comparing people within the same sex category.

Q: What is multivariable adjustment?

Multivariable adjustment accounts for multiple potential confounders at the same time through statistical models. It expands the basic idea behind stratification to continuous factors, non-normal distributions, multiple groups, and several variables simultaneously. Researchers use it when a simple categorical split cannot represent every relevant third factor. The goal remains to isolate the relationship between the exposure and outcome after accounting for measured confounders.

Q: Why was adjustment needed in the fresh-fruit study?

The 2016 New England Journal study examined fresh-fruit consumption and major cardiovascular disease among people across China. People eating more fresh fruit were less likely to have heart attacks, strokes, and other cardiovascular outcomes, but fruit consumption was not randomly assigned. Income, smoking status, and salt preference could be related to both fruit intake and cardiovascular health. Adjustment attempts to separate the fruit association from those alternative explanations.

Q: How does adjustment cut causal lines from confounders?

A causal diagram can show a confounder connected to both the exposure and the outcome. Adjusting for that factor attempts to block its contribution to the observed exposure-outcome relationship. In the fruit example, adjustment for income targets the connection between money and fresh-fruit consumption, while adjustment for smoking status targets the smoking-related line. This helps estimate the relationship that remains after those measured influences are considered.

Q: Can adjustment prove that an observational relationship is causal?

No, adjustment does not guarantee a causal conclusion. It can account only for confounders that were measured and represented adequately in the analysis. Unmeasured or poorly measured confounders may still influence the observed relationship, while linear assumptions can miss thresholds or non-linear patterns. Adjustment may make observational data resemble a randomized comparison, but it cannot replace a well-designed randomized trial.

Summary & Key Takeaways

• Introducing adjustment in research: Observational studies and their interpretation can go wrong, but some methods try to address those problems. Adjustment is the broad process of using mathematics to represent the pure relationship between an exposure and an outcome. It can take several forms, from simple stratification to more complicated multivariable methods. Its purpose is to account for third factors that influence the relationship being studied, although the lesson warns that adjustment may not be as reliable as it first appears.

• Testing the rosé hypothesis: A young researcher notices that people drinking rosé seem shorter than people drinking other beverages and hypothesizes that rosé causes shrinkage. At a local bar, he records each person’s height and whether the person is drinking rosé. After hundreds of observations, his statistical test produces a p-value of .004. He also finds that 30 percent of rosé drinkers are under five foot four inches, compared with 15 percent of non-rosé drinkers.

• Finding the hidden confounder: The observed association does not mean that rosé causes people to shrink. Because the observations did not come from a randomized trial, a third factor could be connected to both beverage choice and height. Sex is proposed as that factor: women may be more likely to drink rosé, and women are, on average, shorter than men. This confounder can therefore create an apparent relationship between rosé drinking and shortness even when no real causal link exists.

• Correcting through stratification: The simplest adjustment method presented is stratification, which repeats the analysis within categories of the potential confounder. When the rosé data are examined only among women, 30 percent of rosé drinkers are short and 30 percent of non-rosé drinkers are short. The original difference disappears within that group. Analyzing women and men separately therefore adjusts for sex and demonstrates how stratification can expose an association created by a third factor.

• Expanding to multiple variables: Statistical methods extend stratification to continuous factors, non-normal distributions, multiple groups, and multiple variables simultaneously. A 2016 New England Journal study of fresh-fruit consumption and major cardiovascular disease in China illustrates the need. People eating more fresh fruit experienced fewer heart attacks, strokes, and other cardiovascular outcomes, but they were not randomized. Income, smoking status, and salt preference could affect both fruit consumption and health, so researchers adjust for these measured confounders while recognizing that adjustment remains incomplete.