What is the difference between confounding and interaction

This would require 96 separate strata to control for all of these variables simultaneously, and as you increase the number of strata, you keep whittling away at the number of people in each stratum, so sample size becomes a major problem, since many of the strata will contain few or no people.

It is possible to minimize confounding by utilizing certain strategies in the design of a study:. There are also analytical techniques that provide a way of adjusting for confounding in the analysis, provided one has information on the status of the confounding factors in the study subjects.

These techniques are:. The term effect modification is applied to situations in which the magnitude of the effect of an exposure of interest differs depending on the level of a third variable.

Reye's syndrome is a rare, but severe condition characterized by the sudden development of brain damage and liver dysfunction after a viral illness. The syndrome is most commonly seen in children between the ages of who have been treated with aspirin while recovering from a viral illness, most commonly chickenpox or influenza.

Fortunately, Reye's syndrome has become very uncommon since aspirin is no longer recommended for routine use in children. While Reye's syndrome can occur in adults, it is distinctly more common in children. Thus, the effect of aspirin treatment for a viral illness is very clearly modified by age. In this situation, computing an overall estimate of association is misleading. One common way of dealing with effect modification is examine the association separately for each level of the third variable.

For example, if one were to calculate the odds ratio for the association between aspirin treatment during a viral infection and development of Reye's syndrome, the odds ratio would be substantially greater in children than in adults.

As another example, suppose a clinical trial is conducted and the drug is shown to result in a statistically significant reduction in total cholesterol. However, suppose that with closer scrutiny of the data, the investigators find that the drug is only effective in subjects with a specific genetic marker and that there is no effect in persons who do not possess the marker. The effect of the treatment is different depending on the presence or absence of the genetic marker. This is an example of effect modification or "statistical interaction".

Consider the following clinical trial conducted to evaluate the efficacy of a new drug to increase HDL cholesterol the "good" cholesterol. One hundred patients are enrolled in the trial and randomized to receive either the new drug or a placebo.

Background characteristics e. Subjects are instructed to take the assigned medication for 8 weeks, at which time their HDL cholesterol is measured again. The results are shown in the table below. On average, the mean HDL levels are 0. Based on their preliminary studies, the investigators had expected a statistically significant increase in HDL cholesterol in the group treated with the new drug, and they wondered whether another variable might be masking the effect of the treatment.

In this study, there are 19 men and 81 women. The table below shows the number and percent of men assigned to each treatment. There is no meaningful difference in the proportions of men assigned to receive the new drug or the placebo, so sex cannot be a confounder here, since it does not differed in the treatment groups.

However, when the data are stratified by sex, they find the following:. This is an example of effect modification by sex, i. In this case there is no apparent effect in women, but there appears to be a moderately large effect in men.

Note, however, that the comparison in men is based on a very small sample size, so this difference should be interpreted cautiously, since it could be the result of random error or confounding. When there is effect modification, analysis of the pooled data can be misleading. In this example, the pooled data men and women combined , shows no effect of treatment. Because there is effect modification by sex, it is important to look at the differences in HDL levels among men and women, considered separately.

In stratified analyses, however, investigators must be careful to ensure that the sample size is adequate to provide a meaningful analysis.

Consider the following hypothetical study comparing hospitalization after a motor vehicle collision for male and female drivers. In this case, the crude analysis suggests an association between male gender and frequency of hospitalization for motor vehicle collisions.

Another good example of effect modification is seen with skin cancers. It is well established that excessive exposure to UV irradiation increases one's risk of skin cancer.

However, the risk of UV-induced skin cancer is 1, times greater in people with xeroderma pigmentosum. This is a rate hereditary defect autosomal recessive in the enzyme system that repairs UV-induced damage to DNA. It is characterized by photosensitivity, pigmentary changes, premature skin aging, and greatly increased susceptibility to malignant tumor development.

If effect modification is present, it is NOT appropriate to use Mantel-Haenszel methods to combine the stratum-specific measures of association into a single pooled measurement. Effect modification is a biological phenomenon that should be described, so the stratum-specific estimates should be reported separately.

In contrast, confounding is a distortion of the true association caused by an imbalance of some other risk factor. Note that in this situation you are only pooling the stratum-specific estimates in order to make a decision about whether confounding is present; you should not report the pooled estimate as an "adjusted" measure of association if there is effect modification.

While the discussion above provides a standard description of effect modification, but on closer scrutiny the concept of effect modification is more complicated than this.

We see two scenarios in which incidence rates in exposed and unexposed individuals are assessed at different ages. Rate ratio and rate difference are both measures of effect, but depending on which we use, our conclusions about effect modification differ. In the first scenario the rate difference remains constant across the spectrum of age, suggesting no effective modification. In the second scenario the rate ratio remains relatively constant, but the rate difference increases with age.

Our conclusion regarding whether or not there is effect modification will depend on which measure of effect we use. Consider also the hypothetical data on the risk of lung cancer in smokers and non-smokers, both with and without exposure to asbestos also adapted from Rothman. First consider the effect of asbestos on the risk associated with smoking.

The risk ratio is 5 both with and without asbestos exposure, suggesting no effect modification. However, the risk difference 4 per , without asbestosis and 40 per , with asbestosis exposure.

This effect measure is clearly modified by asbestos. We can also look at the effect of smoking on the risk associated with asbestos. The risk ratio for asbestos exposure compared to no asbestos exposure is 10 in both smokers and non-smokers, suggesting an absence of effect modification.

However, the risk difference is 45 per , in the presence of smoking, but only 9 per , in the absence of smoking. Thus, the risk ratios suggest no effect modification, but the risk differences suggest substantial effect modification. Rothman argues that this ambiguity regarding effect measure modification and statistical interaction makes it important to make a distinction between statistical interaction which is ambigous and biological interaction which is not ambiguous; it is either present or absent.

Biological interaction between two causes occurs if the effect of one is dependent on the presence of the other. For example, exposure to the measles virus is a component cause of developing measles, but it is dependent on another factor, i. Someone who is immune because of vaccination or having already had measles will not experience any effect from exposure to the measles virus. A discussion of the methods for measuring biological interaction is beyond the scope of this module.

Those who are interested should refer to the discussion in Rothman's excellent text. Unrestrained no seatbelt or air bag :. Restrained with Seatbelt, Air Bag, or Both.

In "Epidemiology - An Introduction" Ken Rothman says the following about this complexity: "The research process of learning about and controlling for confounding can be thought of as a walk through a maze toward a central goal. It is possible to minimize confounding by utilizing certain strategies in the design of a study: Restriction Matching Randomization in intervention studies only There are also analytical techniques that provide a way of adjusting for confounding in the analysis, provided one has information on the status of the confounding factors in the study subjects.

These techniques are: Stratification Multiple variable regression analysis. Lots of resources are saying the same thing. Thank you for brief and clear explanation the action of each and their difference. It make me to think in other way to act up on when I will deal with my analysis. Thank you Kotecha for the nice explanation.

You said that confounders need to be eliminated to prevent distortion of results. Would you eliminate by not including it in the analysis?

Thanks Deevia, your explanation and examples helps me in understanding Effect modification and confounders. Will surely like to keep in touch with you for future. Effect modification is a question related to an exposure: 1 that shows independent of the exposure and outcome 2. Links: 1. A study was conducted to assess the extrapyramidal side effects of a new antipsychotic drug in patients with schizophrenia.

Many of these patients were smokers and some of them were on anticholinergic drugs. What was the role of the anticholinergic drugs in this study? Learn more about the measures of central tendency mean, mode, median and how these need to be critically appraised when reading a paper. Chris set up a student journal club in his first year of Physical Therapy training.

In this blog, he describes how he started the club, how it has changed and expanded throughout his studies, and provides tips and suggested papers to get you started. A brief guide to prevalence and incidence with definitions, explanations and example calculations. What are the key steps in EBM?

Miettinen OS Confounding and effect-modification. Miettinen OS Causal and preventive interdependence. Elementary principles. Pearce N Multistage modelling of lung cancer mortality in asbestos textile workers. Pearce N Analytical implications of epidemiological concepts of interaction.

Pearce N, Checkoway H Case-control studiesusing other diseases as controls: problems of excluding exposure-related diseases. Pearl J Causality: models, reasoning and inference. Phillips CV Quantifying and reporting uncertainty from systematic errors. Robins JM The control of confounding by intermediate variables. Robins JM Data, design, and background knowledge in etiologic inference. Rothman KJ Causes.

Rothman KJ No adjustments are needed for multiple comparisons. Epidemiology —46 CrossRef Google Scholar. Lippincott-Raven, Philadelphia Google Scholar. Saracci R The interactions of tobacco smoking and other agents in cancer etiology. Epidemiologic Reviews — Google Scholar. Schlesselman JJ Assessing effects of confounding variables. American Journal of Epidemiology —8 Google Scholar. Schlesselman JJ Case-control studies: design, conduct, analysis. Siemiatycki J, Thomas DC Biological models and statistical interactions: an example from multistage carcinogenesis.

Steenland K, Thun M Interaction between tobacco smoking and occupational exposures in the causation of lung cancer. Journal of Occupational Medicine — Google Scholar. Thomas DC, Greenland S The relative efficiencies of matched and independent sample designs for case-control studies. Weinberg CR Toward a clearer definition of confounding. Yanagawa T Case-control studies: assessing the effects of a confounding factor. Neil Pearce 1 Sander Greenland 2 3 1.

Personalised recommendations. Cite chapter How to cite?