Confounding factor versus effect modifcation

[Doingmybest86]

Can someone please explain to me the difference between a confounding factor and effect modification. Thanks in advance.

---

Members don't see this ad.

 

[zhopv10]

Tricky concept: so with effect modification the the effect of an exposure on an outcome is modified by another variable. So say you are looking at DVT's and you are checking the effect of estrogen but you have smokers in the mix smoking will modify the outcome and they variable (so modifying estrogen in the smoking group may have a much larger effect). Now confounding is an unforeseen variable that modifies the outcome. Subtle difference but here is the key, you can stratify an effect modifying variable (I.e looking at smokers and non smokers and then look at the effect of estrogen on DVT). On the other hand you cannot stratify out for confounding it is due to unforeseen variables lurking in the background if you will and is a form of bias. You try to remove it with randomization. Effect modification is NOT a bias.

An example of confounding: in a historical prospective study I was doing once we were looking at the effect of receipt of radiation on 5 yr survival in women with breast cancer. Now the raw HR was showing an increased risk in recurrence with radiation. This of course isn't true. So we did multi variate analysis basically stratifying by severity of dz, risk factors etc to see how this changed (stratifying and correcting for known effect modifying variables). In the end however out HR still ended up showing radiation as a risk of recurrence. This is because of course when you were looking at women who received radiation vs those who didn't those who did tended to have worse dz (hence why they were getting it) however, all the reasoning or factors behind why the dz was worse was not/could not be translated into our variables and thus there were unknown variables in the background that we could not adequately control for. That was the confounding, there were confounding variables that we just simply could not account for (we even tried propensity scoring). Hopefully that explanation and real world example helps


Sent from my iPhone using SDN mobile app

 

[zhopv10]

In short: effect modification: a variable that effects the exposure/outcome but you know about it whether that is through precious data or becomes apparent and you can stratify the groups by it. It is not bias and you report it in your data.

Confounding is a variable that is not accounted for that is altering the outcome. It is a form of bias and you try to remove it with randomization etc.

@dempty

Sent from my iPhone using SDN mobile app

 

[teehees]

I think you got it switched at the very end

Confounding is both exposure and outcome . effect modification is just outcome .

 

[zhopv10]

I think you got it switched at the very end

Confounding is both exposure and outcome . effect modification is just outcome .

You sure? It seems like both could just as easily affect the exposure and the outcome? I mean working off the name it would def make sense that it is just the effect haha, but consider this example. In my example above, receipt of radiation and the affect on 5 yr when analyzing recurrence free survival, tumor size was used as a stratifying variable as it effects the outcome, but it also intrinsically affects the exposure too (those patients are more likely to receive radiation). Thus, in that case it seems that the effect modification is plausibly affecting both. I haven't read anything beyond what Uworld had to say about effect modification though so I could very well be off, thoughts?


Sent from my iPhone using SDN mobile app

 

[teehees]

First Aid says clearly that confounding affects both.

A quick look at the other threads discussing this on this forum says that too. Everything else what you said seems right though

Confounding is an unexpected bias. Stratifying can help identify and minimise it.
Effect Modification affects only the outcome and is something thats recognised and adjusted for (like OCP vs Breast Cancer --> Family History)

That has been my understanding of this but like always i'm ready to be wrong

 

[zhopv10]

First Aid says clearly that confounding affects both.

A quick look at the other threads discussing this on this forum says that too. Everything else what you said seems right though

Confounding is an unexpected bias. Stratifying can help identify and minimise it.
Effect Modification affects only the outcome and is something thats recognised and adjusted for (like OCP vs Breast Cancer --> Family History)

That has been my understanding of this but like always i'm ready to be wrong

Agreed on confounding for sure, the effect modification part just seems sorta weird to me but ya I'm no biostatistician that's for sure! Haha either way though I think going off your definitions for the test at least is safe enough and considering effect modification isn't even in FA I'd assume it's not the most high yield thing anyway. Thanks for bringing that point up


Sent from my iPhone using SDN mobile app

 

[amberkre]

I don't know if this will help but when thinking about stats i really like to use examples.

Effect modification: its basically any variable that positively or negatively changes the observed effect of the risk factor on disease status. So that different groups within the population have different risk estimates.

Ex: Your study finds that there is no risk between BRCA2 mutations and breast cancer. But when you stratify the results based on gender you find that men have an increase risk of breast cancer if they have BRCA2 mutation, whereas women have less of a risk. (this is totally made up just an example)

Confounding Variable: the associated between the risk factor and the disease outcome is distorted by another variable

Ex: Your study finds that there is an increased risk of hepatocellular carcinoma in a population of construction workers, and so you conclude that theres must be some environmental exposure in the work place causing hepatocellular carcinoma when in reality there is a hidden "confounding variable" that construction workers are more likely to be alcoholics. The alcoholism is the confounding variable increasing the risk of hepatocellular carcinoma (also this example is totally made up, hope no one takes offense haha)

 

[slz1900]

Confounding doesn't have to be unexpected. It just means that there is a difference in the control vs intervention groups that is affecting the outcome. Ie, let's say you have a trial where you recruit people 18-80 years old to try a new drug and you split them into two groups. If you randomized and 1 group had mostly young people and the other group was mostly old, then age is a confounder. Confounders can also be hidden associations, like 1 group is from a poorer neighborhood compared to the other and therefore lacking a good diet.

Importantly, confounding is actually OK and inevitable in some cases. For instance, epidemiologic studies will report that African Americans have higher rates of sarcoidosis. Obviously this is an extremely confounded variable, since you are comparing two groups that may have different socioeconomic status, live in different areas, etc, and controlling for each confounder and using it as a variable that you would then try and match would be impossible, especially since you don't know what some of the confounders are.

Effect modification means that within a group there is a third variable that influences the outcome. For example, let's say the study with 18-80 year olds didn't show an improvement with drug X compared to placebo, and the average age of both groups is 40. Age is no longer a confounder because both groups have the same average age. However let's say you do a stratified analysis and people 18-30 did respond to drug X while older people didn't. That is effect modification because it is a third variable influencing the outcome. It is not confounding because it is not a difference between the control and intervention groups. Matching cannot prevent effect modification but it can prevent confounding.

tldr: confounding is a difference between control and intervention groups that adds uncertainty to the outcome. Effect modification is a third variable that is influencing the results but that is not due to differences between control and intervention.

I'm sure this was way oversimplified so feel free to correct.

 

[dempty]

Effect modification is commonly referred to as interaction in statistics (if you want to read for further clarification). It means that the relationship between the dependent variable (response variable) and one of the independent variables is altered/moderated by another independent variable (factor, exposure, etc-- possibly more than one). This does not mean that the factor and response are influenced by the effect modifier. An effect modifier merely alters how another independent variable impacts the DV. This is a subtle but important difference. In other words, an effect modifier (say, X2) can be completely independent of some other factor/treatment (say, X1), and X2 can still change the way X1 effects Y (the outcome). An example is that the odds of lung cancer increase/decrease more quickly/slowly in women who smoke compared with men who smoke (made up example, but think that men and women could have different odds ratios-- or you could use age instead of gender in the example). In other words, depending on gender (or age), there is a different effect of smoking on the odds of developing lung cancer. As mentioned before, this isn't a bias.

The easiest way to remember a confounding variable is to understand that it is something we haven't measured (properly or at all) that is muddying our understand of the relationships under examination. The confounder(s) influence the dependent variable and can influence the independent variable(s) as well. Naturally, if you haven't accounted for a confounder, you won't be able truly see the effect that the independent variable has on the DV (because the confounder has an unmeasured effect). Failing to account for this variable can lead to bias in the estimated effects of factors on the response. For example, you're examining how a dosing regimen of antibiotics clears an infection, but you failed to account for patient age (or other variables that probably impact how we respond to infections or uptake drugs). It might appear that one specific regimen is useful, but if we accounted for age, we would see that the influence is age-related instead of the regimen. (In the example, I intended to mention that we did not have a design balanced with respect to age, so it was the younger group (perhaps) that had a better clearance, but we attributed it to the regimen.)

Hope this helps.

Edited for clarification.

 

[zhopv10]

Effect modification is commonly referred to as interaction in statistics (if you want to read for further clarification). It means that the relationship between the dependent variable (response variable) and one of the independent variables is altered/moderated by another independent variable (factor, exposure, etc-- possibly more than one). This does not mean that the factor and response are influenced by the effect modifier. An effect modifier merely alters how another independent variable impacts the DV. This is a subtle but important difference. In other words, an effect modifier (say, X2) can be completely independent of some other factor/treatment (say, X1), and X2 can still change the way X1 effects Y (the outcome). An example is that the odds of lung cancer increase/decrease more quickly/slowly in women who smoke compared with men who smoke (made up example, but think that men and women could have different odds ratios-- or you could use age instead of gender in the example). In other words, depending on gender (or age), there is a different effect of smoking on the odds of developing lung cancer. As mentioned before, this isn't a bias.

The easiest way to remember a confounding variable is to understand that it is something we haven't measured (properly or at all) that is muddying our understand of the relationships under examination. The confounder(s) influence the dependent variable and can influence the independent variable(s) as well. Naturally, if you haven't accounted for a confounder, you won't be able truly see the effect that the independent variable has on the DV (because the confounder has an unmeasured effect). Failing to account for this variable can lead to bias in the estimated effects of factors on the response. For example, you're examining how a dosing regimen of antibiotics clears an infection, but you failed to account for patient age (or other variables that probably impact how we respond to infections or uptake drugs). It might appear that one specific regimen is useful, but if we accounted for age, we would see that the influence is age-related instead of the regimen.

Hope this helps.

Golden, now you put it that way I realized that I'd only really heard of effect modification referred to as interaction (I remember reading up a bit on it when I was figuring out how to run interaction terms in our cox ph analyses). I hadn't put together that they are the same terms and I think that's part of what was confusing me, totally makes more sense now thanks!


Sent from my iPhone using SDN mobile app

 

[dempty]

I don't know if this will help but when thinking about stats i really like to use examples.

Effect modification: its basically any variable that positively or negatively changes the observed effect of the risk factor on disease status. So that different groups within the population have different risk estimates.

Ex: Your study finds that there is no risk between BRCA2 mutations and breast cancer. But when you stratify the results based on gender you find that men have an increase risk of breast cancer if they have BRCA2 mutation, whereas women have less of a risk. (this is totally made up just an example)

Confounding Variable: the associated between the risk factor and the disease outcome is distorted by another variable

Ex: Your study finds that there is an increased risk of hepatocellular carcinoma in a population of construction workers, and so you conclude that theres must be some environmental exposure in the work place causing hepatocellular carcinoma when in reality there is a hidden "confounding variable" that construction workers are more likely to be alcoholics. The alcoholism is the confounding variable increasing the risk of hepatocellular carcinoma (also this example is totally made up, hope no one takes offense haha)

These are some other good examples of an effect modifier vs. a confounder.

 

[Doingmybest86]

Tricky concept: so with effect modification the the effect of an exposure on an outcome is modified by another variable. So say you are looking at DVT's and you are checking the effect of estrogen but you have smokers in the mix smoking will modify the outcome and they variable (so modifying estrogen in the smoking group may have a much larger effect). Now confounding is an unforeseen variable that modifies the outcome. Subtle difference but here is the key, you can stratify an effect modifying variable (I.e looking at smokers and non smokers and then look at the effect of estrogen on DVT). On the other hand you cannot stratify out for confounding it is due to unforeseen variables lurking in the background if you will and is a form of bias. You try to remove it with randomization. Effect modification is NOT a bias.

An example of confounding: in a historical prospective study I was doing once we were looking at the effect of receipt of radiation on 5 yr survival in women with breast cancer. Now the raw HR was showing an increased risk in recurrence with radiation. This of course isn't true. So we did multi variate analysis basically stratifying by severity of dz, risk factors etc to see how this changed (stratifying and correcting for known effect modifying variables). In the end however out HR still ended up showing radiation as a risk of recurrence. This is because of course when you were looking at women who received radiation vs those who didn't those who did tended to have worse dz (hence why they were getting it) however, all the reasoning or factors behind why the dz was worse was not/could not be translated into our variables and thus there were unknown variables in the background that we could not adequately control for. That was the confounding, there were confounding variables that we just simply could not account for (we even tried propensity scoring). Hopefully that explanation and real world example helps


Sent from my iPhone using SDN mobile app

It did help. Thank you for the explanation.

 

[Doingmybest86]

Effect modification is commonly referred to as interaction in statistics (if you want to read for further clarification). It means that the relationship between the dependent variable (response variable) and one of the independent variables is altered/moderated by another independent variable (factor, exposure, etc-- possibly more than one). This does not mean that the factor and response are influenced by the effect modifier. An effect modifier merely alters how another independent variable impacts the DV. This is a subtle but important difference. In other words, an effect modifier (say, X2) can be completely independent of some other factor/treatment (say, X1), and X2 can still change the way X1 effects Y (the outcome). An example is that the odds of lung cancer increase/decrease more quickly/slowly in women who smoke compared with men who smoke (made up example, but think that men and women could have different odds ratios-- or you could use age instead of gender in the example). In other words, depending on gender (or age), there is a different effect of smoking on the odds of developing lung cancer. As mentioned before, this isn't a bias.

The easiest way to remember a confounding variable is to understand that it is something we haven't measured (properly or at all) that is muddying our understand of the relationships under examination. The confounder(s) influence the dependent variable and can influence the independent variable(s) as well. Naturally, if you haven't accounted for a confounder, you won't be able truly see the effect that the independent variable has on the DV (because the confounder has an unmeasured effect). Failing to account for this variable can lead to bias in the estimated effects of factors on the response. For example, you're examining how a dosing regimen of antibiotics clears an infection, but you failed to account for patient age (or other variables that probably impact how we respond to infections or uptake drugs). It might appear that one specific regimen is useful, but if we accounted for age, we would see that the influence is age-related instead of the regimen. (In the example, I intended to mention that we did not have a design balanced with respect to age, so it was the younger group (perhaps) that had a better clearance, but we attributed it to the regimen.)

Hope this helps.

Edited for clarification.

Thank!