Suppressions can be defined as “a variable which increases the predictive validity of another variable (or set of variables) by its inclusion in a regression equation,” a suppression effect would be present when the direct and indirect effects of an independent variable on a dependent variable have opposite signs.
I run a logistic regression with number of selected confounders, all these confounders are important to include in model (some confounders are statistically important and others are important from the medical point of view). I got a situation in which the relationship between my independent variable and a dependent variable (adjusted OR) becomes larger from the unadjusted one. By analyze each confounder alone, I found that number of these confounders act as suppressions.
The question now, how should we deal with this result? Should we accept the result as it? Or maybe better to exclude the suppressions in my model?
Thanks all
Hi,
When researcher do regression analysis, the expectation from adding any new variable to the regression equation is to get higher R, but if R decrease, then this is called suppression. Therefore, the researcher have to think critically on why the suppressor made that effect. Mostly there is interaction with other variable in the equation. As this phenomenon is rare, I would recommend to report the occurrence of suppression, but for the analysis purpose to take the suppressor off the regression equation.
Hi,
When researcher do regression analysis, the expectation from adding any new variable to the regression equation is to get higher R, but if R decrease, then this is called suppression. Therefore, the researcher have to think critically on why the suppressor made that effect. Mostly there is interaction with other variable in the equation. As this phenomenon is rare, I would recommend to report the occurrence of suppression, but for the analysis purpose to take the suppressor off the regression equation.
A suppression situation occurs when the prediction of the DV improves by adding a variable which is uncorrelated with the DV but related to another IV or set of IVs. The addition of the suppressor to the regression equation usually increases the B coefficient of the previously suppressed predictor or predictors.
One of the best papers addressing this topic is Tzelgov & Hanik (1991) (attached). This work shows in detail that there are several types of suppression situations.
I do not regard suppression as all that unusual or unexpected. If you more clearly delineate an exposure it is not surprising that the effect size increases -you have more clearly identified who is affected. If this an observational study there is an obvious danger of over- interpretation (especially if you have tried multiple potential suppressors) and you need to be clear that this is the case and invite studies that are designed to confirm that this really occurs.
see https://www.researchgate.net/publication/236671039_Epidemiology_An_Introduction
and the chapter on analysis in
https://www.researchgate.net/publication/234015155_Health_disease_and_society_a_critical_medical_geography
I I would expect the R-squared (or some version of it) to go up with the inclusion of the suppressor - as it is a better model .
For some debate see
http://andrewgelman.com/2013/05/26/how-to-understand-coefficients-that-reverse-sign-when-you-start-controlling-for-things/
Book Health, Disease and Society: A Critical Medical Geography
Book Epidemiology: An Introduction
Thanks all for these valuable replies
I did a lot reading about this topic, however, I cant find a direct answer for my question. In my data (observational study), the unadjusted relation between my IV and DV is positive with significant P value, by adjusting to confounders the relation still positive (sig P value) but with higher OR. I believe that all the included confounders are important and exclude any one may lead to obvious bias.
For example: My IV (memory) and the DV (fracture), alcohol one of these confounders that acts as suppression, I believe that such confounder is important to include from the medical point of view even if alcohol not statistically related with the memory or fracture.
Can you explain why you do not trust the results you have found? You presumably had a reason for including these supressor variables and you found that they made a difference? So what is the problem behind your question?
Hi Kelvyn, Thank you very much for your answer
I trust these results as I only included what I think its important, I just ask if this is a normal situation to report a higher OR for the whole model than the unadjusted one in a scientific paper.
I agree with Kelvyn, if you had good reasons to include these predictors and they influenced the model, than this could be a valuable insight. But you should check all other relevant diagnostics and if all assumptions for the regression model are met. If the diagnostics give no rise for concern and all assumptions are met, then these are good arguments for your model. Maybe you should also consider bootstrapping methods.
From a practical point of view, I would report the whole model (since you argued that you had good reasons to include the specific predictors). If you want to make it clear and transparent for the reader and reviewers, report both (which could cause length problems) so that everyone can draw their own conclusions.
Hi Usama,
if your explanatory variables are related someone you could increase the quality of your model by including interactions in the model, also.
I would always be careful with the use of the R² in interpreting the accuracy of your model. Usually integrating more variables in your model might increase your prediction accuracy, but you might as well lose power in your tests. Statistics like the AIC and BIC give you a hint how many variables to include because they punish for overfitting a model. You should run a series of tests for different models and see the results for the test statistics. If you receive results you believe are plausible as well, that's all you can ask for.
From Practical experience, the addition of suppressor variables depending on the type of suppressor e.g. zero/classical, reciprocal, and negative suppressor (See Lancaster, 1999) tend to have different effect on the predictive ability of other predictor parameters and the model explained variation. For example in a data i analyzed, there were two type of suppressor variable one having zero relation with criterion or dependent variables and the other having strong non-zero relation with criterion variable. The zero relation variables i labelled zero/classical suppressor while the non-zero variable i labelled, reciprocal suppressor, because it relationship with the criterion is positively significant while at the same time having positive relation in both direction.
Solution:
1. Use the zero/classical suppressor predictor as a moderating variable rather than a direct relationship with the criterion variables. This will reduce incidence of statistical suppression
2. Include an instrument variable - a variable that relate to only one other variables more than orders - for the reciprocal suppressor predictors, since its is possible the error term is correlated with the independent variable causing reciprocity or reverse causality
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression