your question needs more detail. However, overall, matching is a good way to avoid confounders!

There are two way to control confounding, by design and analysis. By design includes randomization, inclusion and exclusion criteria; by analysis includes stratified and matched analysis. Matching is more efficient to control confounding if the sample size is small. It allowed balance distribution of data in each strata, estimate of the OR more stable, smaller standard error, narrower confidence interval. However, you should also consider the disadvantages, which may be time consuming, can be expensive, and we can't always find the exact match.

Thank You Jagdish and Surya. I appreciate it!

If the sample size large enough, multivariate regressions should be used. I often use stratification as a first step in data analysis.

I agree. Thank you.

You can do both. In the design, matching is perfect. In statistical analysis, you can use stratification to reveal confounding, and also use multivariate regressions (e.g., Logistic regression model) to control the effect of confounders.

Thanks Lu!

I agree with colleagues. Design and regression is the best tools to control for confounder

Yes - stratification is good for one or two variables but then becomes impractical. Multivariate regression is good for multiple confounders but you do face the interesting issues of what variables to include in models (ie how to detect a confounder), how to detect collinearity, how to detect interaction of one variable with another -- and how to avoid using in the model as confounders variables that are actually in the causal pathway. All quite tricky and not well described in full in any textbook I know..

I apreciate your response.

Hi Pierre,

Stratification means developing strata for either sampling in the design or in the analysis. For example, if you want to look at males and females in equal numbers. So, you sample 50 males and 50 females, regardless of the population of males and females.

Matching is a form of stratification. It can either be in the design or the analysis. Matching is the pairing of one or more controls to each case on the basis of their similarity with regards to various variables.

So, matching induces confounding in the sample, when there might be none, or at least non-collapsibility in the populations of interest. You must use stratified analysis.

Keep in mind that the limitation of matching involves issues such as the probability that it cannot estimate association of disease and matching factor in this case which equal gender.

Good luck and keep up the good work.

Iniekem (Nikky).

Thank you.

Hi Pierre,

Matching is good when you know or suspect in advance what your confounders are. For example if death is your outcome, an obvious confounder will be old age, so matching for age at the design stage already is a good idea. But avoid overmatching (i.e. matching for too many variables), because you may end up missing associations. For example if your outcome is associated with gender and you have matched for gender, then you will never see the association.

Hello

That depends on your objective, for good a maitrise of factors of confusion I will advise to you to make a stratification as well as possible go until the multiple regression with a model by putting at it the probable factors of confusion and to see how the model comportera(if your project is with explanatory aiming) if your research is with exploratory aiming I advise myself you to make an automatic selection of your predictors.

Excuse me for my bad English I am spirit to learn

Merci: Thank you.

By using analysis method, Propensity score method is a good way for bias reduction. It calculates the conditional probability given the covariates, which can be used to balance the covariates in the two groups, and so reduce this bias.

Thank you.

Hey Pierre,

You have heard it all. Take note of the design and regression analysis advice. Those options are my best ones. I just want to add a blessing: May it be well with you, whether you agree or not.

Thank you!

I think it's too late to commemt, but stratification is more informative than matching or counfound control. It makes no assumptions. Matching and regression model assumes that the risk is evenly distributed in the controlled factor. Such situation is not always the case. In all cases, you must have enough sample in each stratum of the controlled and the studied factors.

Hello Hedley,

Thank you for contributing to my question. I appreciate it.

Matching has no advantage over Analysis ( Stratification) using miltiple regression methods. In fact matching can cause difficulties in chooseing study participants

Thank you.

There is a caveat for stratification: if you have poorly reviewed the literature, it could cause you to stratify groups that would have different etiologic factors. For example if you were to stratify ages 0-18 for certain childhood cancers, you would in fact be grouping age ranges where there could be a very different burden of disease at each yearly range.

Another example would be colon cancer, where if you grouped ages 40-59, you would be combining an age range with relatively low (40-49), with high burdens of disease.

Stratification, as the previous authors noted is more flexible overall, but matching also has it's uses.

Pierre,

Confounder; also known as a third variable, usually distorts the relationship between an independent (exposure) and a dependent (outcome) variable. The distortion can then lead to erroneous conclusions. For example, if a researcher wants to study the effect of alcohol on driving, several variables should be taken into consideration: Speed, quantity of alcohol consumed with a given period of time (minutes, hours), gender, and age. In this case, age and gender could confound the outcome because younger drivers (boys, especially but girls tend to over speed as well, even without drinking. However, alcohol consumption beyond normal capacity can also increase the probability of being involved in a motor vehicle accident (MVA). The question here is what would increase the risk for an MVA? Would it be gender, age, alcohol, or overspending?

To get the correct answer, confounding must be controlled for. Aschengrau and Seage (2009) posited that confounding can be controlled through the design and analytical stages. For the design stage, Aschengrau and Seage prescribed randomization, restriction, and matching of the dependent variables. This means that all study participants or samples should be randomly selected to reduce the possibility of chance. For example; age, a known confounder, should be restricted to a narrow width such as 20-25 years old, 26-30 year olds, and so on. Age restriction ensures equal chances of obtaining similar scores. Similarly, gender can also be adjusted for by matching females to females and males to males.

Using the same principles, as in the design stage, confounding can be controlled for at the analytical stage through standardization (age, gender, and race), stratified analysis, and multivariate analysis Aschengrau and Seage (2009)

Thanks,

Ben

References

Aschengrau, A and Seage, G.R (2009). Essential of epidemiology in public health. Sudbury, MA: Jones and Bartlett

Thanks Ben! Appreciate this elaborate answer! Stratification and matching are both used.

To answer this question it is important to know what kind of studies are you conducting? Because matching as a mean to eliminate confoundings is suitable for case controlled studies mainly, while stratified analysis is suitable for controlling confounding in all incident studies, including case-control and cohort studies, cross-sectional comparative studies and experimental studies. Therefore, I think stratified analysis is the best way to deal with predictor variables in terms of confounding.

Even if stratification or matching can control for confounding, they both have a common major disadvantage: they do not provide a mean to measure the potential influence of this confounder on the association of interest.

I therefore prefer adjusting for confounders at the analytical level. This also provides a possibility to check for interaction (co-factor modulates the effect of an exposure on an outcome). This however requires correctly estimating sample size beforehand for such analysis to have sufficient power.

I think in always situations, stratification is better than matching.

Thanks Fahmi, Paul, and Amir. I used Matching, and all was well.

Depends on the type of confounding. For confounding by indication, (eg in pharmacoepi, physician selects Drug A over Drug B due to a higher perceived benefit or a lower perceived risk based on patient-level risk factors. Propensity score matching is best, but requires sophisticated analytic skills.

For multiple confounders, regression models are useful to adjust for confounders

For few confounders, stratification will adjust but you need the numbers with which to populate each stratum

I stand by my statement try matching on three variables and see how long it takes you to get a control group enrolled  

To control confounding you should determine the study design. The experimental studies have several methods to control confounding: to keep factors constant, to include a control group, random assignment, and statistically control for confounding variables.

To reduce the effect of confounding variables, a researcher can resort to three  phases:

Phase one: at the phase of study design we match for age and sex but no more except in few situations. Over-matching make the choice of controls difficult.

Phase two: At design and analysis , we can use stratification

Phase three: At the analysis we can use stratification and  multivariate  analysis. 

The danger in confounding variables is the overlooking of their effect by researchers.

Though too late to comment .. but recently had to deal with a similar question for a fellow reviewer and I found a nice article by Pourhoseingholi et al, 2012, which I would like to share with fellow researchers. Article How to control confounding effects by statistical analysis