In the Journals

Study: Propensity scores are useful to control measured confounders

Propensity scores are useful in controlling for measured confounding in observational studies of medical treatments or risk factors, but neither propensity score methods nor multivariable regression analysis can control for unmeasured confounding, according to a recently published paper.

In research on propensity scores, how these can be used to control for confounding and their advantages and disadvantages, Eduoard L. Fu, MD, of the department of clinical epidemiology at Leiden University Medical Center in the Netherlands, and colleagues demonstrated different propensity score methods on data from the Netherlands Cooperative Study on the Adequacy of Dialysis, which included 1,955 patients. The researchers used Cox proportional hazards regression analysis to assess correlations between treatment modalities, including propensity score adjustment, propensity score weighting and multivariable Cox regression analysis.

The authors reported that both propensity score matching and propensity score weighting distinguish between the design and analysis steps of a study. Confounding is eradicated through balancing of the confounders (design step), and then the treatment effect is directly estimated (analysis step).

Additionally, inspection of the propensity score distribution may reveal areas of nonoverlap between treated and untreated groups. These areas of nonoverlap may include patients with an absolute indication or contraindication for treatment. The authors noted propensity score matching would exclude these subjects from analysis, as no comparison can be made with these patients. Propensity score weighting may allow for identification of these subjects, because they receive large weights.

Researchers also found propensity score methods may be chosen over multivariable regression analysis if the number of events is low in relation to the number of confounders.

They also noted that with multivariable regression analysis, the association between each individual confounder and the outcome must be modeled properly to adjust for confounding.
Lastly, the study authors reported that despite comparisons with randomized clinical trials, propensity score methods are not able to control for unmeasured confounding.

“Whether [propensity score] PS methods should be used instead of multivariable regression depends on the properties of the data at hand and which treatment effect the researcher wants to estimate,” the researchers wrote. “Neither of these methods, however, controls for unmeasured or unknown confounding.” - by Jennifer Byrne

 

Disclosures: The researchers report no relevant financial disclosures.

Propensity scores are useful in controlling for measured confounding in observational studies of medical treatments or risk factors, but neither propensity score methods nor multivariable regression analysis can control for unmeasured confounding, according to a recently published paper.

In research on propensity scores, how these can be used to control for confounding and their advantages and disadvantages, Eduoard L. Fu, MD, of the department of clinical epidemiology at Leiden University Medical Center in the Netherlands, and colleagues demonstrated different propensity score methods on data from the Netherlands Cooperative Study on the Adequacy of Dialysis, which included 1,955 patients. The researchers used Cox proportional hazards regression analysis to assess correlations between treatment modalities, including propensity score adjustment, propensity score weighting and multivariable Cox regression analysis.

The authors reported that both propensity score matching and propensity score weighting distinguish between the design and analysis steps of a study. Confounding is eradicated through balancing of the confounders (design step), and then the treatment effect is directly estimated (analysis step).

Additionally, inspection of the propensity score distribution may reveal areas of nonoverlap between treated and untreated groups. These areas of nonoverlap may include patients with an absolute indication or contraindication for treatment. The authors noted propensity score matching would exclude these subjects from analysis, as no comparison can be made with these patients. Propensity score weighting may allow for identification of these subjects, because they receive large weights.

Researchers also found propensity score methods may be chosen over multivariable regression analysis if the number of events is low in relation to the number of confounders.

They also noted that with multivariable regression analysis, the association between each individual confounder and the outcome must be modeled properly to adjust for confounding.
Lastly, the study authors reported that despite comparisons with randomized clinical trials, propensity score methods are not able to control for unmeasured confounding.

“Whether [propensity score] PS methods should be used instead of multivariable regression depends on the properties of the data at hand and which treatment effect the researcher wants to estimate,” the researchers wrote. “Neither of these methods, however, controls for unmeasured or unknown confounding.” - by Jennifer Byrne

 

Disclosures: The researchers report no relevant financial disclosures.