IGX

In 2009 epidemiologists at Harvard 
Medical School developed a way for scientists to account for the invisible connections that were confounding their studies. The new approach uses an algorithm that automatically identifies and adjusts for confounders as well as or better than the most knowledgeable scientist can, says Jeremy Rassen, one of the algorithm’s creators.

Called the high-dimensional propensity score algorithm (hd-PS), it is a tool for improving not randomized clinical trials but broader observational studies, in which researchers watch a large pool of participants and look for correlations—like the fact that wine drinkers live longer than other drinkers. Observational studies are cheaper and easier than clinical trials. Unfortunately, the data they yield are rife with confounder problems, but researchers can improve the data by adjusting for suspected confounders and removing the bias they introduce. In the recent observational study on wine and longevity, for instance, after the researchers accounted for smoking, gender, and activity level, they found that beer and hard liquor were just as life-extending as wine.

The use of a randomized controlled trial has advantages when one is attempting to discern a direct cause and effect relationship between an intervention and an outcome. The purpose of randomization is to ensure balance between study arms of both known and unknown causes of bias and confounding. Subsequently, observed differences in outcomes between randomized groups are assumed to be due to the treatment assignment. When treatment is not randomized, however, the factors that affect the decision to use a particular treatment can also affect the observed outcome. Therefore, ignoring the conditions that led to the use of a specific treatment will run the risk of creating a biased estimate of the effect of the treatment on the patient outcome. For example, we may want to examine the use of norepinephrine in patients in a retrospective heart failure registry and its relationship to inpatient mortality. For each encounter, the physician deciding to use norepinephrine would make this decision based on heart rate, blood pressure, code status and other clinical factors present on examination. We can reasonably expect that some combination of these same clinical factors would also be associated with patient mortality. During analysis, we can quantify how likely it wasthat a patient with a certain pattern of characteristics would have received norepinephrine. There are a number of statistical maneuvers available that can address this issue, including Heckman two-step estimators, treatment effect models and instrumental variable models [45]. However, many of these techniques require assumptions regarding the relationship between the predictor variables, the choice of treatment assignment, and the outcome under study that may not hold up in real-world situations.

An alternative technique that accommodates the interplay between predictors, treatment and outcomes is the calculation of propensity scores [46]. A propensity score is, in its simplest conception, the probability of assignment to a treatment based on the observed covariates. Using the above example, the propensity score for norepinephrine use would be the probability that a patient received norepinephrine, given their systolic blood pressure, heart rate, code status and other clinical factors. The propensity score is therefore a summary statistic that replaces the collection of covariates that confound the decision to proceed with one treatment instead of the other, and the score itself is then used as if it were the only covariate affecting the relationship between treatment assignment and outcome [47]. Therefore, the cause and effect relationships can be more accurately described.

To continue the example, a regression model of the likelihood of inpatient death in patients that were given norepinephrine versus nitroglycerin, after adjustment of the model for the propensity score, will give a much cleaner estimate of the true relationship between the treatment and the outcome, assuming that the propensity score is properly calculated and specified. It is critical to note that propensity scores can only be generated from those covariates that have been observed, and more importantly, recorded. It is this fact that will maintain the superiority of randomized trials, where feasible, in the ability to delineate causal relationships. Theoretically, randomization will distribute both known and unknown confounding covariates equally between groups when a large enough sample is utilized. By definition, the unknown covariates cannot be controlled for in registry work. Sensitivity analyses can be conducted to determine to what degree the observed treatment effect varies when one manipulates the odds of treatment group assignment. I would refer readers interested in a detailed discussion of the mechanics of sensitivity analysis in the setting of propensity score modeling to Guo and Fraser [45].