Discussion Questions

• ■
  

  How would you classify this study design?

  
  
• ■
  

  What information do we have on the hospitals and the integrity of the database?

  
  
• ■
  

  How would you describe the analytic approach?

  
  
• ■
  

  What do the tables and figure tell us?

  
  
• ■
  

  How might the information in this study be used?

  
  
• ■
  

  Do other methods exist for predicting outcomes?

Introduction

Third- and fourth-degree perineal lacerations are associated with short-term orbidity, such as blood loss and pain, and long-term morbidity; specifically, pelvic floor dysfunction. Thus, the ability to predict them would be important in obstetrics, as patients at high risk might be candidates for cesarean section. In a new study, researchers used a novel statistical method to identify women who were most susceptible to serious injury. First, they pinpointed independent risk factors. Then they determined which “toxic constellations” of risk factors most increased the likelihood of these lacerations. While many existing studies examine individual risk factors, few group them, making this work especially noteworthy.

George A. Macones, MD, MSCE , Associate Editor

Study Design

Macones: How would you classify this study design?

Cahill: I view this as a retrospective cohort study. The authors had access to a computerized database of over 25,000 deliveries in the Washington, DC-Baltimore area. They looked at a number of exposures—many of which have been described previously—and how they may be associated with perineal laceration.

Macones: What information do we have on the hospitals and the integrity of the database?

Cahill: The authors did a good job in describing this. One university hospital and 3 academic community teaching hospitals were included. Three of the 4 hospitals were regional referral centers. I believe the data were retrieved from the electronic medical record that is used in this health system. Generally, there is a difference between the quality of datasets used for clinical and research purposes, although in this case, the data seemed fairly robust. There were some missing data, but as expected, it is for parameters that aren’t of much clinical use—such as height. I think for the most part, the exposures and outcomes were clear and unambiguous.

Macones: How would you describe the analytic approach?

Stamilio: First, I think it is important to understand the basic premise of what the authors did. The introduction really explains this well. Essentially, in observational studies, we often generate models that look for possible associations between an exposure and an outcome. Adjusted odds ratios generally serve as the output for those models. While those are fine, they don’t tell us anything about whether a factor actually predicts what will happen to the patient. This paper really focused on the prediction of 3rd- and 4th-degree lacerations. In many ways, prediction studies like this have much more clinical utility than standard association studies.

The authors started with a standard multivariate association analysis. More interestingly, they next used classification and regression trees (CART), which have been in use for a number of years but are seldom used in obstetrical literature. The CART technique is also known as recursive partitioning. Essentially, CART takes individual risk factors, starting with the most important one, and splits the dataset on the presence or absence of that risk factor. There are no assumptions going in about cut-points for continuous variables; CART also finds the best cut-point and splits the variables accordingly. This process is repeated until no variables with discriminating power are left. It is very different from what we usually do in terms of modeling.