Regression

Prediction

Research questions focused on prediction take two main forms. They may seek to identify any factors that may influence the detection of a health outcome, or they may seek to identify which of the factors are most predictive of development of the health outcome in affected individuals.

An example of the first use of predictive analysis is the identification of victims of intimate partner violence. A study may be done to identify indicators of partner violence victimization for women seen in a primary care setting. In such a study, a group of factors is found to identify victims. These include injuries, multiple nonspecific physical symptoms (e.g., pain, fatigue, headaches, diarrhea), and psychiatric diseases (e.g., depression, anxiety, post-traumatic stress disorder) as well as characteristics of the victim (e.g., young), perpetrator (e.g., young, excessive alcohol use), and relationship (e.g., wife makes more money than husband). Many of these factors are common among women seen in primary care (e.g., young age, depression). The more characteristics a woman has that were identified as predictive in regression, the more likely that she is a victim of partner violence. This analysis had no interest in identifying the ‘best’ predictor, but in understanding what factors, alone or in combination, predict partner violence so that these factors can be communicated to both physicians and patients. Some of the factors so identified may be outcomes rather than causes of partner violence, but this is not a concern when the purpose of the study is to identify potential victims of partner violence rather than make causal statements about it.

The goal of the second type of predictive model is to identify the most important factors that predict the outcome. For example, it is known that infection with Hepatitis C virus is a risk factor for development of liver cancer. However, not all individuals who are infected with Hepatitis C virus develop this cancer. Now that we can test for Hepatitis C, it may help understand what factors best predict liver cancer among those infected. It is then necessary to examine these other factors, which may include coinfection with Hepatitis B virus, gender, viral genotype, liver enzyme level, use of alcohol or tobacco, and environmental and occupational exposures. Using predictive modeling, we can identify which factors best predict development of liver cancer and then monitor the group with these characteristics more carefully to identify early disease and focus care to more aggressively reduce risk of liver cancer. Here, we are interested in identifying factors that are predictive of liver cancer, not how the factors are associated with or cause cancer.

...