Does EPV predict etiology?
 Revised: 2002-05-10

Question: Does etiologic predictive value (EPV) really predict etiology?

EPV links detectable signs (symptoms, clinical findings or other signs detectable with some kind of device) with test results. Before we can evaluate the test we must start by defining an appropriate disease. This definition must link the sign(s) with a possible etiologic agent. For example a definition could be: "A sore throat caused by group A beta-hemolytic streptococci (GABHS)".
    The key issue is if our defined disease exists. The formulae for EPV cannot know if GABHS causes a sore throat or if a sore throat increase the prevalence of GABHS. (Here one must remember that when we talk about increasing the prevalence of GABHS we mean increased proportion of individuals harboring GABHS, not an increase in the number of GABHS in each individual).
    Thus, the outcome of EPV is an estimate of a statistical correlation. Other facts must decide if this correlation exists because the agent causes the sign or if the sign facilitate the presence of the agent. There are several ways to obtain such evidence, the best way is in a randomized controlled intervention study.

Question: If EPV does not directly predict etiology then what is it for?

There are a number of areas where EPV could add useful information:

  1. In many situations intervention studies establishing a causal relation between an etiologic agent and a sign exists. If we want to adjust previously calculated predictive values of a test to also include knowledge about carriers then EPV is the only alternative. A similar situation is if a new test is developed. If asymptomatic carriers exists then this new test should be evaluated with EPV.
     
  2. If the duration of our defined disease is short then a a correlation between a sign and an agent will imply that the agent actually causes the sign. However, there is an uncertainty in this assumption that we cannot measure. The only way to estimate this uncertainty is by performing an intervention study.
     
  3. EPV could be very useful as a screening to decide which signs might be of interest to investigate further in a randomized controlled intervention study. Let us consider a situation with 7 different signs that are assumed (but not known) to be caused by a specific etiologic agent. This agent can be detected with a test. However, it is also known that this etiologic agent might be present without causing any signs.
        We want to investigate if there is a causal relationship between the agent and any of these signs. However, we cannot perform an intervention study testing all signs at the same time. If we begin by collecting test outcome among individuals with and without these signs we can estimate if there seems to be a correlation between the agent and any of the signs (or combination of signs). If no correlation is found it indicates that an intervention study will not add useful information. If we find a correlation we can predict the combination of signs that is most interesting to investigate in an intervention study.
        Just testing if a correlation exists between any of several signs and findings of a possible etiologic agent could easily be done using correlation-regression. In that kind of situation we would have a single group of individuals where all individuals had at least one sign. However, if carriers should be considered then we are dealing with two groups of individuals, individuals with a sign and healthy controls. This is a situation that correlation-regression cannot analyze. In the (rather unusual) situation where the prevalence of the sign is very common in a population then it could be possible to analyze one large randomly selected sample of individuals with regression-correlation. This single group would contain both individuals with and without the sign. However, in this situation EPV would offer a more easily interpreted outcome than correlation-regression.

Other WebPages of interest

Other pages with subjects that might be of interest is:

(You can click on these links to quickly see the pages)

Ronny Gunnarsson MD PhD
Department of Primary Health Care
Göteborg University
SWEDEN

Back to Main menu