Lecture 9 - Screening

What is the predictive value of a test?

• The probability that the test results are positive in a patient that they have the disease

• Most useful metric for diagnostic tests

• The main factor influencing test accuracy is the population

What does the example of blood tests for Alzheimer's represent?

• >90% accurate screening tests are useless if they are not targeted to specific high-risk populations

• A study came up with a new test for Alzheimer’s through a set of 10 lipids from peripheral blood that predicted mild cognitive impairment or Alzheimer’s within a 2-3 year timeframe with over 90% accuracy

  • A test with 90% accuracy can be incorrect 92% of the time

  • Even if you assume that 5% of people will get the disease, a good estimate for people aged over 60 - the PPV is still only 32%

What is the problem with test accuracy?

• Highly accurate tests have poor predictive value if the disease is rare and given to low-risk populations

• Most diseases have a prevalence on <1% population

• Best use of tests = targeted population

• Need to make the test representative of the population

What is the connection between test accuracy and breast cancer screening

• In women 50 to 59 years old with average risk, breast cancer prevalence = 2.5%

• Screening mammography:

  • Sensitivity = 77.3%

  • Specificity = 98.7%

  • PPV = 4.8%

• If women are screened annually for 10 years, 50% probability of a false positive

What are the types of multiple tests?

• Sequential testing

• Simultaneous testing

What is sequential testing?

• One test after another

• Ex. Diabetes

  • Test 1 Blood Sugar: Those who tested TP and FP move onto test 2

  • Test 2 Glucose Tolerance Test

    • Net sensitivity and specificity

What is simultaneous testing?

• Running multiple tests at once, either applying different diagnostic tests to the same subjects

• Increases net sensitivity

• Decreases net specificity

What is selection bias?

• Non-responders in studies can bias results

• Ronmart et al. 1999 - Swedish study of ~9000 people mailed questionnaire on respiratory diseases

  • 85% response rate (typically unheard of)

  • Non-responders were more likely to be:

    • Smokers

    • Manual workers

    • Suffering from respiratory disease

  • Hence, underestimation of the prevalence of respiratory diseases

  • Survey can be seen as inaccurate due to self-reporting flaws

• Undermines conclusion of the study

• All studies select a sample of the population

What is information bias?

• Differential vs. non-differential

  • Differences in reporting data

• Ex. The RR of the development of breast cancer in relation to previous induced abortions is reported

  • The Southeastern region has the highest RR, probably due to underreporting abortions due to it being a conservative region

• Ex. Circumcision → Men reporting they did not have a circumcision when they actually did because they did not know what that meant

What are some types and sources of information bias?

• Abstracting records

• Interviewing

• Surrogate interviews

• Surveillance

• Recall

• Reporting

What are the approaches to handling confounding?

In designing and carrying out the study:

• Individual matching

• Group matching

In the analysis of data:

• Stratification

• Adjustment

How does stratification give insight into confounding factors?

• By isolating the relationship between an exposure and an outcome within subgroups where the potential confounder is held constant

• Age is the most common confounding factor for human disease

• Age distribution in test vs. control populations must be well-matched

• For diagnostic tests/screening, age is a common way of IDing a high-risk population

What do the examples of lung cancer rates by smoking status and degree of urbanization, and RR of developing cancer of the esophagus in relation to smoking and drinking habits exhibit?

• How multiple factors interact in causing a disease

• When looking at variables, they could be independent, but could also be synergistic risk

What are the 3 questions to determine whether an interaction is or isn’t present?

1. Is there an association

2. If so, is it due to confounding?

3. Is there an association equally strong in strata formed on the basis of a third variable?

• No → Interaction present

• Yes → Interaction not present

What is the additive model for incidence rates and attributable risks for groups exposed to neither risk factor or one or two risk factors?

• The sum of the baseline incidence and the excess incidences due to each exposure, and the attributable risk (AR) is the simple difference in incidence rates

• Indicates no synergistic effect

What is the multiplicative model for incidence rates and attributable risks for groups exposed to neither risk factor or one or two risk factors?

• The combined effect is the sum of their individual relative risks

• Indicates synergistic effect

How do you determine if data is synergistic?

By comparing the observed combined effect of two or more agents to their expected additive effect