Epi lecture 4 - the 2x2 contingency table

How to calculate accuracy

(TP + TN)/total Overall ability of the test to produce correct results

How to calculate specificity (Sp)

- based on "knowns" - true negatives over known negatives - Sp = TN/(FP + TN)

How to calculate sensitivity (Sn)

- based on "knowns" - true positives over known positives - Sn = TP/(TP + FN)

How to calculate population prevalence (P-Prev)

- based on "knowns" - so: P-Prev = total known positives/grand total - aka: P-Prev = (TP + FN)/total

How to calculate negative predictive value (NPV)

- based on "tests" - true negatives over test negatives - NPV = TN/(FN + TN)

How to calculate positive predictive value (PPV)

- based on "tests" - true positives over test positives - PPV = TP/(TP + FP)

How to calculate apparent (test) prevalence (A-Prev)

- based on "tests" - so: A-Prev = total test positives/grand total - aka: A-Prev = (TP + FP)/total

What happens as prevalence approaches 0%?

- false positives go up - any positive test is likely a false positive (because prevalence is near 0%) - this is how USDA tracks you down when you're calling all your TB tests negatives

Specificity in words

Ability of a test to correctly identify the known negative animals

Sensitivity in words

Ability of a test to correctly identify the known positive animals

2x2 tables are based on...

Bayes's theorem

Example: TB testing has a very low Sn (50%) and a Sp of 98%. Why do we use this test if the sensitivity is so low?

Because TB has a very LOW prevalence, which makes the NPV very high (100%)

2x2 table is used for...

Binary data, clinical yes or no decisions

A lower population prevalence would do what to the PPV of the same test?

Decrease it (have more false positives), increase the NPV Opposite is also true = if increase prevalence, *NPV will go down, PPV will go up* Example: FeLV test, when he plugged in the numbers the prevalence was 31%, but the actual prevalence of FeLV in the shelter is 1-2%, so by fixing the prevalence, the PPV goes down and NPV goes up. This means when the test is negative you can say with confidence that the animal does not have the dz, but when it is positive, you should confirm with another test.

T or F - population prevalence does not change the Sn, Sp or PPV, NPV of the test

False - does not change Sn, Sp BUT IT DOES change PPV, NPV (what you want to tell your clients about)

NPV in words

Given a negative test result, the probability that an animal doesn't have the disease

PPV in words

Given a positive test result, the probability that an animal really has the disease *This is what you care for as a clinician*

Sensitivity and specificity tell you...

How well a test works for giving you useful information

Increased sensitivity affects the ______________. What does this tell you?

NPV - can *rule out* a disease Sn(out) = rules it out

Specificity affects the...

PPV - can *rule in* a disease Sp(in) = rules it in

What would be the Sn, Sp, PPV, NPV of a perfect test? Of a worthless test?

Perfect test = 100% everything Worthless test = 50% everything (not bigger than chance)

Test validation studies must be conducted on ____________________

Real animal populations - must make sure one is very careful choosing the population or one may introduce bias (inaccurate results) into the test

Test results or clinical decisions go on which side of the table? What about the known disease status?

Rows - total *test* positives and negatives Columns - total *"known"* positives and negatives

Analytical Sn/Sp

Same calculations, how it works in the lab

ROC analysis

Taking continuous tests results and identifying a middle point that gives the best Sn and Sp possible for that test

Sn and Sp are ______________________. PPV and NPV are _________________ changed by ___________________.

Test parameters Interpretation parameters Prevalence

How do we determine what the known is?

Tricky - we use "gold standards" or multiple tests, hard to assess when all tests have possibility of missing disease, but we do our best

Diagnostic Sn/Sp

What we care about as clinicians, how it works in a real population