test sensitivity
among all the patients with the disease, the proportion that will have a positive test result
= TP/(TP + FN)
- high sensitivity = few false negatives; we want this for screening tests
- doesn’t rule out false positives
SNNOUT- a highly sensitive test when negative rules out the disease
test specificity
among all the patients without the disease, the proportion that will have a negative test result
= TN/(TN + FP)
- high specificity = few false positives; we want this for diagnostic tests
- doesn’t rule out false negatives
SPPIN- a highly specific test when positive will help rule in disease
positive predictive value (PPV)
among all patients with a positive test result, the proportion of patients who
have the disease
= TP/(TP + FP)
increases with increasing disease prevalence
negative predictive value (NPV)
among all the patients with a negative test result, the proportion of patients who
do not have the disease
= TN/(TN + FN)
decreases with increase in disease prevalence
test accuracy
ability of a test to differentiate people with and without the disease correctly
= (TP + TN) / (TP + TN + FP + FN)
If you increase the positive cutoff value for the test (move cutoff from L to R), it will
- decrease sensitivity
- increase specificity
- increase positive predictive value (PPV)
- decrease negative predictive value (NPV)
internal validity vs external validity
internal validity: extent to which the observed results represent the truth in the population
- independent, blind comparison with reference standard of diagnosis?
- was reference standard applied regardless of index diagnostic test result?
external validity: extent to which the study results apply to similar patients in a different setting (IRL)
- affordable, available, accurate?
- study patients similar to patient in question?
- how current is the study we are analyzing?
Lead-time bias
- systemic error
- overestimation of survival duration among screen-detected cases (compared to those dx by signs and symptoms) when survival is measured from diagnosis
- occurs when follow-up of groups does not begin at comparable stages in condition
Ex: interventions for cancer pts detected by screening cannot be compared with interventions for pts whose disease is first detected by clinical examination at a later stage of the disease
Length-time bias / Lag-time bias
- overestimation of survival duration among screen-detected cases due to relative excess of slowly-progressing cases
- occurs when prevalent (new + old) rather than incident (new) cases are included in a case-control study
- screening programs are more likely to detect slower progressing diseases, which have better prognosis <– aggressive diseases generally have a shorter asymptomatic period
Ex: screening program may show falsely improved survival when compared to a cohort that includes a wider spectrum of disease
Overdiagnosis bias
- occurs when a lesion (that would generally be read as a malignant tumor) is so indolent that it would never go on to cause problems for the patient
- extreme form of length-biased sampling
Ex:
1. cancers that are so benign that they have virtually no growth potential or they might spontaneously regress
2. cancers that grow so slowly the pt would die of another competing cause of death first
Selection bias
- systemic error
- intervention groups differ in measured/unmeasured baseline characteristics with respect to prognosis
- due to the way participants were selected for the study or assigned to their study groups
- all case control studies are vulnerable to this selection bias, as the investigator “selects” the cases and controls
Ex: the estimated effect of cigarette smoking on lung cancer will be biased if study participants are
volunteers
Other types of selection bias:
- referall bias
- attrition bias
- volunteer effect
Referral bias
- type of selection bias
- difference that occurs because a pt was referred to specialty care or tertiary care as opposed to primary care
- outcomes are different for these patients
Other biases
- non-response bias: non-respondents are different from survey respondents
- response bias: systematic error due to differences between those who volunteer and those who don’t
- information bias: wrong recording of individual factors (measurement bias, interviewer bias, verification bias, misclassification bias, observer bias)
- ecologic fallacy: assumes all members of a group exhibit characteristics of the group at large
- Hawthorne effect: people behave differently when they know they’re being watched
- placebo effect
- recall bias: systematic error due to differences in accuracy of recall
- healthy worker effect: workers often exhibit lower overall death rates than general population because severely ill/disable persons tend to not be working
- confounding: occurs when association between exposure and outcome is accounted for by other variables
Regression to the mean
- as subjects are followed over time, their average score falls, whether they receive active treatment or not
- refers to the noise introduced by within individual variation
- when repeated measurements are performed, high values tend to decrease and low values tend to increase
number needed to harm (NNH)
number of individuals who need to be exposed to a certain risk factor before one person develops an outcome
NNH = 1/AR
AR = absolute risk
If NNH = 18, that means for every 18 patients there will be 1 who is harmed by the treatment
number needed to treat (NNT)
number of individuals that must be treated, in a particular time period, for one person to benefit from treatment
NNT = 1/ARR
ARR = absolute risk reduction
