Observational Studies: Usmle Endpoint Statistics
Observational Studies: Usmle Endpoint Statistics
Observational Studies: Usmle Endpoint Statistics
OBSERVATIONAL STUDIES
CASE-CONTROL STUDY
Compares a group of people with disease to a group without disease. Retrospective study.
Looks to see if odds of prior exposure or risk factor differs by disease state.
Asks, “What happened?”
Measures:
Odds ratio (OR).
Patients with COPD had higher odds of a history of smoking than those without
COPD.
Subjects are not randomly selected but rather explicitly recruited with (cases) and without
(control) disease. Control are selected who do not have disease of interest, and regardless
of exposure status to risk factor.
Cannot assess incidence or prevalence of disease. Very useful for studying conditions
with very low incidence or prevalence.
Can help determine causal relationships.
COHORT STUDY
Compares a group with a given exposure or risk factor to a group without such exposure.
Looks to see if exposure or risk factor is associated with later development of disease.
Can be:
Prospective (asks, ―Who will develop disease?‖) or
Retrospective (asks, ―Who developed the disease [exposed vs nonexposed]?‖).
Measures:
Relative risk (RR).
―Smokers had a higher risk of developing COPD than nonsmokers.‖
Can determine incidence and causal relationships.
ADOPTION STUDY
Compares siblings raised by biological vs adoptive parents.
Measures heritability and influence of environmental factors.
META-ANALYSIS
Groups the results of several trials (ideally high-quality randomized controlled trials) to
increase statistical power and provide an overall estimate of the effect of an exposure on
an outcome.
ECOLOGICAL STUDY
The unit of analysis in this study is populations not individuals.
Studied using population data.
CLINICAL TRIAL
Experimental study involving humans.
Compares therapeutic benefits of 2 or more treatments, or of treatment and placebo.
Study quality improves when study is randomized, controlled, and double-blinded (ie,
neither patient nor doctor knows whether the patient is in the treatment or control group).
Triple-blind refers to the additional blinding of the researchers analyzing the data.
Four phases (―Does the drug SWIM?‖).
LIKELIHOOD RATIO
LRs can be multiplied with pretest odds of disease to estimate posttest odds.
LR+ > 10 and/ or LR– < 0.1 are easy-to-remember indicators of a very useful diagnostic
test.
UW: Positive and negative predictive values are influenced by disease prevalence.
QUANTIFYING RISK
Definitions and formulas are based on the classic 2
× 2 or contingency table.
RELATIVE RISK
Typically used in cohort studies.
Risk of developing disease in the exposed group divided by risk in the unexposed group
Example:
If 21% of smokers develop lung cancer vs 1% of nonsmokers, RR = 21/1 = 21.
For rare diseases (low prevalence), OR approximates RR.
RR = 1 no association between exposure and disease.
RR > 1 exposure associated with ↑ disease occurrence.
RR < 1 exposure associated with ↓ disease occurrence.
In low prevalence situations, RR = OR.
When the disease is rare, a and c represent small quantities. Therefore, a is
negligible compared to b, and c is negligible compared to d; this results in a
ATTRIBUTABLE RISK
The excess risk in an exposed population that can be explained by exposure to a
particular risk factor.
The difference in risk between exposed and
unexposed groups.
The proportion of disease occurrences that are
attributable to the exposure.
Example:
If risk of lung cancer in smokers is 21% and risk in nonsmokers is 1%, then 20%
of the lung cancer risk in smokers is attributable to smoking.
AR = =
INCIDENCE VS PREVALENCE
PRECISION VS ACCURACY
PRECISION (RELIABILITY)
The consistency and reproducibility of a test gives similar or very close results on
repeat measurements.
The absence of random variation in a test.
Random error ↓ precision in a test.
↑ Precision ↓ standard deviation.
↑ Precision ↑ statistical power (1 − β).
ACCURACY (VALIDITY)
The trueness of test measurements.
The absence of systematic error or bias in a test.
Systematic error ↓ accuracy in a test.
ATTRITION BIAS
Loss to follow-up occurs disproportionately between the exposed and unexposed groups.
Attrition bias is a form of selection bias.
Attrition bias does not occur when the losses happen randomly between the exposed and
unexposed groups as this simply leads to a smaller study population.
RECALL BIAS
Awareness of disorder alters recall by subjects.
Potential problem for retrospective studies such as case-control studies, particularly
when questionnaires are used to inquire about distant past exposure.
Examples:
Patients with disease recall exposure after learning of similar cases.
STRATEGY TO REDUCE BIAS
Decrease time from exposure to follow-up.
MEASUREMENT BIAS
Information is gathered in a systemically distorted manner.
Examples:
Association between HPV and cervical cancer not observed when using non-
standardized classifications.
Hawthorne effect—participants change their behavior in response to their
awareness of being observed.
STRATEGY TO REDUCE BIAS
Use objective, standardized, and previously tested methods of data collection that
are planned ahead of time.
Use placebo group.
PROCEDURE BIAS
Subjects in different groups are not treated the same.
Examples:
Patients in treatment group spend more time in highly specialized hospital units
STRATEGY TO REDUCE BIAS
Blinding and use of placebo reduce influence of participants and researchers on
procedures and interpretation of outcomes as neither are aware of group
allocation.
OBSERVER-EXPECTANCY BIAS
The investigator's decision is affected by prior knowledge of the exposure status.
Researcher’s belief in the efficacy of a treatment changes the outcome of that treatment
(aka Pygmalion effect; self-fulfilling prophecy).
In the classic classroom experiment that first described the Pygmalion effect; a group of
students wererandomly assigned high intelligence quotient (IQ) scores; their teachers
were then told of these artificial results and had higher expectations of this group as a
result. The students with the randomly assigned high IQ scores actually performed better,
likely because their teachers unconsciously behaved in a manner that would facilitate
their success.
Examples:
If observer expects treatment group to show signs of recovery, then he is more
likely to document positive outcomes.
CONFOUNDING BIAS
When a factor is related to both the exposure and outcome, but not on the causal
pathway factor distorts or confuses effect of exposure on outcome.
Examples:
Pulmonary disease is more common in coal workers than the general population;
however, people who work in coal mines also smoke more frequently than the
general population.
STRATEGY TO REDUCE BIAS
Multiple/repeated studies.
Crossover studies (subjects act as their own controls).
Matching (patients with similar characteristics in both treatment and control
groups).
Stratified analysis.
Restriction.
Randomization.
Effect modification is present when the effect of the main exposure on the outcome is
modified by the presence of another variable Effect modification is not a bias.
LEAD-TIME BIAS
Early detection is confused with ↑ survival.
Examples:
Apparent prolongation of survival after applying a screening test that detects a
disease earlier than it would have been otherwise detected but without any real
effect on prognosis.
STRATEGY TO REDUCE BIAS
Measure ―back-end‖ survival (adjust survival according to the severity of disease
at the time of diagnosis).
STATISTICAL DISTRIBUTION
MEASURES OF CENTRAL TENDENCY
Mean = (sum of values)/(total number of values).
Most affected by outliers (extreme values).
Median = middle value of a list of data sorted from least to greatest.
If there is an even number of values, the median will be the average of the middle
two values.
Mode = most common value.
Least affected by outliers.
MEASURES OF DISPERSION
Standard deviation:
A measure of the degree of dispersion from the
mean.
When SD is small, data points tend to have
minimal variation and are tightly clustered around the mean.
In contrast, a large SD implies that the data points are spread out over a large
range.
The SD is calculated such that 68% of all values lie within 1 SD from the mean.
The remaining 32% of values lie outside of 1 SD, with 16% of these above and
16% below 1 SD from the mean.
In addition, 95% of all values are within 2 SDs from the mean and 99.7% are
within 3 SDs.
Standard error:
Shows how closely sample means are related to population means.
= SD/ √n.
NORMAL DISTRIBUTION
Gaussian, also called bell-shaped.
Mean = median = mode.
NON-NORMAL DISTRIBUTIONS
Bimodal:
Suggests two different populations (eg, metabolic polymorphism such as fast vs
slow acetylators; age at onset of Hodgkin lymphoma; suicide rate by age).
Positive skew:
Typically, mean > median > mode.
Asymmetry with longer tail on right.
Negative skew:
Typically, mean < median < mode.
Asymmetry with longer tail on left.
STATISTICAL HYPOTHESES
Null (H0):
Hypothesis of no difference or
relationship (eg, there is no association
between the disease and the risk factor
in the population).
Alternative (H1):
Hypothesis of some difference or
relationship (eg, there is some
association between the disease and the risk factor in the population).
CORRECT RESULT
Stating that there is an effect or difference when one exists (null hypothesis rejected in
favor of alternative hypothesis).
Stating that there is not an effect or difference when none exists (null hypothesis not
rejected).
INCORRECT RESULT
Type I error (α):
Stating that there is an effect or difference when none exists (null hypothesis
incorrectly rejected in favor of alternative hypothesis).
α is the probability of making a
type I error.
p is judged against a preset α
level of significance (usually
0.05).
If p < 0.05, then there is less than a 5% chance that the data will show something
that is not really there.
Also known as false-positive error.
CONFIDENCE INTERVAL
A Cl is a range that can be interpreted as follows:
If the study were repeated 100 times, the results obtained would lie within that
range in 95 out of the 100 times.
All CIs have a null value. For example, the null value would correspond to 0 mm
Hg (no difference between the SBP in cocoa intake and control groups).
If the Cl crosses the null value, then there is no statistically significant difference
between the groups.
For example if the 95% Cl is [-2.7, -1.3], so, it is entirely negative (does not cross
0) the result is considered statistically significant.
When the 95% CI does not include the null value, this gives a corresponding p-value
<0.05 and the association between exposure and outcome is considered statistically
significant.
A p-value <0.05 reflects that there is a very low probability that the result was due to
chance alone; formally, the p-value is the probability of observing a given (or more
extreme) result due to chance alone assuming that the null hypothesis is true.
A statistically significant 95% CI corresponds to a p-value <0.05.
A statistically significant 99% CI corresponds to a p-value <0.01.
ANOVA
Checks differences between means of 3 or more groups. 3 words: ANalysis Of VAriance.
Example:
Assess for differences in mean blood pressure among 3 sample populations
grouped by exercise status (eg, never exercise, exercise occasionally, or exercise
frequently).
CHI-SQUARE (Χ²)
Checks differences between 2 or more (nominal variables) of categorical outcomes
(not mean values). Pronounce Chi-tegorical.
Example:
In order to test the efficacy of a new drug, compare the number of recovered
patients given the drug with those who were not. Chi-square features nominal data
only, and any number of groups (2×2, 2×3, 3×3, etc.).
Which statistical test will most likely be used to analyze the data?
Comparing the blood sugar levels of husbands and wives t-test.
Comparing the number of staff who do or do not call in sick for each of 3 different
nursing shifts Chi-square test. Staff either call in sick or do not (nominal variable)
over 3 shifts (nominal variable). Two nominal variables with a 2 × 3 design, chi-square.
Relationship between time spent on studying and test score Pearson correlation.
A researcher believes that boys with same-sex siblings are more likely to have higher
testosterone levels t-test.
Twenty rats are coated with margarine and 20 with butter as part of a study to explore
the carcinogenic effects of oleo. Chi-square test. Margarine versus butter (nominal),
cancer versus no cancer (nominal). Therefore, chi-square.
Comparison of passing and failure rates at each of 3 test sites Chi-square test.
Passing versus failure (nominal), 3 sites (nominal).
Comparison of actual measured test scores for students at each of 3 test sites
ANOVA.
MEDICINE IN SOCIETY
PRIMARY, SECONDARY & TERTIARY PREVENTION
MORTALITY REDUCTION
Tobacco use, especially smoking, is the single most preventable cause of death and
disease in the United States.
CANCER FACTS:
The most common cancers among women living in the United States during 2011 were
breast cancer, lung cancer and colon cancer.
Breast cancer is the most common cancer among women in the United States, but its
mortality rate is moderately low.
Despite the fact that the incidence of lung cancer is much lower than that for breast
cancer, mortality from lung cancer is higher. Lung cancer has emerged as the leading
cause of cancer mortality in women. Increased mortality is due to ↑ tobacco use.