Transcript EBPP_Lecture_3

Evidence Based Practice in
Psychology – Lecture 3
May 31, 2007
Basic Concepts in Epidemiology
Basic Components of Epidemiological
Research





Health Outcome
Explanation
Key Variables: Exposure, disease, control
variables
Key Methods: Surveys, interviews, samples,
laboratories
Key Designs: Clinical trials, cross-sectional,
case-control, cohort
Design Considerations



Experimental: RCT’s
Observational: descriptive and analytic
Directionality
–
–
–

Forward: Cohorts and RCT’s
Backward: Case-control
Neither: Cross-sectional
Timing
–
–
Prospective: health outcome occurs after study begins
Retrospective: health outcome occurs before study begins
More on RCT’s


May be used to test preventative or
therapeutic interventions
Key features:
–
–
–
–
Randomization (control)
Blinding (minimizing bias)
Ethical concerns (stopping rules)
ITT analysis
More on Cohort Studies





Best living example: Framingham Heart Study (n = 5100, examined every 2
years)
Information about risk factors and disease states collected
Prospective analysis of health outcomes as a function of risk factors
Cohort studies may also be retrospective
Advantages:
–
–
–
–

Forward directionality
Exposure, not disease status affects selection, so relatively free of bias
Useful for examining relatively rare exposures
Retrospective study can be inexpensive and quick (e.g., based on employment
records or death certificates)
Disadvantages:
–
–
–
Attrition due to migration, lack of participation, withdrawal and death
Inefficient for studying rare disease with long latency
Exposed might be followed more closely than nonexposed, creating spurious
exposure-disease relationship
More on Case-Control Studies



Subjects selected on the basis of their disease status; first
selects cases of a particular disease, then controls without the
disease (preferably from same population)
Issue: selection of controls
Advantages:
–
–

Good for studying chronic or rare diseases with long latency
periods
Require smaller sample sizes than other designs
Disadvantages
–
–
–
Don’t allow several diseases to be evaluated, as do cohort studies
Don’t allow disease risk to be estimated directly because they
work backward from disease to exposure
More susceptible to bias
Measures of Disease Frequency






Rate
Proportion
Risk (favored for RCT’s and cohort studies)
Odds (favored for case-control, retrospective
studies)
Prevalence
Incidence
Incidence & Prevalence



Incidence: NEW cases of a disease that develop over a period
of time; useful for identifying risk factors and disease etiology;
estimated from RCT’s and cohort studies
Prevalence: EXISTING cases of a disease at a particular point
in time or over a period of time; estimated from cross-sectional
or case-control studies; useful for planning health care services
(demand for healthcare)
P = I x D, where D = duration
Additional Formulae



P = C/N, where C=existing cases, and N =
steady-state population size
CI = I/N, where CI = cumulative incidence
(measures risk), I = new (incident) cases; N =
size of disease-free cohort
IR = I/PT, where IR = incidence rate, and
PT=accumulated person-time information
IR = I/PT = 5 new cases/25 disease free person years = .20
Rate



Definition: measure of how rapidly health
events (e.g., new diagnosis of disease,
death) are occurring in a population of
interest
Instantaneous: rate at a particular point
Average: rate over time (preferred)
Risk





Probability that an individual will develop or die from a given disease,
or will develop a health status change over a specified followup period
Assumes that the individual does not have the disease at the start of
the followup and does not die from any other cause during the followup
0< risk < 1
Necessary to give the followup period over which risk is to be
predicted (e.g., 24 months, etc.)
Risk factors
–
–

Attribute (e.g., genetic susceptibility, age, sex, etc.)
Exposure (e.g., nutrition, toxicity, injury, etc.)
Risk does not have to refer to disease – could refer to any symptom,
side effect, etc. as long as information relevant to such events are
measured
Risk measures




Attributable risk (AR): Risk in exposed group – risk in nonexposed group
Absolute risk reduction (ARR): similar to AR, but in response to an
intervention; it indicates the reduction in risk associated with exposure to an
intervention
Etiologic fraction (Population AR): proportion of all cases of a disease that
are attributable to an exposure or risk. Proportion of disease in the population
that would be eliminated if the risk factor was eliminated or prevented
Relative Risk (RR): Value between 0 and ∞ that indicates the strength of the
risk factor and disease outcome.
–


Calculated by: Risk(exposed)/Risk(unexposed)
Exposure Odds Ratio (OR): estimate of RR derived from a case-control study;
similar to relative risk when disease is rare
Number Needed to Treat (NNT): number of individuals that would need to be
treated to prevent one adverse outcome in that group of similar individuals at
risk of the problem. Establishes benefit of an intervention compared to doing
nothing. NNT is the reciprocal of AR or ARR
Number Needed to Treat
NNT = 1 / |ARR|
NNT’s for interventions should be relatively low, for preventative studies a
little higher
In a randomized controlled trial looking into the long-term outcome for stroke patients treated in
stroke units (SU) compared with patients treated in general wards (GW), the mortality rate 5
years after the onset of stroke was 59.1% in the patients treated in SU and 70.9% in those
treated in the GW. How many patients need to be treated in stroke units to prevent one
additional death? (Stroke 1997; 28:1861-6)
NNT = 1 / |.709-.591| = 1 / .118 = 8.5 or 9
Number Needed to Harm
NNH = 1 / |ARR|, where “risk” is of adverse side effects
NNH’s for interventions should be relatively high, at least compared to
NNT’s, the higher for more deleterious side effects.
In a randomized clinical trial of a drug for movement disorder in Parkinson’s disease, a certain
number of adverse effects were recognized. In the treated group 140 of 539 (25.97%) patients
developed a clinically measurable memory problem when assessed a year later, while in the
nontreated group, 104 of 513 (20.27%) developed such a disorder.
NNH = 1 / |.260-.203| = 1 / .118 = 17.5 or 18
This means that 18 patients would have to be exposed to the drug to produce one additional
case of memory disorder that would not have appeared naturally in the untreated group
Risk v. Rate



Risk required in studies predicting change in
health status for individual, or in prognostic
studies; rate not useful at the individual level
Risk usually preferred because it is easier to
interpret
Sometimes risk is difficult to measure
(population studies)
Measures of Effect – Risk Ratio
Smokers with heart attacks followed over a 5 year period.
Smoke
Quit
Total
Died
27
14
41
Survived
48
67
115
Total
75
81
156
5-year death risk:
Smokers: 27/75 = 0.36
Quitters: 14/81 = 0.17
Estimated RR = .36/.17 = 2.1
AR= 0.36 – 0.17 = 0.19
Measures of Effect – Odds Ratio
Case-control study – outbreak of GI disease at resort;
cases had diarrhea, controls stayed at resort but did
not
Cases
Controls
Total
Ate raw
Hambg
Did not eat
raw hambg
Total
17
7
24
20
26
46
37
33
70
Odds = P/(1-P)
Cases = .46/(1-.46) = 0.85
Controls = .21/(1-.21) = 0.27
Odds Ratio = .85/.27 = 3.2
Proportions:
0.46
0.21
Alternatively: OR = (a x d)/(b x c)
OR = (17x26)/(7x20) = 3.2
Diagnostic Testing Issues





How well tests perform relative to a “gold standard”
is critical to establishing their empirical basis
Often, the issues of cost-effectiveness and risk play
a role in test selection
Study design, not just test statistics, is important
Cross validation is key
STARD initiative (next slide)
–
http://www.consort-statement.org/stardstatement.htm#flow
Brunswik Lens Model of Clinical Prediction
mTBI
Positive
(Abnormal)
19
Negative
(Normal)
17
Totals
36
No mTBI
a
b
c
d
TOTALS
10
29
276
293
286
322
Sensitivity = a/(a+c) = 19/36 = .53
Specificity = d/(b+d) = 276/286 = .97
Positive Predictive Power = a/(a+b) = 19/29 = .67 (also known as PTL+)
Negative Predictive Power = d/(c+d) = 276/293 = .94
Post-Test Likelihood given Negative Result = 1-NPP = .06
Prevalence = (a+c)/(a+b+c+d) = 36/322 = .11 (pretest probability of d/o)
Pre-test Odds = PTP/(1-PTP) = .11/.89 = .12 (.12:1)
Likelihood Ratio of Positive Test = [a/(a+c)]/[b/(b+d)] = Sens/(1-Spec) =
.53/.03 = 17.67 (17.67:1)
Likelihood Ratio of Negative Test = [c/(a+c)]/d/(b+d)] = (1-Sens)/Spec =
.47/.97 = .48 (.48:1)
Pre-test Odds x LR+ = PPV
Pre-Test Odds x LR- = 1-NPV
Diagnostic Odds Ratio = LR+/LR- = 17.67/.48 = 36.81
SnNout and SpPin


If a test has extremely high Sensitivity and
LR+ (say >20), a Negative test result pretty
much rules out the target disorder.
If t test has high Specificity and the LR- is
very low (say <.05), a Positive test rules in
the target disorder
ROC Analysis
BNP > 76
BNP > 18
Sens = 26/40 = .65; Sens = 35/40 = .88
Spec = 75/86 = .87; Spec = 29/86 = .34
Evaluating Studies of Tests

Was an appropriate spectrum of patients included?
–

All patients subjected to a Gold Standard?
–

(Verification Bias)
Was there an independent, "blind" comparison with a
Gold Standard?
–

(Spectrum Bias)
Observer Bias; Differential Reference Bias
Methods described so you could repeat test?