Transcript Document

Demystifying statistical language
in clinical research reports
Tuesdays with Faculty: An EBP Evening Series
February 27, 2007
David M. Thompson PT
email: [email protected]
web: http://moon.ouhsc.edu/dthompso/
• A
• B
• C
Orienting to statistical procedures
RESEARCH QUESTION and OUTCOME OF INTEREST
LEVEL OF MEASUREMENT OF OUTCOME VARIABLE
Nominal
Ordinal
Interval or Ratio
SOURCE OF SAMPLE
Independent or correlated observations
Number of Samples (1,2, more than 2)
ASSUMED DISTRIBUTION OF OUTCOME VARIABLE
“Parametric” procedures for outcomes with known or
assumed distributions
Non-parametric procedures are “distribution free.”
INFERENTIAL FOCUS
Estimation: point estimate and confidence interval.
Hypothesis testing: p-values and associated null
hypotheses
Locate the key research questions
•
•
•
•
Patient
Intervention
Comparison
Outcome
PICO (McMaster University)
Example
PICO (McMaster University)
• P – adults with shoulder pain due
to impingement
• I – Therapeutic exercise
• C – Rest and NSAIDs
• O - improved function
Types of questions
Defining question type facilitates
search for information
• Therapy / Intervention
• Diagnosis
• Etiology
• Prognosis
“User’s Guides”
How is outcome measured?
• Count
• Proportion
• Continuous
– test score
– BMI
– blood pressure
• Time to event
– disease progression
– return to work
Statistics match
outcome’s level of measurement
• Count
• Proportion
(between-group differences)
• Time to event
(median times by group)
• Continuous
(differences in means)
Nominal
Number of
groups
Observations
independent
or correlated
Binomial
1 sample
2 samples
X2
Goodness
of fit (GOF)
Independent Chi-square
Fisher’s
exact
correlated or McNemar
matched
X2
Independent
Chi-square
k samples
correlated or Cochran Q
matched
Kappa
Association
Regression
and
Modeling
Level of Measurement
Ordinal
Interval or Ratio
Distrib.
Distrib.
unknown
Assumed
normal
Kolmogoro Test for
One
v-Smirnov
symmetry
sample t
WilcoxonMannWhitney
Wilcoxon
signed rank
test
KruskalWallis one
way AOV
Friedman
AOV by
ranks
Permutation Indep. t
tests
Permutation Paired t
tests
Analysis
of
variance
ANOVA on
repeated
measures
/ mixed
models
Spearman rank order
correlation coefficient
Kendall tau
Kendall W
Logistic regression
Poisson regression
Generalized estimating
equations
Linear
regression
Online sources for evidence
• Pubmed
http://www.ncbi.nlm.nih.gov/entrez/quer
y.fcgi
• OUHSC library
http://library.ouhsc.edu
Inference
• Estimation
• Hypotheses testing
Estimation
• Point estimate
• typically an unbiased estimator of a
population quantity
• Interval estimate
• 95 % confidence interval (CI) typically
center on point estimate
“plus or minus”
[(z or t) * SE of point estimate]
Hypothesis Testing
• Determine if results are compatible with
assumption that null hypothesis is true.
• Null is typically an assumption of NO
difference, no association, no effect.
p values
• pr(“of obtaining a test statistic of this
value or larger” | H0 is true)
OR
• pr(you obtained this sample | H0)
• Test statistic is based on observations,
and on assumption that null is TRUE.
• Statistic’s non-significance CANNOT
IMPLY that the null is true (especially
when power is low).
Errors Associated with Hypothesis Tests
• Type I
• Rejecting a null hypothesis that is true
•  = p(reject Ho | Ho)
• Type II
• Failing to reject null when alternative
hypothesis is true OR
• Failing to reject false null
•  = p(fail to reject H0 | Ha)
Power and Sample Size
•  = p(fail to reject H0 | Ha)
• 1- = p(reject H0 | Ha) = POWER
• A test’s power is its probability of making the
correct decision (rejecting the null
hypothesis) when a specific alternate
hypothesis is true
Power and sample size
• POWER=1- = p(reject H0 | Ha)
• a function of:
• Ha, a specific, stated alternative
hypothesis, so requires specification of
effect size
• the known or estimated variability.
• Estimates of variability depend in turn on
sample size. Large samples provide more
precise estimates of variability, and so
also provide greater power.
• http://moon.ouhsc.edu/dthompso/CDM/powe
r/hypoth.htm
Power and sample size calculations
• All calculations require an estimate
of variability.
POWER
SAMPLE
SIZE
EFFECT
SIZE
Levels of evidence
Modified after: SUNY Downstate Medical Center, Medical Research
Library of Brooklyn. (2005). A guide to research methods: The
evidence pyramid. Retrieved January 4, 2006 from
http://servers.medlib.hscbklyn.edu/ebm/2100.htm.
Online resources for
evidence-based practice
http://moon.ouhsc.edu/dthompso/CDM/ebplinks.htm