Date post: | 16-Mar-2019 |
Category: |
Documents |
Upload: | truongquynh |
View: | 227 times |
Download: | 0 times |
STATISTICAL ANALYSIS, INTERPRETATION AND
PRESENTATION
(issues in medical statistics!)
1,2Robin Christensen, BSc, MSc, PhD
1Senior Biostatistician;
Head of Musculoskeletal Statistics Unit
The Parker Institute
2Associate Professor of Statistics in Medicine;
Institute of Sports Science and Clinical Biomechanics,
University of Southern Denmark
How likely is it, that…… (”Perceiving Probability”)
• You will experience an ”Aeroplane Crash”?
• You will win in the Danish ”Lotto”?
• You will have a myocardial infarction?
• You will have a fracture within the next year?
• You will have Diarrhoea when holidaying in Greece?
• You will fall asleep during this stat sessions? 2
Probability: P =
• From a set of playing cards → I can predict a specific card! (?)
(P = 1/52 = 0.019)
• We would say the null hypothesis (H0) is: ”of course you cannot (stupid idiot)”
• Alternatively (HA): ”Wow – it really works (for you)”
• Thus, if we reject the null hypothesis (H0) – we are likely to believe that ”this guy is efficacious”
• However, more trials are needed to confirm this finding! 3
P < 0.05 (5%)
- Likely to be trustworthy? • Removing all the black cards (same set of playing cards) →
”I can predict a specific card”! • We would say the null hypothesis (H0) is: ”of course you cannot (stupid idiot)”
• HA: ”Wow – He really did it” (having implications for public health) • Happy to reject the null hypothesis (H0) – we believe you
(P < 0.05) • More trials are needed to confirm this finding!
(P = 1/26 = 0.038)
4
WHICH COMPARISONS SHOULD BE MADE?
Hypothesis:
”the number of responders (e.g. ”Staying Alive”) will be higher on the new experimental drug compared to control”
i.e.
the ”chance” that a patient responds will be higher on the new experimental drug (pE) compared to control (pC)
The statistical hypothesis:
H0: pE = pC
If P<0.05 (two-sided) then we might assume
HA: pE ≠ pC
Sampling and probability Estimating the number of obese individuals in a sample!
Obs# BMI 1 18.2 2 20.1 3 22.1 4 24.0 5 27.6 6 28.4 7 30.1 8 32.8 9 38.0
10 42.0
Category Count %
Normal 4 40%
Overweight 2 20%
Obese:class I 2 20%
Obese:class II 1 10%
Obese:class III 1 10% 6
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
Normal Overweight Obese, class I Obese, class II Obese, class III
Category
Perc
enta
ge in
sam
ple
Sampling and probability
Median = 28 kg/m2
4/10 are obese
(40%)
7
Sampling and probability • We infer about a population based on a small subset of the
population (= the sample)
• The sample should be representative of the population from which it is drawn AND for which the determination is being made
• If the sample is not representative of the population, there will be bias in the statistical results - leading to misleading conclusions
• We would expect 40% of individuals in the bigger population to be obese! (?)
• NO – we would NOT!!! (the sample was on patients awaiting Total Knee Alloplasty) 8
Steering group:
Dr Doug Altman
Dr David Moher
Dr Kenneth F. Schulz
Dr John Hoey
http://www.equator-network.org
WHATEVER YOU DO – REMEMBER TO: Include the study design in the title
Explicit design in title #1:
I DON’T HAVE A CLUE HOW TO TREAT MY PATIENT:
A Randomised Controlled Trial
Explicit design in title #2:
THE HEAVIER - THE SMARTER: A Cross-Sectional Study
Explicit design in title #3:
MALE PHYSICIANS EARN MORE MONEY DURING THEIR CAREER THAN
FEMALE PHYSICIANS: A Cohort Study
Explicit design in title #4:
WHY DO SOME OF MY PATIENTS DIE?: A Case-Control Study
Explicit design in title #5:
DO I KNOW WHAT I AM DOING?: A Reliability Study
Explicit design in title #6:
ASK FOR A SECOND OPINION: An Agreement Study
Explicit design in title #7:
DO I KNOW WHICH PATIENT IS REALLY ILL?: A Diagnostic Test Accuracy Study
Explicit design in title #8:
I HAVE READ ALL THESE TRIALS ON THE TOPIC - WHAT’S THE OVERALL EVIDENCE?:
A Meta-Analysis of Randomised Controlled Trials
Three aspects of clinical practice Diagnosis
Therapy
Monitoring: outcomes
Inadequate = wrong diagnosis
Unclear (e.g., no patient impact)
Adequate = important
Explicit evaluation of the importance of outcomes (e.g. OA)
• It is essential to differentiate the critical outcomes from the important ones.
1 ---
3 --- 4 --- 5 --- 6 --- 7 --- 8 ---
2 ---
9 --- Importance of end points Critical for decision making
Important but not critical for decision making
Not important for decision making – of lower importance to patients
Mortality Disablement Need for TKA/THA Pain, Disability, HRQoL
X-Ray (JSW/JSN)
http://www.gradeworkinggroup.org/
ROM/MRI/US/CRP Biomarkers…..
Different types of outcome variables (NOCS)
• Nominal data
• Continuous data
• Ordinal data
Responder / Non-Responder; Death / Alive
Body weight (kg); Muscle strength (Nm);
LDL Cholesterol (mmol/L); Diastolic BP (mmHg)
• Survival data Visual Analogue Scale (0-100 mm VAS)
Likert scales (e.g., No pain, Some pain, Moderate pain, Severe pain)
Time to drop-out (still-in-study: yes/no)
Essential Outcome Measures: Trials
• Nominal data
• Continuous data
• Ordinal data
Responder: yes / no
Weight change: Δkg
Health Related Quality of Life: SF-36 (PCS/MCS 0-100)
Change in KOOS-Pain: ΔScore (-100 to 100)
No pain, Some pain, Moderate pain, Severe pain No pain vs Pain
Statistical Analysis
• Descriptive statistics
• Statistical inference
- Describing the entire sample from statistical estimates
- Presenting ”sufficient” estimates capturing the data
- Unbiased estimates from Unbiased samples
- Point estimation
- Establishing the precision (e.g., 95% CIs)
- Hypothesis test(s)
Statistical models in epidemiology
• Causal models: involves a randomisation technique
(excluding selection bias and confounding by indication)
Causal vs. Empirical models
• Empirical models: estimation from observed data
(includes selection bias and confounders)
• Case-Control studies
• Cross-Sectional studies
• Cohort (longitudinal) studies
Statistical models in epidemiology Case-Control studies
• Compares exposures between Disease (+) vs Disease (-)
• Cases (+) and Controls (-) are representative of a population of interest
• Controls (-) should represent the population from which the Cases (+) arose
Statistical models in epidemiology Cross-Sectional studies
• Assess all individuals at the same point in time
• Prevalence of exposures, risk factors, or disease symptoms
• Can indentify potential causal associations; e.g. correlations between variables
Statistical models in epidemiology Cohort (longitudinal) studies
• Follow people over time
• Obtain information about people and their exposures, let time pass, and then assess the occurrence or status of the outcome
• Common: Make contrast between individuals who are exposed and not exposed
• Prospective cohort studies are more reliable than retrospective cohort studies. . . . .
Illustrating Cross-Sectional studies
Time (years)
Outcome measure
µ (t = 0)
π (t = 0)
Illustrating a Cohort study
Time (years)
Outcome measure
µ (t = 0)
π (t = 0)
µ (t = after)
π (t = after)
Illustrating Person-Years
Time (years)
Number of Persons in Cohorts (no.)
Incidents/Person-Years
Time (years)
Number of Persons in Cohort (no.)
x xx
x
x
x xx
x
x x x
x
x xx
x
Reader’s guide to critical appraisal of cohort studies • What comparison is being made? • Does the comparison make clinical sense? • What are the potential selection biases? • Has there been a systematic effort to identify and measure potential confounders? • Is there information on how the potential confounders are distributed between the comparison groups? • What methods are used to assess differences in the distribution of potential confounders? • Are the analytical strategies clearly described? • Do different analytical strategies used yield consistent results? • Are the results plausible?
What is bias?
• A systematic error, or deviation from the truth, in results or inferences
http://www.cochrane-handbook.org/
• Biases can operate in either direction: different biases can lead to underestimation or overestimation of the true intervention effect
• Some are small (and trivial compared with the observed effect) and some are substantial (so that an apparent finding may be entirely due to bias)
• More rigorous studies are more likely to yield results that are closer to the truth.
Bias?
-20 -10 0 10 20
Effect of Intervention
High-Risk of Bias Low-Risk of BiasCombined High-Risk of Bias Combined Low-Risk of Bias
Favors Intervention Favors Placebo
Bias?
-20 -10 0 10 20
Effect of Intervention
High-Risk of Bias Low-Risk of BiasCombined High-Risk of Bias Combined Low-Risk of Bias
Favors Intervention Favors Placebo
X2
X11 X1
X8
X7 X18
X4
X3 X16 X22
X12
X9 X10
X21
X20 X19
X17
X15 X5 X6
X25
X14
X23 X13
X24
Randomized Controlled Trial
X2
X11 X1
X8
X7 X18
X4
X3 X16 X22
X12
X9 X10
X21
X20 X19
X17
X15 X5 X6
X25
X14
X23 X13
X24
EXPERIMENTAL
CONTROL
Randomized Controlled Trial
X2
X11 X1
X8
X7 X18
X4
X3 X16 X22
X12
X9 X10
X21
X20 X19
X17
X15 X5 X6
X25
X14
X23 X13
X24
EXPERIMENTAL
CONTROL
NO!
– THIS IS NOT AN RCT!
Randomized Controlled Trial
• PROTOCOL (i.e. Objective & primary outcome) • www.CLINICALTRIALS.gov
• INFORMED CONSENT
• (BASELINE MEASUREMENTs)
Randomized Controlled Trial
Evidence Synthesis
’The PICO framework’
• Patients
• Intervention(s)
• Comparison(s)
• Outcome(s)
Osteoarthritis (i.e. ACR crit.)
Exercise (add-on: Concomitant med.)
Nothing (add-on: Concomitant med.)
Patient important outcome?
Clinician important outcome?
X1
X8
X7 X18
X4
X3 X16 X22
X12
X9 X10
X21
X20 X19
X17
X15 X5 X6
X25 X24
ELIGIBLE PATIENTS Randomized Controlled Trial
X1
X4
X6 X7
X11
X12 X13 X14
X2
X5 X3
X8
X10 X9
X15
X16 X17 X18
X19 X20
ELIGIBLE PATIENTS - Included Randomized Controlled Trial
X1
X4
X6 X7
X11
X12 X13 X14
X2
X5 X3
X8
X10 X9
X15
X16 X17 X18
X19 X20
ELIGIBLE PATIENTS - Randomized Randomized Controlled Trial
Randomized Controlled Trial (RCT)
Randomized Controlled Trial (RCT)
Net benefit = 0.75 – 0.25 Net Benefit = 0.50 point
(ie, 50%point benefit compared to placebo; NNT=2)
Randomized Controlled Trial (RCT)
Statistical inference –
Unbiased samples….
CO
NSO
RT 2010 checklist of inform
ation to include w
hen reporting a randomised trial
INTRODUCTION
METHODS
RESULTS
DISCUSSION
Risk of bias: RCTs http://www.cochrane-handbook.org/
(1) Sequence generation
(2) Allocation concealment
(3) Blinding of participants, personnel and outcome assessors
(4) Incomplete outcome data
(5) Selective outcome reporting
(6) Other sources of bias Christensen and Bliddal Arthritis Research & Therapy 2010, 12:105
Risk of bias http://www.cochrane-handbook.org/
Sequence generation:
“Describe the method used to generate the allocation sequence in sufficient detail to allow an assessment of whether it should produce comparable groups”
Was the allocation sequence adequately generated?
Adequate
Unclear
Inadequate
INCLUDED PATIENTS - Randomized
We generated the two comparison groups using simple randomization,
with an equal allocation ratio (1:1), by referring to a table of random numbers
Random Code: {0, 1, 1, 1, 1,
0, 1, 0, 0, 0,
1, 0, 0, 1, 1,
0, 1, 0, 0, 1}
Randomized Controlled Trial
Risk of bias http://www.cochrane-handbook.org/
Allocation concealment:
“Describe the method used to conceal the allocation sequence in sufficient detail to determine whether intervention allocations could have been foreseen in advance of, or during, enrolment.”
Was allocation adequately concealed?
Adequate
Unclear
Inadequate
Allocation sequence concealment
• Central randomization
• Sequentially numbered drug containers • Sequentially numbered, opaque, sealed envelopes
Randomized Controlled Trial
Til afdelingens sekretærer!
Vedr. Projekt KF-13-11-08
Hvis du ser en mulig deltager i konsultationen - der vil udfylde ’informed consent’ – da allokér patienten til gruppen som angivet t.h. for kaffemaskinen. Tak!
MVH
Dr. Kokren
Randomized Controlled Trial:
The Parker Institute:
Musculoskeletal Statistics Unit
Til afdelingens sekretærer!
Vedr. Projekt KF-13-11-08
Hvis du ser en mulig deltager i konsultationen - der vil udfylde ’informed consent’ – da allokér patienten til gruppen som angivet t.h. for kaffemaskinen. Tak!
MVH
Dr. Kokren Concealed allocation?
Risk of bias http://www.cochrane-handbook.org/
Sequence generation:
Was the allocation sequence adequately generated?
Allocation concealment: Was allocation adequately concealed?
Adequate
Unclear
Inadequate
Adequate
Unclear
Inadequate
Blinding of participants, personnel and outcome assessors
When considering the risk of bias from lack of blinding it is important to consider specifically:
1. who was and was not blinded;
2. risk of bias in actual outcomes due to lack of blinding during the study (e.g. due to co-intervention or differential behaviour);
3. risk of bias in outcome assessments (considering how subjective or objective an outcome is)
Risk of bias http://www.cochrane-handbook.org/
Blinding of participants, personnel and outcome assessors:
“Describe all measures used, if any, to blind study participants and personnel from knowledge of which intervention a participant received. Provide any information relating to whether the intended blinding was effective.”
Was knowledge of the allocated intervention adequately prevented during the study?
Adequate
Unclear
Inadequate
Handling: Incomplete outcome data
• All envelopes opened ~ Intention-to-treat population
• All randomized included in the analyses
• Use an Intention-to-Treat analysis (Non-responder analysis most appropriate on average)
Randomized Controlled Trial
Statistical inference - applied
100 100
92 80
Handling: Incomplete outcome data
n[E] n[C] N[E] N[C] p[E] p[C]ITT 25 25 100 100 25.0% 25.0%PP #1 25 25 67 90 37.3% 27.8%PP#2 25 25 90 67 27.8% 37.3%Modified ITT 25 25 95 95 26.3% 26.3%
i.e., the results may vary according to something not being the treatment
Randomized Controlled Trial
Incomplete outcome data
0.0%
10.0%
20.0%
30.0%
40.0%
50.0%
60.0%
70.0%
80.0%
90.0%
100.0%
0.0% 10.0% 20.0% 30.0% 40.0% 50.0%
Attrition (drop-out) Rate [%]
Res
pons
e R
ate
(%)
True effect!
Biased estimate!
Bias
LUNDEX ( L ) simulation scene
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
0 12 24 36 48
Duration (Months)
Prop
ortio
n re
spon
ding
After 6 months~ 50% responds and sustains (ITT)
N = 1000
n = 950 n = 900 n = 800
n = 700 n = 600
LUNDEX
Risk of bias http://www.cochrane-handbook.org/
Incomplete outcome data:
“Describe the completeness of outcome data for each main outcome, including attrition and exclusions from the analysis. State whether attrition and exclusions were reported, the numbers in each intervention group (compared with total randomized participants), reasons for attrition/exclusions where reported, and any re-inclusions in analyses performed by the review authors.”
Were incomplete outcome data adequately addressed?
Adequate
Unclear
Inadequate
Selective outcome reporting • Global pain score for index joint
• Pain on walking for index joint
• Western Ontario McMaster Universities (WOMAC) Pain subscale
• Lequesne index
• Pain in index joint during activities other than walking
Selective outcome reporting • Global pain score for index joint
• Pain on walking for index joint
• Western Ontario McMaster Universities (WOMAC) Pain subscale
• Lequesne index
• Pain in index joint during activities other than walking
P = 0.04
P = 0.01
P = 0.09
P = 0.17
P = 0.06
Risk of bias http://www.cochrane-handbook.org/
Selective outcome reporting:
“State how the possibility of selective outcome reporting was examined by the review authors, and what was found.”
Are reports of the study free of suggestion of selective outcome reporting?
Adequate
Unclear
Inadequate
http://www.equator-network.org
http://www.lean.org/
Principles of Lean Thinking (Five steps guiding the implementation of lean)
Specify value from the standpoint of the end customer by product family
Identify all the steps in the value stream for each product family, eliminating whenever possible those steps that do not create value
Make the value-creating steps occur in tight sequence so the product will flow smoothly toward the customer
MAPPING THE VALUE STREAM: Reporting Health Research TITLE (Identify value)
INTRODUCTION (Setting the scene) • Background (What we know!)
• Rationale (Why this is important!)
• Objective (Specific aim!) METHODS (What we anticipate)
• Participants (incl/excl criteria)
• Interventions (comparison(s))
• Outcomes (primary/secondary)
• Sample size (how many)
• Statistical methods (hypotheses)
PROTOCOL (Identify ”the customer”)
RESULTS !!?
• Participant flow
• Recruitment
• Baseline data (ITT population)
• Outcomes and estimation
• Ancillary analyses (post hoc?)
DISCUSSION (What happened?)
• Interpretation
• Generalizability
• Overall evidence
• Implications for practice
• Implications for research
SUBMIT? (Customer/Consumer)
TITLE
INTRODUCTION
METHODS
PROTOCOL
RESULTS
DISCUSSION
SUBMIT
MAPPING THE VALUE STREAM
PUBLISH
(Register)
Population
Stikprøve (N) Stikprøve (N)
Diagram over det teoretiske udfaldsrum for et statistisk velfunderet design (modificeret efter Lund & Røgind, 2004)
Overblik over Stikprøven (N)
Her kan man med rette lave RCT
Intervention (I) Kontrol (K)
Hypotese test:
I = K
Estimation (konfidensinterval)
(I – K ) ± 95% KI
π µ