Transcript STA 101: Properly setting up and designing a clinical

STA 101: Properly Setting up and
Designing a Clinical Research
Study Including Power Analysis
for Proper Patient Numbers
Lecturer: Dr. Daisy Dai
Department of Medical Research
1
Who are biostatisticians?

Ashley Sherman





Phone: 816-701-1347
[email protected]

Daisy Dai


Phone: 816-701-5233
Email: [email protected]


Consultation
Experimental design
and sampling plan
Collaboration in
presentation and
publication of
studies
Education
Research
2
Statistical Courses


SPSS 201: Using SPSS to
perform statistical tests I
(Sep 23rd)

SPSS 202: Using SPSS to
perform statistical tests II


SPSS 204: Using SPSS to
manage data

SPSS 203: Summarize data
with tables and graphs

STA 101: Properly Setting up
and Designing a Clinical
Research Study Including
Power Analysis for Proper
Patient Numbers (July 16th)
STA 102: Commonly Used
Statistical Tests in Medical
Research - Part I (Aug. 20th)
STA 103: Commonly Used
Statistical Nonparametric
Tests in Medical Research Part II (Nov. 5th)
3
Statistics on Scope



Daisy’s statistics website is located in
“Research” tab under scope main page.
Link:
http://www.childrensmercy.org/content/
view.aspx?id=9740
The most useful categories are “SPSS”,
“Useful links” and “Course”.
4
Why do we need sample
size/power calculation in medical
research?



Grant application/IRB study protocol
Peer reviewed journal publication
Journal review
5
Medical Research

Clinical Trials



Intervention or therapeutic
Preventative
Retrospective Studies
6
Statistics

Descriptive Statistics



Methods to organize and summarize information
Mean, median, max, min, frequency and
proportions, etc. that summarize sample
demographics
Inferential Statistics

Methods to draw conclusions about a population
based on information obtained from a sample of
the population
7
Population
Sampling Plan
Conclusion
Inferential
Statistics
Sample
Descriptive
Statistics
8
Information Collections
1.
Historical Data


2.
Census


3.
Pro: Convenient; Save a lot of work
Con: Outdated; Different Objectives and Designs;
Unknown Detailed Information
Pro: reliable, accurate and comprehensive (e.g.
Population census)
Con: Time consuming; requiring more resources; difficult to
investigate all subjects in the population
Sampling


Pro: Efficient; Less risky; exploratory; informative
Caveats: Selection bias; misinterpretation; design flaw
9
Misconducts in Sampling



A clinical foundation used the average weight of a sample of
professional football players to make an inference about the
average weight of all adult males.
A local newspaper estimated the median income of California
residents by sampling the incomes of Beverly Hills residents.
Before the presidential election in 1936, the Literary Digest
magazine conducted an opinion poll and predicted that Alfred
Landon, the Republican candidate, would win the election.
However, Franklin Roosevelt, the democratic candidate, won by
the greatest landslide in the history of presidential elections!
10
Why do we need sampling
plan?

Warrant Research Ethics.


Improve Research Efficiency.


Too many participants could put more subjects under risk.
A un-planned study with too many participants may take
longer to finish and require more resources but miss the
early opportunity to publish interesting findings.
Deliver Reliable Information.

A study without sufficient subjects may lose evidence to
demonstrate potential effects, which could waste resources
or generate misleading information to readers.
11
Protocol – Surgical resection for
patients with gastric cancer

“Sample size calculation were based on a prestudy survey of 26 surgeons, which indicated
that the baseline 5-year survival rate of D1
surgery was expected to be 20%, and an
improvement in survival to 34% (14%
chance) with D2 resection would be a realistic
expectation. Thus 400 patients (200 in each
arm) were to be randomized, providing 90%
power to detect such a difference with pvalue<0.05. ” [1]
[1] Cushieri et al. (1999) Patient survival after D1 and D2 resections for gastric cancer: long-term results of the MRC
randomized surgical trial. Surgical Co-operative Group.
12
Protocol – Steroid or cyclosporine
for oral lichen planus

“It is anticipated that in patients taking
topical steroids, the response rate at 1 month
will be approximately 60%. It is anticipated
that this may be raised to as much as 80% in
those receiving cyclosporine. With two-sided
test size 5%, power 80%, then the
corresponding number of patients required is
approximately 200.” [2]
[2] Poon et al. (2006) A randomized controlled trial to compare steroid with cyclosporine for the topical treatment of
oral lichen planus
13
Three Steps to Calculate Sample
Size



Step 1: Establish study design and
study objectives.
Step 2: Select the outcome variables.
Step 3: Collect information and
determine sample size.
14
Key Elements in Sample Size
Calculation



The level of statistical significance.
The anticipated clinical difference
between treatment groups.
The chance of detecting the anticipated
clinical difference.
15
Statistical Testing Procedures
1.
Null Hypothesis
-
2.
Ho: Mean_Treatment=Mean_Control
Alternative Hypothesis
-
-
Ha: Mean_Treatment ≠ Mean_Control (Two-sided Test)
Ha: Mean_Treatment > Mean_Control (One-sided Test)
Ha: Mean_Treatment < Mean_Control (One-sided Test)
3.
Calculate statistics
4.
Make Inference
-
If P-value > 0.05, then Ho holds
If P-value < 0.05, then Ha holds
16
Two Types of Decision Errors




Action:
Support
Ho
Type I error ( )
The probability of claiming a
significant difference between two
treatments that are actually in
parity.
Usually  = 0.05
Action:
Support
Ha
Fact:
Ho is true

Type I
error
Fact:
Ha is true
Type II
Error

Type II error (1- )


The probability of failing to
differentiate two treatments.
Ideally, 1-  0.2.
17
Effect Size (  )

The standardized difference between
means of two treatments:
T   C


18
Software

Commercial software: nQuery Advisor 7.0



Product Website:
http://www.statsol.ie/index.php?pageID=2
User Guide
http://www.statsol.ie/documents/nQ70_version2_
manual.pdf
Free software: PS 3.0

http://biostat.mc.vanderbilt.edu/twiki/bin/view/Ma
in/PowerSampleSize
19
Compare means in two groups
Control
Test
20
Case Study: Asthma Control Test



An asthma control Test has been conducted to
develop a patient-based tool for identifying patients
with poorly controlled asthma.
Mean of total ACT score for the poorly controlled
group (Control) is 15 and mean of total ACT score for
the well controlled group (Test) is 21. Assume the
standard deviation of total ACT score is 4.
The effect size between Control and Test is

T  C 21  15

 1.5

4
21
22
nQuery Advisor
23
Compare Means
24
Compare Means
25
1.0 0.98
0.9
0.85
Power
0.8
0.7
0.6
9
0.5
4
6
8
15
10
12
14
16
Sample Size Per Group
26
Compare proportions in two
groups
Control
Test
27
Case Study: Asthma Control Test



A researcher is interested to compare allergic
asthmatic patients versus non-allergic asthmatic
patients in response to an antihistamine treatment.
After treatments, patients will evaluate their asthma
status as 0-very bad, 1-bad, 2-good and 3-very
good.
This researcher needs to find out the sample size and
power of a study that hypothesizes 80% of allergic
cohort versus 60% of non-allergic cohort will be in
good or very good status.
28
nQuery Advisor - Proportion
29
Compare Proportions
30
Compare Proportions
31
Agreement Test (Kappa Score)
32
Case Study: Helmet Cure



Children with flat head
syndrome will wear helmet to
keep their head in shape. The
diagnosis and severity of flat
head vary by physicians.
A study is planned to compare
the rating consistency among
physicians.
Assume that 50% of reviewed
cases will be diagnosed as flat
head syndrome. The null
hypothesis assumes only 0.4
(slight) degree of agreement
between two physicians. The
alternative hypothesis assumes
0.7 (strong) degree of
agreement.
33
nQuery Advisor
34
Assess agreement
35
Assess Agreement
36
By Julius Sim and Chris Wright
37
Sample Size Calculation for Nonparametric tests
38
What is non-parametric test?


Tests that are
distribution free.
Compare medians
rather than mean.




Wilcoxon Signed
Rank Test
Wilcoxon Rank Sum
Test
Kruskall Wallis Test
We will cover these
tests in details with
more examples in
STA103.
39
Case Study: Seroxatene
A studied was conducted to evaluate whether a new antidepressant, Seroxatene has a benefit of pain relief. Patients
(n=28) with MRI-confirmed disk herniation and symptomatic leg
pain were enrolled and randomly assigned to receive Seroxatene
or a placebo for 8 weeks. At the end of the study, patients were
asked to provide a overall rating of their pain, relative to
baseline.
Deterioration
Marked
Moderate
Slight
No
Change
-3
-2
-1
0
Improvement
Sight
Moderate
Marked
1
2
3
40
Pain Relieving Scores
------- Placebo Group -------
------- Seroxatene Group -------
ID
Score
ID
Score
ID
Score
ID
Score
1
3
15
0
2
0
16
-1
4
-1
18
-1
3
2
17
2
7
2
19
-3
5
3
20
-3
6
3
21
3
8
-2
22
3
10
1
24
0
12
3
26
2
14
3
27
-1
9
3
23
-2
11
-2
25
1
13
1
28
0
41
Histograms of Pain Scores
Group: Placebo
Group: Seroxatene
2.5
6
5
2.0
Frequency
Frequency
4
1.5
3
1.0
2
0.5
1
Mean = 1.12
Std. Dev. = 2.029
N = 16
Mean = 0.08
Std. Dev. = 1.975
N = 12
0
0.0
-3
-2
-1
0
PainScore
1
2
3
-3
-2
-1
0
1
2
3
PainScore
42
Sample Size Calculation for
Nonparametric Tests


Although the non-parametric tests do
not reply on distribution, the
corresponding sample size calculation is
based on distribution.
A general rule of thumb is to compute
the sample size required for a t test and
add 15%.
43
Practicalities

More than one primary outcome

Internal pilot studies

More than two groups
44
Rules of Thumb



The level of significance needs to be
determined beforehand.
One can balance the testing sensitivity and
resources by appropriately choose sample
size and power.
Feel free to consult statisticians if you have
questions. Here we discussed some principles
in sample size calculation. More sophisticated
methods are available for experimenters.
45
Summary




Review research ethics.
Avoid research misconducts.
Raise awareness in statistical sampling and
design.
Learn basic sample size and power calculation
for means, proportions and agreement.
46
Thank You
For more information, visit my website
http://www.childrensmercy.org/content/vi
ew.aspx?id=9740
Or go to Scope -> Research -> Statistics

47