Biostatistics: Study Design Peter D. Christenson Biostatistician Summer Fellowship Program July 2, 2004
Biostatistics: Study Design
Peter D. Christenson
Biostatistician
Summer Fellowship Program July 2, 2004
Outline
• Example
• Statistical Issues in Research Studies
• Typical Flow of Data in Research Studies
• Biostatistical Resources at LA BioMed and GCRC
• Size and Power of Research Studies
Example : Design Issues
Statistical Aspects of Research Projects
• Target population / sample / generalizability.• Quantification of hypotheses, case
definitions, endpoints. • Control of bias; confounding.• Comparison/control group.• Randomization, blinding.• Justification of study size (power, precision,
other); screened, enrolled, completed.• Use of data from non-completers.• Methods of analysis.• Mid-study analyses.
Typical Flow of Data in Research Studies
Reports
Spreadsheets
Statistics Software
Graphics Software
SourceDocuments
Database
CRFs
Database is the hub: export to applications
Biostatistical Resources at REI and GCRC
• Biostatistician: Peter Christenson
– Study design, analysis of data
• Biostatistics short courses: 6 weeks 2x/yr
• GCRC computer laboratory in RB-3
– Statistical, graphics, database software
– Contact Angel at 781-3601 for key code
• Webpage: http://gcrc.humc.edu/Biostat
NCSS: Basic intuitive statistics package in GCRC computer lab; has power module
SPSS: More advanced statistics package in GCRC lab
SAS: Advanced professional statistics package in GCRC lab
Sigma Plot: Scientific publication graphics software in GCRC lab
nQuery: Professional study size / power software in GCRC lab
http://gcrc.humc.edu/Biostat
www.statsoft.com/textbook/stathome.html
Good general statistics book by a software vendor.
www.StatCrunch.com
NSF-funded software development.
Not a download; use online from web browsers
www.stat.uiowa.edu/~rlenth/PowerOnline Study Size / Power Calculator
Statistical Aspects of Research Projects
• Target population / sample / generalizability.• Quantification of hypotheses, case
definitions, endpoints. • Control of bias; confounding.• Comparison/control group.• Randomization, blinding.• Justification of study size (power, precision,
other); screened, enrolled, completed.• Use of data from non-completers.• Methods of analysis.• Mid-study analyses.
Randomization
• Helps assure attributability of treatment effects.
• Blocked randomization assures approximate chronologic equality of numbers of subjects in each treatment group.
• Recruiters must not have access to randomization list.
• List can be created with a random number generator in software (e.g., Excel, NCSS), printed tables in stat texts, pick slips out of a hat.
Study Size / Power : Definition
• Power is the probability of declaring a treatment effect from the limited number of study subjects, if there really is an effect of a specified magnitude (say 10) among all persons to whom we are generalizing.[ Similar to diagnostic sensitivity. ]
• Power is not the probability that an effect (say 10) observed in the study will be “significant”.
Study Size / Power : Confusion
Reviewer comment on a protocol:
“… there may not be a large enough sample to see the effect size required for a successful outcome. Power calculations indicate that the study is looking for a 65% reduction in incidence of … [disease]. Wouldn’t it also be of interest if there were only a 50% or 40% reduction, thus requiring smaller numbers and making the trial more feasible?”
Investigator response was very polite.
Study Size / Power : Issues
• Power will be different for each outcome.
• Power depends on the statistical method.
• Five factors including power are inter-related. Fixing four of these determines the fifth:– Study size– Power– p-value cutoff (level of significance, e.g.,
0.05)– Magnitude of treatment effect to be
detected– Heterogeneity among subjects (std dev)
Study Size / Power : Example
• “The primary outcomes for the hydrocortisone trial are changes in mean MAP and vasopressor use from the 12 hours prior to initiation of randomized treatment to the 96 hours after initiation.”
• Mean changes in placebo subjects will be compared with hydrocortisone subjects using a two sample t-test.
Project #10038: Dan Kelly & Pejman Cohan
Hypopituitarism after Moderate and Severe Head Injury
Study Size / Power : Example Cont’dThe following table presents detectable differences, with p=0.05 and 80% power, for different study sizes.
Total Number
of Subjects
Detectable Difference in
Change in Mean MAP (mm Hg)(1)
Detectable Difference in
Change in Mean Number
of Vasopressors(2)
20 10.9 0.77 40 7.4 0.49 60 6.0 0.39 80 5.2 0.34
100 4.6 0.30 120 4.2 0.27
Thus, with a total of the planned 80 subjects, we are 80% sure to detect (p<0.05) group differences if treatments actually differ by at least 5.2 mm Hg in MAP
change, or by a mean 0.34 change in number of vasopressors.
Study Size / Power : Example Cont’dPilot data: SD=8.16 for ΔMAP in 36 subjects.
For p-value<0.05, power=80%, N=40/group, the detectable Δ of 5.2 in the previous table is found as:
Study Size / Power : Summary
• Power analysis assures that effects of a specified magnitude can be detected.
• For comparing means, need (pilot) data on variability of subjects for the outcome measure. [E.g., Std dev from previous study.]
• Comparing rates (%s) does not require pilot variability data. Use if no pilot data is available.
• Helps support (superiority) studies with negative conclusions.
• To prove no effect (non-inferiority), use an equivalency study design.