Session 4 - Analysis and reporting

Report
Session 4: Analysis and reporting
Managing missing data
Rob Coe (CEM, Durham)
Developing a statistical analysis plan
Hannah Buckley (York Trials Unit)
Panel on EEF reporting and data archiving
Jonathan Sharples, Camilla Nevill, Steve Higgins
and Andrew Bibby
Managing missing data
Rob Coe
EEF Evaluators Conference, York, 2 June 2014
The problem
 Only if everyone responds to everything is it still
a randomised trial
– Any non-response (post-randomisation) → not an RCT
 It may not matter (much)
∂ if
– Response propensity is unrelated to outcome
– Non-response is low
 Lack of ‘middle ground’ solutions
– Mostly people either ignore or use very complex stats
3
What problem are we trying to solve?
 We want to estimate the distribution of likely
effects of [an intervention] in [a population]
– Typically represented by an effect size and CI
∂
 Missing data may introduce
bias and
uncertainty
– Point estimate effect size different from observed
– Probability distribution for ES (CI) widens
4
What kinds of analysis are
feasible to reduce the risk of
bias from missing data?
5
Vocabulary
Missing Completely at Random
(MCAR)
– Response propensity is unrelated
to outcome
Missing at Random (MAR)∂
– Missing responses can be
perfectly predicted from observed
data
Ignore
missingness
Statistics:
IWP, MI
Missing Not at Random (MNAR)
– We can’t be sure that either of the
above apply
6
??
“When data are missing not at random, no
method of obtaining unbiased estimates exists
that does not incorporate the mechanism of
non-random missingness,
∂ which is nearly
always unknown. Some evidence, however,
shows that the use of a method that is valid
under missing at random can provide some
reduction in bias.”
Bell et al, BMJ 2013
7
Recommendations
1. Plan for dealing with missing data should be in
protocol before trial starts
2. Where attrition likely, use randomly allocated
differential effort to get outcomes
3. Report should clearly state the proportion of
outcomes lost to follow∂ up in each arm
4. Report should explore (with evidence) the
reasons for missing data
5. Conduct simple sensitivity analyses for strength
of relationship between
Outcome score and missingness
Treatment/Outcome interaction and missingness
8
If attrition is not low (>5%?)
6. Model outcome response propensity from
observed variables
7. Conduct MAR analyses
∂
• Inverse weighted probabilities
• Multiple imputation
8. Explicitly evaluate plausibility of MAR
assumptions (with evidence)
9
∂
10
∂
11
Useful references





Bell, M. L., Kenward, M. G., Fairclough, D. L., & Horton, N. J.
(2013). Differential dropout and bias in randomised controlled
trials: when it matters and when it may not. BMJ: British Medical
Journal, 346:e8668. http://www.bmj.com/content/346/bmj.e8668
Graham, J. W. (2009). Missing data analysis: Making it work in
the real world. Annual review of psychology, 60, 549-576.
National Research Council. The Prevention and Treatment of
Missing Data in Clinical Trials.
∂ Washington, DC: The National
Academies Press, 2010.
http://www.nap.edu/catalog.php?record_id=12955
Shadish, W. R., Hu, X., Glaser, R. R., Kownacki, R., & Wong, S.
(1998). A method for exploring the effects of attrition in
randomized experiments with dichotomous outcomes.
Psychological Methods, 3(1), 3.
www.missingdata.org.uk
12
Developing a statistical analysis
plan (SAP)
Hannah Buckley
York Trials Unit
[email protected]
June 2014
Overview
• What is a SAP?
• When is a SAP developed?
• Why is a SAP needed?
• What should be included in a SAP?
What is a SAP?
• Pre-specifies analyses
• Expands on the analysis section of a
protocol
• Provides technical information
When is a SAP developed?
• After protocol finalised
• Before final data received
• Written in the future tense
Why create a SAP
• Pre-specify analyses
• Think through potential pitfalls
• Benefit to other analysts
What should be in a SAP?
ACTIVITY
• What do you think should be covered in a
SAP?
• Sort the cards into two piles
What should be in a SAP?
ACTIVITY DISCUSSION
• Which topics do you think do not need to
be covered in a SAP?
• Are there any topics which you were
unsure about?
What should be in a SAP?
ACTIVITY
1. Which of the cards cover key background
information and which are related to
analysis?
2. Which order would you deal with the
topics in?
The structure of a SAP
Setting the scene
• Restate study objectives
• Study design
• Sample size
• Randomisation methods
The structure of a SAP
Description of outcomes
• Primary outcome
• Secondary outcome(s)
• When outcomes will be measured
• Why outcomes chosen
The structure of a SAP
Analysis - overview
• Analysis set (ITT)
• Software package
• Significance levels
• Blankets statements on confidence
intervals, effect sizes or similar
• Methods for handling missing data
The structure of a SAP
Analysis methods
• Baseline data
• Primary analysis
• Secondary analyses
• Subgroup analyses
• Sensitivity analyses
Conclusions
• Producing a SAP is good practice
• Can help avoid problems in analysis
• Finalised before final data received
• Fairly detailed
• Flexible but should cover key points
References and resources
References
• ICH E9 ‘Statistical principles for clinical trials’
http://www.ich.org/products/guidelines/efficacy/article/efficacyguidelines.html
Resources
• PSI ‘Guidelines for standard operating procedures for good
statistical practice in clinical research’
www.psiweb.org/docs/gsop.pdf
Thank you!
Any questions or discussion points?
EEF reporting and data archiving
Jonathan Sharples (EEF)
Camilla Nevill (EEF)
Steve Higgins (Durham) - Chair
Andrew Bibby (FFT)
The reporting process and publication of
results on EEF’s website
Jonathan Sharples (EEF)
Classifying the security of findings from EEF
evaluations
Camilla Nevill (EEF)
Group
Number
of pupils
Effect size
Estimated months’
progress
Literacy intervention
550
0.10 (0.03, 0.18)
+2
www.educationendowmentfoundation.org.uk/evaluation
Evidence strength
Example Appendix: Chatterbooks
Rating
1. Design
2. Power
(MDES)
3. Attrition
4. Balance
5. Threats
to validity
5
Fair and clear experimental
design (RCT)
< 0.2
< 10%
Well-balanced on
observables
No threats to
validity
4
Fair and clear experimental
design (RCT, RDD)
< 0.3
< 20%
3
Well-matched comparison
(quasi-experiment)
< 0.4
< 30%
2
Matched comparison
(quasi-experiment)
< 0.5
< 40%
1
Comparison group with
poor or no matching
< 0.6
< 50%
0
No comparator
> 0.6
> 50%
Some
threats
Imbalanced on
observables
Significant
threats
Combining the results of evaluations with the
meta-analysis in the Teaching and Learning
Toolkit
Steve Higgins (Durham)
Archiving EEF project data
Andrew Bibby
Prior to archiving…
1. Include permission for linking and archiving in consent forms
2. Retain pupil identifiers
3. Label values and variables
4. Save Syntax or Do files

similar documents