Statistical Aspects of Grant

Proposals
Dr Laura Gray
With thanks to Sarah Lewis and Victoria Owen for some
slides and ideas!

Me Laura Gray
0.2 RDS (Weds), 0.8 Diabetes Research
Previously
1 year statistician Cancer Research UK, Birmingham
5 years statistician Stroke Medicine, Nottingham

PhD
Improving the design and analysis of stroke trials

Statistical aspects of grants

Study design
Sample size
Sample
Outcome variables
Proposed statistical analysis

Study design
Meta A

More subject
to bias

RCT

Quasi Experimental

Observational Studies
Cohort>Case-control>Cross-sectional

Expert opinion / Case studies

Study design
The design fits the question
Methodologically sound
Feasible
Practicalities should be discussed

Observational or Experimental?
In an observational study the researcher
has no control over who is exposed to what
e.g. the researcher cannot decide who smokes
cigarettes and who does not, they can only observe

In an experiment, the researcher is in control

of who is exposed to what participants are
allocated to treatment

Prospective or Retrospective?
Prospective
data collected about subsequent events
temporal associations clearer, but more expensive

Retrospective
data collected about events in the past
cheaper, but may rely on memory

Longitudinal or Cross-sectional?
Longitudinal

data are collected at more than one time-point

powerful, but time-consuming

Cross-sectional

data are collected at one time-point

cheaper, but can only infer association (not
causation)

Case control study

Things to consider CC study

How to select cases

How to select a comparable control group

Selection bias
Matching
Analysis needs to take matching into account

Recall bias
Assessment bias
Confounding
Age

Confounding

Cohort Study

Things to consider Cohort

How to get comparable groups apart from exposure of
interest
Length of follow up
Losses
How to deal with changes in situation
Not suitable for rare events

RCT
2 arm parallel trial
Test

Population

Measure
outcome

Sample
Control

Randomise

Things to consider - RCT

Control group standard care / placebo

Blinding
Randomisation
Patient level
Clusters if contamination
Length of follow up

Inclusion and exclusion criteria

Sample size
Sample size is a fundamental part of any research
design

Too small may fail to detect clinically important

differences
Too big more patients, more costly

However, too many pts is better than too few

Allow for attrition (lost data)
Should take into account response rates
Sample size calculations may not be needed for pilot
studies

Copyright 2009 BMJ Publishing Group Ltd.

Dennis, C-L et al. BMJ 2009;338:a3064

Problems with sample size

No sample size calculation!
A previous study used 150 patients
A sample size calculation is not provided as no previous information to
base one on
The throughput of the clinic is 50 patients a year, its a two year trial,
10% may refuse to take part therefore 90 patients to be included

No justification given for numbers used in calculation

Calculation cannot be replicated with information given
No justification for one-sided calculations
Sample size doesnt relate to the primary outcome

Sample

Sample
Where do they come from?

Have they explained why these are an appropriate

group?
How will subjects be selected?
Inclusion/exclusion criteria

Is your sample size achievable using this population?

Outcome / Explanatory variables

What are you measuring?
How?
When?
What type of data?

Is this a valid measurement?

Is it measuring what you think it is?

Is this a reliable measurement?

If you repeated the measure on average would you get the same
results?
Can you assume that taking measurements over the telephone are
comparable to those collected by post?

Is inter-rater reliability an issue?

How many variables of interest?

Proposed statistical analysis

Adequate detail?
Data will be analysed using SPSS
Data will be analysed using multivariate methods

Does the method of analysis match the design?

Cluster randomised trials should take into account the effect of
the cluster

Does the method of analysis match the type of data

being collected?
Ordinal data being analysed as continuous data

Multiple testing
Repeated measures data where a statistical test is performed at
each time point

Use Statisticians!!
From the RfPB guidance notes
Support and advice from trials methodologists is
crucial
Statistical support is another essential. A last minute
phone call to a statistician for a power calculation is
unlikely to be enough
.how far the team has thought through exactly
how they will go about data analysis. Detail on all of
these is crucial.
Research Design Service, are a first port of call