Title: Reverend Bayes Sample Sizes and Statistical Analysis
1Reverend BayesSample Sizesand Statistical
Analysis
2Statisticians
- Statisticians are generally power mad, that is
they want to minimise uncertainty around any
effect estimate. - Two camps Frequentists and Bayesians.
3Frequentists
- The philosophy underlying most of our statistics
is frequentist. - Frequentists produce the null hypothesis. The
experiment is set up to prove the null
hypothesis of no difference. - See themselves as more objective and
scientific than Bayesians.
4Frequentist Null hypothesis
- This is nonsense!
- If we truly believed in the null hypothesis we
would not undertake a trial. We would just chose
the cheapest treatment or give treatments
according to patient preferences. - We usually have an idea about the likely effect
of a treatment.
5Reverend Bayes
- A minister who lived in the 18th Century but
dabbled in statistics. - Produced Bayes theorem, which includes prior
beliefs in statistical calculations. - Was not published till 20 years after his death
- truly publish or perish.
6God and statistics
- Frequentists believe the truth is out there and
we are getting sample estimates of the truth. - Bayes believed only God can know the truth and
as mere mortals we can only gain probability
estimates of the truth, which is why he developed
Bayes theorum.
7 Bayesians
- Until recently Bayes approach only used in
diagnostic testing in health research. - Widely used in other areas.
- Not widely used partly because of computational
difficulties but also many think it is
unscientific. - More recently computational problems have been
largely solved and increased interest in using
the method.
8Bayesian statistics
- The Bayesian approach is attractive as it is
similar to everyday decision making. - One uses prior experience to make a judgement and
use new data to inform future decisions.
9Bayesians vs Frequentists
- When we seek to observe a 50 increase or
decrease in essence this is a Bayesian approach
as we have a prior belief that A may be 50 more
effective than B. - If we had a belief in the null hypothesis then
the sample size would be infinite to prove no
difference.
10Prior beliefs
- Bayesians want to be more explicit about prior
beliefs and include these in a design and
analysis. - Data would have to be particularly strong to
overturn a prior belief or weaker to confirm.
11Bayesian Problems
- Bayesians argue that one should keep doing a
study until the confidence in the results are
credible enough to stop the trial. - Problem that one cannot really plan a trial
unless we have a prior sample size.
12Not Scientific
- Prior beliefs may be so incorrect that they could
mislead research. Strong prior belief was HRT
prevented heart disease. Shown to be untrue.
Small trials showing this to be a fallacy would
not overturn this strong belief.
13GRIT Trial Bayesian trial
- The design of this trial included prior beliefs
on the effectiveness of early or late delivery of
babies. - Data were analysed every 6 months (without p
values) and presented to clinicians in order for
them to change their minds and either randomise
more patients or stop randomising.
14Bayesian analysis
- Expect to see more studies using Bayesian methods
in the future. - Rapid area of statistical and economic research.
15Statistical Outcomes
- Two measures of effect.
- Dichotomous yes/no dead/alive passed/failed.
- Continuous blood pressure weight exam scores.
16Binary outcomes
- Basically in a RCT we can compare the percentages
in the two groups. - If the percentages are significantly different
this is due to the intervention.
17Continuous outcomes
- Scores, such as blood pressure, quality of life,
test scores are compared. Usually the mean
scores are compared although sometimes the
medians are used. - Usually, mean scores have a normal or near
normal distribution.
18(No Transcript)
19Standard deviation
- This is calculated by taking the differences of
individual scores from the mean squaring these
differences and dividing by the number of
observations. - The square root is the SD of this.
20Effect sizes
- The effect size is the difference between means
divided by the standard deviation. - If students in Group A have a mean score of 60 vs
50 in Group B and the standard deviation is 20
the effect size is 0.5 (10/20). - Few new health care treatments get effect sizes
GREATER than 0.5.
21Relative Risks etc
- Binary outcomes are often described in relative
risk or odds ratios. Relative risk is if
10/100 events in group A versus 5/100 in group B.
A vs B RR 2 (10/5) B vs A RR 0.5 (5/10). - Odds ratios produce similar results for rare
events. - Confidence intervals passing through 1 not
statistically significant.
22Sample sizes for trials
- The bigger the better size matters in trials.
- Most trials approach sample size estimation using
a frequentist approach.
23Background
- Many trials are underpowered that is they are
too small to detect a difference that is
important. - This is commonly referred to as a Type II error.
- At least 30 of trials published in major general
journals are underpowered. - This is worse among other journals.
24Meta-analysis of Hip Protectors (Ranked by Size)
Energy absorbing or unknow types
Community
Community
Nursing Home
Shell type Protectors
Community
Community
25Hip protector trials
- All trials (bar ours of course) were underpowered
to detect large (e.g., 50) reductions in hip
fractures. - Small positive trials tended to be published
giving an overestimated effect of benefit.
26Sample size estimation
- Text books usually recommend the following
approach to sample size estimation. - Define a clinically important difference in
outcome between treatments - Design an experiment that is sufficiently large
to show that such a difference is statistically
significant.
27Clinical Significance
- The first problem is definition of what is
clinically significant. This is usually
unclear. - Any difference of death, for example, is pretty
clinically significant. - To power a trial to reduce mortality by 1 death
would require an almost infinitely large study.
28Epidemiological Significance
- A more common justification of sample size is
observed effect sizes from epidemiological
studies (which may be overestimates). - Or from meta-analyses of smaller trials (which
again may over-estimate due to publication bias).
29Statistical significance
- What is statistical significance? Tradition in
medical research states that p 0.05 or lower is
significant. Difference between p 0.05 and p
0.06 is trivial, one is significant and the other
is not. - Other disciplines, economics, sometimes use p
0.10.
30P values
- Originally Pearson constructed p values as a
guide not as a cut off. The idea was that given
what was known about a treatment (side-effects
etc) the p value would add extra information as
to whether one should accept the finding. - But p value 0.05 has become set in stone.
31Fallacy of P values
- If there is a treatment effect that is not
statistically significant p 0.20 and the null
hypothesis is accepted (I.e. there is no
difference) you would have only a 20 chance of
being correct and 80 of making the wrong
decision. - Really one should go for a treatment that the
data favours irrespective of the p value.
32Significance
- BOTH clinical and statistical significance are
often arbitary constructs. - Economic significance can be less arbitrary.
- One can ascertain an economic difference that
makes sense. - To demonstrate cost neutrality is a significant
endpoint.
33Economic Significance
- For example, a randomised trial of two methods of
endometrial resection was powered to detect a 15
difference in satisfaction. - Important clinical outcome was re-treatment
rates. - An economic difference of significance was about
8 in retreatment rates as this would be cost
saving.
Torgerson Campbell BMJ 2000697.
34Endometrial Resection
- The trial was only sufficiently powerful to show
a 12 difference in retreatment rates. - Trial showed a 4 difference (95 CI of 4 to
11) but could not exclude an 8 difference.
Pinion et al. BMJ 1994309979-83.
35Forget theory
- What normally happens is Clinician says to
statistician I can get 70 patients in a trial in
a year. - Stato says needs to be bigger clinician has a
couple of mates who can add 140 more.
Statistician calculates difference that 210
participants can show.
36What should be done?
- For a continuous outcome (e.g. Quality of Life,
blood pressure) we should aim to detect AT LEAST
half a standardize effect size, which needs 128
participants. - Ideally we need to detect a somewhat smaller
difference. - For dichotomous outcome we should have enough
power to detect a halving or doubling.
37Attrition and clustering
- Do not forget to boost sample size to take into
account loss to follow-up. - Depending on patient group this might range from
5-30. - Finally, if it is a cluster trial total sample
size needs to be inflated.
38How to calculate a sample size
- This is easy. Lots of tables or programmes will
do this. For continuous outcomes a simple
formulae is take standardised difference and
divide the square of this into 32 (80 power) or
42 (90 power). - E.g., 0.5 squared is 0.25 32/0.25 128 or
42/0.25 168.
39For binary outcomes
- Look at sample size tables or use programme, but
rule of thumb about 800 is needed for 80 power
to show 10 difference between 40 and 50 or 50
and 60. To see 5 difference quadruple sample
size.
40Cluster trials
- For cluster trials we need to inflate the sample
size to take into account the ICC of the
clusters. 1(cluster size X ICC) design
effect. - For example, a RCT of adult literacy classes mean
size 8. ICC from a previous trial shows ICC of
reading 0.3.
41Cluster sample size
- We want to detect 0.5 difference which for an
individual RCT 128 for 80 power. Cluster size
8 take 1 7. - 7 x 0.3 2.1 1 3.1 397 participants or t
50 clusters of a mean of 8 per cluster.
42Analysis
- The first analysis that many people do is compare
groups at baseline. - Typical many comparisons are made, for example, a
paper of a trial in the most recent JAMA (Feb 4,
2004) shows this typical baseline comparison
table.
43Baseline Tests (n 24 tests)
44Baseline testing
- Of the 24 comparisons 3 were statistically
significant (I.e, p lt 0.05). - What should we do with this information?
- Has randomisation failed?
- It is useless information and an exercise in
futility.
45Baseline testing
- Assuming randomisation has not been subverted,
which in this case looks unlikely, then any
differences will have occurred by chance they
are random differences.
46What is wrong with baseline testing?
- Baseline testing will ALWAYS throw up chance
differences. This can mislead the credulous into
believing there is something wrong with the
study. Also it can mislead some statisticians
into correcting these baseline imbalances in
the analysis.
47Baseline variables What should be done?
- Before the study starts specify in advance
important co-variates to be used in the analysis
(e.g., centre, age) and adjust for these
IRRESPECTIVE of whether or not randomisation
balances them out.
48Interim Data Analysis
- This is where the trial is analysed BEFORE
completion. - This is done usually for ethical reasons so that
a trial can be stopped early if there is an
overwhelming benefit or harm. - Womens Health Initiative trial undertook an
interim analysis and the trial was stopped
because of harm.
49Dangers of Interim Analysis
- Sample size calculations assume 1 analysis.
Repeated looks at the data WILL showed a
significant differences, by chance, even when no
difference exists. - The temptation is to stop the trial early when a
statistical significance is achieved. - This could be a chance finding.
50Interim Analysis
- To avoid premature stopping of a trial interim
analyses are usually undertaken by an independent
committee with experience trialists. - Statistical significance is adjusted to take
repeated looks of data into account (so p 0.01
is significant rather than p 0.05).
51Analysis
- All point estimates should be bounded by
confidence intervals as well as the exact p
value. A single principal analysis should be
stated in advance (e.g., the primary outcome was
a reduction in ALL fractures) secondary analysis
are for research interest only.
52Summary
- Sample size estimation is EASY. The difficult
bit is determining the likely effect size to
inform the calculations. - Analyses are more straightforward from RCTs than
non-RCTs because you do not need to adjust for
baseline co-variates.