Title: New Developments in Confidence Intervals That Improve Result Reporting: Confidence Levels, Clinical Significance Curves and Risk-Benefit Contours.
1New Developments in Confidence Intervals That
Improve Result Reporting Confidence Levels,
Clinical Significance Curves and Risk-Benefit
Contours.
- Dr. Thomas P. Shakespeare
- MBBS, FRANZCR, FAMS, MPH, GradDipMed(ClinEpi)
2Critical References
- For a full discussion please refer to our
original article in The Lancet - Shakespeare TP, Gebski VJ, Veness MJ, Simes J.
Improving interpretation of clinical studies by
use of confidence levels, clinical significance
curves, and risk-benefit contours. Lancet. 2001
357 134953. - Also you can download a free Confidence
Calculator for our methods www.theshakespeares.co
m/confidence_calculator.html
3Objectives
- To understand the limitations and potential
misinterpretation of p-values and 95 confidence
intervals. - To understand how new methods can improve
statistical analysis and result reporting. - To understand how to calculate confidence levels,
clinical significance curves and risk-benefit
contours. - To understand when it is appropriate to use these
new methods for analyzing and reporting study
results.
4Problems when reporting results
- 1. P values and confidence intervals are often
misinterpreted. - 2. They do not answer our basic clinical
questions - How likely is it that a clinically relevant
benefit or detriment is present? - How confident are we that a benefit is not
outweighed by unacceptable toxicity?
5The solution
- Develop tools that improve result reporting
- Confidence levels
- Clinical significance curves
- Risk-Benefit contours.
- Their advantages
- Prevent misinterpretation
- Answer our clinical questions
- Improve the decision-making process.
6An exampleof methods to report results
- WHO Melanoma Study (Cascinelli et al, Lancet
1998 351 793-96) - 252 patients with truncal melanoma ?1.5mm thick
- Randomized to immediate nodal dissection or
observation (and delayed dissection if required).
7 WHO Melanoma Study
- Results
- 5 year survival favoured immediate nodal
dissection 61.7 vs 51.3 , HR 0.72 - 95 CI 0.49-1.04, not significant p0.07
- Authorsconclusion
- Immediate nodal dissection had no impact on
survival, and should not be used. The results
have been misinterpreted!
8 What information is in the 95 CI?
- 95 CI for the hazard ratio is 0.49-1.04
- Thus we can be 95 confident that the true hazard
ratio lies within these limits, based on this
study. - How likely is it that a survival benefit exists,
or does not exist, based on this data? - Confidence intervals cant tell us, however a
confidence level can.
995 Confidence intervalWHO melanoma study. Point
estimate 0.72, SE 0.192
detrimental
beneficial
1.04
0.49
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
2.5
!
!
!
!
!
!
!
!
!
2.5
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
95
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
Relative survival benefit (hazard ratio)
10Confidence level for any benefit
- We need to determine how much confidence lies
below 1.00 (HR lt 1 indicates a survival benefit).
- From the WHO study, the point estimate for
survival was 0.72, with a standard error of 0.192
(extrapolated from the original publication).
11Confidence level for any benefit
- 1. Calculate the confidence interval around the
hazard ratio with an upper limit of 1.00 (93 CI
in this example) - 2. Calculate how much confidence lies below this
interval (half of 7 3.5) - 3. Add the two percentages (93 3.5 96.5)
- Thus there is 96.5 confidence that a
survival benefit exists. - This is very high despite the lack of
significance!
12Confidence level for any benefit(Point estimate
0.72)
Detriment
Benefit
96.5
1.00
0.52
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
3.5
!
!
!
!
!
!
!
!
!
!
!
!
!
3.5
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
93
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
Relative survival benefit (hazard ratio)
13Other magnitudes of benefit
- Confidence levels can be determined for any
benefit or detriment of interest. - What if my patient is only interested in a 3
benefit or more? - We use the same methods but set an upper value of
0.97 - We are 94 certain that dissection results in a
survival benefit of 3 or more.
14Confidence level for minimum 3 benefit(Point
estimate 0.72)
Clinically
No Relevant
relevant benefit
94
benefit/detriment
0.53
0.97
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
6
!
!
6
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
88
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
0.4
0.5
0.6
0.8
0.9
1
1.1
1.2
0.7
Relative survival benefit (hazard ratio)
15The WHO results revisited
- HR for survival 0.72, not significant (95 CI
0.49-1.04, p0.07) - However
- 96.5 confidence that a benefit exists
- 94 confidence that the benefit is 3 or more
- Thus a clinically relevant benefit is probable,
and further studies are required to confirm it.
16The WHO results revisited
- 96.5 confidence that a benefit exists
- 94 confidence that the benefit is 3 or more
- Thus a clinically relevant benefit is probable,
and further studies are required to confirm it. - This is in contradistinction to the authors
conclusion. - Confidence levels may have avoided
misinterpretation, and provided more clinically
relevant information.
17Confidence levels improve result interpretation
- Confidence levels give us the level of
confidence, likelihood or probability that a
benefit exists, and tell us whether the benefit
is clinically relevant. - They are more useful than P values and confidence
intervals. - Confidence levels have been used to analyze
meta-analyses and clinical studies.
18Clinical Significance Curves
- Individuals may accept different benefit
thresholds before using a new therapy. - We can provide confidence levels for any
threshold of benefit or detriment. - These can be combined to produce a Clinical
Significance Curve (CSC).
19CSC for survival in WHO study
- Individuals can select an acceptable benefit
threshold and determine the level of confidence
associated with it. - For example if a clinician is only interested in
a benefit of 15 or more, we can see that there
is only 81 confidence that such a benefit
exists. - CSCs provide clinically relevant information to
individual clinicians.
20Clinical significance curve for WHO study
100
X
!
!
!
X
90
significance level (97.5)
X
80
Any benefit 96.5 confidence ? 3 benefit 94
confidence ? 15 benefit 81 confidence ? 28
benefit 50 confidence
70
Confidence level ()
60
50
X
40
30
20
10
0
0
5
10
15
20
25
30
35
40
45
50
55
60
Relative survival benefit ()
21Risk-Benefit Contours
- CSCs can be constructed for any comparison eg
OS, local control, toxicity. Alternatively, the
information can be displayed in tabular form, at
say 5 increments of difference. - We can combine CSCs (eg survival with toxicity),
to form Risk-Benefit Contours (RBCs). - RBCs allow us to calculate the confidence
associated with acceptable risk-benefit scenarios.
22An example
- Intergroup study 0099 (Al-Sarraf et al, JCO
1998). - Chemoradiotherapy vs radiotherapy for Stage III
and IV NPC - Results
- 3 year survival 78 v 47 (p0.005)
- Grade 3 or 4 acute toxicity 76 v 50.
23CSC for survival
- High levels of confidence for large survival
benefits (eg up to 20 or so). - Statistically significant survival benefits.
24Clinical significance curve for Intergroup
study(Absolute 3 year survival 78 v 47)
100
X
90
significance level (97.5)
80
Any benefit 99.997 confidence Min 16 benefit
97.5 confidence Min 18 benefit 95 confidence
70
Confidence level ()
60
50
40
30
20
10
0
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
0
10
20
30
40
50
60
Absolute survival benefit ()
25CSC for toxicity
- High level of confidence that
chemoradiotherapy causes excess acute grade 3/4
toxicity.
26Clinical significance curve for Intergroup
study(Acute G3/4 toxicity 76 v 50)
100
X
90
significance level (97.5)
80
Any toxicity ? 99.95 confidence Min 10.4 ?
97.5 confidence Min 13 ? 95 confidence
70
60
Confidence level ()
50
40
30
27Risk-Benefit Contours
- Formed by combining the 2 curves.
- To use RBCs, the individual clinician first
determines an acceptable risk-benefit scenario,
then reads off the corresponding confidence
associated with it.
2860
99.95
99.99
50
99.90
99.00
97.00
95.00
85.00
40
Maximum toxicity detriment ()
90.00
80.00
70.00
60.00
50.00
30
25.00
20
0
5
10
15
20
25
30
35
Minimum survival benefit ()
29Doctor As risk-benefit scenario
- Doctor A believes the new treatment (chemoRT)
is more worthwhile than RT alone if - at least an extra 5 out of his next 100 patients
will be cured due to the new treatment - AND
- at most an extra 50 of these 100 patients will
experience G3/4 toxicity due to the new
treatment. - We can see that there is 99.9 confidence that
the above scenario exists, based on the study.
3060
99.99
99.90
99.95
Doctor As scenario
50
99.90
99.00
97.00
95.00
Maximum toxicity detriment ()
40
85.00
90.00
80.00
70.00
60.00
30
50.00
25.00
20
0
5
10
15
20
25
30
35
Minimum survival benefit ()
31Doctor Bs risk-benefit scenario
- Doctor B will only accept a scenario in which
- at least an extra 5 out of his next 100 patients
will be cured due to the new treatment - AND
- at most an extra 30 of these 100 patients will
experience G3/4 toxicity due to the new
treatment. - We can see that there is only 70 confidence that
this scenario exists. - Doctor B is not very confident. Risk-benefit
contours have aided the decision-making process.
3260
99.99
99.95
Doctor As scenario
50
99.90
99.90
97.00
95.00
40
90.00
Maximum toxicity detriment ()
85.00
80.00
70.00
60.00
50.00
30
Doctor Bs scenario
25.00
20
10
5
15
20
25
30
35
0
Minimum survival benefit ()
33Conclusion
- P values and confidence intervals are too easily
misinterpreted, and do not answer our simple
clinical questions. - By using Confidence Levels, Clinical Significance
Curves and Risk-Benefit Contours, we can - Gain a better understanding of our results
- Avoid misinterpretation
- Aid the decision-making process.
34Confidence Calculator
- To calculate confidence levels, only takes a hand
calculator and z table of normal values. - Clinical significance curves and risk-benefit
contours can be calculated with standard
statistical software. - Alternatively a dedicated Confidence calculator
can be used. - A free calculator is available from
www.theshakespeares.com/confidence_calculator.html
35References
- Please see attached notes page for a full list
of references used in this lecture.