The analysis of contingency tables
(measures of effect and agreement
July 2019
Fatemeh soleimany { soleimany2013@ut.ac.ir}
Amirhossein Aliakbar {amirhossein.aliakbar@gmail.com}
‫توافقی‬ ‫جداول‬ ‫تحلیل‬3(‫توافق‬ ‫و‬ ‫تاثیر‬ ‫معیارهای‬)
‫آذر‬ ‫ایران‬ ‫متابولیسم‬ ‫و‬ ‫دیابت‬ ‫مجله‬-‫دی‬1393‫دوره‬14‫صفحه‬ ‫دوم‬ ‫جلد‬75‫تا‬92
Topics stored from:
3
 Mohammad asghari jafar abadi Associate Professor of Biostatistics, Tabriz
university of medical science
 Seyede Momeneh Mohammadi , Phd candidate of anatomical scienses at Medical
faculty/Tabriz University of Medical Science
 Akbar Soltani, Associate Professor of Endocrinology, Tehran University of Medical Sciences
Objectives ( first part)
 To develop approaches for disease prevention
 Attributable risk/fraction
4
Case-control
Attribute Fraction
(AF)
Attribute fraction for
exposed group( AFE)
Attribute fraction for
population( AFP)
Attribute Risk (AR)
Attribute risk
reduction
(ARR)
Relative Risk
Reduction
(RRR)
Number
needed to treat
(NNT)
Objectives (second part)
5
reliability
Quantitative data
ICC
Bland’s and
Altman
Lin’s coefficient
corellation
Qualitative data
Kappa and
Kohen’s Kappa
Fleiss Kappa
Impact criteria
 AR: Attributable Risk
 AF: Attributable Fraction
 AFE: Attributable Fraction for Exposed Group
 AFP: Attributable Fraction for Population
AR/AF To develop approaches for disease prevention
Overall risk
Risk among non-
exposed
AR =
0
20
40
60
80
100
+ -
Excess
Risk
AR = [ (a+c) / n ] – [ c / ( c+d) ]
a b
c d
AF (attribute fraction )
AF= AR ÷ total risk
AF = AR ÷ 𝑎 + 𝑐 ÷ 𝑛 = 𝑎 + 𝑐 ÷ 𝑛 − 𝑐 ÷ 𝑐 + 𝑑 ÷ [ 𝑎 + 𝑐 ÷ 𝑛 ]
𝒑𝒆 = [ 𝒂 + 𝒃 ÷ 𝒏] 𝑨𝑭 = 𝒑𝒆 𝑹𝑹 − 𝟏 ÷ [ 𝟏 + 𝒑𝒆 𝑹𝑹 − 𝟏 ]
𝑷𝒄 = [ 𝒂 ÷ 𝒂 + 𝒄 ] 𝑨𝑭 = 𝒑𝒄 𝑹𝑹 − 𝟏 ÷ [ 𝑹𝑹 − 𝟏 ]
a b
c d
Relative risk (RR) =
𝑎
𝑎+𝑏
𝑐
𝑐+𝑑
Odds-ratio (OR) =
𝑎𝑑
𝑏𝑐
yes no
yes a b
no c d
event
exposure
Death rate Total number
Number of
living people
Number of
death
0.342 284 187 97 Have used
0.27 832 607 225 Non user
0.289 1116 794 322 total
Frequency of mortality in colorectal cancer patients in terms of the variables of smoking status
 AR = %28.9 – %27.0 = %1.9 ( %1.9 of the risk of mortality from colon cancer can be attributed to smoking )
 𝐀𝐅 =
𝐴𝑅
𝑜𝑣𝑒𝑟𝑎𝑙𝑙 𝑟𝑖𝑠𝑘
=
1.9
28.9
= 0.063 or 6.3% (In comparison with the whole community, almost 6% of death from
colon cancer can be attributed to cigarette smoking.
Overall risk
Probability of
risk in non-
exposed
AFE =
𝑅𝐷
𝑡ℎ𝑒 𝑟𝑖𝑠𝑘 𝑖𝑛 𝑒𝑥𝑝𝑜𝑠𝑒𝑑 𝑔𝑟𝑜𝑢𝑝
=
𝑅𝑅−1
𝑅𝑅
=
( 0.342 −0.27 )
0.342
= 0.208 (By eliminating smoking in the exposure group, 21% of
deaths from smoking can be prevented. )
𝑹𝑹 =
97
284
225
832
= 1.263
Confidence Interval for AFE: ( 0.04 , 0.35 ) ( In comparison with smokers, Minimum and maximum
fraction of mortality attributed to cigarette smoking of colon cancer patients is in the range )
Death rate Total number
Number of
living people
Number of
death
0.342 284 187 97 Have used
0.27 832 607 225 Non user
0.289 1116 794 322 total
Rate of
smokers
Total number
smoking
Number of non-
smoking ones
Number of
smokers
0.671003367Exposed
0.202559446113Non exposed
0.273659479180Total
0.1520.0690.372The rate of
AR = [(a+c) ÷n]-[c÷(c+d)] = [(67+113) ÷ 659]-[113 ÷(113+446)] = (27.3-20.2) = 7.1 (7.1% of the risk of
being smoker can be attributed to peers’insistence )
AF = AR ÷ total risk = (7.1 ÷ 27.3) = 0.26 (In comparison with whole population, 26% of smoking can be
attributed to peer’s insistence)
OR = (a × d) ÷ (b × c) = (67 × 446) ÷ (33×113) = 8.01
AFE = [OR-1] ÷[OR] = (8.01-1) ÷8.01 = 0.88 (By eliminating the effect of peers’insistence, 8.8 percent of
smoking may be prevented in exposed group)
Confidence Interval for AFE: ( 0.80 , 0.92 ) {Therefore, in the statistical
population with 95% confidence, For those with a peer's insistence,
minimum and maximum fraction attributed to peer's insistence is in the
domain above.}
AFP = AFE × percent exposed among cases = 0.875 × (67 ÷180) = 0.326
{By eliminating the effect of peers’insistence, 8.8 percent of smoking may be
prevented in whole population)
ARR and RRR in randomized clinical trial
 ARR: Absolute Risk Reduction
 RRR: Relative Risk Reduction
 NNT: Number Needed to Treat
 NNH: number needed to harm
 ARR= NNT if ………
 ARR =
𝑐
𝑐+𝑑
−
𝑎
𝑎+𝑏
=
21
21+9
+
11
11+16
= 0.70 − 0.41 = 0.29 (the contribution of omega-3 in treatment of neuropathy
is only 29%.)
NNT =
1
𝐴𝑅𝑅
=
1
0.29
=3.4 ≈3 (it's expected that using this supplement in about 3 patients, will improve one of them. )
RRR ==
(
𝑐
𝑐+𝑑
)−(
𝑎
𝑎+𝑏
)
(
𝑐
𝑐+𝑑
)
=
21
21+9
−(
11
11+16
)
(
11
11+16
)
=
0.29
0.41
= 0.71 (The relative reduction in neuropathy induced by intervention was 71%
compared with the placebo group.)
The risk of
neuropathy
Chance of
recovery
Total
number
neuropathy
Non normalnormal
0.300.7030921Omega 3
0.590.41271611Placebo
 NNH= Number needed to harm
Experimental
group ( E)
Control group
(C)
total
Events (E) EE = 75 CE = 100 115
Non-events
(N)
EN = 75 CN = 150 285
Total subjects
(S)
ES = EE + EN
= 150
CS = CE+ CN
= 250
400
Event rate
(ER)
EER = EE/ ES
= 0.5 or 50%
CER = CE/ CS
= 0.4 or 40%
equation variable Abbr. Value
EER-CER Absolute risk
increase
ARI 0.1 or 10%
(EER –CER)
/ CER
Relative risk
increase
RRI 0.25 or 25%
1 / (EER –
CER)
Number
needed to
harm
NNH 10
EER/CER Risk ratio RR 1.25
(EE / EN) /
(CE/CN)
Odds ratio OR 1.5
(EER –CER)
/ EER
Attributable
fraction
among
exposed
AFE 0.2
Agreement measures
Golden standard
ICC and SEM
Bland and Altman’s limit of agreement
Lin’s concordance correlation coefficient
Cohen's kappa and Weighted Cohen’s Kappa
ROC
𝐿𝑅+
and 𝐿𝑅−
 PPV: Positive predictive value
 NPV: Negative predictive value
Diagnosed sick Diagnosed healthy
Sick
True positive False negative
healthy
False positive True negative
Prior and Posterior probability
 If LR > 1 then Posterior probability > Prior probability disease
 If LR < 1 then Posterior probability < Prior probability
Sensitivity =
𝑇𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
𝑇𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒
=
23
23+13
= 0.64 → 64%
Specificity =
𝑇𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒
𝑇𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
=
51
51+13
= 0.80 → 80%
PPV =
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑓𝑎𝑠𝑙𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
=
23
23+13
= 0.64 → 64%
NPV =
𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒+𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒
=
53
53+13
= 0.80 → 80%
GDM
Total number Health Sick
36 13 23 +
64 51 13 -
100 64 36 Total number
 𝑳𝑹+=
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦
1−𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑡𝑦
=
𝑜𝑑𝑑𝑠 𝑜𝑓 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
𝑜𝑑𝑑𝑠 𝑜𝑓 𝑓𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
=
23
36
13
64
= 3.15 (The chance to see a real patient to a false
patient in a positive test result is approximately 3 times of the average effect size)
 𝑳𝑹−
=
1−𝑆𝑒𝑛𝑠𝑖𝑣𝑖𝑡𝑦
𝑆𝑝𝑒𝑐𝑖𝑣𝑖𝑡𝑦
=
𝑜𝑑𝑑𝑠 𝑜𝑓 𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒
𝑜𝑑𝑑𝑠 𝑜𝑓 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒
=
13
36
56
64
= 0.45 (The chance to see a healthy, false patient to
actual normal one, in the negative test results is almost half of the effect size.)
GDM
Total number Health Sick
36 13 23 +
64 51 13 -
100 64 36 Total number
Index Value
Lower
bound
Upper
bound
prevalence 36% 27% 46.2%
Sensitivity 63.9% 46.2% 79.2%
Specificity 79.7% 67.8% 88.7%
PPV 63.9% 46.2% 79.2%
NPV 79.7% 67.8% 88.7%
𝐿𝑅+ 3.15% 1.83 5.42
𝐿𝑅− 0.45% 0.29 0.71
Note in practice: it’s required to calculate
confidence interval for specificity, Sensivity,
𝐿𝑅+ , 𝐿𝑅− , NPV and PPV.
Prior probability of disease= 36 ÷ 100 = 0.36 ( equal for everyone and means prevalence)
prior chance of being actual patient = Pre-test-odds (+)= Pre-test- probability / (1 – Pre-test- probability )
= 0.36÷ 1 − 0.36 = 0.56
(posterior chance of being actual patient) Post test Odds (+)=Pretest odds × LR+ = 0.56 × 3.15 = 1.77
(posterior probability of being actual patient) Post Probability (+)=(Pre test Odds × LR+) / (1 + (Pre test Odds ×
LR+
)) = 1.77÷(1+1.77) = 0.64
(posterior chance of being false patient) Post test odds (-)= Pretest odds × LR−
= 0.56 ×0.45 = 0.26
(posterior probability of being false patient) Post Probability (-)=(Pre test Odds × LR−
) / (1 + (Pre test Odds ×
LR−
)) = 0.26÷(1+0.26) = 0.20
ICC (Intra class correlation coefficient (ICC) for one way and two
way models)
Missing data are omitted in a list wise way.
When considering which form of ICC is appropriate for an actual set of data, one has
take several decisions (Shrout & Fleiss, 1979):
 1. Should only the subjects be considered as random effects (’"oneway"’model) or are subjects and
raters randomly chosen from a bigger pool of persons (’"twoway"’ model).
 2. If differences in judges’ mean ratings are of interest, interrater ’"agreement"’ instead of
’"consistency"’ should be computed.
 3. If the unit of analysis is a mean of several ratings, unit should be changed to ’"average"’. In most
cases, however, single values (unit=’"single"’) are regarded.
Standard Error of Measurement(SEM)
SEM Reliability
SEM = SD× √(1-ICC)
How can I trust
you?
Quantitative measures
The concordance correlation coefficient combines measures of both precision and accuracy to
determine how far the observed data deviate from the line of perfect concordance (that is, the
line at 45 degrees on a square scatter plot). Lin's coefficient increases in value as a function of
the nearness of the data's reduced major axis to the line of perfect concordance (the accuracy of
the data) and of the tightness of the data about its reduced major axis (the precision of the data)
Lin's concordance correlation coefficient
<0.9 poor
0.9 to 0.95 moderate
0.95 to 0.99 substantial
> 0.99
Almost
perfect
Lin’s coefficient
Bland and Altman's limit of agreement (Tukey mean-difference
plot)
 They advocated the use of a graphical method to plot the difference scores of two
or more than two measurements against the mean for each subject and argued that
if the new method agrees sufficiently well with the old, the old may be replaced.
 BlandAltmanLeh and Blandr
ICC (Intraclass Correlation Coefficient)
 is a descriptive statistics that can be used when quantitative measurements are
made on units that are organized into groups. It describes how strongly units in the
same group resemble each other.
90 /
100 9 /
10
Cohen’s Kappa and weighted Kappa for two raters
Inter observer variation can be measured in any situation in which two or more independent observers
are evaluating the same thing.
Sometimes, we are more interested in the agreement across major categories in which there is
meaningful difference.
Strongly
disagree
disagreeNo IdeaagreeStrongly agree
Strongly agree
agree
No Idea
disagree
Strongly
disagree
Colour lost means weight lost
Commonly used weights
Weighted-Kappa requires specific weight, for instance:
 1-
𝒊 −𝒋
𝒌−𝟏
 𝟏 −
𝒊−𝒋
𝒌−𝟏
𝟐
 arbitrary weights
{ I for rows – J for columns
K for maximum number of instruments}
The nearer
apples,
the bigger
weights
totalNo
agreement
Somewhat
agreement
Perfect
agreement
( two
instrument)
204106Perfect
appointment
356209Somewhat
appointment
130504040No appointment
185607055total
Weighted kappa based on first weight
=0.094
And based on second weight = 0.07
Weak
agreement
The power of different Kappa coefficients in determining the amount of agreement
among observers or referees
Agreement reliabilityKappa statistic
amount
WeakLess than 0
low0 – 0.2
Less than average0.21 – 0.4
average0.41 – 0.6
good0.6 – 0.8
excellent0.81 - 1
Is there any
confidence interval for
kappa? Is it valuable?
references
 Understanding Interobserver Agreement:The Kappa Statistic. Anthony J. Viera, MD; Joanne M. Garrett, PhD. Family
Medicine.
 Bland J, Altman D (1986). Statistical methods for assessing agreement between two methods of clinical measurement. The
Lancet 327: 307 - 310.
 Bland J, Altman D (1999). Measuring agreement in method comparison studies. Statistical Methods in Medical Research 8:
135 - 160.
 P. S. Myles, J. Cui, I. Using the Bland–Altman method to measure agreement with repeated measures, BJA: British Journal
ofAnaesthesia, Volume 99, Issue 3, September 2007, Pages 309–311,
 Bland J, Altman D (2007). Agreement between methods of measurement with multiple observations per individual. Journal
of Biopharmaceutical Statistics 17: 571 - 582. (Corrects the formula quoted in the 1999 paper).
 Bradley E, Blackwood L (1989). Comparing paired data: a simultaneous test for means and variances. American
Statistician 43: 234 - 235.
 Burdick RK, Graybill FA (1992). Confidence Intervals on Variance Components. New York: Dekker.
 Dunn G (2004). Statistical Evaluation of Measurement Errors: Design and Analysis of Reliability Studies. London: Arnold.
Thank you

Statistical Journal club

  • 1.
    The analysis ofcontingency tables (measures of effect and agreement July 2019 Fatemeh soleimany { soleimany2013@ut.ac.ir} Amirhossein Aliakbar {amirhossein.aliakbar@gmail.com}
  • 2.
    ‫توافقی‬ ‫جداول‬ ‫تحلیل‬3(‫توافق‬‫و‬ ‫تاثیر‬ ‫معیارهای‬) ‫آذر‬ ‫ایران‬ ‫متابولیسم‬ ‫و‬ ‫دیابت‬ ‫مجله‬-‫دی‬1393‫دوره‬14‫صفحه‬ ‫دوم‬ ‫جلد‬75‫تا‬92 Topics stored from:
  • 3.
    3  Mohammad asgharijafar abadi Associate Professor of Biostatistics, Tabriz university of medical science  Seyede Momeneh Mohammadi , Phd candidate of anatomical scienses at Medical faculty/Tabriz University of Medical Science  Akbar Soltani, Associate Professor of Endocrinology, Tehran University of Medical Sciences
  • 4.
    Objectives ( firstpart)  To develop approaches for disease prevention  Attributable risk/fraction 4 Case-control Attribute Fraction (AF) Attribute fraction for exposed group( AFE) Attribute fraction for population( AFP) Attribute Risk (AR) Attribute risk reduction (ARR) Relative Risk Reduction (RRR) Number needed to treat (NNT)
  • 5.
    Objectives (second part) 5 reliability Quantitativedata ICC Bland’s and Altman Lin’s coefficient corellation Qualitative data Kappa and Kohen’s Kappa Fleiss Kappa
  • 6.
    Impact criteria  AR:Attributable Risk  AF: Attributable Fraction  AFE: Attributable Fraction for Exposed Group  AFP: Attributable Fraction for Population
  • 7.
    AR/AF To developapproaches for disease prevention Overall risk Risk among non- exposed AR = 0 20 40 60 80 100 + - Excess Risk AR = [ (a+c) / n ] – [ c / ( c+d) ] a b c d
  • 8.
    AF (attribute fraction) AF= AR ÷ total risk AF = AR ÷ 𝑎 + 𝑐 ÷ 𝑛 = 𝑎 + 𝑐 ÷ 𝑛 − 𝑐 ÷ 𝑐 + 𝑑 ÷ [ 𝑎 + 𝑐 ÷ 𝑛 ] 𝒑𝒆 = [ 𝒂 + 𝒃 ÷ 𝒏] 𝑨𝑭 = 𝒑𝒆 𝑹𝑹 − 𝟏 ÷ [ 𝟏 + 𝒑𝒆 𝑹𝑹 − 𝟏 ] 𝑷𝒄 = [ 𝒂 ÷ 𝒂 + 𝒄 ] 𝑨𝑭 = 𝒑𝒄 𝑹𝑹 − 𝟏 ÷ [ 𝑹𝑹 − 𝟏 ] a b c d
  • 9.
    Relative risk (RR)= 𝑎 𝑎+𝑏 𝑐 𝑐+𝑑 Odds-ratio (OR) = 𝑎𝑑 𝑏𝑐 yes no yes a b no c d event exposure
  • 10.
    Death rate Totalnumber Number of living people Number of death 0.342 284 187 97 Have used 0.27 832 607 225 Non user 0.289 1116 794 322 total Frequency of mortality in colorectal cancer patients in terms of the variables of smoking status  AR = %28.9 – %27.0 = %1.9 ( %1.9 of the risk of mortality from colon cancer can be attributed to smoking )  𝐀𝐅 = 𝐴𝑅 𝑜𝑣𝑒𝑟𝑎𝑙𝑙 𝑟𝑖𝑠𝑘 = 1.9 28.9 = 0.063 or 6.3% (In comparison with the whole community, almost 6% of death from colon cancer can be attributed to cigarette smoking. Overall risk Probability of risk in non- exposed
  • 11.
    AFE = 𝑅𝐷 𝑡ℎ𝑒 𝑟𝑖𝑠𝑘𝑖𝑛 𝑒𝑥𝑝𝑜𝑠𝑒𝑑 𝑔𝑟𝑜𝑢𝑝 = 𝑅𝑅−1 𝑅𝑅 = ( 0.342 −0.27 ) 0.342 = 0.208 (By eliminating smoking in the exposure group, 21% of deaths from smoking can be prevented. ) 𝑹𝑹 = 97 284 225 832 = 1.263 Confidence Interval for AFE: ( 0.04 , 0.35 ) ( In comparison with smokers, Minimum and maximum fraction of mortality attributed to cigarette smoking of colon cancer patients is in the range ) Death rate Total number Number of living people Number of death 0.342 284 187 97 Have used 0.27 832 607 225 Non user 0.289 1116 794 322 total
  • 12.
    Rate of smokers Total number smoking Numberof non- smoking ones Number of smokers 0.671003367Exposed 0.202559446113Non exposed 0.273659479180Total 0.1520.0690.372The rate of AR = [(a+c) ÷n]-[c÷(c+d)] = [(67+113) ÷ 659]-[113 ÷(113+446)] = (27.3-20.2) = 7.1 (7.1% of the risk of being smoker can be attributed to peers’insistence ) AF = AR ÷ total risk = (7.1 ÷ 27.3) = 0.26 (In comparison with whole population, 26% of smoking can be attributed to peer’s insistence) OR = (a × d) ÷ (b × c) = (67 × 446) ÷ (33×113) = 8.01 AFE = [OR-1] ÷[OR] = (8.01-1) ÷8.01 = 0.88 (By eliminating the effect of peers’insistence, 8.8 percent of smoking may be prevented in exposed group)
  • 13.
    Confidence Interval forAFE: ( 0.80 , 0.92 ) {Therefore, in the statistical population with 95% confidence, For those with a peer's insistence, minimum and maximum fraction attributed to peer's insistence is in the domain above.} AFP = AFE × percent exposed among cases = 0.875 × (67 ÷180) = 0.326 {By eliminating the effect of peers’insistence, 8.8 percent of smoking may be prevented in whole population)
  • 14.
    ARR and RRRin randomized clinical trial  ARR: Absolute Risk Reduction  RRR: Relative Risk Reduction  NNT: Number Needed to Treat  NNH: number needed to harm  ARR= NNT if ………
  • 15.
     ARR = 𝑐 𝑐+𝑑 − 𝑎 𝑎+𝑏 = 21 21+9 + 11 11+16 =0.70 − 0.41 = 0.29 (the contribution of omega-3 in treatment of neuropathy is only 29%.) NNT = 1 𝐴𝑅𝑅 = 1 0.29 =3.4 ≈3 (it's expected that using this supplement in about 3 patients, will improve one of them. ) RRR == ( 𝑐 𝑐+𝑑 )−( 𝑎 𝑎+𝑏 ) ( 𝑐 𝑐+𝑑 ) = 21 21+9 −( 11 11+16 ) ( 11 11+16 ) = 0.29 0.41 = 0.71 (The relative reduction in neuropathy induced by intervention was 71% compared with the placebo group.) The risk of neuropathy Chance of recovery Total number neuropathy Non normalnormal 0.300.7030921Omega 3 0.590.41271611Placebo
  • 16.
     NNH= Numberneeded to harm Experimental group ( E) Control group (C) total Events (E) EE = 75 CE = 100 115 Non-events (N) EN = 75 CN = 150 285 Total subjects (S) ES = EE + EN = 150 CS = CE+ CN = 250 400 Event rate (ER) EER = EE/ ES = 0.5 or 50% CER = CE/ CS = 0.4 or 40% equation variable Abbr. Value EER-CER Absolute risk increase ARI 0.1 or 10% (EER –CER) / CER Relative risk increase RRI 0.25 or 25% 1 / (EER – CER) Number needed to harm NNH 10 EER/CER Risk ratio RR 1.25 (EE / EN) / (CE/CN) Odds ratio OR 1.5 (EER –CER) / EER Attributable fraction among exposed AFE 0.2
  • 17.
    Agreement measures Golden standard ICCand SEM Bland and Altman’s limit of agreement Lin’s concordance correlation coefficient Cohen's kappa and Weighted Cohen’s Kappa ROC
  • 18.
    𝐿𝑅+ and 𝐿𝑅−  PPV:Positive predictive value  NPV: Negative predictive value Diagnosed sick Diagnosed healthy Sick True positive False negative healthy False positive True negative
  • 19.
    Prior and Posteriorprobability  If LR > 1 then Posterior probability > Prior probability disease  If LR < 1 then Posterior probability < Prior probability
  • 20.
    Sensitivity = 𝑇𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑇𝑟𝑢𝑒𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 = 23 23+13 = 0.64 → 64% Specificity = 𝑇𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑇𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 = 51 51+13 = 0.80 → 80% PPV = 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑓𝑎𝑠𝑙𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 = 23 23+13 = 0.64 → 64% NPV = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒+𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡ℎ𝑒 𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 = 53 53+13 = 0.80 → 80% GDM Total number Health Sick 36 13 23 + 64 51 13 - 100 64 36 Total number
  • 21.
     𝑳𝑹+= 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 1−𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑡𝑦 = 𝑜𝑑𝑑𝑠 𝑜𝑓𝑡𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 𝑜𝑑𝑑𝑠 𝑜𝑓 𝑓𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒 = 23 36 13 64 = 3.15 (The chance to see a real patient to a false patient in a positive test result is approximately 3 times of the average effect size)  𝑳𝑹− = 1−𝑆𝑒𝑛𝑠𝑖𝑣𝑖𝑡𝑦 𝑆𝑝𝑒𝑐𝑖𝑣𝑖𝑡𝑦 = 𝑜𝑑𝑑𝑠 𝑜𝑓 𝑓𝑎𝑙𝑠𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 𝑜𝑑𝑑𝑠 𝑜𝑓 𝑡𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒 = 13 36 56 64 = 0.45 (The chance to see a healthy, false patient to actual normal one, in the negative test results is almost half of the effect size.) GDM Total number Health Sick 36 13 23 + 64 51 13 - 100 64 36 Total number
  • 22.
    Index Value Lower bound Upper bound prevalence 36%27% 46.2% Sensitivity 63.9% 46.2% 79.2% Specificity 79.7% 67.8% 88.7% PPV 63.9% 46.2% 79.2% NPV 79.7% 67.8% 88.7% 𝐿𝑅+ 3.15% 1.83 5.42 𝐿𝑅− 0.45% 0.29 0.71 Note in practice: it’s required to calculate confidence interval for specificity, Sensivity, 𝐿𝑅+ , 𝐿𝑅− , NPV and PPV.
  • 23.
    Prior probability ofdisease= 36 ÷ 100 = 0.36 ( equal for everyone and means prevalence) prior chance of being actual patient = Pre-test-odds (+)= Pre-test- probability / (1 – Pre-test- probability ) = 0.36÷ 1 − 0.36 = 0.56 (posterior chance of being actual patient) Post test Odds (+)=Pretest odds × LR+ = 0.56 × 3.15 = 1.77 (posterior probability of being actual patient) Post Probability (+)=(Pre test Odds × LR+) / (1 + (Pre test Odds × LR+ )) = 1.77÷(1+1.77) = 0.64 (posterior chance of being false patient) Post test odds (-)= Pretest odds × LR− = 0.56 ×0.45 = 0.26 (posterior probability of being false patient) Post Probability (-)=(Pre test Odds × LR− ) / (1 + (Pre test Odds × LR− )) = 0.26÷(1+0.26) = 0.20
  • 24.
    ICC (Intra classcorrelation coefficient (ICC) for one way and two way models) Missing data are omitted in a list wise way. When considering which form of ICC is appropriate for an actual set of data, one has take several decisions (Shrout & Fleiss, 1979):  1. Should only the subjects be considered as random effects (’"oneway"’model) or are subjects and raters randomly chosen from a bigger pool of persons (’"twoway"’ model).  2. If differences in judges’ mean ratings are of interest, interrater ’"agreement"’ instead of ’"consistency"’ should be computed.  3. If the unit of analysis is a mean of several ratings, unit should be changed to ’"average"’. In most cases, however, single values (unit=’"single"’) are regarded.
  • 25.
    Standard Error ofMeasurement(SEM) SEM Reliability SEM = SD× √(1-ICC) How can I trust you?
  • 26.
    Quantitative measures The concordancecorrelation coefficient combines measures of both precision and accuracy to determine how far the observed data deviate from the line of perfect concordance (that is, the line at 45 degrees on a square scatter plot). Lin's coefficient increases in value as a function of the nearness of the data's reduced major axis to the line of perfect concordance (the accuracy of the data) and of the tightness of the data about its reduced major axis (the precision of the data) Lin's concordance correlation coefficient <0.9 poor 0.9 to 0.95 moderate 0.95 to 0.99 substantial > 0.99 Almost perfect
  • 27.
  • 28.
    Bland and Altman'slimit of agreement (Tukey mean-difference plot)  They advocated the use of a graphical method to plot the difference scores of two or more than two measurements against the mean for each subject and argued that if the new method agrees sufficiently well with the old, the old may be replaced.  BlandAltmanLeh and Blandr
  • 30.
    ICC (Intraclass CorrelationCoefficient)  is a descriptive statistics that can be used when quantitative measurements are made on units that are organized into groups. It describes how strongly units in the same group resemble each other. 90 / 100 9 / 10
  • 32.
    Cohen’s Kappa andweighted Kappa for two raters Inter observer variation can be measured in any situation in which two or more independent observers are evaluating the same thing. Sometimes, we are more interested in the agreement across major categories in which there is meaningful difference. Strongly disagree disagreeNo IdeaagreeStrongly agree Strongly agree agree No Idea disagree Strongly disagree Colour lost means weight lost
  • 34.
    Commonly used weights Weighted-Kapparequires specific weight, for instance:  1- 𝒊 −𝒋 𝒌−𝟏  𝟏 − 𝒊−𝒋 𝒌−𝟏 𝟐  arbitrary weights { I for rows – J for columns K for maximum number of instruments} The nearer apples, the bigger weights
  • 35.
  • 36.
    The power ofdifferent Kappa coefficients in determining the amount of agreement among observers or referees Agreement reliabilityKappa statistic amount WeakLess than 0 low0 – 0.2 Less than average0.21 – 0.4 average0.41 – 0.6 good0.6 – 0.8 excellent0.81 - 1 Is there any confidence interval for kappa? Is it valuable?
  • 37.
    references  Understanding InterobserverAgreement:The Kappa Statistic. Anthony J. Viera, MD; Joanne M. Garrett, PhD. Family Medicine.  Bland J, Altman D (1986). Statistical methods for assessing agreement between two methods of clinical measurement. The Lancet 327: 307 - 310.  Bland J, Altman D (1999). Measuring agreement in method comparison studies. Statistical Methods in Medical Research 8: 135 - 160.  P. S. Myles, J. Cui, I. Using the Bland–Altman method to measure agreement with repeated measures, BJA: British Journal ofAnaesthesia, Volume 99, Issue 3, September 2007, Pages 309–311,  Bland J, Altman D (2007). Agreement between methods of measurement with multiple observations per individual. Journal of Biopharmaceutical Statistics 17: 571 - 582. (Corrects the formula quoted in the 1999 paper).  Bradley E, Blackwood L (1989). Comparing paired data: a simultaneous test for means and variances. American Statistician 43: 234 - 235.  Burdick RK, Graybill FA (1992). Confidence Intervals on Variance Components. New York: Dekker.  Dunn G (2004). Statistical Evaluation of Measurement Errors: Design and Analysis of Reliability Studies. London: Arnold.
  • 38.