Biostatistics

Primary MMed (Anaesthesia)

2006
Writing a study protocol
 introduction
 research question

 current knowledge

 research hypothesis

 research objective
Writing a study protocol
 methodology
 type of study


experiments
 prospective (randomized, control)

 observational

 prospective

 retrospective

 audit
Writing a study protocol
 methodology
 sample size calculation

 guided by expected difference in either

proportion or numerical data, standard deviation

of known data, α and β values, and intended power
of study
 plan to recruit more subjects in case of drop outs

 size of drop outs will affect the basis of

assumptions and the initial sample size

calculation, and the research hypothesis may
show no significant difference (P>0.05)
Writing a study protocol
 methodology
 patient consent

 inclusion criteria

 exclusion criteria, restriction

 process of randomization

 procedure (how to go about carrying out the

data collection)
 control group

 test group(s)
Writing a study protocol
 methodology
 monitoring of the patients during the period of

procedure

routine - heart rate, blood pressure, oxygen
saturation
 data / observations


side effects of test protocol
 rescue therapy

 safety of patient

 treatment plan

 criteria for withdrawal from study

Writing a study protocol
 statistical analysis
 the data will be subjected to a test for

normality
 statement about treatment of normal or

nonparametric distribution of data

 normal distribution


expressed as mean and standard deviation,
with 95% confidence interval

t-test for comparison of means obtained from
2 groups of data

analysis of variance test for comparison of
means obtained from more than 2 groups

Χ2 test for discrete data
Writing a study protocol
 statistical analysis
 non-normal distribution


expressed as median and range, with limits of
25th and 75th percentiles
 Mann-Whitney U test for analysis of data

from 2 groups
 Kruskal-Wallis test for analysis of data from

more than 2 groups

 state the P value and decide on the significance

level, conventionally it is P<0.05

 submit protocol for ethics committee approval
The null hypothesis

H0

2006
Null hypothesis
 study hypothesis
 investigator conducting a study usually has a

theory in mind
 however, very difficult to prove the hypothesis

 simpler to disprove a hypothesis than proving it

 null hypothesis
 differences observed is not due to exposure to

factor, and is by chance

 always phrased in the negative and that is why

it is termed null
Types of research

Longitudinal studies
Cross-sectional studies

2006
Types of research - longitudinal study
 investigates a process over time
 the effect of external factors on human

subjects
 3 types
 clinical trial, a cohort study, case-control study

 prospective or retrospective studies

 in prospective studies, subjects are grouped

according to ‘exposure’ to some factor

 in retrospective studies, subjects are grouped

according to outcome, the ‘exposure’ effect is

then determined retrospectively
Types of research - cross-sectional
study
 describes a phenomenon fixed in time

 description of staging system for cancer

 laboratory studies of biological processes
Comparison
longitudinal studies cross sectional studies
 prospective studies  disease description
 according to  diagnosis and staging

“exposure”  abnormal ranges


randomised  disease severity

 non-randomised  disease processes

 observational

studies
 retrospective studies
 according to

outcome
 determine

“exposure”
Randomised clinical trial
 randomisation
 is a procedure in which the play of chance

enters into the assignment of a subject to the

alternatives (control and test groups) under
investigation, so that the assignment cannot be
predicted in advance
 tends to produce study groups comparable in

unknown as well as known factors likely to

influence outcome apart from the actual
treatment being given itself
 guarantees that the probabilities obtained from

statistical tests will be valid

Randomised clinical trial - designs
 parallel designs
 one group receives the test treatment, and one

group the control

 cross-over designs
 the subjects receive both the test and the

control treatments in a randomised order

 each subject acts as own control, allowing paired

or matched analysis, and provides an estimate of

the difference between test and control
 useful in chronic disease that remain stable over

time, such as diabetes and asthma, where the

purpose of treatment is palliative, not cure
Designs of a randomised clinical
trial: parallel, cross-over
Randomisation

Assessment
Test

Control

A parallel design clinical trial

Randomisation

Assessment

Assessment
Test Control

Control Test

A two period cross-over design clinical trial

Problems of cross-over design
 cross-over effect
 possibility that the effect of the particular

treatment used in the first period will carry

over to the second period, and may interfere
with how the treatment scheduled for the
second period will act, thus affecting the final
comparison between the two treatments
 to allow for this possibility, a washout period,

in which no treatment is given, should be

included between successive treatment
periods
Problems of cross-over design
 disease may not remain stable over the trial
period
 subject drop-outs
 more will occur in this design trials than in

parallel design trials, due to extended

treatment period
Non randomised studies
 historical controls
 when randomisation is not possible


problem of bias selection already occurring -
those who did not receive transplants may be
more ill or may not have satisfied the criteria
 survival of patients who received heart

transplants and patients who did not

Non randomised studies
 pre-test–post-test studies
 a group of individuals are measured, then

subjected to treatment or intervention, and

then measured again
 purpose of the study is to study the size of the

effect of treatment or intervention (e.g.

campaign)
 major problem is ascribing the change in

measurement to the treatment since other

factors may also have changed in that interval
Cohort
 a cohort is a component of a population identified
so that its characteristics can be ascertained as it
ages through time
 designated groups of persons either born in a

certain year or traced over a period of time

(who ever worked in a factory)
Cohort studies
 cohort study
 one in which subsets of a defined population can be

identified who have been exposed (or will be

exposed) to a factor which may influence the
probability of occurrence of an outcome (given
disease)
 usually confined to studies determining and

investigating aetiological factors and do not allocate

the equivalent of treatments
 also for post-marketing surveillance, comparing

adverse effects of new drug with alternative

treatment
 may be referred to as follow-up, longitudinal or

prospective study
 often termed observational studies, since they

observe the progress of individuals over time

Progress of a cohort study

With With disease

Population

disease Exposed
Without disease
Without
disease
With disease
Not
exposed
Without disease

Time
Problems with cohort study
 exposure to factor may be by natural selection or
up to the individuals’ decision
 bias may influence the measure of interest

 other associated factors may also influence

measure of interest
 e.g. cohort study of cardiovascular risk in men

sterilised by vasectomy
 e.g. incidence of breast cancer with

consumption of alcohol
Problems with cohort study
 size of the study
 the required size of a cohort study depends not

only on the size of the risk being investigated

but also the incidence of the particular
condition under investigation
 cohort studies not suitable for investigating

aetiological factors in rare diseases

Problems with cohort study
 problems with interpretation
 bias pool of study subjects


when cohort is made up of employed
individuals, the risk of dying in the first few
years of follow up is less than in general
population, this is known as healthy worker
effect; people who are sick are less likely to
be employed
 incomplete representation


e.g. people who respond (to questionnaires)
and people who are lost to follow up
Case-control studies
 case-control study
 also known as a case-reference study or

retrospective study
 starts with identification of persons with the

disease (or other outcome variable) of interest,

and a suitable control group of persons without
the disease
 the relationship of a risk factor to the

disease is examined by comparing the two

groups with regard to how frequently the risk
factor is present
Case-control studies

Exposed
Subjects with
disease (case)
Not exposed

Exposed
Subjects without
disease (control)
Not exposed

Time
(Retrospective)
Case-control studies
 designs
 matched design


control subjects can be chosen to match
individual cases for certain important
variables such as age, gender and weight
 unmatched design

 controls can be a sample from a suitable non-

diseased population
Case-control studies
 selection of controls
 not required that the control group are alike in

every aspect to the cases, usually 2 or 3

variables which presumably will influence
outcome are matched, such as age, gender,
social class
 main purpose is to control for confounding

variables that might influence the case-control

comparison
 confounding arises when the effects of two

processes are not separated, e.g. disease

related to 2 exposure factors
Case-control studies
 selection of controls
 matching can be wasteful if matching criteria

leads to many available controls being discarded

because they fail the matching criteria
 if controls are too closely matched to their

respective cases, the relative risk may be

underestimated
Limitations of case-control studies
 ascertainment of exposure relies on previously
recorded data or on memory, and it is difficult to
ensure lack of bias between the cases and
controls
 the cases may be more motivated to recall

possible risk factors

 difficulty with selection of suitable control group
 a major source of criticism
Cross-sectional studies
 subjects are included without reference to either
their exposure or their disease
 usually deals with exposures that do not change,
such as blood type, or chronic smoking habit
 resembles a case-control study except that the
number of cases are not known in advance, but are
simply the prevalent cases at the time of survey
Cross-sectional studies
 sampling methods
 quota sample


to ensure that the sample is representative
of general population in say, age, gender and
social class structure
 not recommended in medical research

 grab or convenience sample


only subjects who are available to the
interviewer can be questioned
Cross-sectional studies
 problems
 bias in the type of responders and non-

responders
 exposures have to be determined by a

retrospective history
Calculation of sample size

2006
Calculation of sample size
 consider
 control group response

 the anticipated benefit (of the treatment)

 significance level

 power
Control group response
 it is first necessary to postulate the response of
the control group patients
 denoted by π1, to distinguish it from the value that
will be obtained from the trial, denoted p1
 experience of other studies may provide π1
The anticipated benefit
 it is also necessary to postulate the size of the
anticipated response in treatment group patients,
denoted by π2, to distinguish it from the value
that will be obtained from the trial, denoted p2
 anticipated benefit δ = π2 - π1
Type I error
 the error of incorrectly rejecting the null
hypothesis when it (the hypothesis) is true
 the error of concluding that the differences

seen in the result is significant when in fact it

is not
 wrongly accepting that differences in the

results as significant when there is no

difference
 designated α

 equivalent to the false positive rate (1-specificity)

One or two-sided
 null hypothesis (H°) is that there is no significant
difference, and chance has occurred
 no assumption about the direction of change or

variation
 alternate hypothesis states that the difference is
real, further that it is due to some specific factor,
 where no direction of change is specified both

ends of the distribution curve are important, and

the test of significance is two-sided, or two-
tailed
 where the direction is specified, then only one tail

of the curve is relevant, and the test of

significance is one-sided, or one-tailed
One or two-sided
 the critical value is the value of a test statistic at
which we decide to accept or reject H°
 critical value for a one-sided test at

significance p, will be equivalent to that for a

two-sided test at 2p
 one-sided p = 0.025 / two-sided p = 0.05

 thus, it is tempting to use one-sided tests as the

significance is greater, but the decision should be
made before the data is collected, not after the
direction of change is observed and should be
clearly stated when presenting results
One or two-sided
Frequency

μ=0

P<5% (two-sided) P<2.5% (one-sided)

1.96σ
z
Significance level
 the significance level, α, is the probability of
making a Type I error and is set before the test
is carried out
 in most cases, it will be two-sided
Significance level
 denoted by the letter P, represents the
probability of the observed value being due solely
to chance variation
 can be interpreted as the probability of obtaining
the observed difference, or one more extreme, if
the null hypothesis is true
 the smaller the value of P, the less likely the
variation is to be due to chance and the stronger
the evidence for rejecting the null hypothesis
Significance level
 most scientific work, by accepted convention,
rejects the null hypothesis at P < 0.05
 probability of the observed value being due

solely to chance is < 0.05 (or < 1 in 20)


this means that we shall reject the null
hypothesis on 5% of occasions, when it is in
fact true, i.e. there was simply a chance
variation and that the 2 treatment are
equally effective
Significance levels
 α
 the probability that a
random variable, (Normally
distributed with mean = 0
and standard deviation = 1)
will be greater than z or α/2 α/2
less than -z
 z –zα +zα
 the value on the horizontal
axis of a Normal
distribution corresponding
to the probability α
Significance level
 if α is 0.05, then the corresponding z is 1.96
 to link z with the corresponding α, we write

z0.05 = 1.96
 zα is the value along the axis of a Normal

distribution
 thus

 0.05/2 = 0.025 is to the left of z = –1.96


0.025 is to the right of z = +1.96
Type II error
 error in incorrectly accepting the null hypothesis
of no difference between treatments, when it
(the hypothesis) is in fact false (and should be
rejected)
 accepting the null hypothesis when there should

be significant difference in the results

 accepting that the differences seen in the

result is not statistically significant and making

the conclusion P>0.05
 the probability of making a type II error is
designated β
 equivalent to the false negative rate (1-
sensitivity)
Type II error
 2 main factors produce β error
 chance alone


unusual data sample which does not support a
difference
 statistical methods can produce incorrect

conclusions
 too small a sample size


the smaller n, the greater must be the real
difference before statistical difference may
be shown
 usually set at a value of β = 0.2 or 20%
Power
 the probability that a study can predict a
difference, when a real difference actually exists,
is termed the statistical power of the study
 it is the probability of rejecting the null
hypothesis when it (the hypothesis) is false
 first decide how much false negative or type II
error (β) rate is reasonable
 power equals 1-β

 the higher the power of the study, the smaller the

difference which may be detected
Statistical Testing Errors
Real difference
No Yes
Significance No effect 1-α Type II error
tests seen β-error
Effect Type I error Power
seen α-error 1-β
Calculation of sample size
 calculations depend on a function (zα + z2β)2, where
zα and z2β are the ordinates for the normal
distribution
 the value of z for the corresponding α and 2β

are read off from the Table of Normal

distribution
Table to assist in sample size calculations, α = 0.05
β Power (1-β) z2β zα (zα +z2β)2
0.3 0.7 0.524 1.960 6.172
0.2 0.8 0.842 1.960 7.849
0.1 0.9 1.282 1.960 10.507
Calculation of sample size
 comparison of proportions
 wish to detect a difference in proportions

δ = π2 - π1
e.g. response rate to placebo and treatment
drug
 for Χ2 test, the number in each group should be
at least
m = (zα + z2β)2{π1(1-π1) + π2(1-π2)}
δ2
Calculation of sample size
 comparison of means (unpaired data)
 wish to detect a difference in means

δ = μ2 - μ1
 the number in each group should be at least
m = 2(zα + z2β)2 σ2
δ2
where σ is presumed to be the same with both
drugs
Calculation of sample size
 comparison of means (paired data)
 with cross over trial

 the number in each group should be at least

m = (zα + z2β)2 σw2

δ2
where σw is the standard deviation of the paired
difference between treatment
Randomisation

2006
Methods of randomisation
 simple randomisation
 either manually by tossing coin or throwing a

six-sided die or from table of random numbers

 a good method in large trials but

 does not guarantee equal numbers of patients

in each of the two groups

 in smaller trials there is a high chance of

getting notable imbalance between the groups

 groups of different sizes
!
 presence of allocation bias
Methods of randomisation
 random permuted block (RPB) randomisation
 a method for ensuring that group sizes never get

too far out of balance and avoids assigning

different numbers to each study group
 combinations (e.g. AABB, ABBA) obtained from

blocked randomisation can be assigned numbers

(say 1-6), the sequence of the combinations can
then be dictated by numbers from table of random
numbers
 a potential problem with the method is that if the

block length becomes know, the method is

predictable and selection bias can arise
 randomly varying block length can help
Methods of randomisation
 stratified randomisation
 if there are important prognostic factors which, if

they were distributed unequally between the

treatment groups, would give rise to a serious bias,
then it may be prudent to intervene in the
randomization process to ensure balance between
these factors to ensure balanced treatment
allocation for patients within each group or centre
 in stratification, RPBs are used within each stratum

defined by the prognostic factors

 stratification can be cumbersome if there are too

many prognostic factors and minimization is a method

which is can provide balance in a less cumbersome
way
Methods of randomisation
 unequal randomization (1:2, 2:1 for sample sizes of
test : control groups)
 although maximum power can be obtained when

the allocations to the groups are in the ratio 1:1,

the loss in power is slight if the ratio departs
only slightly from 1
 there can be practical advantages to unequal

allocation, which might be worth considering in

some applications.
Methods of randomisation
 carrying out randomisation
 randomisation list should be prepared and held

by a person not involved in the investigation and

not the investigator determining patient
eligibility
 this person serves as a check for the trial

 when a patient is confirmed as eligible for the

trial, randomisation is then revealed


over the telephone
 by opening sequentially numbered envelopes
Data description

2006
Types of data
 qualitative data (varying by description)
 nominal data

 ordered categorical or ranked data or ordinal

data
 numerical or quantitative data (varying by number)
 numerical discrete data (differ only by fixed

amount)
 numerical continuous data (differ by any

amount)
Qualitative data
 nominal data
 data that one can name or describe


not measured but simply counted
 expressed in number (or frequency) or

percentage (or relative frequency)

 2 or more groups of observation


2 groups - gender of patients, did or did not
get exposure to factor
 > 2 groups - blood groups, racial groups,

anaesthetists, surgeons
Qualitative data
 ordered categorical or ranked data
 if there are more than two categories of

classification, it may be possible to order them

in some way or assign ranks to categories to
facilitate statistical analysis
 e.g. ordinal data : mild, moderate, severe

- x xxx
xx
Postoperative
bleeding (ml)

- xx xxxx
xx xxx
xxx
- xxxx xx
xxx
x xx
- xxx
x
A(1) B(2) C(3)
Vascular surgeon
Quantitative data
 numerical discrete
 data consists of counts


number of general anaesthetics and regional
anaesthetics performed in this hospital in a
year
 number of anaesthetic trainees passing the

Final MMed in the past 5 years

Quantitative data - numerical continuous
 such data are measurements that can take any
value in a given range
 age, estimated blood volume, hourly blood loss

 interval scale,
 position of zero is arbitrary, a difference

between two measurements has meaning, but

not their ratio
 meaning may change if a ratio or percentage is

applied to a different scale, e.g. 10% increase in

body temperature in Celsius and Fahrenheit
scales
Quantitative data - numerical continuous
 ratio scale,
 value of zero has real meaning, negatives are invalid

 meaning does not change when ratio or percentage is

applied to different units of measure, e.g. 10% increase

in body weight in kilograms or pounds
 continuous data is often dichotomised to make nominal
data and then ordered or ranked for statistical analysis
 e.g. diastolic blood pressure which is continuous can be

converted to hypertension (>90 mmHg) or normotension

(≤ 90 mmHg)
Summary of measurements
 nominal
 name, birthplace

 ordinal
 mild | moderate | severe

 interval
 meaningful distance between values

 ratio
 allows study of absolute magnitude
Summarising data

2006
Summarising data for the study
 measurement of location or central tendency
 mean or arithmetic average


interval or ratio of a quantitative variable
 median and quartiles

 interval or ratio, (± ordinal)

 mode

 nominal, ordinal, interval or ratio

 measurement of dispersion or variability

 range or interquartile range

 standard deviation

 measures of symmetry
Measurement of location or
central tendency

2006
Mean or average
 the mean (x, pronounced xbar) or average of n
observations is the sum of the observations, Σx
divided by their number, n (arithmetic average)
sum of all sample values Σx
x= =
size of sample n
 advantage
 mean uses all the data values, is statistically

efficient
 disadvantage
 vulnerable to outliers,

 single observations (not erroneous measurements)

which if excluded from the calculations, have

noticeable influences on the results
Median and quartiles
 lower, median and upper quartiles divide the data
into 4 equal parts
 approximately equal numbers of observations in

the 4 sections (equal only when n is divisible by

4)
 estimation of quartiles
 the data is first ordered from smallest to

largest, and then counting upwards the number

of observations
Median
 median or middle quartile
 value above and below which half the

measurements fall
 for odd number of observations, is the

observation at the centre of the ordering

 for even number of observations, is the average

of the ‘middle’ two observations

 advantage
 not affected by outliers

 disadvantage
 not statistically efficient as it does not make

use of all the individual data values

Lower and upper quartiles
 practical method of calculating lower or upper
quartiles are by the stem-and-leaf plot
 observations

10,13,20,20,22,22,23,24,25,25,27,28,30,30,30,
31,31,32,32,33,34,35,35,36,37,38,38,39,39,41,4
1,41,42,42,43,43,44,44,46,47,48,50,50,51,52,54
2 1 03
10 2 0022345578
17 3 00011223455678899
12 4 111223344678
5 5 00124
46
Percentile
 25th percentile
 the value above which 75 percent of the

observed cases fall and below which 25 percent

of the observed cases fall
 50th percentile
 the median, the value above and below which

half of the observed values of a variable fall

 75th percentile
 the value above which 25 percent of the

observed values of a variable fall and below

which 75 percent of the observed values of a
variable fall
Mode
 this is the value that occurs most frequently, or if
the data is grouped, the grouping with the higher
frequency
 not much use in statistical analysis as its value
depends on the accuracy with which the data are
measured
 bimodal distribution describes a distribution with
two peaks in it
Measure of dispersion or
variability

2006
Measure of dispersion or variability
 range
 difference between minimum & maximum values

 interquartile range
 (largest value 3rd quartile) - (largest value 1st

quartile)
 standard deviation
 degrees of freedom
Range or interquartile range
 range
 is given as the smallest and the largest

observations
 vulnerable to outliers

 interquartile range
 the distance between the 25th and 75th

percentile -

Postoperative
bleeding (ml)
 not vulnerable to outliers -
-
 displayed as box-whisker plots -
-
-
A B C
Surgeon
Standard deviation, s
 a measure of dispersion, i.e. how far variables are
away from their mean, often abbreviated as SD
 expressed in the same units of measurement as

the observations
Σ(x-x)
√
2
s=
n-1
 the value Σ(x-x)2 is interpreted as
 from each x value subtract the mean x,

 square this difference,

 then add each of the n squared differences

 n-1 (or the degree of freedom)

 compensates for small sample sizes (n < 30) and

higher probability of falling outside the SD

Standard deviation, s
 s reflects the variability in the data
 if x’s are widely scattered about x, then s would

be large
 variance
 a measure of the dispersion of values about the

mean
 the square of the standard deviation
s2 = Σ(x-x)2

 coefficient of variation n-1

 expresses the SD as a percentage of the

sample mean
 c.v. = s × 100%

x
Standard deviation, s
 for data with a normal distribution, there is
 a 68.26% chance that the actual value will be

within 1 standard deviation above or one

standard deviation below the mean value
 a 95.45% chance that the actual value will be

within 2 standard deviations

 a 99.7% chance that the actual value will be

within 3 standard deviations

-
x
68% 34.13% (1SD) 34.13% (1SD)
47.73% (2SD) 47.73% (2SD)
49.85% (3SD) 49.85% (3SD)
Measures of symmetry
 symmetric distribution
 if the distribution is symmetric then the median

and mean will be close

 data expressed as mean and standard deviation

 skewed distribution
 a distribution is skewed to the right (left) if

the longer tail is to the right (left)

 data expressed as median and interquartile

range
 mean and standard deviation are sensitive to

the skewness
Skewness
 an index of the degree to which a distribution is
not symmetric, or to which the tail of the
distribution is skewed or extends to the left or
right
 calculation of skewness, sk
sk = 3(mean-median)
SD
 normality can be confirmed if mean and median
are close
 skewness is used, along with the kurtosis statistic,
to assess if a variable is normally distributed
Skewness
 the normal distribution is
symmetric, and has a
skewness value of zero
 a distribution with a
→ +1
significant positive
skewness has a long right
tail
 a distribution with a
-1← significant negative
skewness has a long left
tail
Kurtosis
 a measure of the extent to which observations are
clustered in the tails
 kurtosis can be used, along with the skewness
statistic, to assess whether a variable is normally
distributed
 for samples from a normal distribution, the values
of kurtosis will fluctuate around 0
Kurtosis
 for a normal distribution,
the value of the kurtosis
statistic is 0
 if a variable has a negative
kurtosis, its distribution
→ -ve -ve ←
has lighter tails than a
normal distribution
 if a variable has a positive
kurtosis, a larger
+ve ← → +ve proportion of cases fall into
the tails of the distribution
than into those of a normal
distribution
Normal probability plot
 normality of
observation can be
confirmed from the

Observation
x
x
Normal probability plot x
x
x
x
x
x
x
x
x

Normal ordinates, z
Normality
 bell-shaped distribution of a continuous variable
symmetrical about its mean
 median and mean will be close

 skewness value is zero sk = 3(mean-median)

 kurtosis value is zero SD
 normal probability plot
Mean or median?
 mean and median
 convey different impressions of the location of

the data
 both give useful information

 if the distribution is symmetric, mean is a better

summary statistic
 if the distribution is skewed, the median is less
influenced by the tails
 for nominal or ordered categorical data, mean is
the proportion of each group
Generating data from sample to
population

2006
Populations and samples
 a population is a theoretical concept used to
describe an entire group (any collection of people,
objects, events, or observations)
 this is usually too large and cumbersome to

study so investigation is usually restricted to

one or more samples drawn from the study
population
Populations and samples
 samples are taken from populations to provide
estimates of the population parameters
 the purpose of summarising the behaviour of a

particular group is usually to draw some

inference about a wider population of which the
group is a sample, such as determining the
reference normal range
 statistics describes the sample
 parameters describe characteristics of the
population
Populations and samples
 to allow true inferences about the study
population from a sample there are a number of
conditions,
 the study population must be clearly defined

 every individual in the population must have an

equal chance of being included in the sample, i.e.

a random sample
 random does not refer to the sample, but the

manner in which it was selected


the opposite of random sampling is purposive
sampling, i.e. every 2nd patient
Sampling errors
 sampling errors
 the smaller the sample size, the greater the

error
 the greater the variability of the observations,

the greater the error

 non-sampling errors
 these do not necessarily decrease as the sample

size increases
 result in bias or systematic distortion of the

results
Central Limit Theorem
 definition: even when the variable is not normally
distributed the sample mean will tend to be
normally distributed
 if random samples of n measurements are
repeatedly drawn from a population with a finite
mean μ, and a standard deviation σ, then when n is
large, the relative frequency histogram for the
(repeated) sample means will tend to be
distributed normally
Normal distribution

2006
Normal distribution
 bell-shaped distribution of a continuous variable
symmetrical about its mean μ
 described by
σ
 population mean, μ

 population standard deviation, σ

 bell is tall and narrow for small standard

deviation, and short and wide for large standard

deviation
 a skewed distribution can be transformed into
Normal distribution shape by taking the logarithm
of the measurements or working with the square
root of the observations
Standard Normal Distribution

Frequency
μ=0

2.5% σ=1 2.5%

95%
0 1.96σ
Scoring of fine motor skills

impaired Frequency

severely borderline
impaired
average above average

+1 SD
-1.5 SD

+2 SD
-1 SD
-2 SD

68.26%
IQ Curve
Normal distribution
 mathematical property of the Normal distribution
 68.26% of the distribution lies between μ ± 1σ

 95.45% of the distribution lies between μ ± 2σ

 μ – 1.96 × σ and μ + 1.96 × σ (for exactly

95%)
 99% of the distribution lies between μ ± 3σ


μ – 2.58 × σ and μ + 2.58 × σ

α/2 1.96σ 1.96σ α/2

1-α
Normal distribution
 in practice, the parameters μ and σ must be estimated
from the sample data
 for this purpose, a random sample from the

population is first taken

 if the sample is taken from a Normal distribution,

and provided that the sample is not too small,

similarly, approximately 95% of the collected data
will be within
x – 1.96 × s to x + 1.96 × s
 1.96 is the 5% percentage point of the normal
distribution
 2.58 is the 1% percentage point of the normal
distribution
Standard error SD(x)/√n
 the standard deviation of the sampling
distribution for a statistic
 can apply to: mean, difference between means,

skewness, kurtosis, Pearson correlation,

regression coefficient, proportion, difference
between proportion
 a measure of how much the value of a test
statistic may vary from sample to sample
Standard error of the mean
 standard deviation of the mean, SD(x) or
standard error of mean SE(x) or SE
 defines the precision with which a mean is

estimated
 SE(x) or SD(x) = SD(x)/√n or s/√n
Exercise: Calculation of SE
 Mean (x) alanine aminopeptidase value for 25
subjects is 1.0 U; SD is 0.3 U
 SE = SD/√n
= 0.3/√25
= 0.3/5
= 0.06
Comparing s and SE
standard deviation, s standard error of mean
 is a measure of the  is a measure of the
variability between uncertainty in the
individuals with sample statistic,
respect to the
 always refers to an
measurement under
consideration estimate of a
 describes the sample population parameter
 the larger the sample

size, the smaller the

standard error of the
mean
Standard error of skewness
 a measure of the variability of the skewness
statistic
 examine how far it is from zero by dividing the
measure of skewness by its standard error
 the larger the absolute value of this quotient,

the less reasonable it is to assume that the

variable comes from a distribution with zero
skewness, such as the normal distribution
Standard error of kurtosis
 a measure of the variability of the kurtosis
statistic
 examine how far it is from zero by dividing the
measure of kurtosis by its standard error (SE
Kurt)
 the larger the absolute value of this quotient,

the less reasonable it is to assume that the

variable comes from a distribution with zero
kurtosis, such as the normal distribution.
Confidence interval
 gives an estimated range of values which is likely
to include an unknown population parameter, the
estimated range being calculated from a given set
of sample data
 if independent samples are taken repeatedly from
the same population, and a confidence interval
calculated for each sample, then a certain
percentage (confidence level) of the intervals will
include the unknown population parameter
 confidence intervals are usually calculated so that
this percentage is 95%, but 90%, 99%, 99.9%
confidence intervals for the unknown parameter
can be produced
Confidence interval
 the width of the confidence interval gives us some
idea about how uncertain we are about the
unknown parameter
 a very wide interval may indicate that more data
should be collected before anything very definite
can be said about the parameter
 confidence intervals are more informative than
the simple results of hypothesis tests (where H0 is
rejected or not) since they provide a range of
plausible values for the unknown parameter
Confidence limits (xx, yy)
 the lower and upper boundaries or values of a
confidence interval
 the values which define the range of a

confidence interval
Confidence interval for a mean
 define a range of values within which the
population mean μ is likely to lie
 that is, a range of values that is likely to cover

the true but unknown population mean value

(say, 95% of time)
 95% confidence interval for a large sample (>60)
 95% CI = x ± 1.96 × s/√n where s/√n is SE(x)

α/2 1.96σ 1.96σ α/2

95% confidence interval

Confidence interval for a mean
 the upper and lower values are the 95%
confidence limits
 a reported CI from a particular study may or may
not include the actual population mean
 but if the study were to be repeated 100 times,

of the 100 resulting 95% CI, we would expect

95 of these to include the population mean
Confidence interval for a mean
 small samples
 less precise statements about population

parameters can be made than with large samples

 x and s will not always be necessarily close to μ

and σ, respectively
 sample size is already taken into account in the

calculation of the standard deviation of the mean,

SD(x), using √n, i.e. s/√n
 of practical importance only when the sample size is

very small (less than 15) and when the distribution

in the population is extremely non-normal
Exercise: Calculation of CI
 Mean (x)- alanine aminopeptidase value for 25
subjects is 1.0 U; SD is 0.3 U
 Calculate 95% CI
 95% CI = x- + (1.96 x SE) where SE = SD/√n
= 1.0 + (1.96 x 0.3/√25)
= 1.0 + (1.96 x 0.06)
= 1.0 + (0.1176)
 95% CI is from 0.8824 to 1.1176
 lower and upper boundaries of 95% CI = (0.8824,
1.1176)
t distribution

2006
t distribution
 t distribution with (n–1) degrees of freedom
 introduced by WS Gossett, who used the pen-

name ‘Student’, and is often called Student’s t

distribution
 like the normal distribution,
 the t distribution is a symmetrical bell-shaped

distribution with a mean of zero, but is more

spread out, having longer tails
 the exact shape of the t distribution
 depends on the degree of freedom (d.f.), n–1, of

the standard deviation s

 degrees of freedom refers to the number of

observations completely free to vary, the fewer

the degrees of freedom, the more the t
distribution is spread out
Normal distribution, t-distribution

α/2 α/2 α/2 α/2

–zα +zα –tα +tα

Confidence interval using t distribution
 confidence interval
 is calculated using t’, the appropriate percentage

point of the t distribution with (n–1) degrees of

freedom
 small sample CI = x ± (t’ × s/√n)

 for small degrees of freedom,

 the percentage points of the t distribution are

larger in value than the corresponding percentage

points of the normal distribution
 because sample standard deviation s may be a poor

estimate of the population σ, and when this

uncertainty is taken into account, the resultant CI is
wider
t-Test
 reflecting this increased conservatism, the
critical value for the t-test

Significance t-Test z-Test

P<0.05 2.26 1.96
P<0.01 3.25 2.58
Statistical tests

the type of test used depends upon

the sample size

2006
Non-parametric tests Use Parametric tests
Wilcoxon signed rank Test of Paired t test
test difference
between paired observations
Wilcoxon rank sum test Comparison of 2
Mann-Whitney U test groups 2-sample t test
Kruskal-Wallis one-way Comparison of
analysis of variance several groups One-way analysis of
Spearman rank Measure of variance
correlation association between 2 Pearson correlation
variables
Χ2 goodness of fit test Comparison of an
observed
frequency
distribution with a theoretical one
Tests statistics
 a statistic derived from sample data, used to
measure the difference between the observed
data and what would be expected under the null
hypothesis:
 z-statistics

 principal of the standard normal deviate

 t-statistics

 small samples, with limited degrees of

freedom
 Χ2-statistics

 categorical or qualitative variables

Normal test, z test
 the Normal test or z test requires that,
 the sample size is large (n > 30)

 the population standard deviation, σ is known

 the variable is assumed to be normally

distributed
 commonly the population σ is unknown,

however it is possible to use the sample

standard deviation, s as an estimate of σ
Large sample
 if the sample size is large, n > 30
 then the sample standard deviation, s, is

considered to be an adequate estimate of the

population σ
 thus, the standard error of the sample mean

becomes,
SE(x) = (s/√n)
 under these circumstances the z-test can again

be used to test the significance of the

difference between the population mean μ and
the sample mean x'
 assuming the population fits a normal

distribution
Small samples
 if n < 30
 the sample standard deviation, s, is not an

adequate estimate of the population σ

 Student's t-test is employed (Gosset in 1908)

 the t-distribution describes a series of curves,

 dependent upon the number of degrees of

freedom

as for the normal distribution, these are
symmetrical with a mean μ = 0
Paired samples t-test
 attributes and demographic data, disease
condition are matched to make two groups of
subjects as similar as possible
 the two groups
 can be two groups of subjects in a matched

case-control study
 can be of the same subjects observed before

and after a treatment as in a cross-over trial

 this test is a statistical test of the null
hypothesis that two population means are equal
 any observed differences between the groups, if
statistically significant, can be attributed to the
variable of interest
Paired t-test
 also known as the related test, or matched test
 paired t = x /(s/√n), d.f. = n-1
 the corresponding P value or significance level,

is obtained from the Table of percentage point

for Student’s t distribution
 95% CI = x ± (t0.05 × s/√n)
One sample t-test
 tests whether a sample mean is different from
some specified value, μ, which need not be zero
 t = (x-μ) / (s/√n), d.f. = n–1
Two sample or unpaired t-test
 also known as independent sample, or unrelated
test, the t test is used for small samples (n<30)
 for analysing data in 2 groups of subjects in a
parallel group clinical trial or the unmatched case-
control study
 requires that the population distributions are
normal
 when comparing 2 means, the validity of the t test
also depends on the equality of the 2 population
standard deviations
Two sample or unpaired t-test
 the standard error for the difference between
the means,
SE (x1-x2) = s√(1/n1 + 1/n2)
 where s is the common standard deviation and is

derived from s1 and s2

 the t value is calculated as
 t = (x1-x2) / s√(1/n1 + 1/n2), d.f. = n1 + n2 - 2
 confidence interval is
 CI = (x1-x2) ± (t’× SE (x1-x2))
Small samples, unequal SD
 first approach is to seek a suitable change of
scale to remedy this, so that the t test can be
used
 taking logarithms of the individual values

 alternatives are to use a non-parametric test or to

use either the Fisher-Behrens or the Welch tests
Comparison of several means

analysis of variance

2006
Analysis of variance
 the t-test is generalised to more than 2 groups by
means of a technique termed analysis of variance
 for this method, there are both between- and
within-groups degrees of freedom
 the between-groups and within-groups degrees

of freedom are quoted in this order for every

case
 depending on the number of factor(s) included

for analysis

one-way analysis of variance
 two-way analysis of variance
One-way analysis of variance
 used to compare the means of several groups
 one-way analysis of variance is used when the
subgroups to be compared are defined by just one
factor
 e.g. comparison of means between different

socioeconomic classes, or different ethnic

groups, or by a disease process
 this method assesses how much of the overall
variation in the data is attributed to differences
between the group means, and comparing this with
the amount attributable to differences within
group
Two-way analysis of variance
 used when the data are classified in 2 ways
 e.g. by age-group and gender

 balanced and unbalance study designs

 a balanced design if there are equal numbers of

observations in each group

 an unbalanced design if there are not equal

numbers of observations in each group


multiple regression test can also be applied
Non-normal distribution

non-parametric tests

2006
Non-parametric tests
 non-parametric statistical tests are for analysing
numerical data that make no assumption about the
underlying normality of distribution
 particularly useful when there is obvious non-
normality in a small data set which cannot be
corrected with a suitable transformation
Wilcoxon signed rank test
 a non-parametric statistical test equivalent of the
paired t test, analysing differences between
paired observations
 it makes no assumptions about the shapes of the
distribution of the two variables
 the absolute values of the differences between
the two variables are calculated as (+/-) for each
case and ranked from smallest to largest
 the test statistic is based on the sums of ranks
for negative and positive differences
Wilcoxon signed (+/-) rank test
 procedure
 the absolute values of the differences between the

paired observation are calculated (+/-) for each

case and ranked from smallest to largest
 exclude any differences which are zero, then rank

in order, ignoring signs (i.e. + or –)

 2 pairs having the same difference are given the

mean of what would have been their successive

ranks (i.e. 2nd & 3rd would have been ranked as
2.5 & 2.5)
 add up the ranks of positive differences and

negative differences separately

 each of the (+) & (–) ranks is totaled, and the

smaller referred to the Table of critical values for

Wilcoxon matched pairs signed rank test for P value
Wilcoxon rank sum test
 a non-parametric equivalent of the unpaired t test
or two-sample test
 procedure
 rank the observations from both groups

together in ascending order of magnitude


if any of the values are equal, average their
ranks
 add up the ranks in the group with the smaller

sample size
 compare this sum with the critical ranges in the

Table of critical ranges for the Wilcoxon rank

sum test
Mann-Whitney U
 a non-parametric equivalent of the unpaired t test
or two-sample test that two independent samples
come from the same population
 similar approach as Wilcoxon rank sum test with
entirely comparable results
Kruskal Wallis one way analysis of
variance
 a non-parametric equivalent of the one way
analysis of variance for normal distribution

vomiting score 5-
Postoperative
nausea and

4-
3-
2-
1- A B C
Groups
Binomial distribution

2006
Binomial distribution
 data (discrete variables) which can take only 0 or
1 response, such as treatment failure or
treatment success, follow the binomial
distribution providing the underlying population
response rate does not change
 proportion (p) of respondents (to treatment) in
sample
 p = R/n , R is the number of respondents and n

is the total number in the sample

 potential variation about this expectation is

expressed by SD(R)
Binomial distribution
 p can be used to estimate the population response
rate, π
 number of successful response in the population

can be estimated to be nπ
 standard error of p is SE(p) = √(pq/n) where q =
1 - p, is the proportion of unsuccessful responders
 95% CI for π is p + 1.96 × SE(p)
The Chi-squared test for
contingency tables

2006
Contingency table
 expected frequencies yes or no outcome
condition or
intervention A B
C (xC × yA)/n (xC × yB)/n xC
D (xD × yA)/n (xD × yB)/n xD
yA yB Total=n
Contingency table
 can be used to represent
 qualitative variables

 discrete quantitative variables

 continuous quantitative variables whose values

have been grouped

The Χ2 test
 is used to
 examine the association between discrete

(categorical or qualitative) variables

 test whether an observed frequency distribution

differs significantly from a postulated theoretical

one
 test whether there is an association or dependence

between the row variable and the column variable,

or
 test whether the distribution of individuals among

the categories of one variable is independent of

their distribution among the categories of the
other
The Χ2 test
 advantage
 it allows comparison of many more categories,

drawn-up into a contingency table

 the null hypothesis is that any number of

categories have equal chance of any other

factor
 when the table has only 2 rows or 2 columns,

this is equivalent to the comparison of

proportions
 applicable for
 a parallel group clinical trial, an unmatched case-

control study, or a cross-sectional survey

The Χ2 test in 2 × 2 tables
 first calculate the values expected in the 4 cells
of the table assuming the null hypothesis is true
 estimated cell value = row total x column total /

total N
 the Χ2 test is calculated from
, d.f.=21 for(O-E)
a 2 ×2 2 table
Χ =Σ E
 the Χ2 test is valid (Cochran 1954) when n is > 40,
regardless of the expected values, and if less than
20% of the expected values are less than 5 and
none are less than 1
 when n is between 20 and 40, the test is only valid
if all the expected values are at least 5
Yates’ correction for continuity
Factor A
Present Absent Total
Factor B Present a c m
Absent b d n
Total r s N
 Yates’ correction for continuity
 for converting the discrete data, as Χ2

distribution used to calculate the P value is

continuous, like the Normal distribution
Χ c2 = Σ
(|O-E|-½)2
, where || means take the
E value as positive
 for ease of calculation, this is equivalent to

N(|ad-bc|-½N)2
Χc =
2
mnrs
Yates’ correction for continuity
 the use of the continuity correction is always
advisable although it has most effect when the
expected numbers are small
 when the numbers are very small, the Χ2 test is not
a good enough approximation even with a
continuity correction and the Fisher’s exact test
for a 2× 2 table should be used
Fisher’s exact test for a 2 × 2 table
Factor A
Present Absent Total
Factor B Present a c m
Absent b d n
Total r s N
 Fisher’s exact test to determine P value
 to be used if any expected counts in a 2 × 2

table is less than 5, as the P value given by the

Χ2 test is not strictly valid
 the probability of observing the particular table

is m!n!r!s!
N!a!b!c!d!
Larger tables
 chi-squared test can also be applied to larger
tables, generally called r× c tables
 r denotes rows and c denotes columns

(O-E) 2
, d.f.= (r-1)× (c-1)
Χ2 = Σ E
 there is no continuity correction or exact test for
contingency tables larger than 2× 2
 chi-squared test should not be applied to tables
showing only proportions or percentages
Goodness of fit
 apart from using contingency tables, the Χ2-
statistic can be used to see if any observed set of
data follows a particular distribution
 e.g. by calculating the expected frequencies

from say a Poisson distribution, and then

comparing these with the observed data
 the degrees of freedom will be the number of
observations, n - 1
Paired comparison in contingency
tables
 McNemar’s test is a test on a 2x2 classification
table when the difference between paired
proportions is tested
 in cross-over trials and matched-pair case-

control studies
Correlation and linear
regression

2006
Correlation and linear regression
 techniques for dealing with the relationship between 2
or more continuous variables
 correlation
 examines linear association between 2 variables

 not whether one variable predicts another

variable
 strength of the association summarised by the

correlation coefficient
 regression
 examines dependence of one variable, the dependent

variable, on the other, the independent variable

 relationship summarised by a regression equation

consisting of a slope and an intercept

Correlation coefficient, r
 summarises the strength of the association
between 2 variables
 allows testing of the hypothesis that the
population correlation coefficient r is zero i.e.
whether an apparent association between the
variables would have arisen by chance
 Pearson correlation coefficient
 when the correlation coefficient is based on the

original observations
 Spearman rank correlation coefficient
 when it is calculated from the ranks of the data
Correlation coefficient, r
 dimensionless quantity ranging from -1 to +1
 a positive correlation is one in which both

variables increase together

 a negative correlation is one in which one

variable increases as the other decreases

 when variables are exactly linearly related, then

the correlation either equals +1 or -1

 unaffected by units of measurement
Different correlations
y r=1 y r = 0.2
• • • •
•• • • •
•• • • •
•• •
strong positivex weak positive x

y r=0 y r = -0.4
• •• • •
• • ••• • • •• ••
• • • •
• • •
•
x x
no correlation weak negative
Correlation coefficient, r
 should not be used
 if the relationship is non-linear

 in situations where one of the variables is

determined in advance
 should be used with caution
 in the presence of outliers

 when the variables are measured over more

than one distinct group e.g. disease and healthy

groups
y r>0 y r=0 y r<0
• • • •
• • •
• • • • ••
••• •••
••
x x x
Variance explained, r2
 squaring of the correlation coefficient and
multiplying by 100 gives r2 or variance explained,
the proportion of the variance of one variable
explained by the other
 e.g. r of 0.9, r2 = 0.81 x100; approximately 80%

of the variance of one variable can be

accounted/explained/predicted by the other
Test of significance
 Pearson correlation (coefficient, r)
 this is a parametric measure of the degree of

association between 2 numerical variables

Σ(x-x)(y-y)
r=
√{Σ(x-x)2 Σ(y-y)2}
 to test whether this is significantly different
from zero, calculate
 SE(r) = √{(1-r2)/(n-2)}
 and t = r/SE(r)
 and compare this with the Table of t-distribution
with n-2 degrees of freedom for P value
Test of significance
 assumes both variables are random samples and at
least one has a normal distribution
 outlying points away from the main body of the
data suggest the variable may not have a normal
distribution
 in this case it may be better to replace the

observation by their ranks and use the

Spearman rank correlation coefficient
Test of significance
 Spearman rank correlation (correlation, rs)
 this is a non-parametric measure of the degree
of association between 2 numerical variables
 the values of each variable are independently
ranked and the measure is based on the
differences between the pairs of rank of the 2
variables 6Σd 2
rs= 1-
n3-n
 where d is the difference between each pair of
ranks
Test of significance
 Spearman rank correlation
 this correlation will have a value between -1 and

1, and its interpretation is similar to that of the

Pearson correlation coefficient
 the significance of the association is tested by

comparing rs with the critical values of the

Spearman rank correlation coefficient table,
significant if rs is > critical value
Regression
 looking for dependence of one variable, the
dependent variable, on the other, the independent
variable
 relationship summarised by a regression equation
consisting of a slope and an intercept
 the slope represents the amount the dependent

variable increases with unit increase of the

independent variable
 the intercept represents the value of the

dependent variable when the independent variable

is zero
 multiple (multivariate) regression examines
simultaneous relationship between one dependent
variable and a number of independent variable
Linear regression
 assumption
 a change in one variable, x, will lead directly to a

change in another variable, y


e.g. haemoglobin increase with age, not the
other way around
 y variable (resultant change) is termed dependent
variable
 x variable is termed the independent variable
Regression line
 regression equation describes the relationship
between y and x
 y = a + bx

y = α + βx for population parameters

a is the intercept and b is the regression
coefficient Σ(x-x)(y-y)
 calculation of b b = Σ(x-x)2
 calculation of intercept, a = y – bx
 unwise to use equation to predict an outcome
based on extrapolation of observed parameters
Multiple regression
 model of multiple regression
 y = a + b x + b x + ……b x
1 1 2 2 k k
 applications
 to look for relationships between continuous

variables, allowing for a third (possibly

confounding) variable
e.g. predicted haemoglobin = 5.24 + 0.11(age) +
0.097(PCV)
 corresponding t-value is b/SE(b) with d.f. of n

minus the number of estimated parameters

 from these, P value is derived from Table of t

distribution
Multiple regression
 95% confidence intervals
 b ± t SE(b)
0.05
 if the interval includes zero, the conclusion
should be that that relationship between y and
x remains the same whether or not x changes
Degrees of freedom

2006
Degrees of freedom
 the number of degrees of freedom depends on 2
factors
 the number of groups we wish to compare

 the number of parameters we need to estimate

to calculate the standard deviation of the

contrast of interest
Degrees of freedom
 for Χ2 test for comparison of 2 proportions, 1 df
 for t test
 2 sets of degrees of freedom

 one degree of freedom between-groups

 one for within-groups

 for comparing 2 means, 1 df for between

groups, there are also dfs for estimating σ

 for paired data, df = number of subjects

minus 1
 for unpaired data, df = n + n minus 2, that is
1 2
n –1 for each group
Degrees of freedom
 for linear regression
 given n independent pairs of observations, 2

degrees of freedom are removed for the 2

parameters that have been estimated, thus d.f.
= n-2
 for multiple regression
 d.f. are n minus the number of estimated

parameters
Diagnostic tests and
implications

Predictions

2006
Sensitivity
 what is the probability that the test result will be
positive when a disease is present? that is when
the test result is positive, how accurate is it in
relation to overall presence of the disease?
 presence of acute myocardial infarction and

positive T-Troponin test

 restriction of atlanto-occipital joint extension

and difficult laryngoscopy

Specificity
 what is the probability that the test result will be
negative when a disease is absent? that is when
the test result is negative, how accurate is it in
relation to overall absent of the disease?
 absence of acute myocardial infarction and

negative T-Troponin test

 Mallampati class I and II and no difficulty in

laryngoscopy
Prediction model
Occurrence of event/disease
Prediction by test present (D+) absent (D-) Total
positive test (T+) a b a+b
negative test (T-) c d c+d
Total e f g
as fraction or percentage:
prevalence of disease = e/g or P(D+)
sensitivity of a test = a/e or P(T+ given D+)
specificity of a test = d/f or P(T- given D-)
false negative rate = 1- sensitivity = c/e (Type II
error)
false positive rate = 1- specificity = b/f (Type I error)
Sensitivity and specificity
 both are useful statistics because they will yield
consistent results for the diagnostic test in a
variety of patient groups with different disease
prevalences
 important point
 sensitivity and specificity are characteristics of

the test, not the population to which the test is

applied
Predictive value of a test
Occurrence of event/disease
Prediction by test present (D+) absent (D-) Total
positive test (T+) a b a+b
negative test (T-) c d c+d
Total e f g
predictive value of a positive test attempts to calculate the
probability of the patient having the disease (D+) when the
test is positive T+, or P(D+ given T+) = a/(a + b)
predictive value of a negative test attempts to calculate the
probability of the patient not having the disease (D-) when
the test is negative T-, or P(D- given T-) = d/(c + d)
discrimination - overall correct classification rate, how well
the model separates those in D+ and D- = (a+d)/g
false classification rate = (c+b)/g or = 1 - discrimination
Predictive value or sensitivity?
 predictive value is what the clinician wants
 positive test, is the disease present?

 negative test, is the disease absent?

 correct classification rate

 sensitivity is supplied with the statistical test

 disease present, was the test positive?

 no disease, was the test negative?

Risk and Odds and ratios

2006
Measure of Abbrevn Description No effect Total
effect success

Absolute risk ARR Absolute change in risk: the risk of an ARR=0% ARR=initial
reduction event in the control group minus the risk risk
of an event in the treated group; usually
expressed as a percentage
Relative risk RRR Proportion of the risk removed by RRR=0% RRR=100%
reduction treatment: the absolute risk reduction
divided by the initial risk in the control
group; usually expressed as a percentage
Relative risk RR The risk of an event in the treated group RR=1 or RR=0
divided by the risk of an event in the RR=100%
control group; usually expressed as a
decimal proportion, sometimes as a
percentage
Odds ratio OR Odds of an event in the treated group OR=1 OR=0
divided by the odds of an event in the
control group; usually expressed as a
decimal proportion
Number NNT Number of patients who need to be NNT=∝ NNT= 1/
needed to treat treated to prevent one event; this is the initial
reciprocal of the absolute risk reduction risk
(when expressed as a decimal fraction);
it is usually rounded to a whole number
Odds
 given that a subject has or does not have a
disease,
 odds always measures the incidence of event

(exposure) to non event (non-exposure)

+ve cases -ve cases
exposed a b
not exposed c d
total e f
 with the disease, odds of event are a/c;

 without the disease, odds of event are b/d

Odds ratio
 odds ratio compares the incidence of event (of
exposure) to non-event (non-exposure) among 2
groups of subjects with 2 opposing outcomes
 odds ratio is the ratio of two odds
+ve cases -ve cases
exposed a b
not exposed c d
total e f
 odds ratio, OR is (a/c)/(b/d)

= ad/bc
Odds Ratio 95% CI Interpretation
1 includes 1 No association
>1 does not Positive association between
include 1 exposure and outcome at the 5%
significance level (the odds of
exposure is greater in cases
Than in controls)
Negative association between
<1 does not
exposure and outcome at the 5%
include 1 significance level (the odds of
exposure is smaller in cases than
controls)
Association of exposure and
includes 1 outcome is not proven by the
study at the 5% significance
level
Risk
+ve cases –ve cases total
exposed a b a+b
not exposed c d c+d
total e f g
 risk of an event is the probability that an event
will occur within a stated period of time
 the risk of developing the disease within the
follow-up time is
 a/(a+b) for the exposed population

 c/(c+d) for the unexposed population

Relative risk
+ve cases –ve cases total
exposed a b a+b
not exposed c d c+d
total e f g
 the relative risk is a summary of the outcome of a
cohort study
 RR = (a/(a+b))/(c/(c+d))
or a(c+d)/c(a+b)
Relative risk 95% CI Interpretation
1 includes 1 No association
>1 does not Positive association between
exposure and outcome at the 5%
include 1 significance level (outcome is
more likely in the exposed
cohort)
Negative association between
<1 does not exposure and outcome at the 5%
significance level (outcome is
include 1 less likely in the exposed cohort)
Association of exposure and
outcome is not proven by the
includes 1 study at the 5% significance
level
How large was the treatment
effect?
PrOpofol - LignOcaine Trial
+ lignocaine control
# of patients randomized 100 100
# (%) patients with pain 15(15%) 20(20%)
absolute risk (pain) reduction 0.2-0.15 = 0.05
relative risk of having pain 0.15/0.20 = 0.75
relative risk reduction (1-0.75) x 100% = 25%
Depends on the sample size
 sample size of 100 each arm
 95% CI for RRR of 25% is -28.15 to 77.76

 lignocaine + propofol probably no benefit

 sample size of 1000 each arm

 95% CI for RRR of 25% is (8.35 to 41.61)

 confident that true RRR is close to 25%

http://ptwww.cchs.usyd.edu.au/Pedro/CIcalculato
r.xls.
Confidence interval around relative
risk reduction

n=1000/gp

RRR 0f 25%

n=100/gp
-38 9 41 59

-50 -25 0 25 50

Relative risk reduction (%)

Are the likely treatment benefits
worth the potential harm and cost?

Number needed to treat (NNT)

Number needed to harm (NNH)

2006
Number needed to treat …
the number of patients who must
receive an intervention of therapy
during a specific period of time
to prevent
one adverse outcome or
produce one positive outcome
Number needed to treat …
Risk of perioperative AMI NNT
without β-blocker with β-blocker* (1/ARR)
(ARR)
40 year old man 2% 1.8%
1/0.002
(0.2% or 0.002) = 500
70 year old man 40% 36% 1/0.04
(4% or 0.04) = 25
*assuming 10% relative risk reduction (RRR) with
preoperative administration of β-blocker
ARR = absolute risk reduction
NNT = number needed to treat
Number needed to treat …
 if there is a higher probability that a patient will
experience an adverse outcome if we do not treat,
 the more likely the patient will benefit from

treatment and
 the fewer such patients we need to treat to

prevent one adverse outcome

Number to treat
ibuprofen control
number of patients 50 50
at least 50% pain relief over 6 hours 27 10
expressed in percentage 54% 20%

absolute risk reduction = 54%-20% (or 0.54-0.20)

= 34% (or 0.34)
number needed to treat = 1/0.34
= 2.94
or 3
Number to treat
 another way of calculating NNT for the analgesic
trial
NNT = 1/(the proportion of patients with at
least 50% pain relief with analgesic
minus the proportion of patients with
at least 50% pain relief with placebo)
= 1/((27/50) - (10/50))
= 1/(0.54 - 0.20)
= 1/0.34
= 2.9
or 3
Number to treat
 the best NNT would be 1,
 every patient with treatment benefited

 no patient given control benefited

 generally NNTs between 2 and 5 are indicative of

effective treatments, but NNTs of 20, 50 or 100
may be useful for prophylactic treatments, like
interventions to reduce death after heart attack
 relevance of which depends on the intervention
and the consequences
Number to harm
 for adverse effects, the number needed to harm
(NNH) can be calculated in exactly the same way
as an NNT
 for an NNH, large numbers are obviously better
than small numbers, because that means that the
adverse effect occurs with less frequency