Vous êtes sur la page 1sur 17

DETERMINATION OF SAMPLE SIZE

RN Mishra

Professor
Division of Biostatistics

INSTITUTE OF MEDICAL SCIENCES


BANARAS HINDU UNIVERSITY
VARANASI-221005
SITUATIONS IN MEDICAL SCIENCES TO
DETERMINE SAMPLE SIZES
 Problem of estimation
 Proportion (P)
 Mean (M)
 Odds ratio (OR)
 Relative risk (RR)
 Incidence rate (I)
 Problem of comparison (Testing of difference)
 Two proportions
 Two means
 Two incidence rates
 Odds ratio against 1 (Independent, matched cases &
controls)
 Relative risk against 1 (Independent, matched cohorts)
 Non-inferiority and superiority
GENERAL PRINCIPLE

 Problem of estimation: (Proportion = P & Size


of the population is very small)
 n = (Z/2)2 P(1-P)/[d2+ (Z/2)2 P(1-P)/N]
 Where, N is the population size; d is the
anticipated difference considered to estimate
population proportion

 Z/2 at 95% confidence is the ordinate value and


is taken equal to 1.96

 In case the value of P is not available, 0.50


preferred
GENERAL PRINCIPLE

 Problem of estimation: (Proportion = P & Size


of the population is very large)
 n = (Z/2)2 P(1-P)/d2
 Where, d is the anticipated difference
considered to estimate population proportion

 Z/2 at 95% confidence is the ordinate value and


is taken equal to 1.96

 In case the value of P is not available, 0.50


preferred
SAMPLE SIZE DETERMINATION- EXAMPLE
 Problem of estimation: (Proportion = P)
 Example: Let the past literature or a pilot study
provide the value of p=30%. Then to estimate the
population proportion with 95% confidence
having a difference of 2 i.e. researcher is 95%
confident that estimate will lie between 28-32, the
sample size needed would be:

 N = (Z/2)2 P(1-P)/d2
= (1.96)2 30x70/(2)2
= 2017
GENERAL PRINCIPLE

 Problem of estimation: (Proportion = P)


N = (Z/2)2 P(1-P)/ε2
Where, P is the population proportion and ε is the
anticipated relative precision in the estimate.

ε = P x Anticipated relative precision

May be 1% or 5% or 10% but not more than 10%


(However, decided based on the problem)

In case the value of P is not available, 0.50 preferred


SAMPLE SIZE DETERMINATION-EXAMPLE
 Problem of estimation: (Proportion = P)
 Example: Let the past literature or a pilot study
provide the value of p=30%. Then to estimate the
population proportion with 95% confidence
having 90% precision i.e. estimate lies between
27.0-33.0, the sample size needed would be:
ε = 30x10/100 = 3.0
 N = (Z/2)2 P(1-P)/ ε2
= (1.96)2 30x70/(3.0)2
= 897
GENERAL PRINCIPLE
 Problem of estimation: (Mean = M and Population size
is very small)

 N = (Z/2)2 S2/[d2 + (Z/2)2 S2/N]

Where, S is the population SD; N population size


and d is the anticipated difference in the estimate
of the population mean.
SAMPLE SIZE DETERMINATION-EXAMPLE
 Problem of estimation: (Mean = M and Population size
is very large)
 Example: Let the past literature or a pilot study
provide the variability S2 as 4.5 with average as
12.5. Then to estimate the population mean with
95% confidence limit between 11.5 to 13.5, the
sample size needed would be:
N = (Z/2)2 S2/d2
N = (1.96)2 (4.5)2/(1)2
N = 78

 Similarly to have the estimate between 10.5 to 14.5, the


sample size needed would be 20
Problem of comparison

 Two proportions
 Two means

 Two incidence rates

 Odds ratio against 1 (Independent, matched

cases & controls)


 Relative risk against 1 (Independent,
matched cohorts)
Problem of comparison
(Two population proportions)
 = [(0Z) + 1Z1-]
 The combined SE of the estimate is:
(p1q1/n1 + p2q2/n2)
When H0 is true i.e. P1 = P2
or both the populations are same
The best estimate of proportion would be
p = (p1 + p2)/2
Researchers should not be restricted to equal
number of subjects in each group. Let n2 = cn1 i.e.
researcher is interested to a sample size from the
second population which is c times of sample size
from the first population.
Problem of comparison
(Two population proportions)
The combined SE of the estimate under H0 is:
0 = (pq/n + pq/cn) = pq (1+1/c)/n
When c = 1, then 0 would be:
0 = 2pq/n
The combined SE of the estimate under H1 is:
1 = (p1q1/n + p2q2/cn) = (p1q1 + p2q2/c)/n
When c = 1, then 1 would be:
(p1q1 + p2q2)/n
Problem of comparison
(Two population proportions)
Substituting the values of 0 and 1 in general
equation, we get
 = [(0Z) + 1Z1-]
 = [Z pq(1+1/c)/n + Z1- (p1q1 + p2q2/c)/n]
Hence,
n = [Z pq (1+1/c) + Z1- (p1q1 + p2q2/c)]2/2
When equal sample size decided i.e. c=1, then
n = [Z 2p*q* + Z1- (p1q1 + p2q2)]2/2
Problem of comparison
(Two population proportions)
n = [Z 2pq + Z1- (p1q1 + p2q2)]2/ 2
The value of Z is substituted for alternative
hypothesis H1 as P1 > P2 or P1 < P2,
But if alternative hypothesis is P1 ≠ P2, value
of Z/2 instead of Z is taken.
For example at =5%,
Z =1.64 for alternative P1 > P2 or P1 < P2
Z/2 =1.96 for alternative P1  P2
Problem of comparison
(Two Means)
The test are carried under the assumption
of common variance of the two population
So one has to obtain first the common
variance or SD based on small samples of
the two populations which are estimates of
the population variance, if not known for
the population
Let S1 and S2 are the SD of two populations
based on sizes of n1 and n2
Problem of comparison
(Two Means)
The Common SD will be then
S = [(n1 -1) S12 + (n2 -1)S22 ]/(n1+n2-2)

If one has to take unequal samples in both


the groups, i.e. n samples from one
population and cn from other population;
then

n = S2 (1+1/k) [Z+ Z1- ]2 /2


If equal sample sizes are decided, put c=1
THANK YOU ?

Vous aimerez peut-être aussi