P8130: Biosta.s.
cal Methods I
Methods of Inference: One-Sample Propor7ons
Instructor: Vahe Khachadourian, MD, MPH, PhD
Methods of Inference: One Sample Propor6ons
This module focuses on inferen0al methods concerning one sample
propor0ons.
You will learn how to:
• Construct a confidence interval to es;mate a popula;on propor;on.
• Conduct a hypothesis test using the normal approxima;on for a single
propor;on.
• Dis;nguish between approximate and exact methods of inference.
Estimation of Binomial Distribution
Motivation: Suppose we are conducting a ‘yes/no’ survey of a random sample. The overall
goal is to determine the proportion of people who responded ‘yes’ and draw conclusions
for the overall population.
Let 𝑛 be the total number of observations, independent and identically distributed (i.i.d.).
Let 𝑋! , 𝑖 = 1, 2, … , 𝑛, be a random variable that takes the following values:
1, 𝑤𝑖𝑡ℎ 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑝 (𝑖𝑓 𝑡ℎ𝑒 𝑎𝑛𝑠𝑤𝑒𝑟 𝑤𝑎𝑠 ′𝑦𝑒𝑠′)
𝑋! = )
0, 𝑤𝑖𝑡ℎ 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 1 − 𝑝 (𝑖𝑓 𝑡ℎ𝑒 𝑎𝑛𝑠𝑤𝑒𝑟 𝑤𝑎𝑠 ′𝑛𝑜′)
It follows that each 𝑋! ~𝐵𝑒𝑟𝑛𝑜𝑢𝑙𝑙𝑖(1, 𝑝), and that ∑$!"# 𝑋! = 𝑋 ~ 𝐵𝑖𝑛𝑜𝑚𝑖𝑎𝑙 (𝑛, 𝑝).
Es1ma1on of Binomial Distribu1on
∑$!"# 𝑋! = 𝑋 ~ 𝐵𝑖𝑛𝑜𝑚𝑖𝑎𝑙 (𝑛, 𝑝) quantifies the number of successes, i.e., number of ‘yes’ answers
in n samples.
Goal: To estimate the population proportion p, considered fixed, but unknown.
The estimate of p is called 𝑝̂ , the sample proportion, and is given by:
∑$!"# 𝑋! 𝑋
𝑝̂ = =
𝑛 𝑛
Because 𝑋 ~ 𝐵𝑖𝑛𝑜𝑚𝑖𝑎𝑙 𝑛, 𝑝 , we can compute the following:
# #
𝐸 𝑝̂ = 𝐸 𝑋 = 𝑛𝑝 = 𝑝; thus 𝑝̂ is an unbiased estimator for 𝑝.
$ $
1 1 𝑝(1 − 𝑝)
𝑣𝑎𝑟 𝑝̂ = % 𝑣𝑎𝑟 𝑋 = % 𝑛𝑝 1 − 𝑝 =
𝑛 𝑛 𝑛
Normal Approxima1on of Binomial
We showed that 𝑝̂ can be represented as an average of Bernoulli trials, each with mean
𝑝 and variance 𝑝(1 − 𝑝).
Based on the Central-Limit Theorem we can see that the sample proportion 𝑝:̂
%(#'%)
C
𝑝̂ = 𝑋~𝑁(𝑝, $ ),
Under normal approximation.
Notes:
• If 𝑝 is close to 0.5, the Binomial distribution will look almost Normal for only 𝑛 = 10
• If 𝑝 is close to 0.1 or 0.9, the Binomial distribution will look almost Normal for 𝑛 = 50
• If 𝑝 is much closer to 0 or 1, then a normal approximation might not work well.
Normal Approxima1on of Binomial
Interval Es1ma1on: One-Sample Propor1on
!(#$!)
Under normal approximation: 𝑝~𝑁(𝑝,
̂ ).
&
Consider then a standardization of 𝑝:̂
𝑝̂ − 𝑝
~𝑁(0,1)
𝑝(1 − 𝑝)/𝑛
We can derive the 95% confidence interval for a population proportion:
𝑝̂ − 𝑝
𝑃 −1.96 < 𝑧 < 1.96 = 0.95 → 𝑃 −1.96 < < 1.96 = 0.95
𝑝(1 − 𝑝)/𝑛
Interval Estimation: One-Sample Proportion
Problem: The parameter of interest 𝑝 is present in the formula that provides the
precision of our es>mate.
Solu*on: Subs>tute 𝑝 with 𝑝̂ in the denominator to get the 95% CI:
!' #$!' !' #$!'
𝑃 𝑝̂ − 1.96 &
≤ 𝑝 ≤ 𝑝̂ + 1.96 &
= 0.95
In general, a 100 1 − 𝛼 % confidence interval for one popula>on propor>on is
given by:
<; 67<; <; 67<;
𝑝̂ − 𝑧678/: , 𝑝̂ + 𝑧67>/:
= =
Exact Interval Estimation: One-Sample Proportion
What if the normal approximaPon is not valid?
In this case, a small-sample method needs to be applied.
Clopper–Pearson Method for obtaining an EXACT 100 1 − 𝛼 % confidence interval for
the binomial parameter 𝑝 is given by (𝑝) , 𝑝* ), where:
$
𝛼 𝑛 +
𝑃 𝑋 ≥ 𝑥 𝑝 = 𝑝) =1− = K 𝑝) (1 − 𝑝) )$'+
2 𝑘
+",
,
𝛼 𝑛 +
𝑃 𝑋 ≤ 𝑥 𝑝 = 𝑝* = =K 𝑝* (1 − 𝑝* )$'+
2 𝑘
+"-
• The limits of this Exact CI can be found in Tables or computed with soXware.
One-Sample Test for Binomial Proportion
Tests for One-Popula;on Propor;on, Normal Theory Methods
𝐻( : 𝑝 = 𝑝( vs 𝐻# : 𝑝 ≠ 𝑝(
With significance level α pre-specified,
compute the test sta>s>c:
"! $!&
𝑧= ~N(0,1), under 𝐻(
!&(#$!&)/&
Critical Region Critical Region
Reject 𝐻( : if 𝑧 > 𝑧#$*/+
Fail to reject 𝐻( : if |𝑧| ≤ 𝑧#$*/+
P-value: 2×[1 − Φ 𝑧,-.-, ]
One-Sample Exact Test for Binomial Propor6on
If the assumption of normal approximation is not valid, then we use exact probabilities
to calculate the p-value.
If 𝑝̂ ≤ 𝑝-, the p-value = 2×𝑃 ≤ 𝑥 𝑠𝑢𝑐𝑐𝑒𝑠𝑠𝑒𝑠 𝑖𝑛 𝑛 𝑡𝑟𝑖𝑎𝑙𝑠 𝐻- 𝑖𝑠 𝑡𝑟𝑢𝑒
, 𝑛 +
∑
= +"- 𝑝- (1 − 𝑝-)$'+
𝑘
If 𝑝̂ > 𝑝-, the p-value = 2×𝑃 ≥ 𝑥 𝑠𝑢𝑐𝑐𝑒𝑠𝑠𝑒𝑠 𝑖𝑛 𝑛 𝑡𝑟𝑖𝑎𝑙𝑠 𝐻- 𝑖𝑠 𝑡𝑟𝑢𝑒
$ 𝑛 +
∑
= +", 𝑝- (1 − 𝑝-)$'+
𝑘
One-Sample Test for Binomial Propor6on
Tests for One-Popula;on Propor;on, Normal Theory Methods with
Con;nuity Correc;on
𝐻( : 𝑝 = 𝑝( vs 𝐻# : 𝑝 ≠ 𝑝(
With significance level α pre-specified, compute the test statistic:
'
' &$
!$!
()
𝑧= ~N(0,1), under 𝐻(
!&(#$!&)/&
Reject 𝐻( : if 𝑧 > 𝑧#$*/+ ; Fail to reject 𝐻( : if |𝑧| ≤ 𝑧#$*/+
P-value: 2×[1 − Φ 𝑧,-.-, ]
One-Sample Test for Binomial Propor6on
Confidence Interval for One-Population Proportion, Normal Theory
Methods with Continuity Correction
100 1 − 𝛼 % confidence interval for one population proportion is given by:
<; 67<; 6 <; 67<; 6
𝑝̂ − 𝑧67. − , 𝑝̂ + 𝑧67. +
/ = := / = :=
One-Sample Binomial Propor1on
Example: In a survey of 300 randomly selected drivers, 125 claimed that they regularly
wear seat belts.
Can we conclude from these data that half of the population wears seat belts?
Perform a hypothesis test and construct a 95% confidence interval for the population
proportion.
One-Sample Binomial Propor1on
Example Seat Belts: Hypothesis Test
One-Sample Binomial Propor1on
Example Seat Belts: 95% CI
Readings
Rosner, Fundamentals of Biostatistics 8th Edition:
• Chapter 6: Section 6.8
• Chapter 7: Section 7.10
• Chapter 10: Section 10.2