Confidence Intervals
.
We will now return to the idea of using one sample from the population to find an estimate for p (the population proportion)
.
Sometimes a single sample proportion $\big(\hat{p} \big)$ is called a point estimate for p.
.
From our work in the previous section (Distribution of sample proportions), we know that we cannot expect the sample proportion $\big(\hat{p} \big)$ of any one sample to exactly match the population proportion $\big( p \big)$
.
Our goal in this section, is to produce an interval either side of $\hat{p}$, so that we can say that the probability p lies within that interval is a given percentage.
.
This interval either side of $\hat{p}$ is called a Confidence Interval.
.
A Confidence Interval of 95% is very common.
ie, There is a probability of 95% that p lies in the interval:
… …$\Big( \hat{p} - c, \; \hat{p} + c \Big)$ … … for some value of c
.
or
.
… … $\text{Pr} \Big( \hat{p} - c < p < \hat{p} + c \Big) = 0.95$
.
We now assume that $\hat{p}$ is normally distributed with a mean of p
This is a reasonable assumption if the sample size is sufficiently large.
.
For the confidence interval to have a probability of 95%, it will need to be the same width as the width of the 95% area of the normal distribution curve.
.
Referring back to the Standard Normal Distribution (Z), we can use the Inverse Normal Distribution to find a such that:
… … $\text{Pr} \big(-a < z < a \big) = 0.95$
.
… …enter: InvNormCDF, $tail = center, \; p = 0.95,\; \sigma = 1, \; \mu = 0$
… … This gives a result of $a = \pm 1.96$
.
Note: If your calculator can only do left tail calculations, then you will need to find b such that:
… … $\text{Pr} \big( z < b \big) = \dfrac{1-0.95}{2} = 0.025$
… … This should also give a result of $b = -1.96$
.
Therefore,
… … $\text{Pr} \big(-1.96 < z < 1.96 \big) = 0.95$
.
Now we can use the rule to convert z into $\hat{p}$ given that $\mu = p$
.
… … $z = \dfrac{x - \mu}{\sigma} \qquad \Longrightarrow \qquad z = \dfrac{\hat{p} - p}{\sigma}$
.
… … $\text{Pr} \Big(-1.96 < \dfrac{\hat{p} - p}{\sigma} < 1.96 \Big) = 0.95$
.
… …$\text{Pr} \Big(-1.96\sigma < \hat{p} - p < 1.96\sigma \Big) = 0.95$
.
… … $\text{Pr} \Big(-\hat{p} - 1.96\sigma < - p < -\hat{p} + 1.96\sigma \Big) = 0.95$
.
… … $\text{Pr} \Big(\hat{p} + 1.96\sigma > p > \hat{p} - 1.96\sigma \Big) = 0.95$
.
Or:
… … $\text{Pr} \Big(\hat{p} - 1.96\sigma < p < \hat{p} + 1.96\sigma \Big) = 0.95$
.
We know that the theoretical standard deviation is given by: $\sigma = \sqrt{ \dfrac{p(1-p)}{n} }$
.
But in a practical situation, we don't know the actual value of p
.
So we use $\hat{p}$ as an approximate value for p and we get:
.
… … $\sigma \approx \sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n} }$
.
So, the 95% confidence interval is given by:
.
… … $\text{Pr} \left( \hat{p} - 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} < p < \hat{p} + 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} \right) \approx 0.95$
.
or
.
Th 95% Confidence Interval is:
.
… … $\left( \hat{p} - 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} , \; \hat{p} + 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} \right)$
.
Note: the constant 1.96 came from the z-score where the central region was 95%
If we need to find a different confidence interval, we simply replace 1.96 with the new z-score.
.
Example 1
The lengths of stay in hospital among patients is of interest to health planners. A random sample of 100 patients was investigated, and 20 were found to have stayed longer than 7 days.
… a) .. Find a point estimate for p, the proportion of patients who stay in hospital longer than 7 days.
… b) .. Calculate a 95% confidence interval for p.
Solution
… a) .. Find a point estimate for p, the proportion of patients who stay in hospital longer than 7 days.
… … … $\hat{p} = \dfrac{\text{num successes}}{\text{num in sample (n)}}$
.
… … … $\hat{p} = \dfrac{20}{100}$
.
… … … $\hat{p} = 0.2$
.
… b) .. Calculate a 95% confidence interval for p.
.
… … … $\text{Pr} \left( \hat{p} - 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} < p < \hat{p} + 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} \right) \approx 0.95$
.
… … … $\text{Pr} \left( 0.2 - 1.96\sqrt{ \dfrac{ 0.2 \big( 1 - 0.2 \big) }{100}} < p < 0.2 + 1.96\sqrt{ \dfrac{ 0.2 \big( 1 - 0.2 \big) }{100}} \right) \approx 0.95$
.
… … … $\text{Pr} \Big( 0.2 - 0.0784 < p < 0.2 + 0.0784 \Big) \approx 0.95$
.
… … … $\text{Pr} \Big(0.1216 < p < 0.2784 \Big) \approx 0.95$
.
Using the CAS to calculate Confidence Intervals
.
Fortunately, the Classpad can do these calculations for us.
- Go to the Statistics Section, then the Calc menu and select Interval
- Select One-Prop Z Int and tap Next
- For C-Level (Confidence Level) Enter: 0.95
- For X (num of successes in sample) Enter: 20
- For n (size of sample) Enter 100
- Tap Next
The calculator should return:
So the 95% Confidence Interval is: $\Big( 0.1216, \; 0.2784\Big)$
- We are 95% sure that the value of p is in this range
- Where p is the proportion of all patients at the hospital who stay longer than 7 days.
.
Changing the Level of Confidence
.
In general, the C% confidence interval will be given by:
… … $\left( \hat{p} - k\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} , \; \hat{p} + k\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} \right)$
.
Where k is found such that:
… … $Pr \Big( -k < z < k \Big) = \dfrac{C}{100}$
Example 2:
Find the general form of the 68% Confidence Interval
Solution
… … Find the general form of the 68% Confidence Interval
.
… … … Use CAS (inverse normalcdf) to:
.
… … … find k such that $Pr \Big( -k < z < k \Big) = 0.68$
.
… … … returns $k = 0.9945$
.
… … … Hence 68% Confidence Interval is
.
… … $\left( \hat{p} - 0.9945\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} , \; \hat{p} + 0.9945\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} \right)$
.
Example 3
Grow Well are 95% sure that 30% to 40% of shoppers prefer their mulch. What sample size was needed for this level of confidence?
Solution
… … In Notation we have: $Pr\big( 0.3 < p < 0.4\big) = 0.95$
.
… … clearly $\hat{p} = \dfrac{0.3+0.4}{2} = 0.35$
.
… … Compare this information with:
… … $\left( \hat{p} - 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} , \; \hat{p} + 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} \right)$
.
… … And we get that
.
… … … … $\hat{p} - 1.96\sqrt{ \dfrac{ \hat{p} \big( 1 - \hat{p} \big) }{n}} = 0.3$
.
… … … … $0.35 - 1.96\sqrt{ \dfrac{ 0.35 \big( 1 - 0.35 \big) }{n}} = 0.3$
.
… … … … which can be solved for n
.
… … … … $n = 349.59$ … … {round off because n is an integer}
.
… … … … $n = 350$ … … {required sample size}
.