Lesson 6 of 7 · 18 min

The normal approximation for p̂

Recognise that p̂ is approximately normal for large samples, check when the approximation is reasonable, and use it with technology to calculate probabilities and quantiles for sample proportions.

What you will be able to do

  • Write the approximate distribution p^≈N ⁣(p,p(1−p)n)\hat{p} \approx N\!\left(p, \frac{p(1-p)}{n}\right) for a large sample.
  • Check that npnp and n(1−p)n(1-p) are large enough for the approximation to be reasonable.
  • Calculate probabilities and quantiles for p^\hat{p} with technology, and interpret them in context.

Working out P(p^≤0.28)P(\hat{p} \le 0.28) for a sample of 500 from the binomial distribution means adding up 141 binomial probabilities. There is a much quicker way. When you simulated p^\hat{p} for large samples, the display was a familiar shape: a symmetric bell.

Interactive

Watch the bell appear

Run a few thousand samples. How well does the normal curve fit? Now set n=20n = 20 and p=0.05p = 0.05: what goes wrong?

Key idea

The normal approximation

For large random samples, the sample proportion is approximately normally distributed:

p^∼approxN ⁣(p,p(1−p)n),\hat{p} \overset{\text{approx}}{\sim} N\!\left(p, \frac{p(1-p)}{n}\right),

with mean pp and standard deviation p(1−p)n\sqrt{\frac{p(1-p)}{n}}.

The approximation is better for larger nn, and better when pp is near 0.5. A common rule of thumb is that it is reasonable when

np≥10andn(1−p)≥10,np \ge 10 \quad\text{and}\quad n(1-p) \ge 10,

that is, when you expect at least 10 members with the attribute and 10 without. (Some texts use 5; either way, the bigger these are, the better.)

Random samples of 525 are taken from a population with p=0.3p = 0.3. Which is the approximate distribution of p^\hat{p}?

Choose one answer

p^≈N(0.3,0.022)\hat{p} \approx N(0.3, 0.02^2). The shaded area is P(0.27<p^<0.34)P(0.27 < \hat{p} < 0.34), found in the worked example below.

Worked example

A normal probability for p̂

  • Complex familiar
  • Technology active

Suppose 30 % of Queensland households have rooftop solar panels. A random sample of 525 households is surveyed.

(a) Explain why the distribution of p^\hat{p} is approximately normal, and state its mean and standard deviation.

(b) Find the probability that the sample proportion with solar panels is between 0.27 and 0.34, correct to three decimal places.

(c) Interpret your answer in terms of the number of households in the sample.

5 steps

Try it yourself first, then check the working one step at a time.

Method

  1. Check np≥10np \ge 10 and n(1−p)≥10n(1-p) \ge 10.

  2. Write p^≈N ⁣(p,p(1−p)n)\hat{p} \approx N\!\left(p, \frac{p(1-p)}{n}\right) with the numbers in.

  3. Use normal CD with μ=p\mu = p and σ=p(1−p)/n\sigma = \sqrt{p(1-p)/n} (or standardise first).

  4. Round at the end, and interpret in context.

Watch out for

  • Entering the variance, 0.0004, as σ\sigma.

  • Using the counts 141.75 and 178.5 with the proportion's mean and standard deviation.

Directly on the calculator

Enter the mean and standard deviation of p^\hat{p} and the bounds as proportions.

  • Casio: Normal CD with Lower 0.270.27, Upper 0.340.34, σ=0.02\sigma = 0.02, μ=0.3\mu = 0.3.
  • TI: normalcdf(0.27, 0.34, 0.3, 0.02).

Use a very large bound such as 109910^{99} (or −1099-10^{99}) for a one-sided probability.

By standardising first

Convert the bounds to zz-scores with z=p^−pSD(p^)z = \frac{\hat{p} - p}{\mathrm{SD}(\hat{p})}:

z1=0.27−0.30.02=−1.5,z2=0.34−0.30.02=2.z_1 = \frac{0.27 - 0.3}{0.02} = -1.5, \quad z_2 = \frac{0.34 - 0.3}{0.02} = 2.

Then find P(−1.5<Z<2)P(-1.5 < Z < 2) for the standard normal ZZ. Both methods give the same answer.

Suppose 20 % of Australian adults have never flown on a plane. For random samples of 400 adults, use the normal approximation to find P(p^>0.24)P(\hat{p} > 0.24), correct to four decimal places.

Round to 4 decimal places.

Common mistake

Variance in the σ box

Calculators ask for the standard deviation σ\sigma, not the variance. Entering p(1−p)n\frac{p(1-p)}{n} instead of its square root makes the distribution far too narrow, and tail probabilities come out almost 0. If an answer looks absurdly small, check this first.

Worked example

Is the sample consistent with the claim?

  • Complex unfamiliar
  • Technology active

An electricity retailer advertises that "80 % of our customers are satisfied with our service". A consumer group surveys a random sample of 400 of the retailer's customers, and 300 say they are satisfied.

Assuming the retailer's claim is true, find the probability of a sample proportion as low as or lower than the one observed. Use your answer to comment on the claim.

5 steps

Try it yourself first, then check the working one step at a time.

Method

  1. Assume the claimed pp and model p^\hat{p} with the normal approximation.

  2. Find the probability of a result at least as extreme as the one observed.

  3. A very small probability (say below 0.05) means the observation is unusual under the claim, which casts doubt on it.

Watch out for

  • Finding P(p^=0.75)P(\hat{p} = 0.75), which is 0 for a continuous model. Use a tail probability.

  • Concluding the claim is definitely false. A small probability is evidence, not proof.

Exam tip

Show the model

On Paper 2 a bare calculator answer risks the method marks. Write the model and its parameters — "p^≈N(0.8,0.022)\hat{p} \approx N(0.8, 0.02^2), since np=320np = 320 and n(1−p)=80n(1-p) = 80 are large" — then the probability statement, then the answer, rounded as asked.

Note

A note on accuracy

p^\hat{p} is really discrete (it can only equal kn\frac{k}{n}), so the normal model is an approximation. For the large samples where you use it, the error is small. You are not expected to use a continuity correction in this course.

Summary

  • For large samples, p^≈N ⁣(p,p(1−p)n)\hat{p} \approx N\!\left(p, \frac{p(1-p)}{n}\right).

  • Rule of thumb: the approximation is reasonable when npnp and n(1−p)n(1-p) are both at least 10 (some texts say 5).

  • Calculate probabilities with technology using mean pp and standard deviation p(1−p)/n\sqrt{p(1-p)/n} — or standardise and use Z∼N(0,1)Z \sim N(0, 1).

  • A probability that is very small means the observed p^\hat{p} would be unusual if the claimed pp were true.

Finished the lesson?

Try the 2 quick checks above and the lesson completes itself, or mark it complete now.

Lesson 6 of 70% read

Save your progress with a free account