Choosing and critiquing statistical models

Choosing and critiquing statistical models

Spec N3 · AQA A-Level Maths: Revision notes

Studying this course? Keep your progress in one place.

Key facts

  • A binomial model needs a fixed nn, two outcomes, independent trials and a constant pp.
  • Name the condition that fails and explain it in the context of the question.
  • A Normal model suits a continuous, symmetrical, bell-shaped variable.
  • Test a model by comparing its expected number n×P(event)n\times P(\text{event}) with the observed.
  • A good answer: choose the model, check its conditions, comment on any that fail, and state the effect.

Binomial conditions

Only when nn is fixed, each trial has two outcomes, trials are independent and pp is constant.

A binomial model X∼B(n,p)X\sim B(n,p) counts successes in nn trials, with P(X=r)=(nr)pr(1−p)n−rP(X=r)=\binom nr p^r(1-p)^{n-r}, mean npnp and variance np(1−p)np(1-p).

It is only appropriate when all four conditions hold. In an answer, always say which condition holds or fails, and apply it to the context given.

00.050.10.150.20.250.0090.001r = 0r = 1r = 2r = 3r = 4r = 5r = 6r = 7r = 8r = 9r = 10rProbability
P(X = r) for X ~ B(10, 0.3): a fixed number of independent trials with constant probability.

Fixed trials

  • nn decided in advance

Two outcomes

  • Success or failure

Independent

  • One result does not change another

Constant probability

  • Same pp every trial

A coin is flipped until the first head appears, and the number of flips is recorded. Which binomial condition fails?

When binomial fails

Look in the context for a reason why each condition could fail.

Selecting from a small population without replacement changes the proportions left, so independence fails. Sampling without replacement from a large population changes the probabilities very little, so the binomial model is then a good approximation.

Give two parts in an explanation: say which condition fails, then say why in terms of the context.

00.10.20.30.40.0040.01k = 0k = 1k = 2k = 3k = 4k = 5Number of redsProbability
  • Without replacement (8 red in 20)
  • Binomial B(5, 0.4)
Drawing 5 from 20 with 8 red, without replacement, differs from B(5, 0.4): independence fails in a small population.

Not independent

  • Family or friends who behave alike
  • Customers in a queue
  • Small population, no replacement

Probability not constant

  • Pupils at different distances
  • Busy and quiet periods
  • People of different ability

Other failures

  • Number of trials not fixed
  • More than two outcomes

Worked example

A tank holds 5050 fish, 1010 of them undersized. Two are taken without replacement. Compare P(neither undersized)P(\text{neither undersized}) with the binomial value.

Pupils in one class are asked whether they like maths. Why might a binomial model for the number who say yes be inappropriate?

When Normal fits

When the quantity is continuous, single-peaked and roughly symmetrical and bell-shaped.

The Normal model suits a continuous quantity. Useful checks:

  • the mean and median are close; a mean much greater than the median suggests positive skew;
  • almost all values lie within about 33 standard deviations of the mean;
  • the model should not give noticeable probability to impossible values, such as negative times or masses;
  • a histogram is roughly bell-shaped (symmetrical).
152025303540450.020.040.060.080.1xymean 31N(31, 4²)
Normal curve N(31, 4²): symmetrical, with the tail beyond 38 shaded.

A data set has mean 5252 and median 4141. What does this suggest about a Normal model?

Testing a model

Use the model to predict a frequency and compare it with what was observed.

The expected number is n×P(event)n\times P(\text{event}). If the observed value is close, the model agrees with the data; if it is very different, the model is weak in that region. For example, a Normal model that predicts 9%9\% negative incomes for a quantity that cannot be negative is weak in the lower tail.

State the prediction, the observed value, and what the difference tells you about the model.

0510152025PredictedObservedNumber of fish
Fish over 38 cm in a sample of 200: the Normal model predicts about 8, but 27 were observed.

Worked example

The model L∼N(31,42)L\sim N(31,4^2) is used for fish lengths in cm. In a sample of 200200 fish, 2727 are longer than 3838 cm. Does the model fit?

A model predicts 1212 successes in a sample and 1111 are observed. What do you conclude?

Choosing the right model

Counts of successes suggest binomial; continuous measurements suggest Normal.

A count of successes in a fixed number of independent trials with the same probability: binomial. A continuous measurement (time, mass, length) with a symmetrical bell-shaped pattern: Normal. A binomial with large nn and pp not close to 00 or 11 can be approximated by N(np,np(1−p))N(np,np(1-p)).

A good answer gives the model, checks its conditions against the context, comments on any that fail, and says what effect this has on the results. A model is never perfect: say whether it is a reasonable approximation, and what it ignores.

  1. 1

    Choose

    binomial for counts, Normal for continuous data

  2. 2

    Check conditions

    against the context given

  3. 3

    Comment

    say which conditions fail and why

  4. 4

    Effect

    state how the results are affected

Critiquing a model

Which model suits the masses of newborn babies?

Try an exam question

A farmer models the mass of an egg as N(60,42)N(60,4^2) grams. In a sample of 500500 eggs, 3030 weigh more than 6868 g. Find the number of eggs the model predicts to weigh more than 6868 g, and comment on the model.

[4 marks]

That's the notes covered.

Carry on to the next subtopic.

Try a question

Which of the following is an assumption needed for this binomial model?

1 mark, marked against the mark scheme

Revision notes