Sampling
Where sampling is used a random sample is preferable whenever possible so as to minimise the risk of bias. This means that each case in the group is allotted a number, and a published random numbers table (e.g. Altman, 1991) is used to identify the case numbers to include. Pocket calculators and computers can also generate random numbers.
Calculating sample sizes for proportions – an example
A primary care team plans to audit the care of people with hypertension. They have 300 people being treated for the disorder, but do not have time to review all the records. They select one key criterion – ‘those on treatment should have had their blood pressure checked and the result should have been below 150/90 mmHg on 3 occasions in the past 12 months’ – and hope to achieve a performance level of 70%.
They are willing to accept 5% inaccuracy due to sampling – in other words, if their findings give a level of 70%, on 95% of occasions the true value would lie between 65% and 75%. They use the public domain software programme Epi Info to calculate the sample size using these parameters, and the sample required is found to be 155.
Interval sampling
Random sampling assumes that the sample can be drawn from a defined population of users or cases. However, users do not form a static population, and the individuals making up the user population (i.e. those attending clinics, practices or who are admitted to hospital) will change during the audit. In these circumstances, the sample is often determined by intervals of time; for example, people admitted to the coronary care unit from Jan to March inclusive - a reasonable approach if admission rates and the quality of care are unaffected by major seasonal factors.
- How to Select An Audit Sample (UBHT 2005)
- Choosing sample sizes - the scientific approach
- Sample size calculations depend on four variables:
- Size of population
- Degree of accuracy required
- Degree of confidence required
- How often you expect your audit criteria to be met
Table 9 below appears in a number of guides to choosing audit sample sizes and assumes an expected incidence of 50% (i.e. that standards will be met 50% of the time). It gives the sample size you will need in order to be 95% sure (degree of confidence) that the results you obtain from the sample will be within 5% (degree of accuracy) of the results you would have obtained for your whole population if you had collected data on them all. Put another way, there is a 1 in 20 chance that your results won’t be representative.
Population size | Sample size: 95% confidence; +/- 5%) |
Sample size: 90% confidence; +/- 5%) |
---|---|---|
50 | 44 | 42 |
100 | 80 | 73 |
150 | 108 | 97 |
200 | 132 | 114 |
250 | 152 | 130 |
300 | 169 | 143 |
350 | 183 | 153 |
400 | 196 | 162 |
450 | 207 | 169 |
500 | 217 | 176 |
600 | 234 | 187 |
700 | 248 | 195 |
800 | 260 | 202 |
900 | 269 | 208 |
1000 | 278 | 213 |
1500 | 306 | 229 |
2000 | 322 | 238 |
3000 | 341 | 248 |
4000 | 351 | 254 |
5000 | 357 | 257 |
Using this table, if your audit showed that criteria X was met in 56% of cases, you could be 95% sure that criteria X would have been met in somewhere between 51-61% of cases had we looked at the whole population.
Note that sample sizes need to be proportionately smaller as the population size increases; looking at 357 out of 5000 patients giving you results with the same degree of certainty as looking at 44 out of a population of 50 patients. This is because the chance of the results being unrepresentative is dramatically reduced as the population size increases. Remember, sample sizes can vary according to any one of the following:
- The expected incidence of the thing you are auditing
- The confidence level you want (it doesn't have to be 95% - could be 90%, 99% etc)
- The level of accuracy you are prepared to accept (could be 5%, 10%, 1% etc)