Describe the methods of sampling.

Sampling Methods

• Why sample — it saves time, money and effort; permits greater depth and better quality of data from fewer subjects; is the only option where testing is destructive; and gives results quickly enough to be useful
• The key requirement is that the sample must be representative, so that inferences drawn from it apply to the population — and this is achieved by random selection, in which every unit has a known, non-zero probability of being chosen
• Probability (random) sampling methods:
• Simple random sampling — each unit has an equal chance, selected by lottery or random number tables from a complete sampling frame. It is the most straightforward and the standard against which others are judged, but requires a complete list of the population, which is frequently unavailable in field conditions
• Systematic sampling — selecting every kth unit after a random start, where k = population size / sample size. It is simple and convenient, but biased if the list has a periodicity that coincides with the interval — sampling every seventh house on a street of repeating pattern, or every seventh day of the week
• Stratified sampling — the population is divided into homogeneous strata (by age, sex, urban or rural, income) and a random sample is drawn from each
• Its advantage is that it guarantees representation of every stratum, including small ones that simple random sampling might miss entirely, and it increases precision for a given sample size. Proportionate allocation samples each stratum in proportion to its size
• Cluster sampling — the population is divided into naturally occurring clusters (villages, wards, schools), a random sample of clusters is taken, and all units within the selected clusters are studied
• It is used where NO sampling frame of individuals exists and where the population is geographically dispersed, which describes most field surveys
• Its cost is that it is less precise than simple random sampling for the same size, because units within a cluster tend to resemble one another — the design effect, which requires a larger sample to achieve the same precision
• The WHO 30-cluster sampling method for immunisation coverage is the standard applied example — 30 clusters, 7 children in each, giving 210 children, designed to be usable by field staff without a population list
• Multistage sampling — sampling in successive stages (state, then district, then village, then household), which is how large national surveys such as NFHS are conducted
• Multiphase sampling — information collected from the whole sample at the first phase, and from a subsample at later phases, which concentrates expensive investigation where it is most informative
• Non-probability sampling methods, which do not permit valid generalisation but have their uses — convenience sampling (whoever is available, which is the weakest); purposive (judgemental) sampling; quota sampling; and snowball sampling, in which participants recruit others — the only practical method for hidden or stigmatised populations such as injecting drug users or sex workers
• Sampling and non-sampling error behave quite differently:
• Sampling error arises from studying a sample rather than the whole population; it is random, quantifiable (by the standard error and confidence interval) and decreases as the sample size increases
• Non-sampling error arises from faults in measurement, recording, non-response and analysis; it is systematic, not quantifiable, and is not reduced by increasing the sample size — a larger sample simply reproduces the same bias more confidently
• This IS the most important practical point: a big biased study is worse than a small unbiased one, because its narrow confidence interval lends false authority to a wrong answer
• Sample size depends on — the expected prevalence or effect size; the desired precision (allowable error); the confidence level (usually 95%); the power (usually 80 or 90%); the variability of the measure; the design effect in cluster sampling; and an allowance for expected non-response and drop-out
• Increasing the sample reduces sampling error only, and does nothing whatever to bias
• A large biased study is worse than a small unbiased one, since its narrow interval lends authority to a wrong answer
• Cluster sampling exists because there is no list of individuals in most field conditions
• The design effect requires a larger cluster sample, because units within a cluster resemble one another
• The WHO thirty-cluster method takes seven children each, giving two hundred and ten, and needs no population list
• Stratification guarantees small groups are represented, which simple random sampling may miss entirely
• Systematic sampling fails with periodicity in the list, such as every seventh house or every seventh day
• Multistage sampling underlies national surveys such as NFHS, sampling state then district then village then household
• Snowball sampling reaches hidden populations, and is often the only practicable method for stigmatised groups
• Non-probability samples cannot be generalised, however large or convenient they are
• Allow for non-response when calculating size, or the achieved sample falls short of the required one
• A well-drawn sample may beat a poor complete count, since fewer subjects permit better quality data