Chapter 4. Selecting the sample

If an estimate of malnutrition is needed for a relatively small group of children, it is best to examine all of them. In a small population of, say, 2000-3000 people - of whom 18 to 20% may be children below 5 years of age (400-500) - all eligible children should be examined. In larger populations it is usually easier to examine and analyse only a sample of children and to draw conclusions on the probable proportion of malnourished children in the total population.

The first step is to define the population for which the estimate is needed. This study population is also called the sampling universe. The sampling universe may be the child population of one or several refugee camps, of a province, or of a country. The estimate will only be valid for the sampling universe as a whole. If separate estimates are needed for ethnic or geographic subgroups or other subdivisions of the sampling universe, each of them must be treated as a separate universe for which a separate sample must be constructed. Therefore, the smallest subdivision on which information is sought should be determined at the outset.

For emergency assessments several types of sampling are available:

Simple random sampling: the children are chosen at random from a list of all eligible children in the sampling universe. This is the ideal procedure but usually not practicable in an emergency.

Systematic random sampling: children are selected systematically, say every 10th child, from a list of all households. Alternatively, if the average number of preschool children per household is known, a sample of households, say every 10th house or tent, may be taken systematically, and all eligible children in these houses are examined.

Cluster sampling: clusters or groups of households are selected from a list or from a map of all clusters; in each selected cluster a predetermined number of children is selected at random, systematically or sequentially.

Another approach to sampling is stratified sampling, which can be used with any of the above techniques. In stratified sampling the universe is stratified by certain characteristics thought to influence nutritional status: age, sex, social or ethnic group, environment. Each stratum is an independent universe from which samples may be drawn by one of the above-listed methods. Stratified sampling is also used where several areas or camps are to be surveyed and each of them is to be viewed separately.

The choice of sampling method depends mainly upon practical conditions. In settlements and camps, systematic random sampling is the method of choice; in a scattered population cluster sampling may have to be the choice. It must be borne in mind that in cluster sampling the sample size needs to be twice that of systematic random sampling.

SYSTEMATIC RANDOM SAMPLING

It is particularly recommended where the population is concentrated in an organized or structured urban setting or in a refugee camp, and where the total number of households is less than 10 000. Knowledge is required on: 1) the number of households, 2) the average number of children in the 6 months to 100 cm (5 years) group per household, 3) the total population or number of people in the universe. The recommended sample size is 450 children.

This number will ensure with a probability of 95% that the estimated prevalence will be within plus or minus 5 prevalence percent of the true prevalence irrespective of the level of prevalence. A safety margin of about 10% is included. Those wishing to include children up to 110 cm instead of 100 cm should increase the sample size to 500 children.

In practice, in camps as well as in permanent settlements, the sampling unit is the household or dwelling. Taking into account an average household size of A persons and an average proportion P of children of the right age/height in a population, the number of households needed to yield the required number of eligible children is calculated as follows:

Number of households to be visited = 450/(A × P)

For example, if the average household size is 6 persons and the proportion of children under 5 years 0.15 or 15%, then 450/(6 × 0.15) = 500 households should be visited.

If the sampling universe consists of 9000 households, the sampling interval equals 9000/500 = 18. Thus every 18th household is to be visited.

Further examples of calculations and procedures and more details are given in Annex 1.

Over- or underestimation of people, of households, or of the proportion of children will result in a sample that is either too small or too large. This will cause unnecessary delays and loss in precision. Estimates should therefore be as accurate as possible.

Estimates can be improved by doing a rapid count of households when planning the survey. If the number of persons in a camp or a village is known, the number of households can be estimated from a subsample of, say, 30 households; by dividing the total population of these households by 30, an average of the number of persons per household is obtained.

If the percentage of children in the appropriate 6 months to 100 cm group is overestimated, fewer households will be surveyed and the resulting sample of children will be too small. It is therefore better to underestimate than to overestimate the percentage. Information on household composition may be available from previous census data of camp or town residents.

In most developing countries, about 15% of the population will be in the required age-length group. However, in emergencies such as famines or wars this figure may be considerably lower or higher, because infants and children may have died or many adult men may be absent.

CLUSTER SAMPLING

In cluster sampling the sample children are not spread randomly over the population but are lumped into randomly selected “clusters”. It is the usual method for large populations and populations spread over a large area for which only rough estimates of the number of people are available. It may also be an advantage in large or newly established camps where numbers and ages of people are still incompletely known. However, the sample size needed to obtain the same precision is about twice that of a systematic random sample, i.e. 900 children.

This sample size ensures with a probability of 95% that the estimated prevalence will be within plus or minus 5 prevalence percent of the true prevalence, irrespective of the value of the prevalence, and assuming a correction factor of 2 (the “design effect”) for cluster sampling. For reliable results it is important to examine not less than 30 clusters and not less than a total of 900 children.

For a rapid assessment in an emergency, when there is little time for preparatory work, the following sampling procedure is recommended.

The area of interest is divided on a map into sections of about equal size, following as far as possible existing geographic or administrative boundaries. Each section should have at least 300 inhabitants. A systematic sample of 30 clusters is drawn from a list of all sections and their population estimates. The total number of clusters is divided by 30 to obtain the cluster interval k starting from a randomly selected cluster on the list, every kth cluster is selected.

For example, suppose there is a total number of 183 sections. This is divided by 30 to obtain the cluster interval (183 divided by 30 = 6.1). Starting from a randomly drawn section, say section no. 15, every 6th section down the list is chosen until the 30 survey sections, the clusters, are selected. During the survey, the team starts at the centre of the cluster and chooses a direction (for example by spinning a pen on a book). The survey is started at the nearest dwelling in that direction, moving to successive houses until 30 children have been examined. At each dwelling, all eligible children should be examined.

The traditional community-based cluster sampling based on proportional sampling with a list of communities and their populations may not always be feasible in emergencies. If multiple areas or camps are to be surveyed, the most efficient approach may well be to treat each area as a stratum and to conduct systematic random sampling in each.

For another example of cluster sampling see Annex 1.

IMPORTANT CONSIDERATIONS IN SAMPLING

Every effort should be made to obtain the needed data on all children in the sample. For example, in cluster sampling all 30 children from each cluster must be seen and all eligible children in the cluster must be given equal probability of being selected.

Although random numbers are used to select sampling sites, households, and starting points, the selection procedure is never haphazard.

Samples must be selected by a rigid and defined methodology. Once the sample selection has begun, the procedure should not be changed or modified. Children must be selected for examination only by using the selected sampling procedure. Any exception will bias the estimates.

In subsequent surveys to measure changes over time, the same methodology should be used to ensure comparable results.

Only children in the households or family groups selected by the sampling procedure should be examined.

All eligible children between 6 months and 100 cm, in each selected household or family grouping should be examined. If necessary, team members must actively search for eligible but absent children, even if a dwelling is temporarily empty, and include such children in the survey.

If a central examination site is chosen, great care must be taken to ensure that all the selected children arrive at the site. During preliminary household visits children may be numbered sequentially and the number given to the mother on a piece of paper to bring with the child to the examination site. Missing children can then be sought.

In spite of its apparently greater simplicity, in a population that is concentrated in a relatively small area, cluster sampling has no advantage over stratified or systematic sampling with its smaller sample size. Therefore, the sampling method should be carefully chosen.

For each survey the method of sample selection should be documented in writing and included in the report of survey results.