Sampling Procedure
Documentation for Kiribati HIES Sample Selection:
SAMPLE SIZE:
In determining an appropriate sample size for a survey of this nature, numerous factors come into the equation. These include:
a) The degree of accuracy required for key estimates
b) The population size of the country
c) The manner in which the sample is selected
d) Cost or staffing constraints which may exist
e) Whether or not estimates are required for sub-populations
f) The level of variability in the data being collected
Each of these factors have different magnitudes of importance, but the major priority should always be on selecting a sample big enough to produce results of suitable accuracy. Many of these issues are generally known as well - for instance:
· A user group may pre-specify what level of accuracy they may wish to achieve for the survey
· The population of a country can normally be estimated to a reasonable level of accuracy
· The sample selection technique adopted is known
· Cost and staff constraints are generally known, and
· A user group can once again provide information on whether estimates for sub-populations are required.
The one thing that normally isn't known is the degree of variability in the data being collected - this information comes after the survey. This factor is important because if there is not much variability in the data for key estimates, then the sample size does not need to be as large, and vice versa.
Without this sort of information, determining the appropriate sample size for a survey can often involve a bit of guess work. For that reason, based on previous survey experience in other Pacific Island countries, a sample of 10 per cent was considered more than sufficient for Kiribati. An additional 10 per cent of sample was selected to allow for sample loss.
As a result, a sample size of 1,555 households (10 per cent of 13,999, with a 10 per cent top-up) was considered suitable for the survey.
Allocation to "Target Areas"
For the Kiribati HIES, five target areas were identified as sub-populations for which estimates would be desirable. These five areas were:
1) South Tarawa
2) Northern Gilbert
3) Central Gilbert
4) Southern Gilbert
5) Line/Phoenix Islands
Once the sample size of 1,555 had been determined, the next step was determining how the sample should be allocated to each of these target areas in order to produce the required level of accuracy for each area. In order to achieve this, the sample was allocated in such a manner that the expected level of accuracy for each stratum would be similar. The resulting sample allocation can be found in the table below.
STRATIFICATION:
To achieve better representation within each target group, each target group was further stratified by grouping "like" islands. The plan from there was to select an island from each stratum to represent it. As a result, 11 strata were formed, with each of the 23 populated island/atolls of Kiribati allocated to one of these strata. The resulting strata, and islands which make them up, can be found in the following table.
The allocation of the sample to stratum within each target group was achieved by simply allocating the sample proportional to the population for that stratum. For example, for the target group Northern Gilbert, an overall sample size of 323 was desirable. To determine how much of that sample would be allocated to the first stratum which consisted of Makin and Butaritari, the following formula was applied:
n (Makin & Butaritari) = 323 * (889)/(889+1290+867)
= 94
EXCLUDED AREAS:
Although it would be desirable to cover all of Kiribati for this survey, due to cost and time constraints a couple of areas were excluded from the frame before the selections were made. The two areas removed from scope were:
· Banaba
· Kanton
The impact on final estimates is considered to be very small given the small populations on these two islands; 61 households on Banaba, and 9 households on Kanton. This accounts for about 0.5 per cent of the population of Kiribati.
SAMPLE SELECTION TECHNIQUE:
Selection of Islands:
For the stratum with more than one island, an island had to be selected in order to represent that stratum. The process used for this stage of selection was probability proportional to size (pps) sampling, where the size measure was the number of households on the island.
An example of how this process worked can be found below for the Central Gilbert - Group 2. For this stratum, a random number was selected between 0 and 1,005. Given the random number (254), fell within the cumulative number of households for Abemama, then that island was selected.
Selection of Households:
To minimize the travel requirements of interviewers, and thus travel costs, a two stage process was adopted to selected households.
The first stage of the process involved selecting EAs using probability proportional to size (pps) sampling. The required number of EAs to select from each island was simply determined by dividing the number of households to sample on that island by roughly 15, where 15 was the cluster size chosen from each selected EA.
Having selected the EAs in sample, a systematic skip was run through the list of households for each EA to determine which households would be selected in sample. These selections were performed in the NSO, and the lists provided to interviewers in the field. The lists used for this stage of selection were based on the 2005 Population Census, and thus would be out-of-date by a year or so, but given the significant gains in simplifying field logistics, this was the preferred option.