Menu ENN Search
Language: English Fran├žais

Second stage sampling

This question was posted the Assessment and Surveillance forum area and has 11 replies.

» Post a reply

Blessing Mureverwi

Normal user

18 Jan 2020, 02:59

We are doing a SMART survey in province X, which has 20 districts. We plan to have a representative sample in each district, then a weighted provincial prevalence. Our districts are divided into sub-districts, wards then villages. We will use cluster sampling and we hope to use the village as the PSU. However, we will not be able to get a list of households for each village, nor do we have time to list households. Households are not arranged in any order either. What kind of method can we use which ensures that, as much as possible, we select a sample as randomly as possible?

Joseph Duduka

Nutrition Field Monitor

Normal user

18 Jan 2020, 13:22

You could use the EPI sampling technic. We used to use it long time ago in South Sudan but since we started listing households and randomly selecting participants, nobody uses the EPI method anymore. It involves moving at the centre of the a village and then spining a pen/pencil to give the direction to a random household along the chosen direction pointing outwards from the centre of the community to its boundary.

In subsequent steps, which are carried out iteratively, the closest household (door to door) to that determined in the previous step is chosen and checked for compliance with the inclusion criteria. The iterations are repeated until the required number of households is surveyed.

It is not as good as listing households in a village and the randomly selecting a sample. 

Anonymous 2953

Normal user

18 Jan 2020, 16:27

Hello Anonymous 730

In each village (cluster) which is drawn at the first stage of your sample, it is necessary to carry out a count of all the households (if the village is not very large, otherwise it is necessary to make the segmentation) and then to use systematic random drawing on the list of households to obtain the number of sample households.

cordially

Bradley A. Woodruff

Self-employed

Technical expert

19 Jan 2020, 01:26

Dear anonymous 703:

Please do not use the EPI proximity sampling methodology first recommended in the 1980s by the Expanded Program on Immunization of the World Health Organization. It was created for use in small coverage surveys of vaccination for use by very naïve personnel in rural Africa. It has several inherent biases built into the methodology which cannot be completely overcome by the various several adaptations which have been created subsequently. These biases are explained on pages 82 and 83 of the WFP publication “A manual: measuring and interpreting malnutrition and mortality” which is available at  https://www.unhcr.org/45f6abc92.pdf. Although this manual describes the theoretical problems with this sampling methodology, a paper from Ehtiopia demonstrates that these concerns are of practical value and may lead to important biases during actual assessments. See: Comparison of two survey methodologies to assess vaccination coverage. Luman ET, Worku A, Berhane Y, Martin R, Cairns L. Int J Epidemiol. 2007 Jun;36(3):633-41. 

Listing households in a selected village, census enumeration area, or other small geographic unit is not that difficult or time-consuming. We have trained relatively naïve personnel to do such listings in 1 or 2 hours in villages of 200 to 250 households. Primary sampling units larger than this can be segmented into smaller geographic units with relative ease. Only in the most extreme cases is it necessary to resort to the non-random sampling method recommended by the EPI.

Blessing Mureverwi

Normal user

19 Jan 2020, 01:52

Thanks Brad

We will certainly avoid the proximity sampling.However,due to time,weather conditions and terrain,we might not be able to list households.How about the QTR+EPIx method mentioned by Mark Myatt.Would it be a possibility?

Bradley A. Woodruff

Self-employed

Technical expert

19 Jan 2020, 19:38

Dear anonymous 730:

I don’t really know much about the sampling methodology you mention (QTR+EPIx); the only description I can find in this forum from 2012. In that short description, it appears to spread the sample of households throughout the primary sampling unit better than the original EPI method, but it is still subject to many of the same biases.

The QTR+EPIx method, like the original EPI method, chooses the next house by counting houses closest to the house just completed. If dwelling density is greater toward the centre of the primary sampling unit, which is true in many populations, this method will move the sample toward the centre of the primary sampling, thus biasing the sample within each primary sampling unit. In addition, one of the biggest problems with the original EPI method is that it leaves the selection of households up to the survey team members. As a result, less accessible or less desirable households will have a lower probability of selection. For example, if 2 households appear to be equally eligible, the survey team will choose the household without the large, angry dog chained in the front yard. In order to be independent and random, household selection must be as independent as possible of any decision-making by members of the survey data collection teams.

Another problem with the original EPI method and all its adaptations is confusion regarding the basic sampling unit. The EPI method selects houses or dwellings, it does not select households. If a dwelling is selected which contains more than one household, either all households in that dwelling can be included in the sample, or 1 household can be selected at random. In the former case, selection of a single small apartment building will complete the number of households required in that cluster, thus providing a relatively non-representative sample which inflates the design effect. In the latter case, because the probability of selection of a single household chosen from a multiple-household dwelling is different from the probability of selection of a household in a single-household dwelling, sample weighting must be employed during analysis. This complicates data analysis and lowers precision.

I will repeat that we have done household listing in many different circumstances and populations without expenditure of exorbitant time, energy, or resources. I would recommend household listing and simple or systematic random household selection in all population survey assessments except perhaps in urgent rapid assessments during acute humanitarian emergencies.

Blessing Mureverwi

Normal user

19 Jan 2020, 22:14

Thank you.

mark@brixtonhealth.com

Consultant Epidemiologist

Normal user

20 Jan 2020, 09:58

You could use a modified or improved version of the EPI method. I tend to use a modified EPI sampling method in which we do not take immediately neighbouring households but spin a bottle and take the third or fifth household in the random direction. Taking every third or fifth HH in a random direction breaks proximity and spreads the sample wider than a small cluster of dwellings located close to the centre of a community. This is described in:

Bennett, S., Radalowicz, A., Vella, V., & Tomkins, A. (1994). A computer simulation of household sampling schemes for health surveys in developing countries. International Journal of Epidemiology, 23(6), 1282–1291

This reference also goes into some details about the problems with the basic EPI method.

I now use a map-segment-sample (MSS) technique which is a development of the QTR+EPIx method for use in RAM surveys as described in the  the RAM-OP manual::

RAM-OP manual

RAM surveys use a spatial first stage sample but MSS will work just as well with a PPS first stage sample and data can be analyses using ENA for SMART just as would be done in a standard SMART survey.

Anastacia's suggestion to use of census planning information is useful. Central Statistics Bureaus usually have maps and household lists of census tracts / enumeration areas which may be updated on an ongoing basis. I have used these to help with sampling urban areas in RAM and SMART surveys in Ethiopia and RAM surveys in Sierra Leone. Segmentation (as in MSS) or quartering (as in QTR + EPIx) spread the sample across the whole community and provides a form of implicit stratification which captures within-community variance that is lost with a proximity sample. This reduces design effects and increases effective sample size yielding better precision.

I hope this is of some value.

Anastacia Maluki

consultant

Normal user

24 Jan 2020, 06:34

Just to emphasize that there are no short cuts in SMART thats is why it is standardized. You will just have to select the villages and give 2-3 days for household listing. You can use the local area chiefs/village elders to help you as they are well familiar with the numbers of household in specific villages. Other question is does the country do census? if so liase with the planning and statistics departement at the province level and get all the necessary information required. 

Gael BITA

NPO, HKI

Normal user

24 Jan 2020, 07:27

Hello,

Would it not also be possible to estimate the number of households (although less precise), if national statistics make it possible to know the average number of people per household (t), by region and knowing the size of the population (P ) by region, then the number of households in the area to be surveyed will be equal to P / t.

Isn't this recommended?

Thank you

Anastacia Maluki

consultant

Normal user

24 Jan 2020, 07:43

This is not recommended, simply because in reality the villages are not of same population size. There are villages with more houses than others and the principle of second stage samling is to ensure that each household has an equal chance of being selected to reduce sampling bias.

In cases where villages have more than 250 households, segmentation has to be applied.That is dividing the village into parts by use of available features e.g road, river , boundaries etc.

Blessing Mureverwi

Normal user

24 Jan 2020, 11:06

Thank you so much Mark!

If you have any problem posting a response, please contact the moderator at post@en-net.org.

Back to top

» Post a reply