Frequently Asked Questions DHS Survey Design: Sample Size

DHS Survey Design:

Frequently Asked Questions

Most DHS surveys are representative* at the national level, for urban

and rural areas, and for the ﬁrst administrative level subdivisions,

which are usually called regions, zones, provinces, governorates, or

states. There is growing interest in providing DHS data for even

lower administrative levels, such as counties and districts.

*In this document, by “representative at a speciﬁc domain level” we

mean that most of the survey results/indicators can be produced on

the level of that domain with good precision.

What factors determine sample size?

Sample sizes for DHS surveys are based on the number of survey

domains (usually subnational units such as regions), the precision

requirements for priority indicators, and the budget. Generally, The

DHS Program samplers design surveys with fertility and childhood

mortality estimates in mind.

To calculate fertility with an adequate level of precision, for

example, samplers include between 800 women for every

subnational region in countries with a high total fertility rate

(TFR) to 1,000 women in countries with a low TFR. The typical

subnational region is usually administrative level 1. In other words,

a high TFR country with 10 regions needs to interview at least

8,000 women in order to have estimates of fertility and child

mortality with reasonably small standard errors for each of the

10 regions. In countries with more than 10 regions, one option

to keep the sample size down and reduce costs is to group some

regions to form study domains. A total sample size of about 10,000

women is ideal to maintain cost efﬁciency and high data quality.

How precise are these estimates?

All survey sampling strategies are subject to sampling error. The

DHS Program designs samples to provide national and subnational

estimates with a reasonable relative standard error. The larger the

sample size, the smaller the relative standard error on any given

indicator will be. The standard errors at the admin 1 level are

DHS Survey Design: Sample Size

Considerations

Validity:

Increasing sample size is a valid

practice if funding and human

resources are sufﬁcient for a larger

survey. All sampling strategies

are subject to sampling error; the

larger the sample, the smaller the

relative standard error will be.

Impact on cost:

Sample size is the single largest

driver of survey cost, as it impacts

all elements of the survey process

from hiring and training of staff to

processing of data, report writing

and dissemination.

Impact on quality:

Large sample sizes can overburden

the implementing agency and

survey management staff and lead

to poorer data quality because of

the challenges in data collection

and overall survey management.

A large survey requires additional

coordination and leadership

and should be undertaken by an

experienced implementing agency

with robust data quality checks in

place.

The Demographic and Health Surveys Project, implemented by ICF and funded by the United States Agency for International

Development (USAID). www.dhsprogram.com

DHS Survey Design: Frequently Asked Questions about Sample Size

wider than those at the national level; standard errors at the admin 2 level are larger than those at the admin

1 level. For this reason, interpretation of subnational trends and comparison of subnational units should be

undertaken with caution unless the total sample is very large. Appropriate signiﬁcance testing is necessary to

conﬁrm changes over time or true differences between subnational areas.

What options exist to provide data at lower administrative levels?

• Full and shortened questionnaires: Depending on the priority indicators for the survey, a compromise

can sometimes be reached wherein many indicators are available at the second administrative level, while

others are available only for admin 1. For example, in the 2014 Kenya DHS the total sample size was 40,000,

which is very large for a DHS. Many indicators were available for all 47 counties, while those indicators that

require larger sample sizes (TFR, childhood mortality rates) were only available for each of the 8 regions.

Though the total sample size was very large, only half of the households received a full questionnaire, while

the other half received a shortened questionnaire.

• Over-sampling in focus areas: Sometimes stakeholders have data needs that are speciﬁc to a few

districts; perhaps they want to monitor progress in key intervention areas. In this case, those districts can be

over-sampled without expanding the entire sample to the second administrative level. This helps to maintain

a manageable and cost-effective sample size while still meeting speciﬁc data needs.

• Larger sample sizes: It is also possible to implement a standard DHS that is representative at the

second administrative level, which, in some countries, requires a sample size of over 50,000 women. This

can be undertaken only when ﬁnancial and human resources are sufﬁcient to properly manage the survey

without threatening the overall data quality. Very large sample sizes present unique logistical challenges for

training, staff management, data and lab processing, and provision of supplies. Management of this type of

survey requires a highly skilled and organized implementing agency coordinated with external technical

assistance.

• Small area estimation: Another alternative for producing estimates at lower subnational levels relies

on small area estimation (SAE) techniques and statistical modeling using ancillary geographic information.

If these techniques are acceptable, the DHS survey scope can be limited to data collection for admin 1

estimates only and these data along with covariate data from external sources can be used to estimate

indicators at lower administrative levels.

Currently, The DHS Program uses statistical modeling using ancillary geographic information to create

interpolated surfaces for lower-level estimates. The DHS Program standardly produces modeled surfaces for

12 key indicators. These maps are presented at a 5x5km grid scale, which can be aggregated to represent the

needed subnational units along with conﬁdence intervals for the aggregated indicators. This approach does

not provide estimates that are as accurate as high quality survey data and is currently limited to a small set of

indicators. However, if appropriate for the speciﬁc data need, interpolated surfaces may offer an alternative

to very large sample sizes. (References: https://dhsprogram.com/pubs/pdf/SAR15/SAR15.pdf, https://

dhsprogram.com/pubs/pdf/SAR14/SAR14.pdf).