Using Stata for Survey Data Analysis

Example 16 Using “table” to create tables 

---

Minot
 
 
 
 
 
Using Stata for Survey Analysis 
Page 36 
 

Using weights  
What are sampling weights? Sampling weights are used to compensate for under- or over-
representing certain households in a sample, allowing it to reflect the population as a whole. Let‟s 
take a simple example: 
Suppose there are 200,000 households in the population, of which 160,000 live in rural areas. 
A survey is carried out with a random sample of 2000 urban households and 2000 rural 
households. Thus, the population is 80% rural but the sample is only 50% rural.
Suppose per capita income is 2800 in urban areas and 1400 in rural areas. The simple average 
income across households in the sample would be 2100, but we know that the average income 
across the country would be lower than this because rural households (who have lower 
incomes) are under-represented in the sample.
We can compensate for this by using a weighted average, with weights equal to the ratio of 
the population to the sample. The urban weight would be 20 (=40 thousand/2 thousand) and 
the rural weight would be 80 (=160 thousand/2 thousand). The weighted average is 
calculated as (80*1400 + 20*2800)/(20+80) = 0.8*1400 + 0.2*2800 = 1680.
The basic principle is that the sampling weight is the inverse of the probability of selection. Because 
of clustering and sampling, virtually all random-sample surveys must use weights to make estimates 
that are valid for the whole population. Furthermore, the calculation of sums, average, and 
percentages must take into account the sampling weights.
Sampling weights in the BLSS  The calculation of the sampling weights in the BLSS is more 
complicated than the example given above, but the principle is the same. The NSB divided the 
country into seven strata:

Thimphu 

other urban areas in the west 

urban areas in the center 

urban areas in the east 

rural areas in the west 

rural areas in the center, and

rural areas in the east. 
In each urban stratum, the NSB selected 60 blocks and about 10 households in each block, while in 
each rural stratum, NBS selected 30 geogs and interviewed about 20 households in each geog (some 
adjustments were necessary in implementation). Since more than half the BLSS sample is urban, it is 
clear that urban areas were over-sampled (this is common practice in household surveys, based on the 
idea that urban households are more diverse). In the BLSS, the sampling weight is called “weight” in 
the file households.dta. The average value of “weight” is about 10 in urban areas: this means that the 
BLSS interviewed about 10% (1/10) of the urban households and that each sample household 
“represents” about 10 households in the urban population. The average value of “weight” in rural 
areas is about 49. This implies that the BLSS interviewed about 2% (=1/49) of the rural households 
and that each sample household “represents” 49 households in the rural population.