General Household Survey 2003

A two-stage theoretical weighting procedure was done on the GHS 2003. In the first stage primary sampling units (PSU) are selected with probability proportional to size (PPS) from the census population.

Because there were undercounts in some PSUs (because households could not be traced or because of refusals to answer), the weight of each such PSU was adjusted upwards by a factor of nHH/nHH where nHH was the number of households which should have been interviewed and nHH was the number of households actually reached. Then all household weights were adjusted upwards by a further factor equal to the estimated population at the time of the GHS 2003 survey divided by the 1996 Census population estimate, to account for population growth between the 1996 Census (from which the master sample was drawn) and the date of the survey. These doubly adjusted weights are reported as the household weights in the data set. The person weights are derived by further adjusting the household weights in order to reproduce the marginal totals of the estimated population at the time of the 2003 GHS by gender, population group, province and age group. A SAS macro called CALMAR was used for this purpose.

The population estimate was derived by a 'bottom up' (cohort-by-cohort) exponential extrapolation from the 1996 and 2001 censuses. Such an estimate is quite reliable for the total population and the gender, population group and provincial subtotals. It is less reliable for the age distribution. Improved population estimates will become available when Statistics South Africa completes its short-term population projection model. The weights in this and other surveys may be modified in the light of model estimates.

Data revisions

In May 2010 Statistics SA revised the population model to produce mid-year population estimates during 2008 in the light of the findings of the Community Survey 2007 and new HIV/AIDS and mortality data. The new data have been used to adjust the benchmarking for all previous datasets. Weighting and benchmarking were also adjusted for the provincial boundaries that came into effect in December 2006. The new weights mean that the data for the GHS 2002 to GHS 2009 are now comparable. The General Household Survey 2003 data files (version 1.2) contain the new weights.

As a result of new statistical programs used for weighting, which discards records with unspecified values for the benchmarking variables, namely age, sex and population group, it became necessary to impute missing values for these variables. A combination of logical and hot deck imputation methods were used to impute the demographic variables of the whole GHS series from 2002-2009.

A new weighting system was also introduced for the household files as part of the revision process. This was based on household estimates that were developed using the headship ratio methodology. The databases of Census 1996, Census 2001, Community Survey 2007 and the Labour Force Survey 2003, Labour Force Survey 2005, and Quarterly Labour Force (Quarter 3) of 2009 were used to analyse trends and develop models to predict the number of households for each year. The weighting system was based on tables for the expected distribution of household heads for specific age categories, per population group and province.

Missing values and unknown values were excluded from totals used as denominators for the calculation of percentages, unless otherwise specified. Frequency values have been rounded off to the nearest thousand. Population totals in all tables reflect the population and sub-populations as calculated with SAS and rounded off. This will not always correspond exactly with the sum of the preceding rows because all numbers are rounded off to the nearest thousand.