Estimates of Sampling Error
As in any sample survey, the results obtained from the LFS are subject to sampling and non-sampling errors. The non-sampling errors arise as a result of imperfections in data collection, data processing and dissemination. These include errors that are introduced at the preparatory stage; errors committed during data collection including those committed by interviewers and respondents; and processing errors. In order to reduce these non-sampling errors several safeguards were adopted. Careful design of survey instruments, training and supervision of LFS staff deployed in data collection and processing, efficient operating procedures in data cleaning and data management, checking consistency and completeness of the tables that were extracted were some of the more important methods and procedures that were used in the survey. However it is known that non-sampling errors would be the major source of errors in the survey results, not withstanding the measures that were adopted in survey design and implementation. In view of the impracticality of measuring non-sampling errors, the total error calculation in surveys is restricted to calculation of sampling errors.
Sampling errors in surveys occur as a result of limiting the survey observations to a subset rather than the whole population. These errors are related to the sample size selected and sampling design adopted in the survey. In order to maintain these errors within acceptable levels, the efficient sampling design with the sample allocation described in Annex 3 (refer to the external resource of Chapter 12 of the Main report for more information) was adopted.
The sampling error indicates the extent to which an estimate from the survey would vary by chance, because only a sample of enumeration areas is included rather than all the enumeration areas into which the country is divided. The sample size and survey design had determined the magnitude of the sampling errors and in respect of some items the sampling errors were known to be high at the design stage of the survey.
IMPS package that was developed by the US Bureau of the Census was used in processing data from the LFS +CAM. Therefore, it was decided to use CENVAR which is the variance calculation module of the IMPS package to compute sampling errors of key aggregates from the survey. For each specified parameter and domain of estimation, CENVAR produces a tabulated output that provides the following measures.
- the estimated value of the parameter
- the standard error
- the coefficient of variation
- the 95 percent confidence interval
- the design effect (DEFF) and
- the number of observations upon which the estimate is based
It is common to allow an interval of either 2 standard errors or 1.6 standard errors in either direction around an estimate from a given sample as the possible range of sampling error. Under the 2 standard error criterion, the population value as estimated from the sample falls within the indicated range in 95 cases out of 100. Under the 1.6 standard error criterion, the probability drops to 90 cases out of 100 but this is still a reasonable basis for judgment for many analytical purposes. Estimates of sampling errors computed using CENVAR have 95 confidence intervals of 2 standard errors. The sampling errors of key aggregates are provided in Tables 84-91 (refer to the external resource of Chapter 12 of the Main report for more information).
As described in the users guide, CENVAR is designed for the calculation of the variances and uses formulae appropriate for stratified multistage sampling designs. The details of the two stage stratified sampling design used in the LFS +CAM including the stratification into 9 strata and sampling weights had been defined as required by the CENVAR system at the stage when variables corresponding to the sample design were specified. However, certain aspects of the sampling design such as the strong implicit stratification by aimag (province) soum( district) built in to the sampling design through the adoption of stratified circular systematic random sampling could not be specified in the CENVAR system. Thus, the sampling errors computed using the program and produced in the attached tables would probably overstate the width of the actual or true confidence intervals of parameters as well as the design effects of the sampling design.
|
