Between Census Household Information Monitoring and Evaluation System 2000

The NMIS evaluation report suggested that instead of two cycles per year in NMIS one survey be carried out every year with detailed analysis that would have wide-ranging dissemination and plans of data use. In the future, BCHIMES (Between Census Household Information, Monitoring and Evaluation System) will be conducted on a regular basis to generate needed data. The following suggestions were also made in the NMIS evaluation report for the effective design of the sample:

• For every new study, always select a new sample so as to minimise the Hawthorne effect.
• In order to minimise the standard error of the estimate, always try to make the cluster size small, i.e., around 50, as compared to an average cluster size of 120 for the NMIS cycles.

Thus, the new sample design should limit the average cluster size to 50 or smaller and a new sample should be drawn for a new study every time for the minimisation of the Hawthorne effect.

Domains of estimation
A sample design to provide district level estimates was desirable keeping in view the decentralisation programme of the His Majesty's Government of Nepal. However, as the sample size needed for this would be very large and the survey undertaking also huge as well as expensive, it was decided that the size of the survey should provide national as well as some sub-regional estimates. Under the guidance of the Steering Committee as well as the discussion between the CBS personnel and UNICEF led to the conclusion that a minimum of 13 estimates is needed for different geographic areas and these are

1. Five eco-development regions each from the Terai and Hills;
2. Estimates for the Kathmandu Valley; and finally
3. Two estimates for the mountain region, for which the Central, Eastern and Western Mountain regions would be combined as one and the other would be the combination of the Mid-western and Far-western Mountain regions.

Although there are some variations within these mountain regions, regions having comparable characteristics would be combined as one. Since the number of households was the basis of the selection of our sample, we used average size of the household as an indicator to provide the similarity between these combined areas. For example, the average household size was 5.5 in both the Far-western and Mid-western Mountains. Likewise, the average household size for the Eastern, Central and Western Mountains is, respectively, 5.3, 5.0 and 4.8. That is, the average household size was slightly higher in the Far-western and the Mid-western regions and was slightly lower in the others including the Eastern, Central and Western Mountains. In other words, the areas that were combined were quite close in terms of average household size.

Stratification
In domains with urban areas, the stratification was done according to urban/rural residence. Although the urban/rural estimates for these domains would be of interest, it would have increased the sample size considerably. Thus, at this stage, there were no plans to obtain urban/rural estimates for these 13 domains of estimation. Note, however, that the urban/ rural estimates could be available for the national level, as well as for the Hills and Terai. Because the sample was selected separately for each domain, there was a built-in stratification for the Hills, Terai and Mountains as well as the development regions for most of the domains of study.

Estimation of sample size
Estimates of the sample size, to a large extent, depend on the variable under study. As some variables have a larger variation, sample size estimates depend on the variables. To circumvent this problem, statisticians usually resort to estimating the sample size for variables where the largest sample size is needed and use this as the required minimum sample size. Also, because most of the sample survey use the cluster sample approach, it was necessary to make an allowance of about 2 for the design effect. The magic figure of 2 was based on the design effect calculated for different variables in the Nepal Family Health Survey 1996. It was estimated that a sample size of 800 was adequate for most of the variables, taking into account a design effect factor of 2. This sample size of 800 was regarded as the minimum sample size required for the domain of analysis. Since there are 13 domains, a total of 13x800 = 10400 households were required.

Sample frame
The sample frame for this study was the data from the 1991 Census data on Households for VDCs and their wards. When the census was undertaken in 1991 there were only 31 urban areas in Nepal. However, after 1991 Census, the government declared new municipalities. As a result, there are currently 58 municipalities, of which one is a metropolitan city and three are sub-metropolitan cities. The census data was updated to take into account the change in urban areas.

Allocation of the sample
In domains that have urban areas, the urban sample was be allocated proportionately. Urban and rural samples were selected separately using a PPS (Probability Proportional to Size) method. Examples for this are provided in Table A1, page 161 of the Report on the Situation of Women, Children and Households, Between Census Household Information, Monitoring and Evaluation system (BCHIMES), March-May 2000.

The total number of clusters surveyed was 208 with an average cluster size of 50, providing a sample size of nearly 10,400. Likewise, the number of urban clusters will be 27 and the number of rural clusters will be 181. The proportion of urban clusters was 13 percent (See Table A1, Appendix 1 of the Report on the Situation of Women, Children and Households).

Selection procedure used
For any given domain, the districts were arranged according to the code for districts provided by the Central Bureau of Statistics. If the code of a district is lowest, it appears first in the list. Within the district, VDCs are listed in an alphabetical order. For each VDC, there will be nine wards, for which there is data regarding number of households, total population, males and females.

Initially, the number of households in a domain was cumulated. The total number of households in a domain is divided by the number of clusters selected in the domain. This provided the systematic interval. Then, a random number between 1 and the systematic interval was selected for the first selection. Once the first selection was made, the systematic interval was added to that for the second selection and so on, until the last selection for the domain was made. If a domain consisted of urban and rural areas, then the selection was made separately for the urban and rural areas. Obviously, a proportionate allocation of sample was done for urban as well as rural areas within a domain. Note that a cluster size of 50 was used for the purpose of data collection. In fact, a number of wards will have a population well over 50, and in some cases a ward could have a population substantially less than 50. In some cases, some wards may have to be split and other wards merged to provide a cluster size of around 50.

Distribution of the samples
A total of 208 clusters (10,295 households), with 181 rural clusters (87%) and 27 urban clusters (13%s) were selected from 69 districts for the survey. The average cluster size was 50 households per cluster. Since the sample was stratified by region, it is not self-weighting; hence, sample weights were used for reporting national-level results.