Dear Masood,

weights for housholds or individuals are used to reduce the bias of deviations of the distribution of characterisitics (such as the frequencies of sex, income level, etc) observed in the emperical data of a survey sample from the population from which the sample was drawn. In other words, weights are intended "to make the emperical study data more representative". Usually polling firms conduct the surveys and then calculate the weights by unpublished algorithms that take the distribution of characteristics in the population into account.

Unfortunately it is not possible to simply add weights or to transform the old weights to new weights for new strata or clusters. To calculate correct new weights you first need to know the "real", either stratified or clustered, distributions of characteristics in your population (e.g., from census data), and second need to apply a proper algorithm (e.g., iterative proportional fitting http://en.wikipedia.org/wiki/Iterative_proportional_fitting) by a software that can do it for you or by a self-written program.

In general, using these kind of weights can be controversial: if the weights in your empirical sample are all close to "1", this would mean that the drawing of your study sample already yielded in a very good representativeness of the data (so you could perform your analyses also with unweighted data). If a large amount of cases in your sample have weights far distant from "1", this would mean that the survey method yielded in poor representativeness of the data, and that the weighting of cases (individuals or households) artificially correct the distrubutions "somewhere", thus the vaildity of the weighted data can be questionnable.

If these were my data, I would either build new strata or clusters with the originally weighted data and analyse them (when I can trust the survey method and the calculated weights). If you do this with your study data, unweighted and weighted frequencies of characteristics in the new strata or clusters must not differ significantly by your own personal criteria (statistically, they will most likely always differ significantly in large samples).

Or, you just do all your analyses in the new strata/clusters with unweighted data, and discuss the fact in the end, that representativeness of the data is reduced.

I hope this helped a little,

Volker

Whether it is sample survey or census survey? If it is a sample survey then use weight case on for sum of households. For sum of households and sum of individuals the weight cases will be different.