Survey data Analysis- standardization weights

[epidem_delhi]

HI.

I am analysing data collected on a non-PPS community derived sample of males and females to study their tobacco related habits. The sample were selected from three strata, rural, urban and urban-slums. Within each stratum, 250 males and 25 females were selected in each 10 year age group starting 15 years till 65 years (5 groups, yielding a total sample size of 1250 males and 1250 females in each stratum, all in all 7500 individuals interviewed).

The variables of interest are smoking prevalence, mean number of cigarettes smoked, relationship between smoking and place of residence (3 Strata) after adjusting for the effects of gender (2 categories), age (5 categories), education (4 categories) and occupation (4 categories).

We put in normalized probability weights as per census of India reported population distributions. (using the technique of weight = 1/probability of selection, weights normalized to as per the sample size studied)

Thus we estimated the age adjusted population proportions of individuals using cigarettes, for each stratum separately. This was easy. Now we have to calculate the age adjusted mean number of cigarettes smoked across each stratum. Then we have to compare the age adjusted consumption across the three strata.

This becomes tricky because:

• the base of the current analysis will be the cigarettes smokers and not the whole sample.
• The proportion of cigarette smokers varies by age groups.
• the mean number of cigarettes consumed also varies across the age groups.

So do we have to recalculate the weights? If yes, what is the procedure to be followed for calculation of weights. Any help on this would be greatly appreciated.

Subsequently we are planing to use logistic regression to calculate adjusted odds ratios for smoking and place of residence (stratifying variable). Can we use the weights we already calculated for this analysis?

I am basically a doctor with training in biostatistics in Delhi and have consulted bio-statisticians in my college. They don't seem to be having much idea about analysis of such designs.

Thanks
Vivek