Calculate Sampling Weights

Question

I have sampled 300 respondents, stratified by zones and have conducted a survey, prior to carrying a census (ideally would have been the other way round). I would like to compute sampling weights to make the sample more representative of the population. I have heard of two approaches:

1) Generating the base sampling weight (probability of being sampled by zone), and generating weights for each of the demographic variable (gender, age category, education, etc.) and simply multiply the weights;

2) Creating a dummies for each combination of demographic variables (i.e. male aged 35-50 with a bachelor's degree, male aged 50-65 with a bachelor's degree, female aged 18-35 with a PhD, etc.) and compute probability weights for each combination.

The second approach makes more intuitive sense, though extremely labor intensive (I have more than 30 demographic variables!), but would be grateful if anyone from the community could share their thoughts.

Thanks

Answer 1

The second approach is called "post-stratificaiton". It is usually much easier to do than anything else: if it's all digitized, you create the group identifiers in your software of preference, count the total in the census/population, count the total in the sample, and divide one by the other.

The first approach does not do much, and cannot be recommended. It performs one cycle of what is called "raking" where you adjust weights margin by margin until the weights stabilize. To get meaningful weights, you have to repeat those steps in a cycle, and that is more time consuming, and unless you have access to statistical software that you can figure this method out in (Stata and R both have it, but one would need to know what they are doing), you are better off with post-stratification.

As a reference, see https://www.stata.com/bookstore/survey-weights/ although you need approximately three pages from that book. But I felt obliged to put a reference of some kind :)

Answer 2

Interesting question! I am currently learning weighting as well. If I understand correctly, the first approach is called 'calibration' and the second approach is called 'post-stratification'. It seems to me both approaches have their advantages and disadvantages. For post-stratification, it's difficult to implement if auxiliary variables are continuous (if you don't categorise them). I find a paper particular useful: https://www.surveypractice.org/article/2809-post-stratification-or-non-response-adjustment