You noted here that you wanted to "calibrate a sample." Did you mean you wanted calibration weights, as opposed to unadjusted survey weights? Calibration of survey weights, which also involves regression weights, whether apparent or not, is considered, I think, a part of the model-assisted design-based approaches. I included information on that in https://www.researchgate.net/publication/261508465_Use_of_Ratios_for_Estimation_of_Official_Statistics_at_a_Statistical_Agency. On page 16 there, I have the following:
"Calibration adjusts survey weights so that they would perfectly estimate known auxiliary data from that part of the auxiliary data associated with a sample."
I think that is the standard definition.
If this is what you mean, then I don't think age and sex are particularly useful for that approach. The "known auxiliary data" are continuous data to my knowledge, though others on ResearchGate may have more relevant knowledge and experience. Perhaps the reference list from the above linked paper might be helpful, say Brewer, and perhaps Deville & Särndal. The link to Brewer(1999) there goes to an archived, but readily available article provide, open access, by the Canadian Government, and there are a number of references in the reference list for that article.
This may not be what you had in mind, but when it comes to survey statistics, it is what occurs to me, though calibration in analytical chemistry is something else involving regression.
Thank you for your reply. indeed, it is the calibration of the survey weights by taking into account the auxiliary information on the structure of the total population by age group and sex.
Well, you may know a lot more than i do that is relevant here, but from my response above, please note
"...that in https://www.researchgate.net/publication/261508465_Use_of_Ratios_for_Estimation_of_Official_Statistics_at_a_Statistical_Agency ... (on) page 16 there, I have the following:
'Calibration adjusts survey weights so that they would perfectly estimate [a] known auxiliary data [total] from that part of the auxiliary data associated with a sample.'
"I think that is the standard definition."
Could age or gender be such auxiliary data? I think more in terms of continuous data, but I guess these variables could qualify. Would age be a good regressor for the y-variable being sampled? As for gender, wouldn't you just want to stratify by gender?
I could be wrong, but I don't see age or gender as being very helpful with calibration. But I could be wrong. I guess it is something like a "balanced sample" in a model-based approach, where you'd want the mean age of the sample to be the same as the mean age of the population. You might want the proportion of women in the sample to be the same as the proportion of women in the population. So in a model-assisted, but design-based (probability-of-selection-based) sample, perhaps calibration is what you want to effectively consider, say, age as something you want represented by the adjustment of survey weights, to be represented as age is in the population. Similarly with gender.
Did I just talk myself into thinking this might be a good idea?
:-)
I still have not considered the mechanics of using the usual calibration optimization equation here. That would be found in my 2012 paper linked above. Perhaps you should give it a try with a small sample, perhaps from a different set of data you have from a different population in the past, just to see that you can make the mechanics of the situation work, and any other preliminary evaluation you can make. If satisfied, perhaps you should then employ your idea.
I suppose that if you use age as an auxiliary variable, you would adjust weights so that the average age would be matched, but that does not mean much about the distribution of age, if that is of interest.
I fully agree with you that the variables used for calibration must be correlated with the variables of interest. in fact, I put myself in the case of a household survey whose objectives are multiple.