I want to run a multinational logit model on my survey data which has a column for individual weighting.
My model is on household level and i used package nnet and multinom function in R,
How should i incorporate the weighting in the model?
They are individual weighting,should i use them for a model on household level?
Thank you
Much depends on how the sample and the resulting weights were generated. In the simplest circumstance, perhaps the Primary Sampling Unit was households chosen by simple random sampling, and then one adult per household was selected as the respondents. In that case, the weights might have been based solely on the number of adults in the household. Then the respondent in a 2-adult household would have gotten a weight of 2, since s/he represents self and one other household member; respondent in a 3-adult household got a weight of 3; etc. In that circumstance, the weights are only needed if conclusions are to be made about the population of individual adults. If, instead, the goal is conclusions about the population of households, you should use the unweighted data -- or equivalently, create a new column of weights, all equal to 1.0 -- or equivalently, create the new column of weights as the reciprocal of the original weights.
If the sampling was complex (such as disproportionate stratified sampling), the problem is tougher. If the weighting was generated through post-stratification to adjust for non-response bias by aligning the responding individuals with known individual-level population demographics, the problem is harder still. Ideally, you would generate a completely new set of weights, using the demographic characteristics of households rather than individuals. If that's not feasible, you may have to choose between two approximate solutions: either use the unweighted data (perhaps preferred if the disproportions in the stratification are not extreme and non-response bias is arguably modest), or use new weights that are the product of the original weights multiplied by the reciprocal of the number of adults in the household (perhaps preferred if the original PSU was individuals, and/or the household-level demographics and non-response are arguably similar to their individual-level counterparts).
Burke D. Grandjean thank you for the answer.
I translated the methodology of sampling data from Dutch to English:
"The distribution of these questionnaires among the respondents was done in collaboration with the research agency iVOX. The latter has a very large panel of respondents, which means that the representativeness at the level of the Belgian territory can be guaranteed. For two of the three questionnaires, the aim was to collect data from 8,000 adults, but also of 2,000 children. For adults, a stratified random sample was used by the research agency selection was made to reflect the different age groups, the level of education, the gender and the Belgian
region to include. It is important to note that it was not possible to direct children
to interview. The data of the children was therefore supplemented by the parents."
"Weighting Coefficients for Analysis Despite all precautions, the sample obtained will of course never correspond exactly to the Belgian population. As a result, a weighting factor was calculated for each observation (respondent) based on her over-/under-representation in the sample. Two sets of weights were calculated: a first at the level of the person and a second at the level of the personal level of the family. For the calculation of the weights, the CALMAR macro, developed by the INSEE (Institut National de la Statistique et des Etudes Economiques,France). The weighting procedure ensures that the data after weighting matches the population distributions to follow.
Weights persons
The weights of the persons were calculated based on the demographics by region,gender and age on the one hand and diploma on the other. For this, the joint distribution to region, gender and age category (2017) of STATBEL consulted. For the correction to diploma (20-64 year olds) the demographics of the Labor Force Survey (2017) used. The weights were calculated by both the joint distribution by region, sex and age as well as the division by diploma using the CALMAR macro".
The weighting column is like 1.63,1.30,1.57 for each ID.
I am confused now between individual and household weighting.
can I use the method to model without weighting data?
Best Regards,
This says, "Two sets of weights were calculated: a first at the level of the person and a second at the level of the personal level of the family." Despite some vagueness in the wording (and despite the omission of additional information about the second set of weights, along the lines of the additional information about the first set under "Weights persons"), it sounds to me like the second set of weights would be the one to use for a household (or "family") level analysis. Is there a column in your data containing that second set of weights? If so, I think that's what you should use to weight the data. Clearly, both sets of weights incorporate post-stratification for non-response, adjusting the sample demographics to match known population characteristics -- one for the population of adults, and the other, it seems, for the population of households/families.
If you don't have access to that second set of weights, and you don't have the time, knowledge, or information to generate a new set of household-level post-stratification weights yourself, then you have the two optons I mentioned before: use the unweighted data, or divide the existing weights by the number of eligible adults in the respondent's household to approximate a household-level post-stratification weighting factor. There's no way to know for sure which approach would yield the more representative sample results. I myself would probably lean toward the latter method (dividing the person-level weights by the number of adults in the household) -- but only if weights "at the level of the personal level of the family" either are unavailable from the original data set, or are found definitely not to mean household-level weights.
- Badges
- Science topic
- Similar topics
- Statistics
- Survey