In conducting research with the use of questionnaires where some of the respondents are not accessible for certain reason. Some replacements were done randomly while others were purposively. What is the best methodology to be used when responses are to be weighed.
A commonly applied correction technique is a weighting adjustment. It assigns an adjustable weight to each respondent. Under-represented respondents get a weight larger than 1 and over-represented one get a weight lesser than 1. The weighted values are used, instead of the values of the variable for calculation of totals, percentages or means, etc. A weighting adjustment technique can be carried out given proper auxiliary variables are available. You can find out a detailed discussion from the following paper:
Zhang Li-Chun (2002). A method of weighting adjustment for survey data subject to nonignorable nonresponse . Available at https://www.uni-trier.de/fileadmin/fb4/projekte/SurveyStatisticsNet/DRPS2.pdf
Dr. Shamindra,
Many thanks for your valuable contribution and the suggested paper.
It is well appreciated.
Olusegun -
Because you said some respondents are not accessible, you are, as Shamindra indicated, dealing with possibly nonignorable nonresponse, meaning that a simple random sample would miss the differences between the 'mechanism' which generates the accessible cases from that which would be behind the inaccessible cases. Yes, response propensity groups may be a good idea, or some other stratification. Within those groups, if you do have auxiliary data available for the population, then it may be a good idea to use regression within each such group/stratum, so that you may generate values for the estimated variances of the prediction errors. If their square roots are relatively large compared to the 'predicted' values, you want to know that, as any bias not alleviated by the data groupings would mean actual uncertainty is even greater. Such regressions may be relatively easy, interpretable, and helpful.
That would be an imputation approach. With nonresponse, or here, inaccessible data, imputation and weighting adjustment are separate, but related methods. Inaccessible data would more likely indicated nonignorable nonresponse, whereas ignorable nonresponse is not really ignorable, just that simple random sampling might then be an appropriate 'fix.'
Cheers - Jim Knaub
I assume that because you are talking about weighting that you are dealing with quantitative data. If you have a qualitative study, then I think that you may just have to carefully define what part of the population you have been able to study, and what part was not possible to directly study.
Prof. James, Many thanks for sharing from your wealth of knowledge. Your contribution is noted and well appreciated. Olusegun.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Applied Statistics
- Survey Design
- Questionnaire