Overview of the Sample Weights (Options 1 and 2 only)
Sample weights are numerical measurements that are essential to producing and evaluating estimates from sample survey data. They are intended to account for the probability that each respondent came into the sample and the differential effects of nonresponse, imperfect sampling frames, and other forces that affect the composition of the sample. Weights for surveys like the NCD Mobile Phone Survey are typically computed in the following five basic steps.
- Base Weight. A base weight is computed for each respondent as the inverse of the (unconditional) probability that the respondent was randomly selected in the sample. This respondent selection probability is in turn determined as the product of the probabilities for the sampling phases that led to selecting the respondent. The probability of selecting a respondent depends on the probability of selecting the respondent’s MPN and the probability of selection of that respondent in the second-phase strata (i.e., defined by the respondent’s age and sex).
- Adjustment for Multiplicity. Multiplicity occurs when a respondent has a greater probability of selection because the respondent could have been selected more than one time. In this survey, this could occur if a respondent has more than one MPN. The base weight should be adjusted to reflect the increased probability of selection (implies adjustment is <1.00 when respondent uses ≥ 1 MPN). The base weight is multiplied times the inverse of the multiplicity factor.
- Adjustment for Multi-user/Clustering. Clustering occurs when a respondent has a lower probability of selection because one of multiple users of the MPN could have been selected. In this survey, this could occur if a MPN is used by more than one potential eligible respondent. The multiplicity-adjusted weight should be adjusted to reflect the decreased probability of selection as described earlier. This implies adjustment is > 1.00 when the MPN is used by more than one individual who is 18 years and older. The base weight is multiplied by the number in the MPN’s cluster of individuals.
- Nonresponse Adjustment. The multiplicity or clustering-adjusted weight is multiplied times the inverse of the individual-level response rates for the selected eligible sample members who are similar to the respondent with respect to characteristics that (hopefully) correlate with key study outcome measurements and the propensity to respond in the survey.
- Calibration. The adjusted weight is then multiplied times a factor that calibrates the sample to the demographic distribution defined by characteristics that are likely to correlate with key study outcomes (e.g., age, sex, and level of education). These calibrated weights become the final adjusted sample weights that should be used for all analyses of the NCD Mobile Phone Survey data in each participating country.
The weighted distribution of the final adjusted weights with respect to the demographic variables used for calibration will thereby match the population counts with respect to these variables.
As noted in Section 4, the probability of selection for each stage of the sample design should be retained on the final analytic file for each selected household and individual. These factors should be computed and stored, and the process followed in computing them should be carefully documented at the time that the sample in each stage of selection is chosen.