Chapter 3. Replicate Weight Construction
To calculate correct standard errors for comparing estimates, it is crucial to account for the complex survey design. Relying on methods that assume a simple random sample will typically underestimate the true sampling error associated with estimates from the NHTS. Replicate weights are provided to allow you to correctly compute estimates of standard errors using appropriate software such as WesVar or SUDAAN.
Person-level Replicates
Replicate Construction
To create replicate weights, the data were sorted by geographical characteristics, MSA status, and census division. After sorting, 99 replicates were created using the delete-one jackknife method (JK1). The ith replicate was constructed by setting the ith item of every 100 records of the full sample weight equal to zero. This procedure was done separately for both the Pre-9/11 and Post-9/11 data files.
Adjustment ror Delete-One Jackknife Method
After “deleting one,” the totals of each of 99 replicates are not equal to the control totals. To adjust for this, the replicates are multiplied by factors of each of the 99 replicates for both the Pre-9/11 and Post-9/11 data files, separately.
Factor = Population Total / Sum of the ith replicate, where
Population total = 277,208,169
Raking
Again, the raking processes were performed on each of the 99 replicates for both the Pre-9/11 and Post-9/11 person data files, applying the same poststratification process for each of the replicates as for the full sample weight. These raking processes are done in exactly the same manner as described above, using the same eight dimensions and the same control totals.
Long-distance Trip Replicates
Long-distance trip replicate weights are simple functions of the person-level replicate weights described in section 3.1, modified for the purpose of producing annual estimates of the number of person trips taken by the respondents. The long-distance trip replicate weights are equal to the final person-level replicates’ weights multiplied by 365/28.
|