This document discusses applying the jackknife method to cluster sampling. Specifically, it explains that when forming random groups for jackknifing, we work with ultimate clusters rather than individual sampling units. The estimators θ̂(α) are computed by eliminating clusters, not individual units. As an example, it shows how the jackknife would be applied to estimate a population total based on probabilities proportional to size sampling of clusters.
This document discusses applying the jackknife method to cluster sampling. Specifically, it explains that when forming random groups for jackknifing, we work with ultimate clusters rather than individual sampling units. The estimators θ̂(α) are computed by eliminating clusters, not individual units. As an example, it shows how the jackknife would be applied to estimate a population total based on probabilities proportional to size sampling of clusters.
This document discusses applying the jackknife method to cluster sampling. Specifically, it explains that when forming random groups for jackknifing, we work with ultimate clusters rather than individual sampling units. The estimators θ̂(α) are computed by eliminating clusters, not individual units. As an example, it shows how the jackknife would be applied to estimate a population total based on probabilities proportional to size sampling of clusters.
This document discusses applying the jackknife method to cluster sampling. Specifically, it explains that when forming random groups for jackknifing, we work with ultimate clusters rather than individual sampling units. The estimators θ̂(α) are computed by eliminating clusters, not individual units. As an example, it shows how the jackknife would be applied to estimate a population total based on probabilities proportional to size sampling of clusters.
into k random groups of size m h , for h = 1, . . . , L, and let θ̂(α) denote the estimator of θ obtained after removing the α-th group of observations from each stratum. Define the pseudovalues
θ̂α = k θ̂ − (k − 1) θ̂(α) .
Then, the estimator of θ and the estimator of its variance are
Little is known about the relative merits of this method vis-à-vis earlier methods. It does seem that the estimators in (4.5.9) have computational advantages over earlier methods. Unless k is large, however, (4.5.9) may be subject to greater instability.
4.6. Application to Cluster Sampling
Throughout this chapter, we have treated the case where the elementary units and sampling units are identical. We now assume that clusters of elementary units, comprising primary sampling units (PSUs), are selected, possibly with several stages of subsampling occurring within each selected PSU. We con- tinue to let θ̂ be the parent sample estimator of an arbitrary parameter θ. Now, however, N and n denote the number of PSUs in the population and sample, respectively. No new principles are involved in the application of jackknife methodology to clustered samples. When forming k random groups of m units each (n = mk), we simply work with the ultimate clusters rather than the elementary units. The reader will recall from Chapter 2 that the term ultimate cluster refers to the aggregate of elementary units, second-stage units, third-stage units, and so on from the same primary unit. See rule (iii), Section 2.4. The estimator θ̂(α) is then computed from the parent sample after eliminating the α-th random group of ultimate clusters (α = 1, . . . , k). Pseudovalues θ̂α , Quenouille’s estimator θ̄ˆ , and the jackknife variance estimators v1 (θ̄ˆ ) and v2 (θ̄ˆ ) are defined in the usual way. As an illustration, consider the estimator n 1! θ̂ = Ŷ = ŷi / pi (4.6.1) n i=1
of the population total θ = Y based on a pps wr sample of n primaries, where ŷi
denotes an estimator of the total in the i-th selected primary based on sampling at 4.6. Application to Cluster Sampling 183
the second and successive stages. For this problem, the θ̂(α) are defined by m(k−1) ! 1 θ̂(α) = Ŷ(α) = ŷi / pi , (4.6.2) m(k − 1) i=1
where the summation is taken over all selected PSUs not in the α-th group. Quenouille’s estimator is k 1! θ̄ˆ = Ȳˆ = Ŷ(α) , k α=1 θ̂α = Ŷα = k Ŷ − (k − 1) Ŷ(α) ,
and !k 1 v1 (θ̄ˆ ) = (Ŷα − Ȳˆ )2 k(k − 1) α=1
is an unbiased estimator of the variance of θ̄ˆ .
If the sample of PSUs is actually selected without replacement with inclu- sion probabilities πi = npi , then the estimators θ̂ and θ̂(α) take the same form as indicated in (4.6.1) and (4.6.2) for with replacement sampling. In this case, the estimator v1 (θ̄ˆ ) will tend to overestimate the variance of θ̄ˆ . See Section 2.4.5 for some discussion of this point. Also, using a proof similar to Theorem 2.4.5, it can be shown that the bias in v1 (θ̄ˆ ) arises only in the between PSU component of variance. For the example θ̂ = Ŷ , regardless of whether the sample primaries are drawn with or without replacement, the reader will note that θ̄ˆ = θ̂ and v1 (θ̄ˆ ) = v2 (θ̄ˆ ) because the parent sample estimator θ̂ is linear in the ŷi . For nonlinear θ̂, the estimators θ̄ˆ and θ̂ are generally not equal, nor are the estimators of variance v1 (θ̄ˆ ) and v2 (θ̄ˆ ). For the nonlinear case, exact results about the moment properties of the estimators generally are not available. Approximations are possible using the theory of Taylor series expansions. See Chapter 6. Also see Appendix B for some discussion of the asymptotic properties of the estimators. If the sample of PSUs is selected independently within each of L ≥ 2 strata, then we use the rule of ultimate clusters together with the techniques discussed in Section 4.5. The methods are now based upon the estimators θ̂(hi) formed by deleting the i-th ultimate cluster from the h-th stratum. Continuing the example θ̂ = Ŷ , we have " # !L ! L 1 θ̂ = 1 + qh θ̂ − qh θ̂(h·) h=1 h=1 " # L ! L· ! = 1+ qh Ŷ − qh Ŷ(h·) , (4.6.3) h=1 h=1
