Comparison of Simulation Methods Using Historical Data in the U.S. International Price Program M.J. Cho, T-C. Chen, P.A. Bobbitt, J.A. Himelein, S.P. Paben, L.R. Ernst, and J.L. Eltinge U.S. Bureau of Labor Statistics Cho.Moon@bls.gov ICES III Session 14 – June 19, 2007 Disclaimer: The views expressed in this paper are those of the authors and do not necessarily reflect the policies of the BLS Outline I. U.S. International Price Program (IPP) II. Bootstrap Variance Estimator III. Adequacy of Approximations in Simulation Studies IV. Simulation Methods Considered V. Numerical Results VI. Conclusions 2 I. U.S. International Price Program (IPP) A. IPP is the nation’s primary source of information on price trends in the international trade of the U.S. economy Export and Import Price Indexes B. IPP involves population structure, sample design and estimation methods with a high degree of complexity 3 C. IPP Index Aggregation Tree Upper Level Strata Lower Level Strata Classification Groups Weight Groups (Company|CG) Items 4 D. IPP Sampling Design (Imported Goods) 1. Sample frame provided by U.S. Customs Border Protection, and divided into two biannual panels 2. Stratified multistage sampling within panels a. Within a broad product category (stratum), select establishments proportional to trading dollar value b. Within establishment, select detailed product categories (CGs) using systematic PPS c. Within Company|CG, select items 5 E. IPP Index Formula (modified Laspeyres index) LTR t parent t LTRchild Wchild child t 1 LTR child Wchild child t 1 LTR parent LTR STR LTR t t t 1 6 F. Weights 1.Item weight is weight group’s weight divided by number of items 2.Weight group’s weight is a trade dollar value divided by selection probability 3.CG weight is based on trade weights from the Census Bureau 4.Stratum weight is aggregation of CG weights 7 II. Bootstrap Variance Estimator A. Based on Rao, Wu and Yue (1992) 1. Draw nh 1 PSUs (establishments) with replacement from the nh sampled PSUs in each stratum 2. Define bootstrap weights: nh * nhi w whik nh 1 where whik is original sample weights * hik for the item k in PSU i of stratum h, and nhi* is the number of times that the PSU i is selected 8 3. Calculate the price STR using bootstrap weights 4. Repeat steps 1-3 B times 5. Compute the bootstrap variance estimator: 1 ˆ VBT B B b 1 ˆb ˆfull 2 where ˆb is a bootstrap STR estimator and ˆfull is the STR estimator from the original sample 9 B. Properties of VˆBT Under Realistic Approximations to IPP Population Structure, Sample Design and Estimation Methods 1. Over R replications of the “simulation design” compute 1 ˆ ˆ , VBT R R r 1 1 ˆ V VBT R 1 ˆ r , VˆBT r R r 1 VˆBT r - VˆBT 2 10 2. Evaluation Criteria Relative Bias = 1 R 1 r 1 R ˆ ˆ Vˆ BT 1 r Degrees of Freedom = ~ ˆ V VBT 1 2 1 2 VˆBT 2 11 III. A. Adequacy of Approximations in Simulation Studies Notation X ˆ Features of population and design ( X T True, X A Approximation ) Estimator, with cdf F (ˆ, X ) Q F (ˆ, X ) Specific functional of F (ˆ, X ) e.g. bias or MSE 12 B. Adequacy of Simulation & Approximation Qˆ {F{ˆ, X A )} for true Q{F{ˆ, X T )} ? 1. Components to include in X A a. Population structure b. Sampling steps c. Nonresponse d. Weighting e. Variance estimation 2. Approximations for each component in (1) 1. Taylor expansion: under conditions Qˆ {F {ˆ, X A )} Q{F {ˆ, X T )} Q{F {ˆ, X )} X X X X A XT T 2 ˆ, X )} 1 Q { F { X A X T X A XT 2 X X X X* (Replication Error) where X * X A (1 ) X T for 0,1 14 2. Four Cases Case 1: Small first derivatives, small X A XT Case 2: Small first derivatives, large X A XT Case 3: Large first derivative Case 4: Large second derivative 15 IV. Simulation Methods Considered A. Use a Single Method for Selection of Sample Units and Related Weights: Same sample units and weights for each month B. Three Methods to Construct Population of Item-Level STRs 1. Resampling Method 2. Fixed-One-Rate Method 3. CDF Interpolation Method 17 V. A. Numerical Results Selected Strata Stratum Stratum Description P07 Edible vegetables, roots, and tubers P08 Edible fruit and nuts; peel of citrus fruit or melons P09 Coffee, tea, mate and spices P22 Beverages, spirits, and vinegar P61 Articles of apparel and clothing accessories P74 Copper and articles thereof P90 Optical, photographic, measuring and medical instruments Item STR1 of Design Strata Original Variance of STR (P90) Relative Bias of STR (P90) Degrees of Freedom of STR (P90) VI. Conclusions A. For Complex Establishment Surveys Like IPP, Simulation-Based Evaluations of Est Properties Require Consideration of 1. Approximations X A to the true population, design and estimation procedures 2. Adequacy of the resulting approximations Q{F (ˆ, X A )} to the true properties Q{F (ˆ, X T )} 23 B. Future Work 1. For IPP: Other features of the pop and design e.g. Independence of STRs from sample units 2. Consider generalized variance estimator to improve stability 3. Explore the surface defined by Q F (ˆ, X ) in “neighborhoods” of the true X T 24