Partial Ranked Set Sampling Design By Abdul Haq Ph.D. Student, Department of Mathematics and Statistics, University of Canterbury, Christchurch, NZ. 1 Outline • • • • • • Simple random sampling. Ranked set sampling. Examples. Partial ranked set sampling. Simulation and case study. Main findings. 2 Estimate the mean height of Arabidopsis Thaliana (AT) plants 3 AT Population 4 Simple Random Sampling (SRS) 1. Select randomly π units from population. 2. Get careful measurements of selected plants. 3. Estimate population mean and variance based on this sample. A simple random sample of size π = 3 5 Simple random sampling (Estimation of population mean) A simple random sample of size π is drawn with replacement from the population having mean μ and variance σ2 say π1 , π2 , … , ππ , then the sample mean is πSRS 1 = π π ππ π=1 1. πSRS is an unbiased estimator of μ i.e. πΈ πSRS = μ. 1 2. πππ πSRS = π σ2 . 6 Ranked set sampling (Estimation of population mean) • • • Actual measurements are expensive. Ranking of sampling units can be done visually and cheaper. It provides more representative sample. Examples: • Estimating average height of students in NZ university. • Estimating average weight of students in NZ university. • Estimating average milk yield from cows in a farm. • Bilirubin level in jaundiced neonatal babies. 7 Ranked set sampling procedure • • • • • Identify π2 units, randomly. Randomly divide these units into π sets, each of size π. Rank units within each set. Select smallest ranked unit from first set of π units, second smallest ranked unit from second set, and so on, select largest ranked unit from last set. This gives a ranked set sample of size π. The above steps can be repeated for larger samples. 8 First set of π = π units After ranking Second set of π = π units After ranking Third set of π = π units After ranking 9 Diagram π = 3, π = 2, π = ππ = 6. Sample values Cycle 1 2 1 2 3 π111 π121 π131 π211 π221 π231 π311 π321 π331 π112 π122 π132 π212 π222 π232 π312 π322 π332 Now apply the RSS procedure to these 3 sets of 2 cycles. 10 Diagram π = 3, π = 2, π = ππ = 6. Judgment ranks Cycle 1 2 Here π π 1 2 3 ππ π:π π π1 2:3 1 π1 3:3 1 π2 2:3 1 ππ π:π π π2 3:3 1 π3 1:3 1 π3 2:3 1 ππ π:π π ππ π:π π π1 2:3 2 π1 3:3 2 π2 2:3 2 ππ π:π π π2 3:3 2 π3 1:3 2 π3 2:3 2 ππ π:π π π:π π , π π π:π π , … , π π π:π π , π π π:π π is a ranked set sample of size π = 6. Notes: 1. For each measured unit, we need π − 1 units. 2. All measured units are independent. 3. If ranking procedure is uniform for all cycles, then measurements from the same judgment class are i.i.d. but the selected units within each cycle are independent but NOT identically distributed. 11 Some Elementary Results • The population mean can be written as 1 π=π π π=1 π(π:π) . • The RSS mean estimator is 1 πRSS = π π π=1 ππ(π:π) . • πRSS is an unbiased estimator of π and more efficient than πSRS i.e. πΈ πRSS = π. πππ πRSS = πππ πSRS 1 − 2 π π (π(π:π) −π)2 . π=1 12 Partial Ranked Set Sampling (PRSS) Design • • • PRSS scheme is a mixture of both SRS and RSS designs. It involves less number of units compared with RSS. RSS design becomes a special case of PRSS design. PRSS Procedure Step 1: Define a coefficient π such that π = [π‘π], where 0 ≤ π‘ < 0.5. Step 2: firstly select 2π simple random samples each of size one. Step 3: For remaining π − 2π units, identify π − 2π sets each of size π. Apply RSS on these sets. Step 4: Above steps can be repeated π times for large samples. PRSS(π, π) represents PRSS design. 13 Diagram: Partial ranked set sample with 36 units PRSS (6,0) Judgment ranks Cycle 1 1 2 3 4 5 6 πΏπ π:π π π1 2:6 1 π1 3:6 1 π1 4:6 1 π1 5:6 1 π1 6:6 1 π2 1:6 1 πΏπ π:π π π2 3:6 1 π2 4:6 1 π2 5:6 1 π2 6:6 1 π3 1:6 1 π3 2:6 1 πΏπ π:π π π3 4:6 1 π3 5:6 1 π3 6:6 1 π4 1:6 1 π4 2:6 1 π4 3:6 1 πΏπ π:π π π4 5:6 1 π4 6:6 1 π5 1:6 1 π5 2:6 1 π5 3:6 1 π5 4:6 1 πΏπ π:π π π5 6:6 1 π6 1:6 1 π6 2:6 1 π6 3:6 1 π6 4:6 1 π6 5:6 1 πΏπ π:π π 14 Diagram: Partial ranked set sample with 26 units PRSS (6,1) Judgment ranks Cycle 1 2 3 4 5 6 πΏπ 1 π2 1:6 1 πΏπ π:π π π2 3:6 1 π2 4:6 1 π2 5:6 1 π2 6:6 1 π3 1:6 1 π3 2:6 1 πΏπ π:π π π3 4:6 1 π3 5:6 1 π3 6:6 1 π4 1:6 1 π4 2:6 1 π4 3:6 1 πΏπ π:π π π4 5:6 1 π4 6:6 1 π5 1:6 1 π5 2:6 1 π5 3:6 1 π5 4:6 1 πΏπ π:π π π5 6:6 1 πΏπ Diagram: Partial ranked set sample with 16 units PRSS (6,2) Judgment ranks Cycle 1 2 3 4 5 6 πΏ1 πΏπ 1 π3 1:6 1 π3 2:6 1 πΏπ π:π π π3 4:6 1 π3 5:6 1 π3 6:6 1 π4 1:6 1 π4 2:6 1 π4 3:6 1 πΏπ π:π π π4 5:6 1 π4 6:6 1 πΏπ πΏπ 16 Estimation of population mean The PRSS mean estimator is 1 πPRSS = π ππ=1 ππ + Its variance is πππ(πPRSS ) = π−π π=π+1 ππ(π:π) 2ππ2 π2 1 +π + π π=π−π+1 ππ . π−π 2 π=π+1 π(π:π) . For symmetric populations • XPRSS is an unbiased estimator of π. • πππ πPRSS ≤ πππ πSRS . i.e. πππ πPRSS = πππ πSRS − 1 π2 π−2π 2 π=1 (π(π:(π−2π)) −π) . 17 Simulation study: Symmetric populations (perfect ranking) 7 Uniform(0,1) 5 3 4 k=0 k=1 k=2 k=3 1 1 20 30 40 50 10 20 30 Number of units Logistic(0,1) Beta(6,6) 40 50 7 Number of units RSS PRSS PRSS PRSS 5 3 4 k=0 k=1 k=2 k=3 1 2 3 1 2 Relative Efficiency k=0 k=1 k=2 k=3 4 5 6 RSS PRSS PRSS PRSS 6 7 10 Relative Efficiency RSS PRSS PRSS PRSS 2 Relative Efficiency 5 3 4 k=0 k=1 k=2 k=3 2 Relative Efficiency 6 RSS PRSS PRSS PRSS 6 7 Normal(0,1) 10 20 30 Number of units 40 50 10 20 30 40 50 Number of units 18 Simulation study: Asymmetric populations (perfect ranking) 4.0 3.5 3.0 2.0 2.5 k=0 k=1 k=2 k=3 1.0 20 30 40 50 10 20 30 40 Number of units Lognormal(0,1) Gamma(0.5,2) k=0 k=1 k=2 k=3 2.0 2.5 3.0 3.5 RSS PRSS PRSS PRSS 50 1.0 1.5 2.0 1.5 1.0 Relative Efficiency k=0 k=1 k=2 k=3 2.5 3.0 3.5 RSS PRSS PRSS PRSS 4.0 Number of units 4.0 10 Relative Efficiency RSS PRSS PRSS PRSS 1.5 3.0 1.5 2.0 2.5 k=0 k=1 k=2 k=3 1.0 Relative Efficiency 3.5 RSS PRSS PRSS PRSS Weibull(0.5,1) Relative Efficiency 4.0 Exponential(1) 10 20 30 Number of units 40 50 10 20 30 40 50 Number of units 19 Simulation study: Bivariate Normal Distribution (imperfect ranking) Bivariate Normal(0,0,1,1,0.80) 10 20 30 40 2.5 2.0 k=0 k=1 k=2 k=3 1.5 50 10 20 30 40 50 Bivariate Normal(0,0,1,1,0.50) Bivariate Normal(0,0,1,1,0.20) k=0 k=1 k=2 k=3 1.02 1.04 RSS PRSS PRSS PRSS 1.00 1.1 Relative Efficiency k=0 k=1 k=2 k=3 1.2 1.3 1.4 RSS PRSS PRSS PRSS 1.06 Number of units 1.5 Number of units 1.0 Relative Efficiency RSS PRSS PRSS PRSS 1.0 2 3 4 k=0 k=1 k=2 k=3 1 Relative Efficiency 5 RSS PRSS PRSS PRSS Relative Efficiency 6 3.0 Bivariate Normal(0,0,1,1,0.99) 10 20 30 Number of units 40 50 10 20 30 40 50 Number of units 20 An application to Conifer trees data Study variable Auxiliary variable Correlation coefficient π: Height of trees (ft). π: Diameter of trees at chest level (cm). π: 0.908 Relative efficiencies of the estimators of population mean π RSS PRSS PRSS PRSS π=0 π =1 π =2 π =3 4 π (ranking on π) 1.92037 1.38698 _______ _______ 4 π(ranking on π) 1.91247 1.38545 _______ _______ 5 π (ranking on π) 2.21737 1.51641 1.15711 _______ 5 π(ranking on π) 2.20384 1.51553 1.15685 _______ 6 π (ranking on π) 2.52342 1.61253 1.26187 _______ 6 π(ranking on π) 2.49267 1.60968 1.26067 _______ 7 π (ranking on π) 2.80967 1.68997 1.32322 1.10689 7 π(ranking on π) 2.77011 1.68898 1.32021 1.10399 See Platt et al. (1988). 21 Main Findings • • • • PRSS requires less number of units, which helps in saving time and cost. RSS is special case of PRSS design. Mean estimators under PRSS are better than SRS for perfect and imperfect rankings. PRSS can be used as an efficient alternative to SRS design. 22