R package sae: Methodology

Size: px
Start display at page:

Download "R package sae: Methodology"

Transcription

1 R package sae: Methoology Isabel Molina *, Yolana Marhuena March, 2015 Contents 1 Introuction 3 2 Function irect Sampling without replacement Sampling with replacement Function pssynt 5 4 Function ss 6 5 Function eblupfh FH fitting metho ML fitting metho REML fitting metho Function msefh FH fitting metho ML fitting metho REML fitting metho Function eblupsfh ML fitting metho REML fitting metho Function msesfh REML fitting metho ML fitting metho * Department of Statistics, Universia Carlos III e Mari. Aress: C/Mari 126, Getafe (Mari), Spain, Tf: , Fax: , isabel.molina@uc3m.es 1

2 9 Function pbmsesfh Function npbmsesfh Function eblupstfh REML fitting metho Function pbmsestfh Function eblupbhf Function pbmsebhf Function ebbhf Function pbmseebbhf 32 References 33 2

3 1 Introuction The R package sae estimates characteristics of the omains of a population, when a sample is available from the whole population, but not all the target omains have a sufficiently large sample size. This ocument escribes the methoology behin the functions of the package, giving the exact formulas an proceures implemente in each function. The following notation will be use through all this ocument. U enotes the target population, assume to be partitione into D subsets U 1,..., U D calle inistinctly omains or areas, of respective sizes N 1,..., N D. The measurement of the variable of interest for j-th iniviual within -th area is enote Y j an y = (Y 1,..., Y N ) is the vector with the measurements for -th area. The target omain parameters have the form δ = h (y ), = 1,..., D, for known real measurable functions h (). Particular cases of δ are the omain means δ = Ȳ = N 1 N j=1 Y j, = 1,..., D. These parameters are estimate using the information coming from a ranom sample s of size n rawn from U. Here, s is the subsample of size n from omain U, = 1,..., D, where n = D =1 n, an r = U s is the sample complement from omain, of size N n, = 1,..., D. 2 Function irect Function irect estimates the area means Ȳ, = 1,..., D, where, for each area, the estimator of Ȳ is calculate using only the sample ata s from that same omain. The obtaine estimators are unbiase with respect to the sampling esign (esign-unbiase). The call to the function is irect(y, om, sweight, omsize, ata, replace = FALSE) The particular estimator calculate by the function epens on the specifie arguments replace an sweight, relate to the sampling esign. If replace is TRUE, the sampling esign is assume to be with replacement an otherwise without replacement. The sampling weights shoul be introuce through the argument sweight, but when this argument is roppe, the function assumes simple ranom sampling (SRS). Uner SRS, the sizes of the omains N, = 1,..., D (omsize) are unnecessary. 2.1 Sampling without replacement Consier that the sample s is rawn without replacement within omain U, = 1,..., D. Let π j be the inclusion probability of j-th unit from -th omain in the corresponing omain sample s an let w j = π 1 j be the corresponing sampling 3

4 weight. The unbiase estimator of Ȳ is the Horvitz-Thompson estimator, given by ˆȲ DIR = N 1 Y j = N 1 w j Y j. π j s j j s Now let π,jk be the inclusion probability of the pair of units j an k from -th omain in the sample s. The sampling variance is given by DIR V π ( ˆȲ ) = 1 N Y 2 N N j Y j Y k (1 π N 2 j ) + 2 (π,jk π j π k ) π j π j π k. j=1 If π,jk > 0, (j, k), an unbiase estimator of this variance is given by DIR ˆV π ( ˆȲ ) = 1 j (1 π N 2 πj 2 j ) + 2 Y j Y k (π,jk π j π k ) π j s j π k π,jk. (1) j s Y 2 This estimator requires the secon orer inclusion probabilities, but many times these are not available. Then, it is common to fin an approximation of (1). A simple approximation is obtaine by consiering π,jk π j π k, which hols exactly in the case of Poisson sampling. This approximation makes the secon sum in (1) equal to zero an leas to the estimator DIR ˆV π ( ˆȲ ) = 1 N 2 j=1 k=1 k>j k s k>j 1 π j π 2 j s j Writing the approximate unbiase estimator of the variance in terms of the sampling weights w j (sweight), we get the expression provie by function irect when the argument sweight is specifie, DIR ˆV π ( ˆȲ ) = 1 N 2 Y 2 j. j. j s w j (w j 1)Y 2 Uner SRS without replacement, the result of the previous estimator oes not coincie with the usual unbiase estimator. Thus, when the argument sweight is roppe, the function irect assumes SRS without replacement an returns the usual unbiase estimators uner this esign ˆȲ = ȳ = n 1 j s Y j, ˆVπ ( DIR ˆȲ ) = (1 f ) S2, n is the omain quasi- where f = n /N is the omain sampling fraction an S 2 variance S 2 = (n 1) 1 j s (Y j ȳ ) 2. 4

5 2.2 Sampling with replacement Now consier the case in which s is rawn with replacement within omain U, = 1,..., D, with initial selection probabilities P j, j = 1,..., N. In this case, the unbiase estimator of the omain mean is given by ˆȲ DIR = N 1 j s Y j n P j. To obtain this estimator, the argument sweight must contain the vector of weights calculate as w j = (n P j ) 1, j s, = 1,..., D. Using these weights, the unbiase estimator of Ȳ calculate by the function irect with replace=true has exactly the same shape as that one in Section 2.1, i.e. ˆȲ DIR = N 1 j s w j Y j. The sampling variance of this estimator is given by V π ( DIR ˆȲ ) = 1 n N j=1 ( ) 2 Yj N P Ȳ P j. j an using again w j = (n P j ) 1, we obtain the unbiase variance estimator calculate by the function irect, DIR ˆV π ( ˆȲ ) = 1 ( ) 2 Yj DIR ˆȲ = 1 ( f w j Y j ˆȲ ) 2. n N j s P j n j s Uner SRS with replacement, the population sizes N (omsize) are not neee. Thus, when the argument omsize is roppe, the function assumes SRS an calculates the classical unbiase estimators ˆȲ DIR = ȳ = n 1 3 Function pssynt j s Y j, ˆVπ ( DIR ˆȲ ) = n 1 S2. Inirect estimators borrow strength from other omains by making assumptions establishing some homogeneity relationship among omains. The post-stratifie synthetic estimator is a basic inirect estimator. It assumes that ata are istribute into K groups (calle post-strata) that cut across the omains, an such that the within group mean is constant across omains. The groups are assume to have sufficient sample sizes to allow obtaining accurate irect estimates of the group means. These assumptions allow us to estimate a omain mean using a weighte combination of the (reliable) estimates of the group means. The function pssynt calculates post-stratifie synthetic estimates of the omain means Ȳ, = 1,..., D. The call to this function is 5

6 pssynt(y, sweight, ps, omsizebyps, ata) More specifically, the post-stratifie synthetic estimator consiers that there is a grouping variable (ps) which ivies the ata into K post-strata. The population count in the crossing between post-stratum k an omain, N k (omsizebyps), is assume to be known for all k an with N = K k=1 N k. Then, the mean of omain can be calculate as a weighte combination of the means in the crossings of omain with each post-strata Ȳk, k = 1,..., K, as follows Ȳ = 1 N K k=1 N k Ȳ k. (2) Uner the assumption of constant mean across omains within each post-stratum, Ȳ k = Ȳ+k, k = 1,..., K, where Ȳ+k enotes the mean of post-stratum k, an estimator of Ȳ can be obtaine by replacing Ȳ+k = Ȳk in (2) by some irect estimate of Ȳ+k, k = 1,..., K. Thus, we estimate the omain mean Ȳ using all the observations from the poststrata that cut across that omain. To estimate Ȳ+k, we consier the ratio HT estimator, given by ˆȲ +k = Ŷ +k DIR ˆN DIR +k, (3) where Ŷ +k DIR is the irect estimator of the total Y+k DIR in post-stratum k an ˆN +k DIR is the irect estimator of the population count in the same post-stratum, N +k, calculate using the sampling weights w j (sweight) of the units in that post-stratum. Replacing (3) for Ȳk in (2), we obtain the post-stratifie synthetic estimate returne by the function pssynt, ˆȲ SY N = 1 N K N k ˆȲ +k. k=1 4 Function ss The irect estimator of Ȳ is inefficient for a omain with a small sample size. On the other han, the post-stratifie synthetic estimator is biase when the means across omains within a post-stratum are not constant, which is likely to occur in practice. To balance the bias of a synthetic estimator an the instability of a irect estimator, Drew, Singh & Chouhry (1982) propose to take a weighte combination (or composition) of the two, with weight epening on the sample size of the omain. Thus, the function ss estimates the omain means Ȳ, = 1,..., D by a kin of composite estimators calle sample-size epenent estimators. The call to this function is ss(om, sweight, omsize, irect, synthetic, elta = 1, ata) 6

7 As mentione above, the sample-size epenent estimator is obtaine by composition of a irect estimator (irect) an a synthetic estimator (synthetic), both specifie by the user, that is, ˆȲ C = φ ˆȲ DIR SY N + (1 φ ) ˆȲ. The composition weight φ is obtaine as { DIR 1, ˆN δn φ = ; ˆN DIR DIR /(δn ), ˆN < δn, for N known (omsize) an for a given constant δ > 0 (elta). Thus, for a DIR omain with sample size large enough so that the estimate count ˆN is greater than δn, the sample size epenent estimator becomes the irect estimator ˆȲ DIR. Otherwise, it becomes the composition of the irect an the synthetic SY N estimator ˆȲ. The constant δ (elta) controls how much strength to borrow from other omains by attaching more or less weight to the synthetic estimator, with a large value of δ meaning to borrow more strength. 5 Function eblupfh Fay-Herriot (FH) moels were introuce by Fay & Herriot (1979) to obtain small area estimators of meian income in small places in the U.S. These moels are well known in the literature of small area estimation (SAE) an are the basic tool when only aggregate auxiliary ata at the area level are available. The function eblupfh estimates omain characteristics δ = h (y ), = 1,..., D, base on the mentione FH moel, an the call to this function is eblupfh(formula, varir, metho = "REML", MAXITER = 100, PRECISION= , ata) The basic FH moel is efine in two stages. First, since true values δ are DIR not observable, our ata will be the irect estimates ˆδ (left han sie of formula). These estimates have an error an this error might be ifferent for each area because samples sizes in the areas are generally ifferent. Thus, in the first stage, we assume the following moel representing the error of irect estimates, ˆδ DIR in = δ + e, e N(0, ψ ), = 1,..., D, (4) an referre to as sampling moel, where ψ is the sampling variance (varir) DIR of irect estimator ˆδ given δ, assume to be known for all. In a secon stage, true values δ are assume to be linearly relate with a vector of auxiliary variables (right han sie of formula), δ = x β + v, v ii N(0, A), = 1,..., D, (5) 7

8 where v is inepenent of e, = 1,..., D. This is calle the linking moel because it links all the areas through the common moel parameter β. Replacing (5) in (4), we obtain ˆδ DIR = x β + v + e, v ii N(0, A), e in N(0, ψ ), = 1,..., D, or equivalently, ˆδ DIR in N(x β, ψ + A), = 1,..., D. (6) In matrix notation, (6) may be written as y N{Xβ, Σ(A)}, where y = (ˆδ 1 DIR DIR,..., ˆδ D ), X = (x 1,..., x D ) an Σ(A) = iag(a + ψ 1,..., A + ψ D ). The best linear unbiase preictor (BLUP) of δ is given by δ = DIR ˆδ ψ {ˆδDIR x A + ψ β(a) } = {1 B (A)}ˆδ DIR + B (A) x β(a), (7) where B (A) = ψ /(A+ψ ) an β(a) is the maximum likelihoo (ML) estimator of β obtaine from (6) an also the weighte least squares (WLS) estimator of β without normality assumption, given by β(a) = {X Σ 1 (A)X} 1 X Σ 1 (A)y = { D } 1 D (A + ψ ) 1 x x (A + ψ ) 1 x ˆδDIR. (8) =1 =1 Expression (7) shows that δ is obtaine by composition of the irect estimator ˆδ an the regression synthetic estimator x β(a), with more weight DIR attache to the irect estimator when ψ is small relative to the total variance A + ψ, which means that the irect estimator is reliable, an more weight attache to the regression synthetic estimator x β when the irect estimator is not reliable enough an then more strength is require to borrow from the other omains. Since A is unknown, in practice it is replace by a consistent estimator Â. Several estimation methos (metho) for A are consiere incluing a moment estimator calle Fay-Herriot (FH) metho, maximum likelihoo (ML) an restricte (or resiual) ML (REML), see the next subsections. Substituting the obtaine estimator  for A in the BLUP (7), we get the final empirical BLUP (EBLUP) returne by eblupfh, an given by ˆδ = δ (Â) = {1 B (Â)}ˆδ DIR + B (Â)x ˆβ, (9) where ˆβ = β(â). Function eblupfh elivers, together with the estimate moel coefficients, i.e. the components of ˆβ, their asymptotic stanar errors given by the iagonal 8

9 elements of the Fisher information epening on the specifie estimation metho (metho), the Z statistics obtaine by iviing the estimates by their stanar errors, an the p-values of the significance tests. Since for large D, the three possible estimators satisfy ˆβ N { β, I 1 (β) }, where I(β) is the Fisher information, then the Z statistic for a coefficient β j is Z j = ˆβ j / v jj, j = 1,..., p, where v jj is the estimate asymptotic variance of ˆβ j, given by the j-th element in the iagonal of I 1 ( ˆβ). Finally, for the test p-values are obtaine as H 0 : β j = 0 versus H 1 : β j 0, p-value = 2 P (Z > Z j ), where Z is a stanar normal ranom variable. Three ifferent gooness of fit measures are also elivere by function eblupfh. The first one is the estimate log-likelihoo l(â, ˆβ; y), obtaine by replacing the obtaine estimates  an ˆβ in (12). The secon criteria is the Akaike Information Criteria (AIC), given in this case by AIC = 2 l(â, ˆβ; y) + 2(p + 1). Finally, the Bayesian Information Criteria (BIC) is obtaine as 5.1 FH fitting metho BIC = 2 l(â, ˆβ; y) + (p + 1) log(d). FH metho gives an estimator of A base on a moments metho originally propose by Fay & Herriot (1979). The FH estimator is given by ÂF H = max(0, A F H ) with A F H obtaine iteratively as the solution of the following nonlinear equation in A, D =1 (A + ψ ) 1 {ˆδ DIR x β(a)} 2 = D p. (10) This equation is solve using an iterative metho such as the Fisher-scoring algorithm. For this, let us efine S F H (A) = D =1 (A + ψ ) 1 {ˆδ DIR x β(a)} 2 D p. 9

10 By a first orer Taylor expansion of S F H (A F H ) aroun the true A, we get 0 = S F H (A F H ) S F H (A) + S F H(A) A (A F H A). (11) The Fisher-scoring algorithm replaces in this equation, the erivative S F H (A)/ A by its expectation E{ S F H (A)/ A}, which in this case is equal to E { S } F H(A) = A Then, solving for A F H in (11), we get A F H = A + D (A + ψ ) 1. =1 [ { E S }] 1 F H(A) S F H(A). A This algorithm is starte taking A = A (0) F H, an then in each iteration it upates the estimate of A with the upating equation A (k+1) F H { = A(k) F H [E + S } ] 1 F H(A) S F H(A (k) F H A ). A=A (k) F H In the function eblupfh, the starting value is set to A (0) F H = meian(ψ ). It stops either when the number of iterations k > MAXITER where MAXITER can be chosen by the user, or when A (k+1) F H A(k) F H A (k) F H Convergence of the iteration is generally rapi. 5.2 ML fitting metho < PRECISION. The moel parameters A an β can be estimate by ML or REML proceures base on the normal likelihoo (6). In fact, uner regularity conitions, the estimators erive from these two methos (an using the Normal likelihoo) remain consistent at orer O p (D 1/2 ) even without the Normality assumption, for etails see Jiang (1996). ML estimators of A an β are obtaine by maximizing the log-likelihoo, given by l(a, β; y) = c 1 2 log Σ(A) 1 2 (y Xβ) Σ 1 (A)(y Xβ), (12) where c enotes a constant. Taking erivative of l with respect to β an equating to zero, we obtain the equation that gives the ML (or WLS) estimator (8). The 10

11 ML equation for A is obtaine taking erivative of l with respect to A an equating to zero, an is given by D =1 (A + ψ ) 2 {ˆδ DIR x β(a)} 2 = D (A + ψ ) 1. (13) =1 Again, Fisher-scoring algorithm is use to solve this equation. efine as S ML (A) = l(a, β; y)/ A an is given by The score is S ML (A) = D =1 (A + ψ ) 2 {ˆδ DIR x β(a)} 2 D (A + ψ ) 1. =1 The Fisher information for A is obtaine by taking expectation of the negative erivative of S ML (A), an is given by I ML (A) = E { S } ML(A) = 1 A 2 Finally, the upating equation for the ML estimator of A is A (k+1) ML D (A + ψ ) 2. (14) =1 { } 1 = A(k) ML + I ML (A (k) ML ) SML (A (k) ML ). Starting value A (0) ML an stopping criterion are the same as in the FH metho escribe above. If A ML is the estimate obtaine in the last iteration of the algorithm, then the final ML estimate returne by eblupfh is ÂML = max(0, A ML ). 5.3 REML fitting metho The REML estimator of A is obtaine by maximizing the so calle restricte likelihoo, which is the joint p..f. of a vector of D p inepenent contrasts F y of the ata y, where F is an D (D p) full column rank matrix satisfying F F = I D p an F X = 0 (D p) p. The restricte likelihoo of F y oes not epen on β, an the log-restricte likelihoo is It hols that where l R (A; y) = c 1 2 log F Σ(A)F 1 2 y F {F Σ(A)F} 1 F y. F {F Σ(A)F} 1 F = P(A), P(A) = Σ 1 (A) Σ 1 (A)X { X Σ 1 (A)X } 1 X Σ 1 (A). Using this relation, we obtain l R (A; y) = c 1 2 log F Σ(A)F 1 2 y P(A)y. 11

12 The score is efine as S R (A) = l R (A; y)/ A, an is given by S R (A) = 1 2 trace{p(a)} y P 2 (A)y D [ {X = (A + ψ ) 1 trace Σ 1 (A)X } ] 1 X Σ 2 (A)X = {y X β(a)} Σ 2 (A){y X β(a)} = D (A + ψ ) 1 =1 D =1 D =1 x {ˆδ DIR x β(a)} 2 (A + ψ ) 2. { D k=1 (A + ψ k) 1 x k x k} 1 x (A + ψ ) 2 The REML estimator of A is obtaine by solving the non-linear equation S R (A) = 0. Again, application of Fisher-scoring algorithm requires also the Fisher information for A associate with l R (A; y), which is given by I R (A) = { E S } R(A) = 1 A 2 trace{p2 (A)} = 1 2 trace{σ(a) 2 } trace [ {X Σ 1 (A)X} 1 X Σ 3 (A)X ] Finally, the upating equation is trace ( [{X Σ 1 (A)X} 1 X Σ 3 (A)X ] 2 ). { } 1 A (k+1) REML = A(k) REML + I R (A (k) REML ) SR (A (k) REML ). Starting value A (0) REML an stopping criterion are the same as in FH an ML methos. Again, if A REML is the value obtaine in the last iteration, then the REML estimate is finally ÂREML = max(0, A REML ). 6 Function msefh The accuracy of an EBLUP ˆδ is usually assesse by the estimate MSE. Function msefh accompanies the EBLUPs with their corresponing estimate MSEs. The call to this function is msefh(formula, varir, metho = "REML", MAXITER = 100, PRECISION = , ata) where the arguments are exactly those of function eblupfh. 12

13 where Uner moel (4) (5), the MSE of the BLUP for A known is given by MSE( δ ) = E( δ δ ) 2 = g 1 (A) + g 2 (A), g 1 (A) = ψ {1 B (A)}, (15) { D 1 g 2 (A) = {B (A)} 2 x (A + ψ ) 1 x x } x, (16) where g 1 (A) is ue to the preiction of the ranom effect v an is O(1) for large D, an g 2 (A) is ue to the estimation of β an is O(D 1 ). This means that, for large D, a large reuction in MSE over MSE(ˆδ DIR ) = ψ can be obtaine when 1 B (A) = A/(A + ψ ) is small. Uner normality of ranom effects an errors, the MSE of the EBLUP satisfies =1 MSE(ˆδ ) = MSE(ˆδ ) + E(ˆδ δ ) 2 = [g 1 (A) + g 2 (A)] + g 3 (A) + o(d 1 ), where g 3 (A) is the uncertainty arising from the estimation of A, given by (17) g 3 (A) = {B (A)} 2 (A + ψ ) 1 V ( Â), (18) where V (Â) is the asymptotic variance (as D ) of the estimator  of A. Thus, g 3 (A) epens on the choice of estimator  but for the three available fitting methos FH, ML an REML, this term is O(D 1 ) (Prasa & Rao, 1990). The MSE of the EBLUP epens on the true variance A, which is unknown. If we want to have an unbiase estimator of the MSE up to o(d 1 ) terms (or secon orer unbiase), the MSE estimator epens on the metho use to fin  in the EBLUP (metho). The following subsections escribe the MSE estimates returne by msefh for each selecte fitting metho. 6.1 FH fitting metho When using the FH estimator ÂF H, a secon orer unbiase estimator of MSE(ˆδ ) using ÂF H is given by mse F H (ˆδ ) = g 1 (ÂF H) + g 2 (ÂF H) + 2 g 3 (ÂF H) b F H (ÂF H){B (ÂF H)} 2, (19) where, in g 3 (ÂF H), the asymptotic variance is an b F H (A) = V (ÂF H) = see Datta, Rao & Smith (2005). 2D { D } 2 (20) =1 (A + ψ ) 1 [ 2 D { D =1 (A + ψ D } ] 2 ) 2 =1 (A + ψ ) 1 { D } 3, (21) =1 (A + ψ ) 1 13

14 6.2 ML fitting metho When using the ML estimator ÂML of A, a secon orer unbiase estimator of MSE(ˆδ ) was obtaine by Datta & Lahiri (2000) an is given by mse ML (ˆδ ) = g 1 (ÂML) + g 2 (ÂML) + 2g 3 (ÂML) b ML (ÂML) g 1 (ÂML), (22) where the asymptotic variance involve in g 3 (A) is equal to the inverse of the Fisher information given in (14), V (ÂML) = I 1 ML (A) = 2 D =1 (A + ψ, (23) 2 ) [ { D } 1 { trace l=1 (A + ψ D } ] l) 1 x l x l l=1 (A + ψ l) 2 x l x l b ML (A) = D l=1 (A + ψ l) 2 an g 1 (A) = g { } 2 1(A) A = ψ. A + ψ 6.3 REML fitting metho When using the REML estimator ÂREML, a secon orer unbiase estimator of MSE(ˆδ ) is given by mse REML (ˆδ ) = g 1 (ÂREML) + g 2 (ÂREML) + 2g 3 (ÂREML), (24) where V (ÂREML) = V (ÂML) is given in (23), see Datta & Lahiri (2000). 7 Function eblupsfh Function eblupsfh estimates omain parameters δ = h (y ), = 1,..., D, base on a FH moel with spatially correlate area effects. The call to the function is eblupsfh(formula, varir, proxmat, metho = "REML", MAXITER = 100, PRECISION = , ata) The moel consiere by this function is, in matrix notation, y = Xβ + v + e, (25) where y = (ˆδ 1 DIR DIR,..., ˆδ D ) is the vector of irect estimates for the D small areas (left han sie of formula), X = (x 1,..., x D ) is a D p matrix containing in its columns the values of p explanatory variables for the D areas (right han sie of formula), v = (v 1,..., v D ) is the vector of area effects an e = (e 1,..., e D ) is the vector of inepenent sampling errors, inepenent of v, with e N(0 D, Ψ), 14

15 where the covariance matrix Ψ = iag(ψ 1,..., ψ D ) is known (varir contains the iagonal elements). The vector δ = Xβ+v = (δ 1,..., δ D ) collects the target omain parameters. The vector v follows an simultaneously autoregressive (SAR) process with unknown autoregression parameter ρ ( 1, 1) an proximity matrix W, i.e., v = ρwv + u, (26) see Anselin (1988) an Cressie (1993). Moel (25) (26) will be calle hereafter spatial FH (SFH) moel. We assume that the matrix (I D ρw) is non-singular, where I D enotes the D D ientity matrix. Then v can be expresse as v = (I D ρw) 1 u, (27) where u = (u 1,..., u D ) satisfies u N(0 D, A I D ) for A unknown. The matrix W (proxmat) is obtaine from an original proximity matrix W 0, whose iagonal elements area equal to zero an the remaining entries are equal to 1 when the two areas corresponing to the row an the column inices are consiere as neighbor an zero otherwise. Then W is obtaine by row-stanarization of W 0, obtaine by iviing each entry of W 0 by the sum of elements in the same row, see Anselin (1988), Cressie (1993) an Banerjee, Carlin & Gelfan (2004) for more etails on the SAR(1) process with the above parametrization. When W is efine in this fashion, ρ is calle spatial autocorrelation parameter (Banerjee, Carlin & Gelfan, 2004). Hereafter, the vector of variance components will be enote θ = (θ 1, θ 2 ) = (A, ρ). Equation (27) implies that v has mean vector 0 an covariance matrix equal to G(θ) = A {(I D ρw) (I D ρw)} 1. (28) Since e is inepenent of v, the covariance matrix of y is equal to Σ(θ) = G(θ) + Ψ. Combining (25) an (27), the moel is y = Xβ + (I D ρw) 1 u + e (29) The BLUP of δ = x β + v uner moel (29) is calle Spatial BLUP (Petrucci & Salvati, 2006) an is given by δ (θ) = x β(θ) + b G(θ)Σ 1 (θ){y X β(θ)}, (30) where β(θ) = {X Σ 1 (θ)x} 1 X Σ 1 (θ)y is the WLS estimator of β an b is the 1 vector (0,..., 0, 1, 0,..., 0) with 1 in the -th position. The Spatial BLUPs δ (θ), = 1,..., D, epen on the unknown vector of variance components θ = (A, ρ). Replacing a consistent estimator ˆθ = (Â, ˆρ) for θ in (30), we obtain the Spatial EBLUPs ˆδ = δ ( ˆθ), = 1,..., D, returne by function eblupsfh. The following subsections escribe the two fitting methos (metho) for the SFH moel (25) (26) supporte by eblupsfh. 15

16 7.1 ML fitting metho The ML estimator of θ = (A, ρ) is obtaine by maximizing the log-likelihoo of θ given the ata vector y, l(θ; y) = c 1 2 log Σ(θ) 1 2 (y Xβ) Σ 1 (θ)(y Xβ), where c enotes a constant. The Fisher-scoring iterative algorithm is applie to maximize this log-likelihoo. Let S(θ) = (S A (θ), S ρ (θ)) be the scores or erivatives of the log-likelihoo with respect to A an ρ, an let I(θ) be the Fisher information matrix obtaine from l(θ; y), with elements ( ) IA,A (θ) I I(θ) = A,ρ (θ). (31) I ρ,a (θ) I ρ,ρ (θ) The Fisher-scoring algorithm starts with an initial estimate θ (0) = (σ 2(0) u, ρ (0) ) an then at each iteration k, this estimate is upate with the equation The ML equation for β yiels Let us enote an θ (k+1) = θ (k) + I 1 (θ (k) )S(θ (k) ). β(θ) = { X Σ 1 (θ)x } 1 X Σ 1 (θ)y. (32) C(ρ) = (I D ρw) (I D ρw) P(θ) = Σ 1 (θ) Σ 1 (θ)x { X Σ 1 (θ)x } 1 X Σ 1 (θ). (33) Then the erivative of C(ρ) with respect to ρ is C(ρ) ρ = W W + 2ρW W an the erivatives of Σ(θ) with respect to A an ρ are respectively given by Σ(θ) A = C 1 (ρ), Σ(θ) ρ = A C 1 (ρ) C(ρ) ρ C 1 (ρ) R(θ). The scores associate to A an ρ, after replacing (32), are given by S A (θ) = 1 2 trace { Σ 1 (θ)c 1 (ρ) } y P(θ)C 1 (ρ)p(θ)y, S ρ (θ) = 1 2 trace { Σ 1 (θ)r 1 (θ) } y P(θ)R(θ)P(θ)y. 16

17 The elements of the Fisher information matrix are I A,A (θ) = 1 2 trace { Σ 1 (θ)c 1 (ρ)σ 1 (θ)c 1 (ρ) }, (34) I A,ρ (θ) = I ρ,a = 1 2 trace { Σ 1 (θ)r(θ)σ 1 (θ)c 1 (ρ) }, (35) I ρ,ρ (θ) = 1 2 trace { Σ 1 (θ)r(θ)σ 1 (θ)r(θ) }. (36) In the function eblupsfh, the starting value of A is set to A (0) = meian(ψ ). For ρ, we take ρ (0) = 0.5. The algorithm stops either when the number of iterations k > MAXITER where MAXITER can be chosen by the user, or when max { σ 2(k+1) u σ 2(k) u σ 2(k) u 7.2 REML fitting metho }, ρ (k+1) ρ (k) < PRECISION. The REML estimator of θ is obtaine by maximizing the restricte likelihoo efine as in Section 5.3. Uner the SFH moel, the restricte log-likelihoo is given by ρ (k) l R (θ; y) = c 1 2 log F Σ(θ)F 1 2 y P(θ)y, where P(θ) is efine in (33). Using the following properties of the matrix P(θ), P(θ)Σ(θ)P(θ) = P(θ), P(θ) θ r = P(θ) Σ(θ) P(θ), θ r we obtain the scores or erivatives of l R (θ; y) with respect to each element of θ, SA(θ) R = 1 2 trace { P(θ)C 1 (ρ) } y P(θ)C 1 (ρ)p(θ)y, Sρ R (θ) = 1 2 trace {P(θ)R(θ)} y P(θ)R(θ)P(θ)y, Finally, the elements of the Fisher information obtaine from l R (θ; y) are IA,A(θ) R = 1 2 tr{p(θ)c 1 (ρ)p(θ)c 1 (ρ)}, IA,ρ(θ) R = Iρ,A R = 1 2 tr{p(θ)r(θ)p(θ)c 1 (ρ)}, Iρ,ρ(θ) R = 1 2 tr{p(θ)r(θ)p(θ)r(θ)}. The upating equation of the Fisher-scoring algorithm is then θ (k+1) = θ (k) + { I R (θ (k) ) } 1 S R (θ (k) ), 17

18 where S R (θ) = (SA R(θ), SR ρ (θ)) is the scores vector an ( I I R R (θ) = A,A (θ) IA,ρ R (θ) ) Iρ,A R (θ). (37) IR ρ,ρ(θ) is the Fisher information matrix. Starting values an stopping criterion are set the same as in the case of ML estimates. 8 Function msesfh Function msesfh gives estimate MSEs of the Spatial EBLUPs uner the SFH moel (25) (26). The call to the function is msesfh(formula, varir, proxmat, metho = "REML", MAXITER = 100, PRECISION = , ata) which has the same arguments as function eblupsfh. Similarly as in Section 6, uner normality of ranom effects an errors, the MSE of the Spatial EBLUP can be ecompose as MSE{ δ ( ˆθ)} = MSE{ δ (θ)} + E{ δ ( ˆθ) δ (θ)} 2 = [g 1 (θ) + g 2 (θ)] + g 3 (θ), (38) where the first two terms on the right han sie are easily calculate ue to the linearity of the Spatial BLUP δ (θ) in the ata vector y. They are given by { g 1 (θ) = b G(θ) G(θ)Σ 1 (θ)g(θ) } b, (39) { g 2 (θ) = b ID G(θ)Σ 1 (θ) } X(X Σ 1 (θ)x) 1 X { I D Σ 1 (θ)g(θ) } b. (40) However, for the last term g 3 (θ) = E{ δ ( ˆθ) δ (θ)} 2, an exact analytical expression oes not exist ue to the non-linearity of the EBLUP δ ( ˆθ) in y. Uner the basic FH moel (4)-(5) with inepenent ranom effects v (iagonal covariance matrix Σ), Prasa & Rao (1990) obtaine an approximation up to o(d 1 ) terms of g 3 (θ) through Taylor linearization, see Section 6. Their formula can be taken as a naive approximation of the true g 3 (θ) uner the SFH moel (25) (26). Straightforwar application of this formula to moel (25) (26) yiels g P R 3 (θ) = trace { L (θ)σ(θ)l (θ)i 1 (θ) }, (41) where I(θ) is the Fisher information efine in (31) with elements (34) (36) an ( ) b L (θ) = {C 1 (ρ)σ 1 (θ) AC 1 (ρ)σ 1 (θ)c 1 (ρ)σ 1 (θ)} b {R(θ)Σ 1 (θ) AC 1 (ρ)σ 1 (θ)r(θ)σ 1. (θ)} Then the full MSE can be approximate by MSE P R { δ ( ˆθ)} = g 1 (θ) + g 2 (θ) + g P R 3 (θ). (42) 18

19 Singh, Shukla & Kunu (2005) arrive to the same formula (42) for the true MSE uner a SFH moel. However, this formula is not accounting for the extra uncertainty ue to estimation of the spatial autocorrelation parameter ρ. Next subsections escribe the MSE estimates returne by function msesfh, epening on the specifie fitting metho. 8.1 REML fitting metho Let ˆθ REML be the estimator of ˆθ when REML fitting metho is specifie in msesfh. In that case, the function msesfh returns the MSE estimator erive by Singh, Shukla & Kunu (2005) an given by mse SSK [ δ ( ˆθ REML )] = g 1 ( ˆθ REML ) + g 2 ( ˆθ REML ) + 2g3 P R ( ˆθ REML ) g 4 ( ˆθ REML ), (43) where g 1, g 2 an g3 P R are given respectively in (39), (40) an (41), an g 4 (θ) reas g 4 (θ) = r=1 s=1 b Ψ Σ 1 (θ) 2 Σ(θ) θ r θ s Σ 1 (θ)ψ v rs (θ) b, where v rs (θ) enotes the element (r, s) of I(θ) 1, for the Fisher information matrix I(θ) efine in (31) with elements (34) (36). The secon orer erivatives of Σ(θ) are given by 2 Σ(θ) = 0 A 2 D D, 2 Σ(θ) A ρ = 2 Σ(θ) A ρ 2 Σ(θ) ρ 2 = 2AC 1 (θ) C(ρ) ρ 8.2 ML fitting metho = C 1 (θ) C(ρ) ρ C 1 (θ), C 1 (θ) C(ρ) ρ C 1 (θ) 2AC 1 (θ)w WC 1 (θ). Let ˆθ ML be the estimator of ˆθ when ML fitting metho is specifie. In that case, the estimate returne by msesfh reas mse SSK ML { δ ( ˆθ)} = g 1 ( ˆθ ML ) + g 2 ( ˆθ ML ) + 2g3 P R ( ˆθ ML ) g 4 ( ˆθ ML ) b ML( ˆθ ML ) g 1 ( ˆθ ML ). In this expression, g 1, g 2 an g3 P R are efine respectively in (39), (40) an (41), g 1 (θ) = ( g 1 (θ)/ A, g 1 (θ)/ ρ) is the graient of g 1 (θ) with erivatives g 1 (θ) A g 1 (θ) ρ = b { C 1 (ρ) 2 A C 1 (ρ)σ 1 (θ)c 1 (ρ) + A 2 C 1 (ρ)σ 1 (θ)c 1 (ρ)σ 1 (θ)c 1 (ρ) } b, = b { R(θ) 2 A C 1 (ρ)σ 1 (θ)r(θ) A 2 C 1 (ρ)σ 1 (θ)r(θ)σ 1 (θ)c 1 (ρ) } b, 19

20 an finally, b ML ( ˆθ ML ) is the bias of the ML estimator ˆθ ML up to o(d 1 ) terms, given by b ML ( ˆθ) = I 1 ( ˆθ ML )h( ˆθ ML )/2 with h( ˆθ ML ) = (h 1 ( ˆθ ML ), h 2 ( ˆθ ML )) an [ {X h 1 (θ) = trace Σ 1 (θ)x } ] 1 X Σ 1 (θ)c 1 (ρ)σ 1 (θ)x, [ {X h 2 (θ) = trace Σ 1 (θ)x } ] 1 X Σ 1 (θ)r(θ)σ 1 (θ)x. 9 Function pbmsesfh Function pbmsesfh gives parametric bootstrap estimates of the MSEs of the Spatial EBLUPs uner the SFH moel (25) (26) using an extension of González- Manteiga et al. (2008b). The MSE estimators obtaine by this proceure are expecte to be consistent if the moel parameter estimates are consistent. The call to the function is pbmsesfh(formula, varir, proxmat, B = 100, metho = "REML", MAXITER = 100, PRECISION = , ata) where the arguments are those of function eblupsfh, an aitionally the number of bootstrap replicates B. The bootstrap proceure procees as follows: 1) Fit the SFH moel (25) (26) to the initial ata y = (ˆδ 1 DIR,..., obtaining moel parameter estimates ˆθ = (Â, ˆρ) an ˆβ = β( ˆθ). ˆδ DIR D ), 2) Generate a vector t 1 whose elements are D inepenent copies of a N(0, 1). Construct bootstrap vectors u = Â1/2 t 1 an v = (I D ˆρW) 1 u, an calculate the bootstrap quantity of interest δ = X ˆβ + v, by regaring ˆβ an ˆθ as the true values of the moel parameters. 3) Generate a vector t 2 with D inepenent copies of a N(0, 1), inepenently of the generation of t 1, an construct the vector of bootstrap ranom errors e = Ψ 1/2 t 2. 4) Obtain bootstrap ata applying the moel y = δ + e = X ˆβ + v + e. 5) Regaring ˆβ an ˆθ as the true values of β an θ, fit the SFH moel (25) (26) to bootstrap ata y, obtaining estimates of the true ˆβ an ˆθ base on y. For this, first calculate the estimator of ˆβ evaluate at the true value ˆθ, β ( ˆθ) { ˆθ)X} 1 = X Σ 1 ( X Σ 1 ( ˆθ)y ; next, obtain the estimator ˆθ of ˆθ base on y an, finally, the estimator of ˆβ evaluate at ˆθ, that is, β ( ˆθ ). 6) Calculate the bootstrap Spatial BLUP from bootstrap ata y an regaring ˆθ as the true value of θ, δ ( ˆθ) = x β ( ˆθ) + b G( ˆθ)Σ( ˆθ) {y 1 X β ( ˆθ) }. 20

21 Calculate also the bootstrap Spatial EBLUP replacing ˆθ for the true ˆθ, δ ( ˆθ ) = x β ( ˆθ ) + b G( ˆθ )Σ 1 ( ˆθ )[y X β ( ˆθ )]. 7) Repeat steps 2) 6) B times. In b-th bootstrap replication, let δ (b) quantity of interest for -th area, ˆθ (b) the bootstrap estimate of θ, the bootstrap Spatial BLUP an for -th area. δ (b) 8) A parametric bootstrap estimator of g 3 (θ) is g P B 3 ( ˆθ) = B 1 B b=1 be the (b) δ ( ˆθ) ( ˆθ (b) ) the bootstrap Spatial EBLUP { δ (b) ( ˆθ (b) (b) ) δ ˆθ)} 2 (. Function pbmsesfh returns the naive parametric bootstrap estimator of the full MSE, given by mse nap B { δ ( ˆθ)} = B 1 B b=1 { δ (b) ( ˆθ } 2 (b) ) δ (b). (44) Function pbmsesfh also returns a bias-correcte MSE estimate obtaine as in Pfeffermann an Tiller (2005), an given by mse bcp B { δ ( ˆθ)} { = 2 g 1 ( ˆθ) + g 2 ( ˆθ) } + g3 P B ( ˆθ) B 1 10 Function npbmsesfh B b=1 { g 1 ( ˆθ (b) ) + g 2 ( ˆθ } (b) ). (45) Function npbmsesfh gives MSE estimates for the Spatial EBLUPs uner the SFH moel (25) (26), using the nonparametric bootstrap approach of Molina, Salvati & Pratesi (2009). The call to the function is npbmsesfh(formula, varir, proxmat, B = 100, metho = "REML", MAXITER = 100,PRECISION = , ata) where the arguments are the same as in pbmsesfh. The function resamples ranom effects {u 1,..., u D } an errors {e 1,..., e D } from the respective empirical istribution of preicte ranom effects {û 1,..., û D } an resiuals {ˆr 1,..., ˆr D }, DIR where r = ˆδ δ ( ˆθ), = 1,..., D, all previously stanarize. This metho avois the nee for istributional assumptions of u an e ; therefore, it is expecte to be more robust to non-normality of the ranom moel components. Uner moel (25) (26), the BLUPs of u an v are respectively given by ṽ(θ) = G(θ)Σ 1 (θ){y X β(θ)}, ũ(θ) = (I ρw)ṽ(θ), 21

22 an the covariance matrix of ũ(θ) is Σ u (θ) = (I ρw)g(θ)p(θ)g(θ)(i ρw ). Let us efine the vector of resiuals obtaine from the BLUP r(θ) = y X β(θ) ṽ(θ) = (ˆδ 1 DIR δ DIR 1 (θ),..., ˆδ D δ (θ)). It is easy to see that the covariance matrix of r(θ) is Σ r (θ) = Ψ P(θ)Ψ, for P(θ) efine in (33). The covariance matrices Σ u (θ) an Σ r (θ) are not iagonal; hence, the elements of the vectors ũ(θ) an r(θ) are correlate. Inee, both ũ(θ) an r(θ) lie in a subspace of imension D p. Since the methos that resample from the empirical istribution work well uner an ieally ii setup, before resampling a previous stanarization step is crucial. Here û = ũ( ˆθ) an ˆr = r( ˆθ) are transforme to achieve vectors that are as close as possible to be uncorrelate an with unit variance elements. We escribe the stanarization metho only for û, since for ˆr the process is analogous. Let us consier the estimate covariance matrix ˆΣ u = Σ u ( ˆθ). The spectral ecomposition of ˆΣ u is ˆΣ u = Q u u Q u, where u is a iagonal matrix with the D p non-zero eigenvalues of ˆΣ u an Q u is the matrix with the corresponing eigenvectors in the columns. Take 1/2 the square root matrix ˆΣ u = Q u 1/2 u Q u. Squaring this matrix gives a generalize inverse of ˆΣ u. With the obtaine square root, we transform û as û S = ˆΣ 1/2 u û. The covariance matrix of û S is then V (û S ) = Q u Q u, which is close to an ientity matrix. Observe that in the transformation û S = Q u 1/2 u Q uû, the vector Q uû contains the coorinates of û in its principal components, which are uncorrelate an with covariance matrix u. Then multiplying by 1/2 u, these coorinates are stanarize to have unit variance. Finally, this stanarize vector in the space of the principal components is returne to the original space by multiplying by Q u. Thus, the transforme vector û S contains the coorinates of the vector 1/2 u Q uû, with stanar elements, in the original space. The eigenvalues, which are the variances of the uncorrelate principal components, collect better the variability than the iagonals of ˆΣ u. Inee, simulations were inicate that taking simply û S = û / v, where v is the -th iagonal element of ˆΣ u, oes not work well. The final nonparametric bootstrap proceure is obtaine by replacing steps 2) an 3) in the parametric bootstrap 1) 8) of Section 9 by the new steps 2 ) an 3 ) given below: 22

23 2 ) With the estimates ˆθ = (Â, ˆρ) an ˆβ = β( ˆθ) obtaine in step 1), calculate preictors of v an u as follows ˆv = G( ˆθ)Σ( ˆθ) 1 (y X ˆβ), û = (I ˆρW)ˆv = (û 1,..., û m ). ˆΣ 1/2 ˆΣ 1/2 u Then take û S = u û = (û S 1,..., û S D ), where is the square root of the generalize inverse of ˆΣ u obtaine by the spectral ecomposition. It is convenient to re-scale the elements û S so that they have sample mean exactly equal to zero an sample variance Â. This is achieve by the transformation û SS = Â(û S D 1 D l=1 ûs l ) D 1 D =1 (ûs, = 1,..., D. D 1 D l=1 ûs l )2 Construct the vector u = (u 1,..., u D ), whose elements are obtaine by extracting a simple ranom sample with replacement of size D from the set {û SS 1,..., û SS D }. Then obtain v = (I ˆρW) 1 u an calculate the vector of bootstrap target parameters δ = X ˆβ + v = (δ1,..., δ ) 3 ) Compute the vector of resiuals ˆr = y X ˆβ ˆv = (ˆr 1,..., ˆr D ). Stanarize these resiuals as ˆr S 1/2 = ˆΣ r ˆr = (ˆr 1 S,..., ˆr D S ), where ˆΣ r = Ψ P( ˆθ)Ψ 1/2 is the estimate covariance matrix an ˆΣ r is a square root of the generalize inverse erive from the spectral ecomposition of ˆΣr. Again, re-stanarize these values ˆr SS = ˆr S m 1 D l=1 ˆrS l D 1 D =1 (ˆrS, = 1,..., D. D 1 D l=1 ˆrS l )2 Construct r = (r1,..., rd ) by extracting a simple ranom sample with replacement of size D from the set {ˆr 1 SS,..., ˆr D SS}. Then take e = (e 1,..., e D ), where e = ψ1/2 r, = 1,..., D. The function npbmsesfh yiels naive an bias-correcte nonparametric bootstrap estimators mse nanp B { δ ( ˆθ)} an mse bcnp B { δ ( ˆθ)} analogous to (44) an (45) respectively. 11 Function eblupstfh Function eblupstfh gives small area estimators of δ = h (y ), = 1,..., D, uner an extension of the FH moel that takes into account the spatial correlation between neighbor areas an also incorporates historical ata (Marhuena, Molina & Morales, 2013). The area parameter for omain at current time instant T is estimate borrowing strength from the T time instants an from the D omains. The call to the function is 23

24 eblupstfh(formula, D, T, varir, proxmat, moel = "ST", MAXITER = 100, PRECISION = , ata) Let θ t be the target area characteristic for area an time instant t, for = DIR 1,..., D an t = 1,..., T. Let ˆδ t be a irect estimator of δ t (left han sie of formula) an x t a column vector containing the aggregate values of p auxiliary variables relate linearly with δ t (right han sie of formula). The spatio-temporal FH (STFH) moel is state as follows. In the first stage, we assume ˆδ t DIR = δ t + e t, = 1,..., D, t = 1,..., T, (46) where, given δ t, sampling errors e t are assume to be inepenent an normally istribute with variances ψ t known for all an t (varir). In the secon stage, the target parameters for all omains an time points are linke through the moel δ t = x tβ + u 1 + u 2t, = 1,..., D, t = 1,..., T. (47) Here, the vectors of area-time ranom effects (u 21,..., u 2T ) are i.i.. for each area, following an AR(1) process with autocorrelation parameter ρ 2, that is, u 2t = ρ 2 u 2,t 1 + ɛ 2t, ρ 2 < 1, ɛ 2t ii N(0, σ 2 2). (48) The vector of area effects (u 11,..., u 1D ) follows a SAR(1) process with variance parameter σ 2 1, spatial autocorrelation ρ 1 an row-stanarize proximity matrix W = (w,l ) efine as in Section 7 (proxmat), that is, u 1 = ρ 1 l w,l u 1l + ɛ 1, ρ 1 < 1, ɛ 1 ii N(0, σ 2 1), = 1,..., D. (49) Let us efine the following vectors an matrices obtaine by stacking the elements of the moel in columns y = e = col ( col DIR (ˆδ t )), X = col ( col 1 D 1 t T 1 D 1 t T (x t)), col ( col (e t)), u 1 = 1 D 1 t T col (u 1) an u 2 = 1 D col ( col (u 2t). 1 D 1 t T Defining aitionally Z 1 = I D 1T where I D is the D D ientity matrix, 1 T is a vector of ones of size T an is the Kronecker prouct, Z 2 = I n, where n = DT is the total number of observations, u = (u 1, u 2) an Z = (Z 1, Z 2 ), the moel can be expresse as a general linear mixe moel in the form y = Xβ + Zu + e. Let θ = (σ 2 1, ρ 1, σ 2 2, ρ 2 ) be the vector of unknown parameters involve in the covariance matrix of y. Observe that here e N(0 n, Ψ), where 0 n enotes a vector of zeros of size n an Ψ is the iagonal matrix Ψ = iag 1 D (iag 1 t T (ψ t )). 24

25 Moreover, u N{0 n, G(θ)}, where the covariance matrix is the block iagonal matrix G(θ) = iag{σ 2 1Ω 1 (ρ 1 ), σ 2 2Ω 2 (ρ 2 )}, with Ω 1 (ρ 1 ) = { (I D ρ 1 W) (I D ρ 1 W) } 1, (50) Ω 2 (ρ 2 ) = iag 1 D {Ω 2 (ρ 2 )}, 1 ρ 2... ρ T 2 2 ρ T 2 1. Ω 2 (ρ 2 ) = 1 ρ ρ T ρ ρ T ρ2 ρ T 2 1 ρ2 T 2... ρ 2 1 Thus, the covariance matrix of y is given by Σ(θ) = ZG(θ)Z + Ψ. T T, = 1,..., D. (51) The WLS estimator of β an the (componentwise) BLUP of u obtaine by Henerson (1975) are given by β(θ) = { X Σ 1 (θ)x } 1 X Σ 1 (θ)y, ũ(θ) = G(θ)Z Σ 1 (θ){y X β(θ)}. (52) Since u = (u 1, u 2), the secon ientity leas to the BLUPs of u 1 an u 2, respectively given by ũ 1 (θ) = σ 2 1Ω 1 (ρ 1 )Z 1Σ 1 (θ){y X β(θ)}, ũ 2 (θ) = σ 2 2Ω 2 (ρ 2 )Σ 1 (θ){y X β(θ)}. Replacing an estimator ˆθ for θ in previous formulas we obtain ˆβ = β( ˆθ) an the EBLUPs of u 1 an u 2 respectively, û 1 = ũ 1 ( ˆθ) = (û 11,..., û 1D ) an û 2 = ũ 2 ( ˆθ) = (û 211,..., û 2DT ). Finally, the EBLUP of the area characteristic δ t uner the STFH moel (46) (49) returne by function eblupstfh is given by ˆδ t = x t ˆβ + û 1 + û 2t, = 1,..., D, t = 1,..., T. The following subsection escribes the REML moel fitting proceure applie by function eblupstfh to estimate θ an β. Remark 1. Computation of the inverse of the n n matrix Σ(θ) involve in (52) can be too time consuming for large n. This is replace by the inversion of two smaller matrices as follows. Observe that Σ(θ) can be expresse as Σ(θ) = σ 2 1Z 1 Ω 1 (ρ 1 )Z 1 + Γ(θ), 25

26 where Γ(θ) = iag 1 D {Γ (θ)} an Γ (θ) = σ 2 2Ω 2 (ρ 2 ) + iag 1 t T (ψ t ), = 1,..., D. Applying the inversion formula (A + CBD) 1 = A 1 A 1 C(B 1 + DA 1 C) 1 DA 1 (53) with A = Γ(θ), B = σ 2 1Ω 1 (ρ 1 ), C = Z 1 an D = Z 1, we obtain Σ 1 (θ) = Γ 1 (θ) Γ 1 (θ)z 1 { σ 2 1 Ω 1 1 (ρ 1 ) + Z 1Γ 1 (θ)z 1 } 1 Z 1 Γ 1 (θ), where Γ 1 (θ) = iag 1 D {Γ 1 (θ)}. Here, Γ (θ) is inverte using again (53). This proceure only requires inversion of the T T matrix Ω 2 (ρ 2 ) given in (51), which is constant for all, an the D D matrix Ω 1 (ρ 1 ) given in (50) REML fitting metho REML fitting metho maximizes the restricte likelihoo, which is the joint p..f. of a vector of n p linearly inepenent contrasts F y, where F is an n (n p) full column rank matrix satisfying F F = I n p an F X = 0 n p. It hols that F y is inepenent of ˆβ given in (52). Consequently, the p..f. of F y oes not epen on β an is given by { f R (θ; y) = (2π) (n p)/2 X X 1/2 Σ(θ) 1/2 X Σ 1 (θ)x 1/2 exp 1 } 2 y P(θ)y, where P(θ) = Σ 1 (θ) Σ 1 (θ)x { X Σ 1 (θ)x } 1 X Σ 1 (θ). Observe that P(θ) satisfies P(θ)Σ(θ)P(θ) = P(θ) an P(θ)X = 0 n. The REML estimator of θ = (θ 1,..., θ 4 ) = (σ 2 1, ρ 1, σ 2 2, ρ 2 ) is the maximizer of l R (θ; y) = log f R (θ; y). This maximum is compute using the Fisher-scoring algorithm. Let S R (θ) = l R (θ; y)/ θ = (S R 1 (θ),..., S R 4 (θ)) be the scores vector an I R (θ) = E{ 2 l R (θ; y)/ θ θ } = (I R rs(θ)) the Fisher information matrix associate with θ. Using the fact that P(θ) θ r = P(θ) Σ(θ) θ r P(θ), r = 1,..., 4, the first orer partial erivative of l R (θ; y) with respect to θ r is Sr R (θ) = 1 { 2 tr P(θ) Σ(θ) } + 1 θ r 2 y P(θ) Σ(θ) P(θ)y, r = 1,..., 4. θ r The element (r, s) of the Fisher information matrix is the expecte value of the negative secon orer partial erivative of l R (θ; y) with respect to θ r an θ s, which yiels Irs(θ) R = 1 { 2 tr P(θ) Σ(θ) P(θ) Σ(θ) }, r, s = 1,..., 4. θ r θ s 26

27 Then, if θ (k) is the value of the estimator at iteration k, the upating formula of the Fisher-scoring algorithm is given by θ (k+1) = θ (k) + I 1 R (θ(k) )S R (θ (k) ). Finally, the partial erivatives of Σ(θ) with respect to the components of θ, involve in S R (θ) an I R (θ), are given by Σ(θ) σ 2 1 Σ(θ) σ 2 2 = Z 1 Ω 1 (ρ 1 )Z 1, = iag {Ω 2 (ρ 2 )}, 1 D Σ(θ) ρ 1 Σ(θ) ρ 2 = σ1z 2 1 Ω 1 (ρ 1 ) Ω 1 1 (ρ 1 ) { Ω2 (ρ 2 ) = σ 2 2 iag 1 D ρ 2 Ω 1 (ρ 1 )Z ρ 1, 1 }, where an Ω 2 (ρ 2 ) ρ 2 = 1 1 ρ 2 2 Ω 1 1 (ρ 1 ) ρ 1 = W W + 2ρ 1 W W (T 1)ρ T (T 2)ρ T (T 2)ρ2 T (T 1)ρ T ρ 2Ω 2 (ρ 2 ). 1 ρ Function pbmsestfh Function pbmsestfh gives parametric bootstrap MSE estimates for the EBLUPs of the omain parameters uner the STFH moel (46) (49) as in Marhuena, Molina & Morales (2013). The call to the function is pbmsestfh(formula, D, T, varir, proxmat, B = 100, moel = "ST", MAXITER = 100, PRECISION = , ata) where the arguments are the same as in eblupstfh, together with the number of bootstrap replicates B. The parametric bootstrap proceure is escribe below: (1) Using the available ata {(ˆδ t DIR, x t ), t = 1,..., T, = 1,..., D}, fit the STFH moel (46) (49) an obtain moel parameter estimates ˆβ, ˆσ 1, 2 ˆρ 1, ˆσ 2 2 an ˆρ 2. (2) Generate bootstrap area effects {u (b) 1, = 1,..., D}, from the SAR(1) process given in (49), using (ˆσ 1, 2 ˆρ 1 ) as true values of parameters (σ1, 2 ρ 1 ). (3) Inepenently of {u (b) 1 } an inepenently for each, generate bootstrap time effects {u (b) 2t, t = 1,..., T }, from the AR(1) process given in (48), with (ˆσ 2, 2 ˆρ 2 ) acting as true values of parameters (σ2, 2 ρ 2 ). 27

28 (4) Calculate true bootstrap quantities, δ (b) t = x t ˆβ + u (b) 1 + u (b) 2t, t = 1,..., T, = 1,..., D. (5) Generate errors e (b) t sampling moel, in. N(0, ψ t ) an obtain bootstrap ata from the ˆδ DIR (b) t = δ (b) t + e (b) t, t = 1,..., T, = 1,..., D. (6) Using the new bootstrap ata {(ˆδ DIR (b) t, x t ), t = 1,..., T, = 1,..., D}, fit the STFH moel (46) (49) an obtain the bootstrap EBLUPs, ˆδ (b) t = x t ˆβ (b) + û (b) 1 + û (b) 2t, t = 1,..., T, = 1,..., D. (7) Repeat steps (1)-(6) for b = 1,..., B, where B is a large number. (8) The parametric bootstrap MSE estimates returne by function pbmsestfh are given by mse(ˆδ t ) = 1 B B b=1 (ˆδ (b) t ) 2 δ (b) t, t = 1,..., T, = 1,..., D. (54) 13 Function eblupbhf Function eblupbhf estimates the area means Ȳ, = 1,..., D, uner the unit level moel introuce by Battese, Harter & Fuller (1988) (BHF moel). The call to the function is eblupbhf(formula, om, selectom, meanxpop, popnsize, metho = "REML", ata) The function allows to select a subset of omains for estimation through the argument selectom, but ropping this argument it estimates in all omains. Let Y j be the value of the target variable for unit j in omain (left han sie of formula). The BHF moel assumes Y j = x jβ + u + e j, j = 1,..., N, = 1,..., D, u ii N(0, σ 2 u), e j ii N(0, σ 2 e), (55) where x j is a vector containing the values of p explanatory variables for the same unit (right han sie of formula), u is the area ranom effect an e j is the iniviual error, where area effects u an errors e j are inepenent. Let us efine vectors an matrices obtaine by stacking in columns the elements for omain y = col 1 j N (Y j ), X = col 1 j N (x j ), e = col 1 j N (e j ). 28

29 Then, the omain vectors y are inepenent an follow the moel y = X β + u 1 N + e, e in N(0, σ 2 ei N ), = 1,..., D, where u is inepenent of e. Uner this moel, the mean vector an the covariance matrix of y are given by µ = X β an V = σ 2 u1 N 1 N + σ 2 ei N. Consier the ecomposition of y into sample an out-of-sample elements y = (y r, y s ), an the corresponing ecomposition of X an V as ( ) ( ) Xs Vs V X =, V = sr. X r V rs If σu 2 an σe 2 are known, the BLUP of the small area mean Ȳ is given by ( Ȳ = 1 Y j + ) Ỹ j, (56) N j s j r V r where Ỹj = x β j + ũ is the BLUP of Y j. Here, β is the WLS estimator of β an ũ is the BLUP of u, given respectively by ( D ) 1 D β = X V 1 s X X V 1 s y, (57) =1 ũ = γ (ȳ s x β), s (58) where ȳ s = n 1 j s Y j, x s = n 1 j s x j = 1,..., D. an γ = σu/(σ 2 u 2 + σe/n 2 ), Let ˆσ u 2 an ˆσ e 2 be consistent estimators of σu 2 an σe 2 respectively, such as those obtaine by ML or REML. The EBLUP is ( ˆȲ = 1 Y j + ) Ŷ j, N j s j r (59) =1 where Ŷj = x j ˆβ + û is the EBLUP of Y j, ˆβ an û are given respectively by ˆβ = ( D =1 X ˆV 1 s X ) 1 D =1 X ˆV 1 s y (60) û = ˆγ (ȳ s x s ˆβ), (61) with ˆγ = ˆσ 2 u/(ˆσ 2 u + ˆσ 2 e/n ) an ˆV s = ˆσ 2 u1 n 1 n + ˆσ 2 ei n, = 1,..., D. Replacing (60) an (61) in (59), we obtain the expression for the EBLUP of Ȳ returne by function eblupbhf, ˆȲ = f ȳ s + ( X f x s ) ˆβ + (1 f )û, 29

30 where f = n /N is the sampling fraction. Note that the EBLUP requires the vector of population means of the auxiliary variables X (meanxpop) an the population sizes (popnsize) apart from the sample ata (specifie in formula), but the iniviual values of the auxiliary variables for each population unit are not neee. 14 Function pbmsebhf Function pbmsebhf gives a parametric bootstrap MSE estimate for the EBLUP uner the BHF moel (55). The call to the function is pbmsebhf(formula, om, selectom, meanxpop, popnsize, B = 200, metho = "REML", ata) The function applies the parametric bootstrap proceure for finite populations introuce by González-Manteiga et al. (2008a) particularize to the estimation of means. The estimate MSEs are obtaine as follows: 1) Fit the BHF moel (55) to sample ata y s = (y 1s,..., y Ds ) an obtain moel parameter estimates ˆβ, ˆσ 2 u an ˆσ 2 e. 2) Generate bootstrap omain effects as u (b) ii N(0, ˆσ 2 u), = 1,..., D. 3) Generate, inepenently of the ranom effects u (b), bootstrap errors for sample elements e (b) ii j N(0, ˆσ e), 2 j s, an error omain means Ē (b) ii N(0, ˆσ e/n 2 ), = 1,..., D., 4) Compute the true omain means of this bootstrap population, given by Ȳ (b) = X ˆβ + u (b) + Ē (b), = 1,..., D. Observe that computation of Ȳ (b) oes not require the iniviual values x j, for each out-of-sample unit j r. 5) Using the known sample vectors x j, j s, generate the moel responses for sample elements from the moel Let y (b) s Y (b) j = x j ˆβ + u (b) + e (b) j, j s, = 1,..., D. = ((y (b) 1s ),..., (y (b) Ds ) ) be the bootstrap sample ata vector. 6) Fit the BHF moel (55) to bootstrap ata ys (b) (b) EBLUPs ˆȲ, = 1,..., D. an obtain the bootstrap 30

A comparison of small area estimators of counts aligned with direct higher level estimates

A comparison of small area estimators of counts aligned with direct higher level estimates A comparison of small area estimators of counts aligne with irect higher level estimates Giorgio E. Montanari, M. Giovanna Ranalli an Cecilia Vicarelli Abstract Inirect estimators for small areas use auxiliary

More information

Survey Sampling. 1 Design-based Inference. Kosuke Imai Department of Politics, Princeton University. February 19, 2013

Survey Sampling. 1 Design-based Inference. Kosuke Imai Department of Politics, Princeton University. February 19, 2013 Survey Sampling Kosuke Imai Department of Politics, Princeton University February 19, 2013 Survey sampling is one of the most commonly use ata collection methos for social scientists. We begin by escribing

More information

This module is part of the. Memobust Handbook. on Methodology of Modern Business Statistics

This module is part of the. Memobust Handbook. on Methodology of Modern Business Statistics This moule is part of the Memobust Hanbook on Methoology of Moern Business Statistics 26 March 2014 Metho: EBLUP Unit Level for Small Area Estimation Contents General section... 3 1. Summary... 3 2. General

More information

Improving Estimation Accuracy in Nonrandomized Response Questioning Methods by Multiple Answers

Improving Estimation Accuracy in Nonrandomized Response Questioning Methods by Multiple Answers International Journal of Statistics an Probability; Vol 6, No 5; September 207 ISSN 927-7032 E-ISSN 927-7040 Publishe by Canaian Center of Science an Eucation Improving Estimation Accuracy in Nonranomize

More information

Estimating Unemployment for Small Areas in Navarra, Spain

Estimating Unemployment for Small Areas in Navarra, Spain Estimating Unemployment for Small Areas in Navarra, Spain Ugarte, M.D., Militino, A.F., an Goicoa, T. Departamento e Estaística e Investigación Operativa, Universia Pública e Navarra Campus e Arrosaía,

More information

Small Area Estimation: A Review of Methods Based on the Application of Mixed Models. Ayoub Saei, Ray Chambers. Abstract

Small Area Estimation: A Review of Methods Based on the Application of Mixed Models. Ayoub Saei, Ray Chambers. Abstract Small Area Estimation: A Review of Methos Base on the Application of Mixe Moels Ayoub Saei, Ray Chambers Abstract This is the review component of the report on small area estimation theory that was prepare

More information

The Role of Models in Model-Assisted and Model- Dependent Estimation for Domains and Small Areas

The Role of Models in Model-Assisted and Model- Dependent Estimation for Domains and Small Areas The Role of Moels in Moel-Assiste an Moel- Depenent Estimation for Domains an Small Areas Risto Lehtonen University of Helsini Mio Myrsylä University of Pennsylvania Carl-Eri Särnal University of Montreal

More information

Estimation of District Level Poor Households in the State of. Uttar Pradesh in India by Combining NSSO Survey and

Estimation of District Level Poor Households in the State of. Uttar Pradesh in India by Combining NSSO Survey and Int. Statistical Inst.: Proc. 58th Worl Statistical Congress, 2011, Dublin (Session CPS039) p.6567 Estimation of District Level Poor Househols in the State of Uttar Praesh in Inia by Combining NSSO Survey

More information

'HVLJQ &RQVLGHUDWLRQ LQ 0DWHULDO 6HOHFWLRQ 'HVLJQ 6HQVLWLYLW\,1752'8&7,21

'HVLJQ &RQVLGHUDWLRQ LQ 0DWHULDO 6HOHFWLRQ 'HVLJQ 6HQVLWLYLW\,1752'8&7,21 Large amping in a structural material may be either esirable or unesirable, epening on the engineering application at han. For example, amping is a esirable property to the esigner concerne with limiting

More information

Survey-weighted Unit-Level Small Area Estimation

Survey-weighted Unit-Level Small Area Estimation Survey-weighte Unit-Level Small Area Estimation Jan Pablo Burgar an Patricia Dörr Abstract For evience-base regional policy making, geographically ifferentiate estimates of socio-economic inicators are

More information

A Review of Multiple Try MCMC algorithms for Signal Processing

A Review of Multiple Try MCMC algorithms for Signal Processing A Review of Multiple Try MCMC algorithms for Signal Processing Luca Martino Image Processing Lab., Universitat e València (Spain) Universia Carlos III e Mari, Leganes (Spain) Abstract Many applications

More information

Spurious Significance of Treatment Effects in Overfitted Fixed Effect Models Albrecht Ritschl 1 LSE and CEPR. March 2009

Spurious Significance of Treatment Effects in Overfitted Fixed Effect Models Albrecht Ritschl 1 LSE and CEPR. March 2009 Spurious Significance of reatment Effects in Overfitte Fixe Effect Moels Albrecht Ritschl LSE an CEPR March 2009 Introuction Evaluating subsample means across groups an time perios is common in panel stuies

More information

Euler equations for multiple integrals

Euler equations for multiple integrals Euler equations for multiple integrals January 22, 2013 Contents 1 Reminer of multivariable calculus 2 1.1 Vector ifferentiation......................... 2 1.2 Matrix ifferentiation........................

More information

This module is part of the. Memobust Handbook. on Methodology of Modern Business Statistics

This module is part of the. Memobust Handbook. on Methodology of Modern Business Statistics This moule is part of the Memobust Hanbook on Methoology of Moern Business Statistics 26 March 2014 Metho: Balance Sampling for Multi-Way Stratification Contents General section... 3 1. Summary... 3 2.

More information

A COMPARISON OF SMALL AREA AND CALIBRATION ESTIMATORS VIA SIMULATION

A COMPARISON OF SMALL AREA AND CALIBRATION ESTIMATORS VIA SIMULATION SAISICS IN RANSIION new series an SURVEY MEHODOLOGY 133 SAISICS IN RANSIION new series an SURVEY MEHODOLOGY Joint Issue: Small Area Estimation 014 Vol. 17, No. 1, pp. 133 154 A COMPARISON OF SMALL AREA

More information

Small Domains Estimation and Poverty Indicators. Carleton University, Ottawa, Canada

Small Domains Estimation and Poverty Indicators. Carleton University, Ottawa, Canada Small Domains Estimation and Poverty Indicators J. N. K. Rao Carleton University, Ottawa, Canada Invited paper for presentation at the International Seminar Population Estimates and Projections: Methodologies,

More information

Robust Forward Algorithms via PAC-Bayes and Laplace Distributions. ω Q. Pr (y(ω x) < 0) = Pr A k

Robust Forward Algorithms via PAC-Bayes and Laplace Distributions. ω Q. Pr (y(ω x) < 0) = Pr A k A Proof of Lemma 2 B Proof of Lemma 3 Proof: Since the support of LL istributions is R, two such istributions are equivalent absolutely continuous with respect to each other an the ivergence is well-efine

More information

M-Quantile Regression for Binary Data with Application to Small Area Estimation

M-Quantile Regression for Binary Data with Application to Small Area Estimation University of Wollongong Research Online Centre for Statistical & Survey Methoology Working Paper Series Faculty of Engineering an Information Sciences 2012 M-Quantile Regression for Binary Data with Application

More information

7.1 Support Vector Machine

7.1 Support Vector Machine 67577 Intro. to Machine Learning Fall semester, 006/7 Lecture 7: Support Vector Machines an Kernel Functions II Lecturer: Amnon Shashua Scribe: Amnon Shashua 7. Support Vector Machine We return now to

More information

LATTICE-BASED D-OPTIMUM DESIGN FOR FOURIER REGRESSION

LATTICE-BASED D-OPTIMUM DESIGN FOR FOURIER REGRESSION The Annals of Statistics 1997, Vol. 25, No. 6, 2313 2327 LATTICE-BASED D-OPTIMUM DESIGN FOR FOURIER REGRESSION By Eva Riccomagno, 1 Rainer Schwabe 2 an Henry P. Wynn 1 University of Warwick, Technische

More information

Least-Squares Regression on Sparse Spaces

Least-Squares Regression on Sparse Spaces Least-Squares Regression on Sparse Spaces Yuri Grinberg, Mahi Milani Far, Joelle Pineau School of Computer Science McGill University Montreal, Canaa {ygrinb,mmilan1,jpineau}@cs.mcgill.ca 1 Introuction

More information

Tutorial on Maximum Likelyhood Estimation: Parametric Density Estimation

Tutorial on Maximum Likelyhood Estimation: Parametric Density Estimation Tutorial on Maximum Likelyhoo Estimation: Parametric Density Estimation Suhir B Kylasa 03/13/2014 1 Motivation Suppose one wishes to etermine just how biase an unfair coin is. Call the probability of tossing

More information

Diagonalization of Matrices Dr. E. Jacobs

Diagonalization of Matrices Dr. E. Jacobs Diagonalization of Matrices Dr. E. Jacobs One of the very interesting lessons in this course is how certain algebraic techniques can be use to solve ifferential equations. The purpose of these notes is

More information

Package sae. R topics documented: July 8, 2015

Package sae. R topics documented: July 8, 2015 Type Package Title Small Area Estimation Version 1.1 Date 2015-08-07 Author Isabel Molina, Yolanda Marhuenda Package sae July 8, 2015 Maintainer Yolanda Marhuenda Depends stats, nlme,

More information

The derivative of a function f(x) is another function, defined in terms of a limiting expression: f(x + δx) f(x)

The derivative of a function f(x) is another function, defined in terms of a limiting expression: f(x + δx) f(x) Y. D. Chong (2016) MH2801: Complex Methos for the Sciences 1. Derivatives The erivative of a function f(x) is another function, efine in terms of a limiting expression: f (x) f (x) lim x δx 0 f(x + δx)

More information

Separation of Variables

Separation of Variables Physics 342 Lecture 1 Separation of Variables Lecture 1 Physics 342 Quantum Mechanics I Monay, January 25th, 2010 There are three basic mathematical tools we nee, an then we can begin working on the physical

More information

Estimation of Complex Small Area Parameters with Application to Poverty Indicators

Estimation of Complex Small Area Parameters with Application to Poverty Indicators 1 Estimation of Complex Small Area Parameters with Application to Poverty Indicators J.N.K. Rao School of Mathematics and Statistics, Carleton University (Joint work with Isabel Molina from Universidad

More information

Combining Time Series and Cross-sectional Data for Current Employment Statistics Estimates 1

Combining Time Series and Cross-sectional Data for Current Employment Statistics Estimates 1 JSM015 - Surey Research Methos Section Combining Time Series an Cross-sectional Data for Current Employment Statistics Estimates 1 Julie Gershunskaya U.S. Bureau of Labor Statistics, Massachusetts Ae NE,

More information

arxiv: v1 [cs.lg] 22 Mar 2014

arxiv: v1 [cs.lg] 22 Mar 2014 CUR lgorithm with Incomplete Matrix Observation Rong Jin an Shenghuo Zhu Dept. of Computer Science an Engineering, Michigan State University, rongjin@msu.eu NEC Laboratories merica, Inc., zsh@nec-labs.com

More information

Computing Exact Confidence Coefficients of Simultaneous Confidence Intervals for Multinomial Proportions and their Functions

Computing Exact Confidence Coefficients of Simultaneous Confidence Intervals for Multinomial Proportions and their Functions Working Paper 2013:5 Department of Statistics Computing Exact Confience Coefficients of Simultaneous Confience Intervals for Multinomial Proportions an their Functions Shaobo Jin Working Paper 2013:5

More information

19 Eigenvalues, Eigenvectors, Ordinary Differential Equations, and Control

19 Eigenvalues, Eigenvectors, Ordinary Differential Equations, and Control 19 Eigenvalues, Eigenvectors, Orinary Differential Equations, an Control This section introuces eigenvalues an eigenvectors of a matrix, an iscusses the role of the eigenvalues in etermining the behavior

More information

Multi-View Clustering via Canonical Correlation Analysis

Multi-View Clustering via Canonical Correlation Analysis Technical Report TTI-TR-2008-5 Multi-View Clustering via Canonical Correlation Analysis Kamalika Chauhuri UC San Diego Sham M. Kakae Toyota Technological Institute at Chicago ABSTRACT Clustering ata in

More information

Poisson M-quantile regression for small area estimation

Poisson M-quantile regression for small area estimation University of Wollongong Research Online Centre for Statistical & Survey Methoology Working Paper Series Faculty of Engineering an Information Sciences 2013 Poisson M-quantile regression for small area

More information

Parametric Bootstrap Methods for Bias Correction in Linear Mixed Models

Parametric Bootstrap Methods for Bias Correction in Linear Mixed Models CIRJE-F-801 Parametric Bootstrap Methods for Bias Correction in Linear Mixed Models Tatsuya Kubokawa University of Tokyo Bui Nagashima Graduate School of Economics, University of Tokyo April 2011 CIRJE

More information

SYNCHRONOUS SEQUENTIAL CIRCUITS

SYNCHRONOUS SEQUENTIAL CIRCUITS CHAPTER SYNCHRONOUS SEUENTIAL CIRCUITS Registers an counters, two very common synchronous sequential circuits, are introuce in this chapter. Register is a igital circuit for storing information. Contents

More information

NOTES ON EULER-BOOLE SUMMATION (1) f (l 1) (n) f (l 1) (m) + ( 1)k 1 k! B k (y) f (k) (y) dy,

NOTES ON EULER-BOOLE SUMMATION (1) f (l 1) (n) f (l 1) (m) + ( 1)k 1 k! B k (y) f (k) (y) dy, NOTES ON EULER-BOOLE SUMMATION JONATHAN M BORWEIN, NEIL J CALKIN, AND DANTE MANNA Abstract We stuy a connection between Euler-MacLaurin Summation an Boole Summation suggeste in an AMM note from 196, which

More information

ELEC3114 Control Systems 1

ELEC3114 Control Systems 1 ELEC34 Control Systems Linear Systems - Moelling - Some Issues Session 2, 2007 Introuction Linear systems may be represente in a number of ifferent ways. Figure shows the relationship between various representations.

More information

The total derivative. Chapter Lagrangian and Eulerian approaches

The total derivative. Chapter Lagrangian and Eulerian approaches Chapter 5 The total erivative 51 Lagrangian an Eulerian approaches The representation of a flui through scalar or vector fiels means that each physical quantity uner consieration is escribe as a function

More information

Lectures - Week 10 Introduction to Ordinary Differential Equations (ODES) First Order Linear ODEs

Lectures - Week 10 Introduction to Ordinary Differential Equations (ODES) First Order Linear ODEs Lectures - Week 10 Introuction to Orinary Differential Equations (ODES) First Orer Linear ODEs When stuying ODEs we are consiering functions of one inepenent variable, e.g., f(x), where x is the inepenent

More information

Deliverable 2.2. Small Area Estimation of Indicators on Poverty and Social Exclusion

Deliverable 2.2. Small Area Estimation of Indicators on Poverty and Social Exclusion Deliverable 2.2 Small Area Estimation of Inicators on Poverty an Social Exclusion Version: 2011 Risto Lehtonen, Ari Veijanen, Mio Myrsylä an Maria Valaste The project FP7-SSH-2007-217322 AMELI is supporte

More information

Function Spaces. 1 Hilbert Spaces

Function Spaces. 1 Hilbert Spaces Function Spaces A function space is a set of functions F that has some structure. Often a nonparametric regression function or classifier is chosen to lie in some function space, where the assume structure

More information

WEIGHTING A RESAMPLED PARTICLE IN SEQUENTIAL MONTE CARLO. L. Martino, V. Elvira, F. Louzada

WEIGHTING A RESAMPLED PARTICLE IN SEQUENTIAL MONTE CARLO. L. Martino, V. Elvira, F. Louzada WEIGHTIG A RESAMPLED PARTICLE I SEQUETIAL MOTE CARLO L. Martino, V. Elvira, F. Louzaa Dep. of Signal Theory an Communic., Universia Carlos III e Mari, Leganés (Spain). Institute of Mathematical Sciences

More information

Introduction to the Vlasov-Poisson system

Introduction to the Vlasov-Poisson system Introuction to the Vlasov-Poisson system Simone Calogero 1 The Vlasov equation Consier a particle with mass m > 0. Let x(t) R 3 enote the position of the particle at time t R an v(t) = ẋ(t) = x(t)/t its

More information

Hyperbolic Moment Equations Using Quadrature-Based Projection Methods

Hyperbolic Moment Equations Using Quadrature-Based Projection Methods Hyperbolic Moment Equations Using Quarature-Base Projection Methos J. Koellermeier an M. Torrilhon Department of Mathematics, RWTH Aachen University, Aachen, Germany Abstract. Kinetic equations like the

More information

A Modification of the Jarque-Bera Test. for Normality

A Modification of the Jarque-Bera Test. for Normality Int. J. Contemp. Math. Sciences, Vol. 8, 01, no. 17, 84-85 HIKARI Lt, www.m-hikari.com http://x.oi.org/10.1988/ijcms.01.9106 A Moification of the Jarque-Bera Test for Normality Moawa El-Fallah Ab El-Salam

More information

Applications of the Wronskian to ordinary linear differential equations

Applications of the Wronskian to ordinary linear differential equations Physics 116C Fall 2011 Applications of the Wronskian to orinary linear ifferential equations Consier a of n continuous functions y i (x) [i = 1,2,3,...,n], each of which is ifferentiable at least n times.

More information

Optimization of Geometries by Energy Minimization

Optimization of Geometries by Energy Minimization Optimization of Geometries by Energy Minimization by Tracy P. Hamilton Department of Chemistry University of Alabama at Birmingham Birmingham, AL 3594-140 hamilton@uab.eu Copyright Tracy P. Hamilton, 1997.

More information

Likelihood-Based Methods

Likelihood-Based Methods Likelihood-Based Methods Handbook of Spatial Statistics, Chapter 4 Susheela Singh September 22, 2016 OVERVIEW INTRODUCTION MAXIMUM LIKELIHOOD ESTIMATION (ML) RESTRICTED MAXIMUM LIKELIHOOD ESTIMATION (REML)

More information

Influence of weight initialization on multilayer perceptron performance

Influence of weight initialization on multilayer perceptron performance Influence of weight initialization on multilayer perceptron performance M. Karouia (1,2) T. Denœux (1) R. Lengellé (1) (1) Université e Compiègne U.R.A. CNRS 817 Heuiasyc BP 649 - F-66 Compiègne ceex -

More information

Lecture 2: Correlated Topic Model

Lecture 2: Correlated Topic Model Probabilistic Moels for Unsupervise Learning Spring 203 Lecture 2: Correlate Topic Moel Inference for Correlate Topic Moel Yuan Yuan First of all, let us make some claims about the parameters an variables

More information

Calculus of Variations

Calculus of Variations 16.323 Lecture 5 Calculus of Variations Calculus of Variations Most books cover this material well, but Kirk Chapter 4 oes a particularly nice job. x(t) x* x*+ αδx (1) x*- αδx (1) αδx (1) αδx (1) t f t

More information

Chapter 5: Models used in conjunction with sampling. J. Kim, W. Fuller (ISU) Chapter 5: Models used in conjunction with sampling 1 / 70

Chapter 5: Models used in conjunction with sampling. J. Kim, W. Fuller (ISU) Chapter 5: Models used in conjunction with sampling 1 / 70 Chapter 5: Models used in conjunction with sampling J. Kim, W. Fuller (ISU) Chapter 5: Models used in conjunction with sampling 1 / 70 Nonresponse Unit Nonresponse: weight adjustment Item Nonresponse:

More information

Space-time Linear Dispersion Using Coordinate Interleaving

Space-time Linear Dispersion Using Coordinate Interleaving Space-time Linear Dispersion Using Coorinate Interleaving Jinsong Wu an Steven D Blostein Department of Electrical an Computer Engineering Queen s University, Kingston, Ontario, Canaa, K7L3N6 Email: wujs@ieeeorg

More information

Test of Hypotheses in a Time Trend Panel Data Model with Serially Correlated Error Component Disturbances

Test of Hypotheses in a Time Trend Panel Data Model with Serially Correlated Error Component Disturbances HE UNIVERSIY OF EXAS A SAN ANONIO, COLLEGE OF BUSINESS Working Paper SERIES Date September 25, 205 WP # 000ECO-66-205 est of Hypotheses in a ime ren Panel Data Moel with Serially Correlate Error Component

More information

Homework 3 - Solutions

Homework 3 - Solutions Homework 3 - Solutions The Transpose an Partial Transpose. 1 Let { 1, 2,, } be an orthonormal basis for C. The transpose map efine with respect to this basis is a superoperator Γ that acts on an operator

More information

ensembles When working with density operators, we can use this connection to define a generalized Bloch vector: v x Tr x, v y Tr y

ensembles When working with density operators, we can use this connection to define a generalized Bloch vector: v x Tr x, v y Tr y Ph195a lecture notes, 1/3/01 Density operators for spin- 1 ensembles So far in our iscussion of spin- 1 systems, we have restricte our attention to the case of pure states an Hamiltonian evolution. Toay

More information

Topic 7: Convergence of Random Variables

Topic 7: Convergence of Random Variables Topic 7: Convergence of Ranom Variables Course 003, 2016 Page 0 The Inference Problem So far, our starting point has been a given probability space (S, F, P). We now look at how to generate information

More information

UNIFYING PCA AND MULTISCALE APPROACHES TO FAULT DETECTION AND ISOLATION

UNIFYING PCA AND MULTISCALE APPROACHES TO FAULT DETECTION AND ISOLATION UNIFYING AND MULISCALE APPROACHES O FAUL DEECION AND ISOLAION Seongkyu Yoon an John F. MacGregor Dept. Chemical Engineering, McMaster University, Hamilton Ontario Canaa L8S 4L7 yoons@mcmaster.ca macgreg@mcmaster.ca

More information

BAYESIAN ESTIMATION OF THE NUMBER OF PRINCIPAL COMPONENTS

BAYESIAN ESTIMATION OF THE NUMBER OF PRINCIPAL COMPONENTS 4th European Signal Processing Conference EUSIPCO 006, Florence, Italy, September 4-8, 006, copyright by EURASIP BAYESIA ESTIMATIO OF THE UMBER OF PRICIPAL COMPOETS Ab-Krim Seghouane an Anrzej Cichocki

More information

Equilibrium in Queues Under Unknown Service Times and Service Value

Equilibrium in Queues Under Unknown Service Times and Service Value University of Pennsylvania ScholarlyCommons Finance Papers Wharton Faculty Research 1-2014 Equilibrium in Queues Uner Unknown Service Times an Service Value Laurens Debo Senthil K. Veeraraghavan University

More information

under the null hypothesis, the sign test (with continuity correction) rejects H 0 when α n + n 2 2.

under the null hypothesis, the sign test (with continuity correction) rejects H 0 when α n + n 2 2. Assignment 13 Exercise 8.4 For the hypotheses consiere in Examples 8.12 an 8.13, the sign test is base on the statistic N + = #{i : Z i > 0}. Since 2 n(n + /n 1) N(0, 1) 2 uner the null hypothesis, the

More information

Lecture Introduction. 2 Examples of Measure Concentration. 3 The Johnson-Lindenstrauss Lemma. CS-621 Theory Gems November 28, 2012

Lecture Introduction. 2 Examples of Measure Concentration. 3 The Johnson-Lindenstrauss Lemma. CS-621 Theory Gems November 28, 2012 CS-6 Theory Gems November 8, 0 Lecture Lecturer: Alesaner Mąry Scribes: Alhussein Fawzi, Dorina Thanou Introuction Toay, we will briefly iscuss an important technique in probability theory measure concentration

More information

Problems Governed by PDE. Shlomo Ta'asan. Carnegie Mellon University. and. Abstract

Problems Governed by PDE. Shlomo Ta'asan. Carnegie Mellon University. and. Abstract Pseuo-Time Methos for Constraine Optimization Problems Governe by PDE Shlomo Ta'asan Carnegie Mellon University an Institute for Computer Applications in Science an Engineering Abstract In this paper we

More information

Concentration of Measure Inequalities for Compressive Toeplitz Matrices with Applications to Detection and System Identification

Concentration of Measure Inequalities for Compressive Toeplitz Matrices with Applications to Detection and System Identification Concentration of Measure Inequalities for Compressive Toeplitz Matrices with Applications to Detection an System Ientification Borhan M Sananaji, Tyrone L Vincent, an Michael B Wakin Abstract In this paper,

More information

6 Wave equation in spherical polar coordinates

6 Wave equation in spherical polar coordinates 6 Wave equation in spherical polar coorinates We now look at solving problems involving the Laplacian in spherical polar coorinates. The angular epenence of the solutions will be escribe by spherical harmonics.

More information

Entanglement is not very useful for estimating multiple phases

Entanglement is not very useful for estimating multiple phases PHYSICAL REVIEW A 70, 032310 (2004) Entanglement is not very useful for estimating multiple phases Manuel A. Ballester* Department of Mathematics, University of Utrecht, Box 80010, 3508 TA Utrecht, The

More information

Model-based Estimation of Poverty Indicators for Small Areas: Overview. J. N. K. Rao Carleton University, Ottawa, Canada

Model-based Estimation of Poverty Indicators for Small Areas: Overview. J. N. K. Rao Carleton University, Ottawa, Canada Model-based Estimation of Poverty Indicators for Small Areas: Overview J. N. K. Rao Carleton University, Ottawa, Canada Isabel Molina Universidad Carlos III de Madrid, Spain Paper presented at The First

More information

Contextual Effects in Modeling for Small Domains

Contextual Effects in Modeling for Small Domains University of Wollongong Research Online Applied Statistics Education and Research Collaboration (ASEARC) - Conference Papers Faculty of Engineering and Information Sciences 2011 Contextual Effects in

More information

Subspace Estimation from Incomplete Observations: A High-Dimensional Analysis

Subspace Estimation from Incomplete Observations: A High-Dimensional Analysis Subspace Estimation from Incomplete Observations: A High-Dimensional Analysis Chuang Wang, Yonina C. Elar, Fellow, IEEE an Yue M. Lu, Senior Member, IEEE Abstract We present a high-imensional analysis

More information

Homework 2 Solutions EM, Mixture Models, PCA, Dualitys

Homework 2 Solutions EM, Mixture Models, PCA, Dualitys Homewor Solutions EM, Mixture Moels, PCA, Dualitys CMU 0-75: Machine Learning Fall 05 http://www.cs.cmu.eu/~bapoczos/classes/ml075_05fall/ OUT: Oct 5, 05 DUE: Oct 9, 05, 0:0 AM An EM algorithm for a Mixture

More information

A PAC-Bayesian Approach to Spectrally-Normalized Margin Bounds for Neural Networks

A PAC-Bayesian Approach to Spectrally-Normalized Margin Bounds for Neural Networks A PAC-Bayesian Approach to Spectrally-Normalize Margin Bouns for Neural Networks Behnam Neyshabur, Srinah Bhojanapalli, Davi McAllester, Nathan Srebro Toyota Technological Institute at Chicago {bneyshabur,

More information

TEMPORAL AND TIME-FREQUENCY CORRELATION-BASED BLIND SOURCE SEPARATION METHODS. Yannick DEVILLE

TEMPORAL AND TIME-FREQUENCY CORRELATION-BASED BLIND SOURCE SEPARATION METHODS. Yannick DEVILLE TEMPORAL AND TIME-FREQUENCY CORRELATION-BASED BLIND SOURCE SEPARATION METHODS Yannick DEVILLE Université Paul Sabatier Laboratoire Acoustique, Métrologie, Instrumentation Bât. 3RB2, 8 Route e Narbonne,

More information

Math Notes on differentials, the Chain Rule, gradients, directional derivative, and normal vectors

Math Notes on differentials, the Chain Rule, gradients, directional derivative, and normal vectors Math 18.02 Notes on ifferentials, the Chain Rule, graients, irectional erivative, an normal vectors Tangent plane an linear approximation We efine the partial erivatives of f( xy, ) as follows: f f( x+

More information

Linear First-Order Equations

Linear First-Order Equations 5 Linear First-Orer Equations Linear first-orer ifferential equations make up another important class of ifferential equations that commonly arise in applications an are relatively easy to solve (in theory)

More information

THE EFFICIENCIES OF THE SPATIAL MEDIAN AND SPATIAL SIGN COVARIANCE MATRIX FOR ELLIPTICALLY SYMMETRIC DISTRIBUTIONS

THE EFFICIENCIES OF THE SPATIAL MEDIAN AND SPATIAL SIGN COVARIANCE MATRIX FOR ELLIPTICALLY SYMMETRIC DISTRIBUTIONS THE EFFICIENCIES OF THE SPATIAL MEDIAN AND SPATIAL SIGN COVARIANCE MATRIX FOR ELLIPTICALLY SYMMETRIC DISTRIBUTIONS BY ANDREW F. MAGYAR A issertation submitte to the Grauate School New Brunswick Rutgers,

More information

Optimal Signal Detection for False Track Discrimination

Optimal Signal Detection for False Track Discrimination Optimal Signal Detection for False Track Discrimination Thomas Hanselmann Darko Mušicki Dept. of Electrical an Electronic Eng. Dept. of Electrical an Electronic Eng. The University of Melbourne The University

More information

Spectral properties of a near-periodic row-stochastic Leslie matrix

Spectral properties of a near-periodic row-stochastic Leslie matrix Linear Algebra an its Applications 409 2005) 66 86 wwwelseviercom/locate/laa Spectral properties of a near-perioic row-stochastic Leslie matrix Mei-Qin Chen a Xiezhang Li b a Department of Mathematics

More information

ECE 422 Power System Operations & Planning 7 Transient Stability

ECE 422 Power System Operations & Planning 7 Transient Stability ECE 4 Power System Operations & Planning 7 Transient Stability Spring 5 Instructor: Kai Sun References Saaat s Chapter.5 ~. EPRI Tutorial s Chapter 7 Kunur s Chapter 3 Transient Stability The ability of

More information

Robust Low Rank Kernel Embeddings of Multivariate Distributions

Robust Low Rank Kernel Embeddings of Multivariate Distributions Robust Low Rank Kernel Embeings of Multivariate Distributions Le Song, Bo Dai College of Computing, Georgia Institute of Technology lsong@cc.gatech.eu, boai@gatech.eu Abstract Kernel embeing of istributions

More information

inflow outflow Part I. Regular tasks for MAE598/494 Task 1

inflow outflow Part I. Regular tasks for MAE598/494 Task 1 MAE 494/598, Fall 2016 Project #1 (Regular tasks = 20 points) Har copy of report is ue at the start of class on the ue ate. The rules on collaboration will be release separately. Please always follow the

More information

LOGISTIC regression model is often used as the way

LOGISTIC regression model is often used as the way Vol:6, No:9, 22 Secon Orer Amissibilities in Multi-parameter Logistic Regression Moel Chie Obayashi, Hiekazu Tanaka, Yoshiji Takagi International Science Inex, Mathematical an Computational Sciences Vol:6,

More information

Physics 505 Electricity and Magnetism Fall 2003 Prof. G. Raithel. Problem Set 3. 2 (x x ) 2 + (y y ) 2 + (z + z ) 2

Physics 505 Electricity and Magnetism Fall 2003 Prof. G. Raithel. Problem Set 3. 2 (x x ) 2 + (y y ) 2 + (z + z ) 2 Physics 505 Electricity an Magnetism Fall 003 Prof. G. Raithel Problem Set 3 Problem.7 5 Points a): Green s function: Using cartesian coorinates x = (x, y, z), it is G(x, x ) = 1 (x x ) + (y y ) + (z z

More information

Analytic Scaling Formulas for Crossed Laser Acceleration in Vacuum

Analytic Scaling Formulas for Crossed Laser Acceleration in Vacuum October 6, 4 ARDB Note Analytic Scaling Formulas for Crosse Laser Acceleration in Vacuum Robert J. Noble Stanfor Linear Accelerator Center, Stanfor University 575 San Hill Roa, Menlo Park, California 945

More information

Introduction to Markov Processes

Introduction to Markov Processes Introuction to Markov Processes Connexions moule m44014 Zzis law Gustav) Meglicki, Jr Office of the VP for Information Technology Iniana University RCS: Section-2.tex,v 1.24 2012/12/21 18:03:08 gustav

More information

Math 342 Partial Differential Equations «Viktor Grigoryan

Math 342 Partial Differential Equations «Viktor Grigoryan Math 342 Partial Differential Equations «Viktor Grigoryan 6 Wave equation: solution In this lecture we will solve the wave equation on the entire real line x R. This correspons to a string of infinite

More information

Euler Equations: derivation, basic invariants and formulae

Euler Equations: derivation, basic invariants and formulae Euler Equations: erivation, basic invariants an formulae Mat 529, Lesson 1. 1 Derivation The incompressible Euler equations are couple with t u + u u + p = 0, (1) u = 0. (2) The unknown variable is the

More information

Bayesian Estimation of the Entropy of the Multivariate Gaussian

Bayesian Estimation of the Entropy of the Multivariate Gaussian Bayesian Estimation of the Entropy of the Multivariate Gaussian Santosh Srivastava Fre Hutchinson Cancer Research Center Seattle, WA 989, USA Email: ssrivast@fhcrc.org Maya R. Gupta Department of Electrical

More information

Lecture 2 Lagrangian formulation of classical mechanics Mechanics

Lecture 2 Lagrangian formulation of classical mechanics Mechanics Lecture Lagrangian formulation of classical mechanics 70.00 Mechanics Principle of stationary action MATH-GA To specify a motion uniquely in classical mechanics, it suffices to give, at some time t 0,

More information

Gaussian processes with monotonicity information

Gaussian processes with monotonicity information Gaussian processes with monotonicity information Anonymous Author Anonymous Author Unknown Institution Unknown Institution Abstract A metho for using monotonicity information in multivariate Gaussian process

More information

Inter-domain Gaussian Processes for Sparse Inference using Inducing Features

Inter-domain Gaussian Processes for Sparse Inference using Inducing Features Inter-omain Gaussian Processes for Sparse Inference using Inucing Features Miguel Lázaro-Greilla an Aníbal R. Figueiras-Vial Dep. Signal Processing & Communications Universia Carlos III e Mari, SPAIN {miguel,arfv}@tsc.uc3m.es

More information

The Sokhotski-Plemelj Formula

The Sokhotski-Plemelj Formula hysics 25 Winter 208 The Sokhotski-lemelj Formula. The Sokhotski-lemelj formula The Sokhotski-lemelj formula is a relation between the following generalize functions (also calle istributions), ±iǫ = iπ(),

More information

Optimized Schwarz Methods with the Yin-Yang Grid for Shallow Water Equations

Optimized Schwarz Methods with the Yin-Yang Grid for Shallow Water Equations Optimize Schwarz Methos with the Yin-Yang Gri for Shallow Water Equations Abessama Qaouri Recherche en prévision numérique, Atmospheric Science an Technology Directorate, Environment Canaa, Dorval, Québec,

More information

All s Well That Ends Well: Supplementary Proofs

All s Well That Ends Well: Supplementary Proofs All s Well That Ens Well: Guarantee Resolution of Simultaneous Rigi Boy Impact 1:1 All s Well That Ens Well: Supplementary Proofs This ocument complements the paper All s Well That Ens Well: Guarantee

More information

Estimating International Migration on the Base of Small Area Techniques

Estimating International Migration on the Base of Small Area Techniques MPRA Munich Personal RePEc Archive Estimating International Migration on the Base of Small Area echniques Vergil Voineagu an Nicoleta Caragea an Silvia Pisica 2013 Online at http://mpra.ub.uni-muenchen.e/48775/

More information

Multivariate area level models for small area estimation. a

Multivariate area level models for small area estimation. a Multivariate area level models for small area estimation. a a In collaboration with Roberto Benavent Domingo Morales González d.morales@umh.es Universidad Miguel Hernández de Elche Multivariate area level

More information

A New Minimum Description Length

A New Minimum Description Length A New Minimum Description Length Soosan Beheshti, Munther A. Dahleh Laboratory for Information an Decision Systems Massachusetts Institute of Technology soosan@mit.eu,ahleh@lis.mit.eu Abstract The minimum

More information

Relative Entropy and Score Function: New Information Estimation Relationships through Arbitrary Additive Perturbation

Relative Entropy and Score Function: New Information Estimation Relationships through Arbitrary Additive Perturbation Relative Entropy an Score Function: New Information Estimation Relationships through Arbitrary Aitive Perturbation Dongning Guo Department of Electrical Engineering & Computer Science Northwestern University

More information

Optimal CDMA Signatures: A Finite-Step Approach

Optimal CDMA Signatures: A Finite-Step Approach Optimal CDMA Signatures: A Finite-Step Approach Joel A. Tropp Inst. for Comp. Engr. an Sci. (ICES) 1 University Station C000 Austin, TX 7871 jtropp@ices.utexas.eu Inerjit. S. Dhillon Dept. of Comp. Sci.

More information

1 Heisenberg Representation

1 Heisenberg Representation 1 Heisenberg Representation What we have been ealing with so far is calle the Schröinger representation. In this representation, operators are constants an all the time epenence is carrie by the states.

More information

Chapter 2 Lagrangian Modeling

Chapter 2 Lagrangian Modeling Chapter 2 Lagrangian Moeling The basic laws of physics are use to moel every system whether it is electrical, mechanical, hyraulic, or any other energy omain. In mechanics, Newton s laws of motion provie

More information