Share
Embed
Transcript
16 slides
-
Penalized Likelihood Estimation of Trivariate Binary Models Panagiota Filippou 1
Giampiero Marra 2 Rosalba Radice 3 1,2Department of Statistical Science, University College London 3Department of Economics, Mathematics and Statistics, Birkbeck, University of London July 4, 2017 SAM conference, University of Liverpool -
Introduction Aim: estimate a trivariate system of binary regressions which
accounts for several types of covariate effects (e.g., linear, non-linear and spatial effects) residual dependence between the responses. Example: model jointly multiple births (i.e., twins, triplets, etc.) premature birth (i.e., infant was born before completing the 37 gestational week) low birth weight (i.e., infant’s birth weight is ≤ 2500 grams). Trivariate probit models deal with these problems. -
The trivariate probit model y∗ 1i = η1i + ε1i
= v1i γ1 + ˜ N1 ν1=1 s1ν1 (z1ν1i ) + ε1i y∗ 2i = η2i + ε2i = v2i γ2 + ˜ N2 ν2=1 s2ν2 (z2ν2i ) + ε2i y∗ 3i = η3i + ε3i = v3i γ3 + ˜ N3 ν3=1 s3ν3 (z3ν3i ) + ε3i , i = 1, . . . , n. y∗ mi is a latent continuous variable and ymi = 1(y∗ mi > 0), ∀m = 1, 2, 3. vmi γm and ˜ Nm νm=1 smνm (zmνmi ) refer to parametric and non- parametric components. (ε1i , ε2i , ε3i ) = εi iid ∼ N3(0, Σ), Σ = 1 ϑ12 ϑ13 ϑ21 1 ϑ23 ϑ31 ϑ32 1 . -
Estimation Log-likelihood function for the trivariate probit model: (δ) =
n i=1 {y1i y2i y3i log p111i + y1i y2i ˜ y3i log p110i + y1i ˜ y2i y3i log p101i + ˜ y1i y2i y3i log p011i + ˜ y1i ˜ y2i ˜ y3i log p000i + ˜ y1i y2i ˜ y3i log p010i +˜ y1i ˜ y2i y3i log p001i + y1i ˜ y2i ˜ y3i log p100i } , where ˜ ymi = (1−ymi ), p¯ e1 ¯ e2 ¯ e3 = P(y1i = ¯ e1, y2i = ¯ e2, y3i = ¯ e3), ∀¯ em ∈ {0, 1}, ∀m = 1, 2, 3. Simultaneous estimation of all parameters can be achieved by penalized MLE (PMLE) solving the optimization problem ˆ δ := arg min δ − (δ) − 1 2 δ ˜ Sλδ , where δ ˜ Sλδ is a penalty term that controls for the model’s smoothness and depends on the smoothing parameter λ which controls the trade-off between fit and smoothness. -
Simulation results–DGP1 Method Comparison − DGP1 − n = 1000
−0.9 −0.7 −0.5 mvprobit SemiParTRIV ϑ ^ 12 −0.8−0.6−0.4−0.2 mvprobit SemiParTRIV ϑ ^ 13 0.4 0.6 0.8 1.0 mvprobit SemiParTRIV ϑ ^ 23 −1.8 −1.4 γ ^ 31 1.6 2.0 2.4 γ ^ 32 −2.0 −1.5 −1.0 γ ^ 33 −1.4 −1.0 −0.6 γ ^ 21 −1.8 −1.4 −1.0 γ ^ 22 0.5 1.0 1.5 γ ^ 23 1.41.61.82.0 γ ^ 11 0.6 0.8 1.0 1.2 γ ^ 12 −2.0 −1.5 −1.0 γ ^ 13 -
Method Comparison − DGP1 − n = 10000 −0.9 −0.7
−0.5 mvprobit SemiParTRIV ϑ ^ 12 −0.8−0.6−0.4−0.2 mvprobit SemiParTRIV ϑ ^ 13 0.4 0.6 0.8 1.0 mvprobit SemiParTRIV ϑ ^ 23 −1.8 −1.4 γ ^ 31 1.6 2.0 2.4 γ ^ 32 −2.0 −1.5 −1.0 γ ^ 33 −1.4 −1.0 −0.6 γ ^ 21 −1.8 −1.4 −1.0 γ ^ 22 0.5 1.0 1.5 γ ^ 23 1.41.61.82.0 γ ^ 11 0.6 0.8 1.0 1.2 γ ^ 12 −2.0 −1.5 −1.0 γ ^ 13 -
Simulation results–DGP2 The particular choice of the correlation parameters seemed
to be problematic from an estimation perspective: convergence was not achieved in most of the replicates the estimates of the correlation parameters were not close to the true values. n = 1000 n = 10000 Estimator Bias (%) RMSE Bias (%) RMSE ˆ ϑ12 11.36 0.0935 -0.79 0.0262 ˆ ϑ13 13.53 0.1204 1.86 0.0320 ˆ ϑ23 -2.02 0.0567 0.16 0.0129 Table 1: Percentage biases and root mean squared errors of the correlation estimates obtained applying SemiParTRIV(). -
Correlation-based penalty The issue with estimating the correlation parameters can
be addressed by optimization of ˆ δ := arg min δ − (δ) − 1 2 δ ˜ Sλδ − Pλϑ (δ) , where Pλϑ (δ) is a penalty acting on the correlations. In this work, we employ the Ridge, Lasso and Adaptive Lasso ap- proaches. The non-differentiability of the Lasso and Adaptive Lasso can be avoided by employing the local quadratic approximation ap- proach Pλϑ (δ) ≈ 1 2 δ Λλϑ δ, where Λλϑ = 0Q×Q 0Q×3 03×Q Aλϑ . -
Simulation results–DGP2 (correlation-based penalty) n=1000 Estimator Correlation-based penalty Bias (%)
RMSE ˆ ϑ12 Unpenalized 11.36 0.0935 Ridge 0.10 0.0903 Lasso 0.02 0.0835 Adaptive Lasso -0.31 0.0862 ˆ ϑ13 Unpenalized 13.53 0.1204 Ridge 0.13 0.1158 Lasso 0.07 0.1092 Adaptive Lasso 0.03 0.1142 ˆ ϑ23 Unpenalized -2.02 0.0567 Ridge -0.03 0.0551 Lasso -0.02 0.0475 Adaptive Lasso 0.01 0.0428 Table 2: Percentage biases and root mean squared errors of the correlation estimates. -
Case study The data set consists of 61,426 female newborns
which provides details on infant and maternal health, and parental characteris- tics using 2007−2008 birth data from the North Carolina Center of Health Statistics (http://www.schs.state.nc.us/). Aim: analyse jointly multiple births (= 1 if singleton and = 0 if twins, triplets, quadruplets and quintuplets) premature birth (= 1 if infant was born before completing the 37 gestational week and = 0 otherwise) low birth weight (= 1 if infant’s birth weight is ≤ 2500 grams and = 0 otherwise). Parameter estimation was carried out without the need of im- posing a penalty on the correlation coefficients. -
Case study results SemiParTRIV() mvprobit() ˆ ϑ12 -0.7617 (-0.7612, -0.7622)
-0.5191 (-0.5027, -0.5351) ˆ ϑ13 -0.6397 (-0.6390, -0.6402) -0.4277 (-0.4107, -0.4443) ˆ ϑ23 0.7853 ( 0.7850, 0.7856) 0.6796 ( 0.6692, 0.6897) Time 296.26 349.41 Table 3: Correlation parameter estimates obtained using SemiParTRIV() and mvprobit(). Corresponding 95% intervals (CIs) are reported in parentheses. The execution time (in seconds) for each method is reported at the bottom of the table. -
Case study results (continued) 0.5 1.0 1.5 SemiParTRIV 0.5 1.0
1.5 mvprobit Figure 1: Joint probabilities (in %) that birth is multiple, infant’s birth weight is normal and the baby is born full term by county in North Carolina, obtained using SemiParTRIV() and mvprobit(). -
Extending the trivariate probit model Extension I: Modelling unobserved confounders:
Endogeneity of a treatment variable Non-random sample selection of individuals into (or out) of sample. Trivariate Models deal with these problems: Endogenous trivariate model Double sample selection model Endogenous–sample selection model. Extension II: Trivariate Gaussian copula models with arbitrary margins Φ3 Φ−1(F1 (η1i )), Φ−1(F2 (η2i )), Φ−1(F3 (η3i )) , where Fm (ηmi ) can be either the normal, logistic or Gumbel univariate cdf, ∀m = 1, 2, 3. -
Extension III: Trivariate binary models with flexible association parameters: ϑ12,i
= η12,i = v12,i γ12 + ˜ N12 ν12=1 s12ν12 (z12ν12,i ) ϑ13,i = η13,i = v13,i γ13 + ˜ N13 ν13=1 s13ν13 (z13ν13,i ) ϑ23,i = η23,i = v23,i γ23 + ˜ N23 ν23=1 s23ν23 (z23ν23,i ) , i = 1, . . . , n. -
The SemiParTRIV() routine The models can be easily estimated using
the SemiParTRIV() rou- tine in the R package SemiParBIVProbit: out <- SemiParTRIV(formula = f.l, data = dat, margins = marg, Model = mod, penCor = PenFun) where f.l <- list(eqn1, eqn2, eqn3, eqn4, eqn5, eqn6) eqn1 <- y1 ~ v1 + z1; eqn2 <- y2 ~ v2 + z2; eqn3 <- y3 ~ v3 + z3; eqn4 <- ~ v1 + z1; eqn5 <- ~ v2 + z2; eqn6 <- ~ v3 + z3 marg <- c("probit", "logit", "cloglog") mod <- "T"; mod <- "TSS"; mod <- "ESS" PenFun <- "ridge"; PenFun <- "lasso"; PenFun <- "alasso"