ssmodels: A package for fit the sample selection models

1 Introduction

In order to facilitate the adjustment of the sample selection models existing in the literature, we created the ssmodels package. Our package allows the adjustment of the classic Heckman model (Heckman (1976), Heckman (1979)), and the estimation of the parameters of this model via the maximum likelihood method and two-step method, in addition to the adjustment of the Heckman-t models, introduced in the literature by Marchenko and Genton (2012) and the Heckman-Skew model introduced in the literature by Ogundimu and Hutton (2016). We also implemented functions to adjust the generalized version of the Heckman model, introduced by Bastos, Barreto-Souza, and Genton (2022), that allows the inclusion of covariables to the dispersion and correlation parameters and a function to adjust the Heckman-BS model introduced by Bastos and Barreto-Souza (2020) that uses the Birnbaum-Saunders distribution as a joint distribution of the selection and primary regression variables.

2 Real Databases

2.1 MEPS 2001

The MEPS is a set of large-scale surveys of families, individuals and their medical providers (doctors, hospitals, pharmacies, etc.) in the United States. It has data on the health services Americans use, how often they use them, the cost of these services and how they are paid, as well as data on the cost and reach of health insurance available to American workers. The sample is restricted to persons aged between 21 and 64 years and contains a variable response with 3328 observations of outpatient costs, of which 526 (15.8%) correspond to unobserved expenditure values and identified as zero expenditure for adjustment of the models. It also includes the following explanatory variables:

library(ssmodels)
#Leitura do dados MEPS2001
data(MEPS2001)
#tornando visiveis as colunas do data-frame
attach(MEPS2001)
barfill <- "grey"
barlines <- "black"
p1 <- ggplot(MEPS2001,aes(ambexp))+geom_histogram(colour = barlines, fill = barfill)+
  scale_x_continuous(name = "(a) Expenditures Medical",
                     breaks = seq(0, 15000, 2500),
                     limits=c(0, 15000))+
  scale_y_continuous(name = "Count",
                     breaks = seq(0, 800, 100),
                     limits=c(0, 800))
p2 <- ggplot(MEPS2001,aes(lambexp))+geom_histogram(colour = barlines, fill = barfill)+
  scale_x_continuous(name = "(b) Log of Expenditures Medical",
                     breaks = seq(0, 11, 1),
                     limits=c(0, 11))+
  scale_y_continuous(name = "Count",
                     breaks = seq(0, 300, 100),
                     limits=c(0, 300))
grid.arrange(p1, p2, ncol=2)

According Marchenko and Genton (2012) and Zhelonkin, Genton, and Ronchetti (2016) it is natural to fit a sample selection model to such data, since the willingness to spend \((Y_{2}^{*})\) is likely to be related to the expense amount \((Y_{1}^{*}).\) However, after fitting the classic Heckman model and using the Wald, likelihood ratio, or gradient tests for \(H_{0}:\rho=0\) against \(H_{1}:\rho\neq 0,\) the conclusion is that there is not sufficient evidence \((p\textrm{-valor}>0.1)\) to reject \(H_{0},\) that is, there is no bias of selection. Colin and Trivedi (2009) suspected this conclusion and Marchenko and Genton (2012) argued that a more robust model could evidence the selection bias present in the data and rejected the normality hypothesis of the data. However, we understand that the adjustment of the dispersion and/or the correlation can be an alternative to study these data without the need to modify the assumption of normality of the errors, because as our simulations indicate, the estimations of the parameters obtained by the adjustment of the model Heckman can be more severely affected by heteroscedasticity than by incorrect distribution of error terms.

However, due to data asymmetry, we also adjusted the Heckman-BS model. Thus, we consider the following system of equations for classic and generalized Heckman models, Heckman-t and Heckman-Skew:

\[\begin{align} lnambx &= \beta_{1}+ \beta_{2}age + \beta_{3}female + \beta_{4}educ + \beta_{5}blhisp + \beta_{6}totchr + \beta_{7}ins + \epsilon_{1i}, \label{eq_reg_aplic1} \\ dambexp&= \gamma_{1}+\gamma_{2}age + \gamma_{3}female + \gamma_{4}educ + \gamma_{5}blhisp + \gamma_{6}totchr + \gamma_{7}ins + \gamma_{8}income + \epsilon_{2i}, \label{eq_sel_aplic1} \end{align}\] where errors are defined for classic Heckman models, Heckman-t and Heckman-Skew according to \[\begin{align}\label{2:disterro} \begin{pmatrix}\epsilon_{1i}\\ \epsilon_{2i} \end{pmatrix} &\overset{ind.}{\sim} \mathcal{N} \begin{bmatrix} \begin{pmatrix} 0\\ 0 \end{pmatrix}\!\!,& \begin{pmatrix} \sigma^{2} & \rho\sigma \\ \sigma\rho & 1 \end{pmatrix} \end{bmatrix}, i=1,\cdots,n, \end{align}\] and for generalized Heckman models how \[\begin{align}\label{2:disterro1} \begin{pmatrix}\epsilon_{1i}\\ \epsilon_{2i} \end{pmatrix} &\overset{ind.}{\sim} \mathcal{N} \begin{bmatrix} \begin{pmatrix} 0\\ 0 \end{pmatrix}\!\!,& \begin{pmatrix} \sigma_{i}^{2} & \rho_{i}\sigma_{i} \\ \sigma_{i}\rho_{i} & 1 \end{pmatrix} \end{bmatrix}, i=1,\cdots,n, \end{align}\] in such a way that \[\begin{align*} \log{(\sigma_{i})} &= \lambda_{1}+\lambda_{2}age + \lambda_{3}female + \lambda_{4}totchr + \lambda_{5}ins\quad \textrm{and}\\ \rho_{i}&=\tanh{(\kappa_{0})}, i=1,\cdots, 3328. \end{align*}\]

To adjust the Heckman-BS model we consider \[\begin{align*} \log\mu_{1i} &= \beta_{1} + \beta_{2}age + \beta_{3}female + \beta_{4}educ + \beta_{5}blhisp + \beta_{6}totchr + \beta_{7}ins,\\ \log\mu_{2i} &= \gamma_{1}+\gamma_{2}age + \gamma_{3}female + \gamma_{4}educ + \gamma_{5}blhisp + \gamma_{6}totchr + \gamma_{7}ins + \gamma_{8}income,\quad i=1,2,\cdots,3328. \end{align*}\]

and \((ambexp,dambexp)\) is independently distributed with Birnbaum-Saunders (BS) Bivariate parameter distribution \(\mu_{1}>0, \mu_{2}>0, \phi_{1}=\phi>0, \phi_{2}=1\) and \(-1<\rho<1,\) that is, \[(ambexp_{i},dambexp_{i})\stackrel{ind}{\sim}\mbox{ BS}(\mu_{1i},\mu_{2i}, \phi,1,\rho), \mu_{1i}>0, \mu_{2i}>0, \phi>0\quad \textrm{and}\quad -1<\rho<1, i=1,\cdots, n\]

We observed similar values for the parameters estimated by all models, mainly the parameters of the variable of interest. However, the only model that allows a direct interpretation of such parameters is the Heckman-BS model. We can, for example, affirm, based on the result of this model, that keeping the other parameters fixed, changing a unit in age represents an increase of \(24.6\%\) in ambulatory expenses. For the other models, the interpretation is related to the log of outpatient expenses. See the results of the estimates of the parameters below:

selectEq <- dambexp ~ age + female + educ + blhisp + totchr + ins + income
outcomeEq <- lnambx ~ age + female + educ + blhisp + totchr + ins 
outcomeS <- cbind(age,female,totchr,ins)
outcomeC <- 1
outcomeBS <- ambexp ~ age + female + educ + blhisp + totchr + ins 
mCL <- HeckmanCL(selectEq, outcomeEq, data = MEPS2001)
mBS <- HeckmanBS(selectEq, outcomeBS, data = MEPS2001)
mSK <- HeckmanSK(selectEq, outcomeEq, data = MEPS2001,lambda = 1)
mtS <- HeckmantS(selectEq, outcomeEq, data = MEPS2001,df=12)
mGe <- HeckmanGe(selectEq, outcomeEq,outcomeS, outcomeC, data = MEPS2001)
Parameters <- c("Intercept", "age", "female", "educ", "blhisp", "totchr", "ins", "income",
   "Intercept", "age", "female", "educ", "blhisp", "totchr", "ins", "sigma", "age", "female", "totchr", "ins", "rho", "nu", "lambda")
HBS <- round(mBS$coefficients, digits = 3)
HCL <- round(mCL$coefficients, digits = 3)
HSK <- round(mSK$coefficients, digits = 3)
HtS <- round(mtS$coefficients, digits = 3)
HGe <- round(mGe$coefficients, digits = 3)
Results <- data.frame("Parameters"= Parameters,
                      "HeckmanGe" = c(HGe[1:21], "NA", "NA"), 
                      "HeckmanCL" = c(HCL[1:16], "NA", "NA", "NA", "NA", HCL[17], "NA", "NA"),
                      "HeckmanBS" = c(HBS[1:16], "NA", "NA", "NA", "NA", HBS[17], "NA", "NA"), 
                      "HeckmantS" = c(HtS[1:16], "NA", "NA", "NA", "NA", HtS[17:18], "NA" ),
                      "HeckmanSK" = c(HSK[1:16], "NA", "NA", "NA", "NA", HSK[17], "NA", HSK[18]))
kable(Results, format = "html", align = c("c", "c", "c", "c", "c"))
Parameters HeckmanGe HeckmanCL HeckmanBS HeckmantS HeckmanSK
Intercept -0.59 -0.676 0.117 -0.748 -0.899
age 0.094 0.088 0.09 0.099 0.087
female 0.632 0.663 0.722 0.725 0.645
educ 0.055 0.062 0.068 0.065 0.06
blhisp -0.33 -0.364 -0.399 -0.394 -0.351
totchr 0.722 0.797 0.806 0.89 0.775
ins 0.165 0.17 0.179 0.18 0.17
income 0.002 0.003 0.003 0.003 0.003
Intercept 5.862 5.044 5.779 5.206 6.452
age 0.172 0.212 0.246 0.207 0.202
female 0.15 0.348 0.411 0.307 0.287
educ 0.001 0.019 -0.007 0.017 0.013
blhisp -0.135 -0.219 -0.215 -0.193 -0.201
totchr 0.413 0.54 0.589 0.513 0.492
ins -0.109 -0.03 -0.061 -0.052 -0.08
sigma 0.613 1.271 0.703 1.195 1.736
age -0.029 NA NA NA NA
female -0.13 NA NA NA NA
totchr -0.119 NA NA NA NA
ins -0.112 NA NA NA NA
rho -0.742 -0.131 0.273 -0.322 -0.339
nu NA NA NA 12.93 NA
lambda NA NA NA NA -1.662

In addition, there are also the results of the significance of the parameters. The classic Heckman model adjusted to such data indicates no correlation between the variables value spent and the decision to spend. According to Colin and Trivedi (2009), such a result is suspect and should be better analyzed.

summary(mCL)
## 
## --------------------------------------------------------------
##           Classic Heckman Model (Package: ssmodels)           
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 22 
## Log-Likelihood: -5836.219 
## AIC: 11706.44 BIC: 11810.31 
## Number of observations: ( 526 censored and 2802 observed ) 
## 17 free parameters ( df= 3311 ) 
## --------------------------------------------------------------
## Probit selection equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -0.676054   0.194029  -3.484  0.00050 ***
## age          0.087936   0.027421   3.207  0.00135 ** 
## female       0.662665   0.060937  10.875  < 2e-16 ***
## educ         0.061948   0.012029   5.150 2.76e-07 ***
## blhisp      -0.363937   0.061874  -5.882 4.46e-09 ***
## totchr       0.796952   0.071130  11.204  < 2e-16 ***
## ins          0.170137   0.062871   2.706  0.00684 ** 
## income       0.002708   0.001316   2.057  0.03973 *  
## --------------------------------------------------------------
## Outcome equation:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  5.04406    0.22813  22.111  < 2e-16 ***
## age          0.21197    0.02301   9.213  < 2e-16 ***
## female       0.34814    0.06011   5.791 7.64e-09 ***
## educ         0.01872    0.01055   1.774 0.076078 .  
## blhisp      -0.21857    0.05967  -3.663 0.000253 ***
## totchr       0.53992    0.03933  13.727  < 2e-16 ***
## ins         -0.02999    0.05109  -0.587 0.557260    
## --------------------------------------------------------------
## Error terms:
##       Estimate Std. Error t value Pr(>|t|)    
## sigma  1.27102    0.01838  69.157   <2e-16 ***
## rho   -0.13060    0.14708  -0.888    0.375    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------

The generalized Heckman model, on the other hand, indicates that there is the presence of heteroscedasticity in the data and, probably, due to this, the classic Heckman model is not the most suitable for modeling these data. Our model, in addition to indicating the presence of variable dispersion, also indicates a significant correlation between the variables amount spent and the decision to spend.

summary(mGe)
## 
## --------------------------------------------------------------
##        Generalized Heckman Model (Package: ssmodels)          
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 68 
## Log-Likelihood: -5804.973 
## AIC: 11651.95 BIC: 11780.26 
## Number of observations: ( 526 censored and 2802 observed ) 
## 21 free parameters ( df= 3307 ) 
## --------------------------------------------------------------
## Probit selection equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -0.589895   0.184517  -3.197 0.001402 ** 
## age          0.093828   0.025969   3.613 0.000307 ***
## female       0.631622   0.058818  10.739  < 2e-16 ***
## educ         0.055187   0.011359   4.858 1.24e-06 ***
## blhisp      -0.329940   0.058875  -5.604 2.27e-08 ***
## totchr       0.722090   0.068092  10.605  < 2e-16 ***
## ins          0.164968   0.060096   2.745 0.006082 ** 
## income       0.002379   0.001208   1.969 0.049045 *  
## --------------------------------------------------------------
## Outcome equation:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  5.8622175  0.1931949  30.344  < 2e-16 ***
## age          0.1721672  0.0236647   7.275  4.3e-13 ***
## female       0.1498488  0.0554233   2.704  0.00689 ** 
## educ         0.0009664  0.0101716   0.095  0.92431    
## blhisp      -0.1346592  0.0577309  -2.333  0.01973 *  
## totchr       0.4131455  0.0285553  14.468  < 2e-16 ***
## ins         -0.1090664  0.0513537  -2.124  0.03376 *  
## --------------------------------------------------------------
## Dispersion terms:
##            Estimate Std. Error t value Pr(>|t|)    
## interceptS  0.61322    0.06434   9.531  < 2e-16 ***
## age        -0.02863    0.01274  -2.247   0.0247 *  
## female     -0.12974    0.02849  -4.554 5.46e-06 ***
## totchr     -0.11853    0.01849  -6.411 1.66e-10 ***
## ins        -0.11224    0.02792  -4.021 5.93e-05 ***
## --------------------------------------------------------------
## Correlation terms:
##     Estimate Std. Error t value Pr(>|t|)    
## rho  -0.6306     0.1024  -6.159  8.2e-10 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------

On the other hand, as the observed data show asymmetry, the Heckman-BS model can be a good indication to adjust such data. Note that the Heckman-BS model also indicates the presence of a significant correlation. It also presents, as previously mentioned, the advantage of parsimony and the direct interpretation of the relationship of the parameters with the response variable of interest.

summary(mBS)
## 
## --------------------------------------------------------------
##    Birnbaum-Saunders Heckman Model (Package: ssmodels)        
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 26 
## Log-Likelihood: -24470.39 
## AIC: 48974.79 BIC: 49078.66 
## Number of observations: ( 526 censored and 2802 observed ) 
## 17 free parameters ( df= 3311 ) 
## --------------------------------------------------------------
## Probit selection equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  0.117214   0.235265   0.498  0.61836    
## age          0.089502   0.031571   2.835  0.00461 ** 
## female       0.721876   0.069037  10.456  < 2e-16 ***
## educ         0.068319   0.014689   4.651 3.43e-06 ***
## blhisp      -0.399014   0.074350  -5.367 8.57e-08 ***
## totchr       0.805651   0.068209  11.812  < 2e-16 ***
## ins          0.179379   0.073102   2.454  0.01419 *  
## income       0.003425   0.001430   2.395  0.01668 *  
## --------------------------------------------------------------
## Outcome equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  5.778691   0.172658  33.469  < 2e-16 ***
## age          0.246156   0.022143  11.117  < 2e-16 ***
## female       0.410538   0.048236   8.511  < 2e-16 ***
## educ        -0.006632   0.009890  -0.671    0.503    
## blhisp      -0.215212   0.053172  -4.047  5.3e-05 ***
## totchr       0.588751   0.034299  17.165  < 2e-16 ***
## ins         -0.061326   0.049610  -1.236    0.216    
## --------------------------------------------------------------
## Error terms:
##       Estimate Std. Error t value Pr(>|t|)    
## sigma  0.70281    0.02072  33.917  < 2e-16 ***
## rho    0.27305    0.05737   4.759 2.02e-06 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------

The Heckman-t model, even adjusted to the expense log, indicates the presence of deviation from normality, since the degree of freedom parameter is approximately equal to \(13\) and is significant. Such a model also observes the significant correlation parameter.

summary(mtS)
## 
## --------------------------------------------------------------
##       t-Student Heckman Model (Package: ssmodels)             
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 25 
## Log-Likelihood: -5822.076 
## AIC: 11680.15 BIC: 11790.13 
## Number of observations: ( 526 censored and 2802 observed ) 
## 18 free parameters ( df= 3310 ) 
## --------------------------------------------------------------
## Probit selection equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -0.747972   0.207700  -3.601 0.000321 ***
## age          0.098549   0.029746   3.313 0.000933 ***
## female       0.724862   0.068559  10.573  < 2e-16 ***
## educ         0.064839   0.012805   5.063 4.34e-07 ***
## blhisp      -0.393569   0.066529  -5.916 3.64e-09 ***
## totchr       0.890081   0.087252  10.201  < 2e-16 ***
## ins          0.180029   0.068010   2.647 0.008157 ** 
## income       0.002978   0.001447   2.058 0.039636 *  
## --------------------------------------------------------------
## Outcome equation:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  5.20584    0.20961  24.835  < 2e-16 ***
## age          0.20683    0.02260   9.152  < 2e-16 ***
## female       0.30653    0.05640   5.435 5.87e-08 ***
## educ         0.01731    0.01026   1.688 0.091591 .  
## blhisp      -0.19297    0.05775  -3.342 0.000842 ***
## totchr       0.51271    0.03583  14.310  < 2e-16 ***
## ins         -0.05250    0.05048  -1.040 0.298439    
## --------------------------------------------------------------
## Error terms:
##       Estimate Std. Error t value Pr(>|t|)    
## sigma  1.19485    0.02365  50.524  < 2e-16 ***
## rho   -0.32199    0.12210  -2.637   0.0084 ** 
## df    12.93032    2.85903   4.523 6.32e-06 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------

Finally, the Heckman-Skew model also indicates the presence of a significant correlation between the variables of interest and selection. In addition, it also indicates deviation from the normality of the transformed data. It is important to note that without the data transformation, the iterative method of estimating the parameters of the Heckman-Skew model does not converge.

summary(mSK)
## 
## --------------------------------------------------------------
##       Skew Normal Heckman Model (Package: ssmodels)           
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 79 
## Log-Likelihood: -5805.744 
## AIC: 11647.49 BIC: 11757.47 
## Number of observations: ( 526 censored and 2802 observed ) 
## 18 free parameters ( df= 3310 ) 
## --------------------------------------------------------------
## Probit selection equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -0.899099   0.205570  -4.374 1.26e-05 ***
## age          0.086608   0.026563   3.260  0.00112 ** 
## female       0.644807   0.061736  10.445  < 2e-16 ***
## educ         0.060281   0.011774   5.120 3.23e-07 ***
## blhisp      -0.350940   0.060960  -5.757 9.35e-09 ***
## totchr       0.774834   0.072534  10.682  < 2e-16 ***
## ins          0.169602   0.061055   2.778  0.00550 ** 
## income       0.002657   0.001278   2.079  0.03770 *  
## --------------------------------------------------------------
## Outcome equation:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  6.45165    0.22388  28.818  < 2e-16 ***
## age          0.20218    0.02211   9.146  < 2e-16 ***
## female       0.28739    0.05507   5.218 1.92e-07 ***
## educ         0.01334    0.01007   1.325 0.185146    
## blhisp      -0.20148    0.05637  -3.575 0.000356 ***
## totchr       0.49158    0.03589  13.696  < 2e-16 ***
## ins         -0.08050    0.04979  -1.617 0.106042    
## --------------------------------------------------------------
## Error terms:
##        Estimate Std. Error t value Pr(>|t|)    
## sigma    1.7359     0.0470  36.938   <2e-16 ***
## rho     -0.3390     0.1411  -2.402   0.0164 *  
## lambda  -1.6620     0.1434 -11.587   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------

2.2 Nhanes 2003-2004

The US National Health and Nutrition Examination Study (NHANES) is a survey data collected by the US National Center for Health Statistics. The survey data dates back to 1999, where individuals of all ages are interviewed in their home annually and complete the health examination component of the survey. The study variables include demographic variables (e.g. age and annual household income), physical measurements (e.g. BMI – body mass index), health variables (e.g. diabetes status), and lifestyle variables (e.g. smoking status). This data frame contains the following columns:

library(ssmodels)
data(nhanes)
attach(nhanes)
perc <- function(x,data){
    nna <- ifelse(sum(is.na(x))!=0,summary(x)[[7]],0)
    perc <- ifelse(sum(is.na(x))!=0,(nna/length(data$id))*100,0)
   return(perc)
}
Variables <- c("SBP (mm Hg)", "Age (year)", "Gender", "BMI (Kg/$m^{2}$)", "Education (years)", "Race", "Income ($\\$1000$ per year)", "Numbers Obs.")
perc1 <- round(perc(sbp,nhanes), digits = 2)
perc2 <- round(perc(age,nhanes), digits = 2)
perc3 <- round(perc(gender,nhanes), digits = 2)
perc4 <- round(perc(bmi,nhanes), digits = 2)
perc5 <- round(perc(educ,nhanes), digits = 2)
perc6 <- round(perc(race,nhanes), digits = 2)
perc7 <- round(perc(Income,nhanes), digits = 2)
nObs <- length(Income)
Percentage <- c(perc1, perc2, perc3, perc4, perc5, perc6, perc7, nObs)
df <- subset(nhanes, !is.na(sbp))
df <- subset(df, !is.na(bmi))
attach(df)
perc11 <- round(perc(sbp,df), digits = 2)
perc12 <- round(perc(age,df), digits = 2)
perc13 <- round(perc(gender,df), digits = 2)
perc14 <- round(perc(bmi,df), digits = 2)
perc15 <- round(perc(educ,df), digits = 2)
perc16 <- round(perc(race,df), digits = 2)
perc17 <- round(perc(Income,df), digits = 2)
nObs1 <- length(Income)
Percentage1 <- c(perc11, perc12, perc13, perc14, perc15, perc16, perc17, nObs1)
table <- data.frame("Variables" = Variables, "Percentage of Missing"= Percentage, "Without Missing"= Percentage1)
kable(table, format = "html", align = c("c", "c", "c"))
Variables Percentage.of.Missing Without.Missing
SBP (mm Hg) 34.94 0.00
Age (year) 0.00 0.00
Gender 0.00 0.00
BMI (Kg/\(m^{2}\)) 9.91 0.00
Education (years) 17.22 0.00
Race 0.00 0.00
Income (\(\$1000\) per year) 24.41 25.43
Numbers Obs. 9643.00 6193.00
library("ggplot2")
library("gridExtra")
barfill <- "grey"
barlines <- "black"
p1 <- ggplot(df, aes(Income)) + geom_histogram( breaks = seq(0, 10, 0.5), aes(y = ..density..), colour = barlines, fill = barfill)+
  scale_x_continuous(name = "Income",
                   breaks = seq(0, 10, 2),
                   limits=c(0, 10))

p2 <- ggplot(df, aes(log(sbp))) + geom_histogram( colour = barlines, fill = barfill) +
  scale_x_continuous(name = "Log Systolic blood pressure")
grid.arrange(p1, p2, ncol=2)

df$YS <- ifelse(is.na(df$Income),0,1)
df$educ <- ifelse(df$educ<=2,0,1)
df$Income <- ifelse(is.na(df$Income),0,df$Income)
attach(df)
selectionEq <- YS~age+gender+educ+race
outcomeEq   <- log(sbp)~age+gender+educ+bmi+Income
outcomeBS   <- sbp~age+gender+educ+bmi+Income
mCL <- HeckmanCL(selectionEq, outcomeEq, data = df)
mBS <- HeckmanBS(selectionEq, outcomeBS, data = df)
mSK <- HeckmanSK(selectionEq, outcomeEq, data = df, lambda = 0)
mtS <- HeckmantS(selectionEq, outcomeEq, data = df, df = 15)

    Parameters <- c("Intercept", "age", "gender", "educ", "race", "Intercept", "age", "gender", "educ", "bmi", "income", "sigma", "rho", "nu", "lambda")
HBS <- round(mBS$coefficients, digits = 5)
HCL <- round(mCL$coefficients, digits = 5)
HSK <- round(mSK$coefficients, digits = 5)
HtS <- round(mtS$coefficients, digits = 5)
Results <- data.frame("Parameters"= Parameters, 
                      "HeckmanCL" = c(HCL[1:13], "NA", "NA"),
                      "HeckmanBS" = c(HBS[1:13], "NA", "NA"), 
                      "HeckmantS" = c(HtS[1:13], HtS[14], "NA"),
                      "HeckmanSK" = c(HSK[1:13], "NA", HSK[14]))
kable(Results, format = "html", align = c("c", "c", "c", "c", "c"))
Parameters HeckmanCL HeckmanBS HeckmantS HeckmanSK
Intercept 0.44001 1.35518 0.50396 0.83765
age 0.01113 0.01349 0.01079 0.01066
gender 0.13624 0.15098 0.1373 0.12074
educ -0.53594 -0.67172 -0.55715 -0.42047
race -0.07011 -0.09013 -0.07708 -0.04967
Intercept 4.51594 4.52773 4.52781 4.63651
age 0.00448 0.00448 0.00435 0.00481
gender -0.0229 -0.02355 -0.02622 -0.02081
educ -0.04788 -0.0474 -0.04532 -0.05107
bmi 0.00362 0.00361 0.00367 0.00365
income 0.00044 4e-04 0.00049 0.00034
sigma 0.14354 96.68214 0.13023 0.21735
rho 0.77446 0.77208 0.7398 0.90901
nu NA NA 14.40706 NA
lambda NA NA NA -1.62376
summary(mCL)
## 
## --------------------------------------------------------------
##           Classic Heckman Model (Package: ssmodels)           
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 27 
## Log-Likelihood: -216.546 
## AIC: 459.0919 BIC: 546.5972 
## Number of observations: ( 1575 censored and 4618 observed ) 
## 13 free parameters ( df= 6180 ) 
## --------------------------------------------------------------
## Probit selection equation:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  0.4400117  0.0751789   5.853 5.08e-09 ***
## age          0.0111345  0.0008816  12.630  < 2e-16 ***
## gender       0.1362360  0.0344494   3.955 7.75e-05 ***
## educ        -0.5359397  0.0384988 -13.921  < 2e-16 ***
## race        -0.0701069  0.0134076  -5.229 1.76e-07 ***
## --------------------------------------------------------------
## Outcome equation:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  4.516e+00  1.080e-02 418.113  < 2e-16 ***
## age          4.477e-03  9.126e-05  49.064  < 2e-16 ***
## gender      -2.290e-02  4.027e-03  -5.687 1.35e-08 ***
## educ        -4.788e-02  4.864e-03  -9.844  < 2e-16 ***
## bmi          3.623e-03  2.905e-04  12.474  < 2e-16 ***
## Income       4.369e-04  7.609e-04   0.574    0.566    
## --------------------------------------------------------------
## Error terms:
##       Estimate Std. Error t value Pr(>|t|)    
## sigma 0.143541   0.002325   61.74   <2e-16 ***
## rho   0.774464   0.022220   34.85   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------
summary(mtS)
## 
## --------------------------------------------------------------
##       t-Student Heckman Model (Package: ssmodels)             
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 100 
## Log-Likelihood: -172.8373 
## AIC: 373.6746 BIC: 467.911 
## Number of observations: ( 1575 censored and 4618 observed ) 
## 14 free parameters ( df= 6179 ) 
## --------------------------------------------------------------
## Probit selection equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  0.503962   0.079702   6.323 2.74e-10 ***
## age          0.010794   0.001004  10.747  < 2e-16 ***
## gender       0.137302   0.036467   3.765 0.000168 ***
## educ        -0.557147   0.040707 -13.687  < 2e-16 ***
## race        -0.077080   0.014286  -5.395 7.09e-08 ***
## --------------------------------------------------------------
## Outcome equation:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  4.528e+00  1.044e-02 433.844  < 2e-16 ***
## age          4.348e-03  8.977e-05  48.436  < 2e-16 ***
## gender      -2.622e-02  3.844e-03  -6.822  9.8e-12 ***
## educ        -4.532e-02  4.959e-03  -9.139  < 2e-16 ***
## bmi          3.672e-03  2.854e-04  12.864  < 2e-16 ***
## Income       4.896e-04  7.432e-04   0.659     0.51    
## --------------------------------------------------------------
## Error terms:
##        Estimate Std. Error t value Pr(>|t|)    
## sigma  0.130226   0.001353   96.26   <2e-16 ***
## rho    0.739803   0.050805   14.56   <2e-16 ***
## df    14.407064   1.097547   13.13   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------
summary(mBS)
## 
## --------------------------------------------------------------
##    Birnbaum-Saunders Heckman Model (Package: ssmodels)        
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 76 
## Log-Likelihood: -22281.19 
## AIC: 44588.38 BIC: 44675.89 
## Number of observations: ( 1575 censored and 4618 observed ) 
## 13 free parameters ( df= 6180 ) 
## --------------------------------------------------------------
## Probit selection equation:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.355185   0.094721  14.307  < 2e-16 ***
## age          0.013492   0.001039  12.984  < 2e-16 ***
## gender       0.150982   0.043582   3.464 0.000535 ***
## educ        -0.671716   0.049316 -13.621  < 2e-16 ***
## race        -0.090127   0.016965  -5.313 1.12e-07 ***
## --------------------------------------------------------------
## Outcome equation:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  4.528e+00  1.072e-02 422.194  < 2e-16 ***
## age          4.478e-03  9.105e-05  49.182  < 2e-16 ***
## gender      -2.355e-02  4.029e-03  -5.844 5.36e-09 ***
## educ        -4.740e-02  4.851e-03  -9.772  < 2e-16 ***
## bmi          3.610e-03  2.907e-04  12.418  < 2e-16 ***
## Income       3.974e-04  7.621e-04   0.522    0.602    
## --------------------------------------------------------------
## Error terms:
##       Estimate Std. Error t value Pr(>|t|)    
## sigma 96.68214    3.15376   30.66   <2e-16 ***
## rho    0.77208    0.02258   34.20   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------
summary(mSK)
## 
## --------------------------------------------------------------
##       Skew Normal Heckman Model (Package: ssmodels)           
## --------------------------------------------------------------
## --------------------------------------------------------------
## Maximum Likelihood estimation 
## optim function with method BFGS-iterations numbers: 45 
## Log-Likelihood: -202.82 
## AIC: 433.6401 BIC: 527.8765 
## Number of observations: ( 1575 censored and 4618 observed ) 
## 14 free parameters ( df= 6179 ) 
## --------------------------------------------------------------
## Probit selection equation:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  0.8376493  0.0577195  14.512  < 2e-16 ***
## age          0.0106572  0.0006882  15.486  < 2e-16 ***
## gender       0.1207434  0.0271668   4.445 8.96e-06 ***
## educ        -0.4204717  0.0331856 -12.670  < 2e-16 ***
## race        -0.0496700  0.0105806  -4.694 2.73e-06 ***
## --------------------------------------------------------------
## Outcome equation:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  4.6365130  0.0122797 377.575  < 2e-16 ***
## age          0.0048084  0.0001139  42.199  < 2e-16 ***
## gender      -0.0208051  0.0041517  -5.011 5.56e-07 ***
## educ        -0.0510738  0.0048687 -10.490  < 2e-16 ***
## bmi          0.0036457  0.0002900  12.571  < 2e-16 ***
## Income       0.0003417  0.0007569   0.451    0.652    
## --------------------------------------------------------------
## Error terms:
##        Estimate Std. Error t value Pr(>|t|)    
## sigma   0.21735    0.01093  19.885   <2e-16 ***
## rho     0.90901    0.01496  60.747   <2e-16 ***
## lambda -1.62376    0.18321  -8.863   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## --------------------------------------------------------------

References

Bastos, Fernando de Souza, and Wagner Barreto-Souza. 2020. “Birnbaum–Saunders Sample Selection Model.” Journal of Applied Statistics.
Bastos, Fernando de Souza, Wagner Barreto-Souza, and Marc G Genton. 2022. “A Generalized Heckman Model with Varying Sample Selection Bias and Dispersion Parameters.” Statistica Sinica.
Colin, Cameron A, and Pravin K Trivedi. 2009. “Microeconometrics Using STATA.” Lakeway Drive, TX: Stata Press Books.
Heckman, James J. 1976. “The Common Structure of Statistical Models of Truncation, Sample Selection and Limited Dependent Variables and a Simple Estimator for Such Models.” In Annals of Economic and Social Measurement, Volume 5, Number 4, 475–92. NBER.
———. 1979. “Sample Selection Bias as a Specification Error.” Econometrica: Journal of the Econometric Society, 153–61.
Marchenko, Yulia V, and Marc G Genton. 2012. “A Heckman Selection-t Model.” Journal of the American Statistical Association 107 (497): 304–17.
Ogundimu, Emmanuel O, and Jane L Hutton. 2016. “A Sample Selection Model with Skew-Normal Distribution.” Scandinavian Journal of Statistics 43 (1): 172–90.
Zhelonkin, Mikhail, Marc G Genton, and Elvezio Ronchetti. 2016. “Robust Inference in Sample Selection Models.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 78 (4): 805–27.