-
Notifications
You must be signed in to change notification settings - Fork 0
/
README.Rmd
199 lines (155 loc) · 6.99 KB
/
README.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
---
output:
github_document:
html_preview: true
---
<!-- README.md is generated from README.Rmd. Please edit that file -->
```{r, include = FALSE}
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>",
fig.path = "man/figures/README-",
out.width = "100%"
)
```
# semiBRM: Semiparametric Binary Response Models in R
This **R** package **semiBRM** offers an implementation of single-index semiparametric binary response
models, theorized in a seminal work in the semiparametric econometrics literature
[Klein and Spady (1993)](https://doi.org/10.2307/2951556).
This is built upon [Rcpp](http://www.rcpp.org) along with OpenMP, which parallelizes computation of nonparametric
conditional expectations over given data points. This improves computation efficiency enormously in
parameter estimation, which is one of the main features of this package.
The package is still in development, with the goal of making it as easy to employ [Klein and Spady (1993)](https://doi.org/10.2307/2951556) as to run Probit or Logistic regressions in **R**.
## Identifiable Set of Parameters
In a single-index semiparametric approach, not all parameters are identifiable. First of all,
intercept cannot be estimated. Secondly, the coefficients can be estimated as ratios to a "basis"
coefficient. This package sets the coefficient of the first explanatory variable as the basis so
that its coefficient is normalized to 1, and those of the rest of the variables are estimated
conformably to it.
In the context of binary response models, or nonlinear models more broadly, this is not problematic
at all, as the coefficients themselves have little interpretation. As the matter of fact,
the identifiable set of parameters in this approach can correctly estimate the conditional
probability of Pr(Y=1|X), which would be the quantity of interest in many cases.
## Installation
The development version can be installed from GitHub:
``` r
library(devtools)
install_github(repo="hk599/semiBRM")
```
### This version was built with:
* R version 4.0.5 (2021-03-31)
* maxLik 1.4-8
* Rcpp 1.0.6
* RcppArmadillo 0.10.4.0.0
## Example
### Setup
```{r setup, message=FALSE, warning=FALSE}
library(semiBRM)
set.seed(20190815) # for reproduction of results
```
### Data Generating Process
```{r dgp, message=FALSE, warning=FALSE}
## data generating process
N <- 1500L
X1 <- rnorm(N)
X2 <- (X1 + 2*rnorm(N))/sqrt(5) + 1
X3 <- rnorm(N)^2/sqrt(2)
X <- cbind(X1, X2, X3)
beta <- c(2, 2, -1, -1) # this is the original set of coefficients
V <- as.vector(cbind(X, 1)%*%beta)
Y <- ifelse(V >= rnorm(N), 1L, 0L)
```
### Parameter Estimation
```{r estimation, message=FALSE, warning=FALSE}
## estimands: the rescaled coefficients by the first coefficient excluding intercept
coefs_true <- c(1, -.5)
data <- data.frame(Y, X1, X2, X3)
## Klein and Spady (1993): semiparametric approach
semi <- semiBRM(Y ~ X1 + X2 + X3, data = data, control = list(iterlim=50))
coefs_semi <- coef(semi)
## Probit: parametric approach
probit <- glm(Y ~ X1 + X2 + X3, family = binomial(link = "probit"), data = data)
coefs_probit <- probit$coefficients[-1L][-1L]/probit$coefficients[-1L][1L]
## formatted print
{
cat(sprintf(" %7s %7s %7s\n", "True", "Probit", "Semi"))
cat(sprintf("parm 1: %7.4f %7.4f %7.4f\n", coefs_true[1L], coefs_probit[1L], coefs_semi[1L]))
cat(sprintf("parm 2: %7.4f %7.4f %7.4f\n", coefs_true[2L], coefs_probit[2L], coefs_semi[2L]))
}
```
### In-Sample Prediction
```{r in-pred, message=FALSE, warning=FALSE}
## in-sample conditional probability
in_prob_true <- pnorm(V)
in_prob_semi <- predict(semi)
in_prob_probit <- fitted(probit)
## formatted print
target <- sample.int(N, size = 10L)
{
cat(sprintf("%7s %8s %8s %8s %12s\n", "Obs.", "True", "Probit", "Semi", "non.endpoint") )
for (i in target){
cat(sprintf("[%04d]: %8.6f %8.6f %8.6f %12s\n",
i, in_prob_true[i], in_prob_probit[i], in_prob_semi$prob[i], in_prob_semi$non.endpoint[i]))
}
}
```
### Out-of-Sample Prediction
```{r out-pred, message=FALSE, warning=FALSE}
## conditional probability at the means
Xbar <- colMeans(X)
newdata <- as.data.frame(as.list(Xbar))
## predictions
out_prob_true <- pnorm(as.vector(c(Xbar, 1)%*%beta))
out_prob_semi <- predict(semi, newdata, boot.se = TRUE)
out_prob_probit <- pnorm(as.vector(coef(probit)%*%c(1, Xbar)))
## standard errors of Probit
grad <- dnorm(as.vector(coef(probit)%*%c(1, Xbar))) * c(1, Xbar)
out_stde_probit <- sqrt(as.vector(crossprod(grad, vcov(probit))%*%grad))
## formatted print
{
cat(sprintf(" %7s %7s %7s\n", "True", "Probit", "Semi"))
cat(sprintf("Prob. Est.: %7.4f %7.4f %7.4f\n", out_prob_true, out_prob_probit, out_prob_semi$prob))
cat(sprintf(" Std. Err.: %7s %7.4f %7.4f\n", "", out_stde_probit, out_prob_semi$boot.se))
}
```
### Average Marginal Effects
```{r avme, message=FALSE, warning=FALSE}
## marginal effects as difference between conditional probabilities with and without perturbation
delta <- sd(X1) # size of perturbation
me_true <- pnorm(as.vector(cbind(X1+delta, X2, X3, 1)%*%beta)) - pnorm(as.vector(cbind(X, 1)%*%beta))
## average marginal effects
av_me_true <- mean(me_true)
av_me_semi <- MarginalEffects(semi, variable = "X1", delta = delta)
## formatted print
{
cat(sprintf(" %7s %7s \n", "True", "Semi"))
cat(sprintf("Marg.Eff.: %7.4f %7.4f\n", av_me_true, av_me_semi[1L]))
cat(sprintf("Std. Err.: %7s %7.4f\n", "", av_me_semi[2L]))
}
```
### Quantile Marginal Effects
```{r qvme, message=FALSE, warning=FALSE}
## percentile cutoffs
p.cutoffs <- c(1/4, 2/4, 3/4)
## group indicators
q_vals <- quantile(X1, p.cutoffs)
q1 <- ifelse(X1 <= q_vals[1L], 1L, 0L)
q2 <- ifelse(q_vals[1L] < X1 & X1 <= q_vals[2L], 1L, 0L)
q3 <- ifelse(q_vals[2L] < X1 & X1 <= q_vals[3L], 1L, 0L)
q4 <- ifelse(q_vals[3L] < X1, 1L, 0L)
## quantile marginal effects
q_me_true <- c("G1" = mean(me_true[q1==1]), "G2" = mean(me_true[q2==1]),
"G3" = mean(me_true[q3==1]), "G4" = mean(me_true[q4==1]))
q_me_semi <- MarginalEffects(semi, variable = "X1", delta, p.cutoffs)
## formatted print
{
cat(" ", sprintf(" %5s ", paste0("G", 1:4)), fill = TRUE)
cat("True :", sprintf(" %.4f ", q_me_true), fill = TRUE)
cat("Semi :", sprintf(" %.4f ", q_me_semi[,1L]), fill = TRUE)
cat("Std.Err.:", sprintf("(%.4f)", q_me_semi[,2L]), fill = TRUE)
}
```
## References
Eddelbuettel, D., & François, R. (2011). Rcpp: Seamless R and C++ integration. _Journal of Statistical Software_, _40(8)_, 1-18. _[https://dirk.eddelbuettel.com/code/rcpp/Rcpp-introduction.pdf](https://dirk.eddelbuettel.com/code/rcpp/Rcpp-introduction.pdf)_
Klein, R. W., & Spady, R. H. (1993). An Efficient Semiparametric Estimator for Binary Response Models. _Econometrica_, _61(2)_, 387-421. _[https://doi.org/10.2307/2951556](https://doi.org/10.2307/2951556)_.
Klein, R., & Vella, F. (2009). A Semiparametric Model for Binary Response and Continuous Outcomes Under Index Heteroscedasticity. _Journal of Applied Econometrics_, _24(5)_, 735-762. _[https://doi.org/10.1002/jae.1064](https://doi.org/10.1002/jae.1064)_