空间计量经济模型中p.值的计算:summary()和texreg()之间为什么不一致?
我估计了一些空间计量经济模型,其中既包含空间自回归项rho,也包含空间误差项lambda。在尝试传达我的结果时,我使用了空间计量经济模型中p.值的计算:summary()和texreg()之间为什么不一致?,r,spatial,p-value,texreg,R,Spatial,P Value,Texreg,我估计了一些空间计量经济模型,其中既包含空间自回归项rho,也包含空间误差项lambda。在尝试传达我的结果时,我使用了texreg包,它接受我正在使用的sacsarlm模型。但是,我注意到,texreg正在为相同的rho和lambda参数打印p值Texreg似乎正在返回model@LR1$p.value模型对象的插槽 参数rho和lambda的大小不同,标准误差也不同,因此它们不应具有等效的p值。如果我对模型对象调用summary,我会得到唯一的p值,但尽管在str(model)调用中遍历了每
texreg
包,它接受我正在使用的sacsarlm
模型。但是,我注意到,texreg
正在为相同的rho和lambda参数打印p值Texreg
似乎正在返回model@LR1$p.value
模型对象的插槽
参数rho和lambda的大小不同,标准误差也不同,因此它们不应具有等效的p值。如果我对模型对象调用summary,我会得到唯一的p值,但尽管在str(model)
调用中遍历了每个元素,但无法确定这些值存储在模型对象中的什么位置
我的问题有两个:
library(spdep)
library(texreg)
set.seed(42)
W.ran <- matrix(rbinom(100*100, 1, .3),nrow=100)
X <- rnorm(100)
Y <- .2 * X + rnorm(100) + .9*(W.ran %*% X)
W.test <- mat2listw(W.ran)
model <- sacsarlm(Y~X, type = "sacmixed",
listw=W.test, zero.policy=TRUE)
summary(model)
Call:sacsarlm(formula = Y ~ X, listw = W.test, type = "sacmixed",
zero.policy = TRUE)
Residuals:
Min 1Q Median 3Q Max
-2.379283 -0.750922 0.036044 0.675951 2.577148
Type: sacmixed
Coefficients: (asymptotic standard errors)
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.91037455 0.65700059 1.3857 0.1659
X -0.00076362 0.10330510 -0.0074 0.9941
lag.(Intercept) -0.03193863 0.02310075 -1.3826 0.1668
lag.X 0.89764491 0.02231353 40.2287 <2e-16
Rho: -0.0028541
Asymptotic standard error: 0.0059647
z-value: -0.47849, p-value: 0.6323
Lambda: -0.020578
Asymptotic standard error: 0.020057
z-value: -1.026, p-value: 0.3049
LR test value: 288.74, p-value: < 2.22e-16
Log likelihood: -145.4423 for sacmixed model
ML residual variance (sigma squared): 1.0851, (sigma: 1.0417)
Number of observations: 100
Number of parameters estimated: 7
AIC: 304.88, (AIC for lm: 585.63)
screenreg(model)
=================================
Model 1
---------------------------------
(Intercept) 0.91
(0.66)
X -0.00
(0.10)
lag.(Intercept) -0.03
(0.02)
lag.X 0.90 ***
(0.02)
---------------------------------
Num. obs. 100
Parameters 7
AIC (Linear model) 585.63
AIC (Spatial model) 304.88
Log Likelihood -145.44
Wald test: statistic 1.05
Wald test: p-value 0.90
Lambda: statistic -0.02
Lambda: p-value 0.00
Rho: statistic -0.00
Rho: p-value 0.00
=================================
*** p < 0.001, ** p < 0.01, * p < 0.05
库(spdep)
图书馆(texreg)
种子(42)
W.rantexreg
作者在这里。谢谢你抓到这个。如我在回复中所述,texreg
使用extract
方法从任何(当前支持70多种)模型对象类型中检索相关信息。对于sarlm
对象,方法的GOF部分似乎存在故障
以下是该方法当前的外观(从texreg
1.36.13开始):
您可以看到,该函数首先区分不同的子模型(error
vs.sac
/sacmixed
vs.else),然后决定使用哪些标准错误,然后动态计算p值,而不将其保存在任何位置
因此,这也是我们在extract
方法中需要做的,以便获得与spdep
包中的summary
方法相同的结果。我们还需要将其从GOF块移到表的系数块(有关讨论,请参阅下面的注释部分)。以下是我在提取方法中采用他们方法的尝试:
# extension for sarlm objects (spdep package)
extract.sarlm <- function(model, include.nobs = TRUE, include.loglik = TRUE,
include.aic = TRUE, include.lr = TRUE, include.wald = TRUE, ...) {
s <- summary(model, ...)
names <- rownames(s$Coef)
cf <- s$Coef[, 1]
se <- s$Coef[, 2]
p <- s$Coef[, ncol(s$Coef)]
if (model$type != "error") { # include coefficient for autocorrelation term
rho <- model$rho
cf <- c(cf, rho)
names <- c(names, "$\\rho$")
if (!is.null(model$rho.se)) {
if (!is.null(model$adj.se)) {
rho.se <- sqrt((model$rho.se^2) * model$adj.se)
} else {
rho.se <- model$rho.se
}
rho.pval <- 2 * (1 - pnorm(abs(rho / rho.se)))
se <- c(se, rho.se)
p <- c(p, rho.pval)
} else {
se <- c(se, NA)
p <- c(p, NA)
}
}
if (!is.null(model$lambda)) {
cf <-c(cf, model$lambda)
names <- c(names, "$\\lambda$")
if (!is.null(model$lambda.se)) {
if (!is.null(model$adj.se)) {
lambda.se <- sqrt((model$lambda.se^2) * model$adj.se)
} else {
lambda.se <- model$lambda.se
}
lambda.pval <- 2 * (1 - pnorm(abs(model$lambda / lambda.se)))
se <- c(se, lambda.se)
p <- c(p, lambda.pval)
} else {
se <- c(se, NA)
p <- c(p, NA)
}
}
gof <- numeric()
gof.names <- character()
gof.decimal <- logical()
if (include.nobs == TRUE) {
n <- length(s$fitted.values)
param <- s$parameters
gof <- c(gof, n, param)
gof.names <- c(gof.names, "Num.\ obs.", "Parameters")
gof.decimal <- c(gof.decimal, FALSE, FALSE)
}
if (include.loglik == TRUE) {
ll <- s$LL
gof <- c(gof, ll)
gof.names <- c(gof.names, "Log Likelihood")
gof.decimal <- c(gof.decimal, TRUE)
}
if (include.aic == TRUE) {
aic <- AIC(model)
aiclm <- s$AIC_lm.model
gof <- c(gof, aiclm, aic)
gof.names <- c(gof.names, "AIC (Linear model)", "AIC (Spatial model)")
gof.decimal <- c(gof.decimal, TRUE, TRUE)
}
if (include.lr == TRUE && !is.null(s$LR1)) {
gof <- c(gof, s$LR1$statistic[[1]], s$LR1$p.value[[1]])
gof.names <- c(gof.names, "LR test: statistic", "LR test: p-value")
gof.decimal <- c(gof.decimal, TRUE, TRUE)
}
if (include.wald == TRUE && !is.null(model$Wald1)) {
waldstat <- model$Wald1$statistic
waldp <- model$Wald1$p.value
gof <- c(gof, waldstat, waldp)
gof.names <- c(gof.names, "Wald test: statistic", "Wald test: p-value")
gof.decimal <- c(gof.decimal, TRUE, TRUE)
}
tr <- createTexreg(
coef.names = names,
coef = cf,
se = se,
pvalues = p,
gof.names = gof.names,
gof = gof,
gof.decimal = gof.decimal
)
return(tr)
}
setMethod("extract", signature = className("sarlm", "spdep"),
definition = extract.sarlm)
是的,有时R有点模糊-当您调用summary时,您感兴趣的值可能不是从mode对象中提取的,而是由summary函数在此时计算的。这不是一个完整的答案,但您应该查看broom
包,它可以在易于使用的data.frame(通常包括p值等)中提供重要的模型信息。为什么不手动替换texreg
输出中的值呢?它吐出乳胶或html毕竟…我会看看扫帚包,谢谢@jakub。我正在一个Swave文档中生成texreg的latex输出,该文档正在编织到最终产品中,因此手动编辑latex是不现实的。根据位于的完整列表,broom
还不支持这些对象。我喜欢broom
的概念,但我认为texreg
的泛型extract
函数的概念是一样的,主要区别在于提取的输出被交给几个函数来创建回归表(在texreg
中),而不是数据帧(在broom
中)。这是一个非常彻底的答案,谢谢!事实上,我已经为我正在估算的sacsarlm模型编写了(一个不太完整的)提取函数(我想要一些特殊的更改,以及将rho和lambda的位置移到系数上),并以通常的方式计算p值。您的想法是在spdep源代码中查找摘要的确切内容(而不是通过猜测反向工程)更合理。此外,我认为使用提取函数扩展texreg是一个好主意,我非常感谢您以这种方式编写它!很高兴它起到了作用。您认为GOF块中项目的理想顺序是什么?实际上,我将Rho和Lambda移到了顶部的系数列中。Rho是主要的coef我对我的研究很感兴趣,因为我正在用它来捕捉网络分析中的同侪效应。使用这些模型并不少见,如果是这种情况,那么你当然希望空间自相关系数显示为系数。这实际上就是extract.lnam()的方法如果您使用sna软件包中这种空间模型的lnam版本(速度要慢得多,这就是我使用spdep的原因),是否会发生这种情况.我不是说提取函数必须改变才能这样工作。我怀疑一些研究人员在对其他关系建模感兴趣时,正在拟合空间模型来控制空间内生性,而对于他们来说,将其作为GOF可能更有意义。然而,似乎使rho和lambda系数might更灵活:如果你认为它们是GOF,如果它们显示为系数,那没什么大不了的,但是如果你对系数的大小感兴趣,如果它们显示为GOF度量,那就很烦人了。这只是我的意见。这很有意义。我已经将rho和lambda移到系数块并更新了答案。它有趣的是,您正在使用spdep
进行网络分析。您可能还想看看我的tnam
软件包,用于R
(实现时态网络自相关模型)。这是您使用的模型的扩展,用于任意类型的网络相关性以及面板数据(如果需要).除了空间
print.summary.sarlm <- function(x, digits = max(5, .Options$digits - 3),
signif.stars = FALSE, ...)
{
cat("\nCall:", deparse(x$call), sep = "", fill=TRUE)
if (x$type == "error") if (isTRUE(all.equal(x$lambda, x$interval[1])) ||
isTRUE(all.equal(x$lambda, x$interval[2])))
warning("lambda on interval bound - results should not be used")
if (x$type == "lag" || x$type == "mixed")
if (isTRUE(all.equal(x$rho, x$interval[1])) ||
isTRUE(all.equal(x$rho, x$interval[2])))
warning("rho on interval bound - results should not be used")
cat("\nResiduals:\n")
resid <- residuals(x)
nam <- c("Min", "1Q", "Median", "3Q", "Max")
rq <- if (length(dim(resid)) == 2L)
structure(apply(t(resid), 1, quantile), dimnames = list(nam,
dimnames(resid)[[2]]))
else structure(quantile(resid), names = nam)
print(rq, digits = digits, ...)
cat("\nType:", x$type, "\n")
if (x$zero.policy) {
zero.regs <- attr(x, "zero.regs")
if (!is.null(zero.regs))
cat("Regions with no neighbours included:\n",
zero.regs, "\n")
}
if (!is.null(x$coeftitle)) {
cat("Coefficients:", x$coeftitle, "\n")
coefs <- x$Coef
if (!is.null(aliased <- x$aliased) && any(x$aliased)){
cat(" (", table(aliased)["TRUE"],
" not defined because of singularities)\n", sep = "")
cn <- names(aliased)
coefs <- matrix(NA, length(aliased), 4, dimnames = list(cn,
colnames(x$Coef)))
coefs[!aliased, ] <- x$Coef
}
printCoefmat(coefs, signif.stars=signif.stars, digits=digits,
na.print="NA")
}
# res <- LR.sarlm(x, x$lm.model)
res <- x$LR1
pref <- ifelse(x$ase, "Asymptotic", "Approximate (numerical Hessian)")
if (x$type == "error") {
cat("\nLambda: ", format(signif(x$lambda, digits)),
", LR test value: ", format(signif(res$statistic,
digits)), ", p-value: ", format.pval(res$p.value,
digits), "\n", sep="")
if (!is.null(x$lambda.se)) {
if (!is.null(x$adj.se)) {
x$lambda.se <- sqrt((x$lambda.se^2)*x$adj.se)
}
cat(pref, " standard error: ",
format(signif(x$lambda.se, digits)),
ifelse(is.null(x$adj.se), "\n z-value: ",
"\n t-value: "), format(signif((x$lambda/
x$lambda.se), digits)),
", p-value: ", format.pval(2*(1-pnorm(abs(x$lambda/
x$lambda.se))), digits), "\n", sep="")
cat("Wald statistic: ", format(signif(x$Wald1$statistic,
digits)), ", p-value: ", format.pval(x$Wald1$p.value,
digits), "\n", sep="")
}
} else if (x$type == "sac" || x$type == "sacmixed") {
cat("\nRho: ", format(signif(x$rho, digits)), "\n",
sep="")
if (!is.null(x$rho.se)) {
if (!is.null(x$adj.se)) {
x$rho.se <- sqrt((x$rho.se^2)*x$adj.se)
}
cat(pref, " standard error: ",
format(signif(x$rho.se, digits)),
ifelse(is.null(x$adj.se), "\n z-value: ",
"\n t-value: "),
format(signif((x$rho/x$rho.se), digits)),
", p-value: ", format.pval(2 * (1 - pnorm(abs(x$rho/
x$rho.se))), digits), "\n", sep="")
}
cat("Lambda: ", format(signif(x$lambda, digits)), "\n", sep="")
if (!is.null(x$lambda.se)) {
pref <- ifelse(x$ase, "Asymptotic",
"Approximate (numerical Hessian)")
if (!is.null(x$adj.se)) {
x$lambda.se <- sqrt((x$lambda.se^2)*x$adj.se)
}
cat(pref, " standard error: ",
format(signif(x$lambda.se, digits)),
ifelse(is.null(x$adj.se), "\n z-value: ",
"\n t-value: "), format(signif((x$lambda/
x$lambda.se), digits)),
", p-value: ", format.pval(2*(1-pnorm(abs(x$lambda/
x$lambda.se))), digits), "\n", sep="")
}
cat("\nLR test value: ", format(signif(res$statistic, digits)),
", p-value: ", format.pval(res$p.value, digits), "\n",
sep="")
} else {
cat("\nRho: ", format(signif(x$rho, digits)),
", LR test value: ", format(signif(res$statistic, digits)),
", p-value: ", format.pval(res$p.value, digits), "\n",
sep="")
if (!is.null(x$rho.se)) {
if (!is.null(x$adj.se)) {
x$rho.se <- sqrt((x$rho.se^2)*x$adj.se)
}
cat(pref, " standard error: ",
format(signif(x$rho.se, digits)),
ifelse(is.null(x$adj.se), "\n z-value: ",
"\n t-value: "),
format(signif((x$rho/x$rho.se), digits)),
", p-value: ", format.pval(2 * (1 - pnorm(abs(x$rho/
x$rho.se))), digits), "\n", sep="")
}
if (!is.null(x$Wald1)) {
cat("Wald statistic: ", format(signif(x$Wald1$statistic,
digits)), ", p-value: ", format.pval(x$Wald1$p.value,
digits), "\n", sep="")
}
}
cat("\nLog likelihood:", logLik(x), "for", x$type, "model\n")
cat("ML residual variance (sigma squared): ",
format(signif(x$s2, digits)), ", (sigma: ",
format(signif(sqrt(x$s2), digits)), ")\n", sep="")
if (!is.null(x$NK)) cat("Nagelkerke pseudo-R-squared:",
format(signif(x$NK, digits)), "\n")
cat("Number of observations:", length(x$residuals), "\n")
cat("Number of parameters estimated:", x$parameters, "\n")
cat("AIC: ", format(signif(AIC(x), digits)), ", (AIC for lm: ",
format(signif(x$AIC_lm.model, digits)), ")\n", sep="")
if (x$type == "error") {
if (!is.null(x$Haus)) {
cat("Hausman test: ", format(signif(x$Haus$statistic,
digits)), ", df: ", format(x$Haus$parameter),
", p-value: ", format.pval(x$Haus$p.value, digits),
"\n", sep="")
}
}
if ((x$type == "lag" || x$type == "mixed") && x$ase) {
cat("LM test for residual autocorrelation\n")
cat("test value: ", format(signif(x$LMtest, digits)),
", p-value: ", format.pval((1 - pchisq(x$LMtest, 1)),
digits), "\n", sep="")
}
if (x$type != "error" && !is.null(x$LLCoef)) {
cat("\nCoefficients: (log likelihood/likelihood ratio)\n")
printCoefmat(x$LLCoef, signif.stars=signif.stars,
digits=digits, na.print="NA")
}
correl <- x$correlation
if (!is.null(correl)) {
p <- NCOL(correl)
if (p > 1) {
cat("\n", x$correltext, "\n")
correl <- format(round(correl, 2), nsmall = 2,
digits = digits)
correl[!lower.tri(correl)] <- ""
print(correl[-1, -p, drop = FALSE], quote = FALSE)
}
}
cat("\n")
invisible(x)
}
# extension for sarlm objects (spdep package)
extract.sarlm <- function(model, include.nobs = TRUE, include.loglik = TRUE,
include.aic = TRUE, include.lr = TRUE, include.wald = TRUE, ...) {
s <- summary(model, ...)
names <- rownames(s$Coef)
cf <- s$Coef[, 1]
se <- s$Coef[, 2]
p <- s$Coef[, ncol(s$Coef)]
if (model$type != "error") { # include coefficient for autocorrelation term
rho <- model$rho
cf <- c(cf, rho)
names <- c(names, "$\\rho$")
if (!is.null(model$rho.se)) {
if (!is.null(model$adj.se)) {
rho.se <- sqrt((model$rho.se^2) * model$adj.se)
} else {
rho.se <- model$rho.se
}
rho.pval <- 2 * (1 - pnorm(abs(rho / rho.se)))
se <- c(se, rho.se)
p <- c(p, rho.pval)
} else {
se <- c(se, NA)
p <- c(p, NA)
}
}
if (!is.null(model$lambda)) {
cf <-c(cf, model$lambda)
names <- c(names, "$\\lambda$")
if (!is.null(model$lambda.se)) {
if (!is.null(model$adj.se)) {
lambda.se <- sqrt((model$lambda.se^2) * model$adj.se)
} else {
lambda.se <- model$lambda.se
}
lambda.pval <- 2 * (1 - pnorm(abs(model$lambda / lambda.se)))
se <- c(se, lambda.se)
p <- c(p, lambda.pval)
} else {
se <- c(se, NA)
p <- c(p, NA)
}
}
gof <- numeric()
gof.names <- character()
gof.decimal <- logical()
if (include.nobs == TRUE) {
n <- length(s$fitted.values)
param <- s$parameters
gof <- c(gof, n, param)
gof.names <- c(gof.names, "Num.\ obs.", "Parameters")
gof.decimal <- c(gof.decimal, FALSE, FALSE)
}
if (include.loglik == TRUE) {
ll <- s$LL
gof <- c(gof, ll)
gof.names <- c(gof.names, "Log Likelihood")
gof.decimal <- c(gof.decimal, TRUE)
}
if (include.aic == TRUE) {
aic <- AIC(model)
aiclm <- s$AIC_lm.model
gof <- c(gof, aiclm, aic)
gof.names <- c(gof.names, "AIC (Linear model)", "AIC (Spatial model)")
gof.decimal <- c(gof.decimal, TRUE, TRUE)
}
if (include.lr == TRUE && !is.null(s$LR1)) {
gof <- c(gof, s$LR1$statistic[[1]], s$LR1$p.value[[1]])
gof.names <- c(gof.names, "LR test: statistic", "LR test: p-value")
gof.decimal <- c(gof.decimal, TRUE, TRUE)
}
if (include.wald == TRUE && !is.null(model$Wald1)) {
waldstat <- model$Wald1$statistic
waldp <- model$Wald1$p.value
gof <- c(gof, waldstat, waldp)
gof.names <- c(gof.names, "Wald test: statistic", "Wald test: p-value")
gof.decimal <- c(gof.decimal, TRUE, TRUE)
}
tr <- createTexreg(
coef.names = names,
coef = cf,
se = se,
pvalues = p,
gof.names = gof.names,
gof = gof,
gof.decimal = gof.decimal
)
return(tr)
}
setMethod("extract", signature = className("sarlm", "spdep"),
definition = extract.sarlm)
=======================================
Model 1
---------------------------------------
(Intercept) 0.91 (0.66)
X -0.00 (0.10)
lag.(Intercept) -0.03 (0.02)
lag.X 0.90 (0.02) ***
rho -0.00 (0.01)
lambda -0.02 (0.02)
---------------------------------------
Num. obs. 100
Parameters 7
Log Likelihood -145.44
AIC (Linear model) 585.63
AIC (Spatial model) 304.88
LR test: statistic 288.74
LR test: p-value 0.00
=======================================
*** p < 0.001, ** p < 0.01, * p < 0.05