是否可以使用ggplot2绘制gam拟合的平滑组件?

时间:2013-11-01 20:47:10

标签: r ggplot2 gam mgcv

我使用gam包中的mgcv拟合模型并将结果存储在model中,到目前为止,我一直在使用plot(model)查看平滑组件。我最近开始使用ggplot2并喜欢它的输出。所以我想知道,是否可以使用ggplot2绘制这些图形?

以下是一个例子:

x1 = rnorm(1000)
x2 = rnorm(1000)
n = rpois(1000, exp(x1) + x2^2)

model = gam(n ~ s(x1, k=10) + s(x2, k=20), family="poisson")
plot(model, rug=FALSE, select=1)
plot(model, rug=FALSE, select=2)

我对s(x1, k=10)s(x2, k=20)不感兴趣。

部分答案:

我深入研究plot.gammgcv:::plot.mgcv.smooth并构建了我自己的函数,它从平滑的组件中提取预测的效果和标准误差。它不处理plot.gam的所有选项和案例,因此我只将其视为部分解决方案,但它对我来说效果很好。

EvaluateSmooths = function(model, select=NULL, x=NULL, n=100) {
  if (is.null(select)) {
    select = 1:length(model$smooth)
  }
  do.call(rbind, lapply(select, function(i) {
    smooth = model$smooth[[i]]
    data = model$model

    if (is.null(x)) {
      min = min(data[smooth$term])
      max = max(data[smooth$term])
      x = seq(min, max, length=n)
    }
    if (smooth$by == "NA") {
      by.level = "NA"
    } else {
      by.level = smooth$by.level
    }
    range = data.frame(x=x, by=by.level)
    names(range) = c(smooth$term, smooth$by)

    mat = PredictMat(smooth, range)
    par = smooth$first.para:smooth$last.para

    y = mat %*% model$coefficients[par]

    se = sqrt(rowSums(
      (mat %*% model$Vp[par, par, drop = FALSE]) * mat
    ))

    return(data.frame(
      label=smooth$label
      , x.var=smooth$term
      , x.val=x
      , by.var=smooth$by
      , by.val=by.level
      , value = y
      , se = se
    ))
  }))
}

这会返回一个带有平滑组件的“熔化”数据框,因此现在可以使用上面示例中的ggplot

smooths = EvaluateSmooths(model)

ggplot(smooths, aes(x.val, value)) + 
  geom_line() + 
  geom_line(aes(y=value + 2*se), linetype="dashed") + 
  geom_line(aes(y=value - 2*se), linetype="dashed") + 
  facet_grid(. ~ x.var)

如果有人知道在一般情况下允许这样的包裹,我将非常感激。

3 个答案:

答案 0 :(得分:20)

您可以将visreg包与plyr包结合使用。 visreg基本上绘制了你可以使用predict()的任何模型。

library(mgcv)
library(visreg)
library(plyr)
library(ggplot2)

# Estimating gam model:
x1 = rnorm(1000)
x2 = rnorm(1000)
n = rpois(1000, exp(x1) + x2^2)
model = gam(n ~ s(x1, k=10) + s(x2, k=20), family="poisson")

# use plot = FALSE to get plot data from visreg without plotting
plotdata <- visreg(model, type = "contrast", plot = FALSE)

# The output from visreg is a list of the same length as the number of 'x' variables,
#   so we use ldply to pick the objects we want from the each list part and make a dataframe: 
smooths <- ldply(plotdata, function(part)   
  data.frame(Variable = part$meta$x, 
             x=part$fit[[part$meta$x]], 
             smooth=part$fit$visregFit, 
             lower=part$fit$visregLwr, 
             upper=part$fit$visregUpr))

# The ggplot:
ggplot(smooths, aes(x, smooth)) + geom_line() +
  geom_line(aes(y=lower), linetype="dashed") + 
  geom_line(aes(y=upper), linetype="dashed") + 
  facet_grid(. ~ Variable, scales = "free_x")

我们可以把整个东西放到一个函数中,并添加一个选项来显示模型中的残差(res = TRUE):

ggplot.model <- function(model, type="conditional", res=FALSE, 
                       col.line="#7fc97f", col.point="#beaed4", size.line=1, size.point=1) {
  require(visreg)
  require(plyr)
  plotdata <- visreg(model, type = type, plot = FALSE)
  smooths <- ldply(plotdata, function(part)   
    data.frame(Variable = part$meta$x, 
             x=part$fit[[part$meta$x]], 
             smooth=part$fit$visregFit, 
             lower=part$fit$visregLwr, 
             upper=part$fit$visregUpr))
  residuals <- ldply(plotdata, function(part)
    data.frame(Variable = part$meta$x, 
               x=part$res[[part$meta$x]], 
               y=part$res$visregRes))
  if (res)
    ggplot(smooths, aes(x, smooth)) + geom_line(col=col.line, size=size.line) +
      geom_line(aes(y=lower), linetype="dashed", col=col.line, size=size.line) +
      geom_line(aes(y=upper), linetype="dashed", col=col.line, size=size.line) +
      geom_point(data = residuals, aes(x, y), col=col.point, size=size.point) +
      facet_grid(. ~ Variable, scales = "free_x")
  else
    ggplot(smooths, aes(x, smooth)) + geom_line(col=col.line, size=size.line) +
      geom_line(aes(y=lower), linetype="dashed", col=col.line, size=size.line) +
      geom_line(aes(y=upper), linetype="dashed", col=col.line, size=size.line) +
      facet_grid(. ~ Variable, scales = "free_x")
  }

ggplot.model(model)
ggplot.model(model, res=TRUE)

ggplot without residuals ggplot with residuals 颜色是从http://colorbrewer2.org/中挑选出来的。

答案 1 :(得分:4)

仅供参考,visreg可以直接输出gg对象:

visreg(model, "x1", gg=TRUE)

enter image description here

答案 2 :(得分:0)

已更新,允许用户选择要绘制的变量。 将“残差”一词更改为“ res_data”,以避免与residuals函数冲突。

ggplot.model <- function(model, type="conditional", res=FALSE, 
                       col.line="#7fc97f", col.point="#beaed4", size.line=1, size.point=1, no_col = NULL,
                       what = "all") {
  require(visreg)
  require(plyr)
  
  plotdata <- visreg(model, type = type, plot = FALSE)
  smooths <- ldply(plotdata, function(part)   
    data.frame(Variable = part$meta$x, 
             x=part$fit[[part$meta$x]], 
             smooth=part$fit$visregFit, 
             lower=part$fit$visregLwr, 
             upper=part$fit$visregUpr))
  res_data <- ldply(plotdata, function(part)
    data.frame(Variable = part$meta$x, 
               x=part$res[[part$meta$x]], 
               y=part$res$visregRes))
  
   if (what != "all") {
    smooths <- smooths %>%
      filter(lapply(Variable,as.character)%in% what)
    res_data <- res_data%>%
      filter(lapply(Variable,as.character)%in% what)
   }
  
  
  if (res)
    ggplot(smooths, aes(x, smooth)) + geom_line(col=col.line, size=size.line) +
      geom_line(aes(y=lower), linetype="dashed", col=col.line, size=size.line) +
      geom_line(aes(y=upper), linetype="dashed", col=col.line, size=size.line) +
      geom_point(data = res_data, aes(x, y), col=col.point, size=size.point) +
      facet_wrap(. ~ Variable, scales = "free_x", ncol = no_col) + theme_bw()
  else
    ggplot(smooths, aes(x, smooth)) + geom_line(col=col.line, size=size.line) +
      geom_line(aes(y=lower), linetype="dashed", col=col.line, size=size.line) +
      geom_line(aes(y=upper), linetype="dashed", col=col.line, size=size.line) +
      facet_wrap(. ~ Variable, scales = "free_x", ncol=no_col)
}