R ggplot无法使用镶嵌面_包裹和分组绘制平滑gam

我正在尝试使用ggplot绘制一个多面板和多线图，该图具有群体美学以及

facet\u wrap

。但是，当一组数据点太少时，

geom_smooth

会对刻面图中的所有线失败

plot1 <- ggplot(data=df1, 
                aes(x=Year, y=Mean, group=Group2, linetype=Group2, shape=Group2)) +  
  geom_errorbar(aes(ymin=Mean-SE, ymax=Mean+SE), width=0.2) +  
  geom_smooth(method = "gam", se=F, formula = y ~ s(x, k=3), size = 1, colour="black") + 
  geom_point(position=pd, size=2, fill="white") +  
  scale_x_continuous(limits=c(min(df1$Year-0.1), max(df1$Year+0.1)), 
                     breaks=seq(min(df1$Year),max(df1$Year),5)) +  
  facet_wrap(~Group1, scales = "free", ncol=2) +  
  theme_bw() + 
  theme(axis.text.x = element_text(),
        axis.title.x = element_blank(), 
        strip.background = element_blank(), 
        axis.line.x = element_line(colour="black"),
        axis.line.y = element_line(colour="black"), 
        panel.grid.minor = element_blank(), 
        panel.grid.major = element_blank(),
        panel.border = element_blank(), 
        panel.background = element_blank(),
        legend.position="top",
        legend.title = element_blank())
plot(plot1)

策划

---

很明显，有足够的数据在组factorB中为factorC绘制一条平滑的线。有什么想法吗

我认为这相当棘手。在对for

StatSmooth

进行了一些测试和阅读后，我将我的发现总结如下：

观察

geom_smooth（）
如果将绘图分为多个面板，则仅具有不符合要求的数据组的面板受到影响
这不会发生在
公式=y~x
（即默认公式）上
使用默认公式的某些其他方法（例如
“lm”
，
“glm”
）不会出现这种情况，但使用
方法=“黄土”
时会出现这种情况
如果数据组只有一个观察值，则不会发生这种情况

我们可以用一些简化的代码重现上述内容：

# create sample data
n <- 30
set.seed(567)
df.1 <- data.frame( # there is only 1 observation for group == B
  x = rnorm(n), y = rnorm(n),
  group = c(rep("A", n - 1), rep("B", 1)),
  facet = sample(c("X", "Y"), size = n, replace = TRUE))    
set.seed(567)
df.2 <- data.frame( # there are 2 observations for group == B
  x = rnorm(n), y = rnorm(n),
  group = c(rep("A", n - 2), rep("B", 2)),
  facet = sample(c("X", "Y"), size = n, replace = TRUE))

# create base plot
p <- ggplot(df.2, aes(x = x, y = y, color = group)) + 
  geom_point() + theme_bw()

# problem: no smoothed line at all in the entire plot
p + geom_smooth(method = "gam", formula = y ~ s(x, k = 3))

# problem: no smoothed line in the affected panel
p + facet_wrap(~ facet) + 
  geom_smooth(method = "gam", formula = y ~ s(x, k = 3))

# no problem with default formula: smoothed lines in both facet panels
p + facet_wrap(~ facet) + geom_smooth(method = "gam")

# no problem with lm / glm, but problem with loess
p + facet_wrap(~ facet) + geom_smooth(method = "lm")
p + facet_wrap(~ facet) + geom_smooth(method = "glm")
p + facet_wrap(~ facet) + geom_smooth(method = "loess")

# no problem if there's only one observation (instead of two)
p %+% df.1 + geom_smooth(method = "gam", formula = y ~ s(x, k = 3))
p %+% df.1 + facet_wrap(~ facet) + 
  geom_smooth(method = "gam", formula = y ~ s(x, k = 3))

当参数
method=“gam”，formula=y~s（x，k=3）
用于只有2个观测值的数据帧时，会发生以下情况：

model <- do.call(mgcv::gam,
                 args = list(formula = y ~ s(x, k = 3),
                             data = df.2 %>% filter(group == "B" & facet == "X")))

也将返回一个有效的对象，尽管有大量的警告。但是，将其传递给
predictdf
将导致错误：

model <- do.call(stats::loess,
                 args = list(formula = y ~ x,
                             data = df.2 %>% filter(group == "B" & facet == "X")))

result <- ggplot2:::predictdf(
  model, 
  xseq = seq(-2, 1.5, length.out = 80), # pseudo range of x-axis values
  se = FALSE, level = 0.95) # default SE / level parameters

换句话说，如果指定组中只有一个观测值，
StatSmooth
立即返回一个空白数据帧。因此，它永远不会到达代码的后续部分来抛出任何错误

解决方法：
在确定了事情偏离轨道的地方之后，我们可以对
compute\u组
code进行调整（请参见注释和注释部分）：

处理此问题的一个非常简单的方法是将导致传递到
geom\u smooth
的数据中出现问题的行进行子集划分：

库（tidyverse）
df1%group_by（Group1，Group2）%>%filter（n（）>2），#子集
method=“gam”，公式=y~s（x，k=3））+
几何点（）
面_包裹（~Group1）

您只是没有足够的数据用于GAM。既然直线是直的，为什么不使用
lm
<代码>ggplot（数据=df1，aes（x=年，y=平均值，颜色=组2））+几何误差条（aes（ymin=平均值SE，ymax=平均值+SE），宽度=0.2）+几何平滑（方法='lm'，SE=假）+几何点（）+面包（~Group1）
谢谢Alistaire，但显然有四个点足以实现gam平滑，然而，由于另一组中只有两点，它失败了。问题是为什么完整的情节失败了，当然它不应该为这两个点画一个gam平滑，但它肯定应该为另一组画一个gam平滑。为了方便起见，我还提供了摘要数据，因为数据集非常庞大，而且有许多组比较。问题在于未能绘制平滑的曲线，而不是其背后的统计信息！！！你可以创建一个有3个点的GAM，但是作为一个观测者，我非常怀疑任何数据很少的回归，特别是如果它是曲线的话。这就是说，这个问题可以有效地推广到更合理的情况下，平滑需要更多的数据。谢谢大家的努力。阿利斯泰尔创造了一个简单的解决方案，但Z-Lin确实打开了引擎盖，找出了失败的原因和原因。
model <- do.call(mgcv::gam,
                 args = list(formula = y ~ s(x, k = 3),
                             data = df.2 %>% filter(group == "B" & facet == "X")))

model <- do.call(mgcv::gam, # or stats::lm, stats::glm
                 args = list(formula = y ~ x,
                             data = df.2 %>% filter(group == "B" & facet == "X")))

result <- ggplot2:::predictdf(
  model, 
  xseq = seq(-2, 1.5, length.out = 80), # pseudo range of x-axis values
  se = FALSE, level = 0.95) # default SE / level parameters

model <- do.call(stats::loess,
                 args = list(formula = y ~ x,
                             data = df.2 %>% filter(group == "B" & facet == "X")))

result <- ggplot2:::predictdf(
  model, 
  xseq = seq(-2, 1.5, length.out = 80), # pseudo range of x-axis values
  se = FALSE, level = 0.95) # default SE / level parameters

if (length(unique(data$x)) < 2) {
      # Not enough data to perform fit
      return(data.frame())
    }

new.compute_group <- function(
  data, scales, method = "auto", formula = y~x, se = TRUE, n = 80, span = 0.75, 
  fullrange = FALSE, xseq = NULL, level = 0.95, method.args = list(), na.rm = FALSE) {
  if (length(unique(data$x)) < 2) return(data.frame()) 
  if (is.null(data$weight)) data$weight <- 1
  if (is.null(xseq)) {
    if (is.integer(data$x)) {
      if (fullrange) {
        xseq <- scales$x$dimension()
      } else {
        xseq <- sort(unique(data$x))
      }
    } else {
      if (fullrange) {
        range <- scales$x$dimension()
      } else {
        range <- range(data$x, na.rm = TRUE)
      }
      xseq <- seq(range[1], range[2], length.out = n)
    }
  }
  if (identical(method, "loess")) method.args$span <- span 
  if (is.character(method)) method <- match.fun(method)
  base.args <- list(quote(formula), data = quote(data), weights = quote(weight))

  # if modelling fails, return empty data frame
  # model <- do.call(method, c(base.args, method.args))
  model <- try(do.call(method, c(base.args, method.args)))
  if(inherits(model, "try-error")) return(data.frame())

  # if modelling didn't fail, but prediction returns NA,
  # also return empty data frame
  # predictdf(model, xseq, se, level)
  pred <- try(ggplot2:::predictdf(model, xseq, se, level))
  if(inherits(pred, "try-error")) return(data.frame())
  return(pred)
}

# same as stat_smooth() except that it uses stat = StatSmooth2, rather 
# than StatSmooth
stat_smooth_local <- function(
  mapping = NULL, data = NULL, geom = "smooth", position = "identity", ...,
  method = "auto", formula = y ~ x, se = TRUE, n = 80, span = 0.75,
  fullrange = FALSE, level = 0.95, method.args = list(), na.rm = FALSE,
  show.legend = NA, inherit.aes = TRUE) {
  layer(
    data = data, mapping = mapping, stat = StatSmooth2,
    geom = geom, position = position, show.legend = show.legend,
    inherit.aes = inherit.aes,
    params = list(
      method = method, formula = formula, se = se, n = n,
      fullrange = fullrange, level = level, na.rm = na.rm,
      method.args = method.args, span = span, ...
    )
  )
}

# inherit from StatSmooth
StatSmooth2 <- ggproto(
  "StatSmooth2", ggplot2::StatSmooth,
  compute_group = new.compute_group
)

# problem resolved: smoothed line for applicable group in the entire plot
p + stat_smooth_local(method = "gam", formula = y ~ s(x, k = 3))

# problem resolved: smoothed line for applicable group in the affected panel
p + facet_wrap(~ facet) + 
  stat_smooth_local(method = "gam", formula = y ~ s(x, k = 3))

# still no problem with default formula
p + facet_wrap(~ facet) + stat_smooth_local(method = "gam")

# still no problem with lm / glm; problem resolved for loess
p + facet_wrap(~ facet) + stat_smooth_local(method = "lm")
p + facet_wrap(~ facet) + stat_smooth_local(method = "glm")
p + facet_grid(~ facet) + stat_smooth_local(method = "loess")

# still no problem if there's only one observation (instead of two)
p %+% df.1 + stat_smooth_local(method = "gam", formula = y ~ s(x, k = 3))
p %+% df.1 + facet_wrap(~ facet) + 
  stat_smooth_local(method = "gam", formula = y ~ s(x, k = 3))

# showing one pair of contrasts here
cowplot::plot_grid(
  p + facet_wrap(~ facet) + ggtitle("Before") +
    geom_smooth(method = "gam", formula = y ~ s(x, k = 3)),
  p + facet_wrap(~ facet) + ggtitle("After") +
    stat_smooth_local(method = "gam", formula = y ~ s(x, k = 3)),
  nrow = 2
)