如何在ggplot中限制loess线置信区间阴影的范围

I’ve run into this exact issue before—when working with non-negative variables like height, weight, or abundance, seeing loess confidence bands dip below zero just doesn’t make sense. Here are two solid solutions to fix it:

Option 1: Visually Truncate the Confidence Interval (Quick Fix) #

The problem with using scale_y_continuous(limits = c(0, 200)) is that it filters out any data points (and parts of the confidence interval) that fall below 0. Instead, use coord_cartesian() to crop the plot display without removing data:

library(ggplot2)
df <- as.data.frame(rep(1:7, each = 5))
df[,2] <- c(0,1,5,0,6,0,7,2,9,1,1,18,4,2,34,8,18,24,56,12,12,18,24,63,48, 40,70,53,75,98,145,176,59,98,165)
names(df) <- c("x", "y")

ggplot(df, aes(x=x, y=y)) + 
  geom_point() + 
  geom_smooth(method = "loess") + 
  coord_cartesian(ylim = c(0, 200))  # This crops instead of filtering

This keeps the full loess fit (including the parts below 0) but only shows everything from y=0 upwards. The confidence interval shadow will be cleanly cut off at 0 instead of disappearing entirely.

Option 2: Force the Confidence Interval to Stay Above Zero (Model-Level Fix) #

If you want the loess fit itself to never generate values below 0, you’ll need to manually fit the model and adjust the confidence bounds before plotting. Here’s how:

1. Fit the loess model and get predictions with standard errors
2. Calculate the confidence interval, then clamp the lower bound to 0
3. Plot using the adjusted data

library(ggplot2)

# Same data as before
df <- as.data.frame(rep(1:7, each = 5))
df[,2] <- c(0,1,5,0,6,0,7,2,9,1,1,18,4,2,34,8,18,24,56,12,12,18,24,63,48, 40,70,53,75,98,145,176,59,98,165)
names(df) <- c("x", "y")

# Fit loess and get predictions with SE
loess_fit <- loess(y ~ x, data = df)
preds <- predict(loess_fit, se = TRUE)

# Create a dataframe for plotting, clamping lower CI to 0
pred_df <- data.frame(
  x = seq(min(df$x), max(df$x), length.out = 100),  # Smooth x sequence for plotting
  y_pred = predict(loess_fit, newdata = data.frame(x = seq(min(df$x), max(df$x), length.out = 100))),
  se = predict(loess_fit, newdata = data.frame(x = seq(min(df$x), max(df$x), length.out = 100)), se = TRUE)$se.fit
) %>%
  mutate(
    y_low = pmax(y_pred - 1.96 * se, 0),  # Ensure lower bound never goes below 0
    y_high = y_pred + 1.96 * se
  )

# Plot with adjusted CI bounds
ggplot(df, aes(x = x, y = y)) +
  geom_point() +
  geom_ribbon(data = pred_df, aes(y = y_pred, ymin = y_low, ymax = y_high), alpha = 0.2) +
  geom_line(data = pred_df, aes(y = y_pred), color = "#00BFC4") +
  scale_y_continuous(limits = c(0, 200))

This approach ensures that neither the fit line nor the confidence interval ever goes below 0, which is more statistically appropriate for non-negative variables.

Which Option Should You Choose? #

• Use Option 1 if you just need a quick visual fix and don’t mind that the underlying model still generates negative values (they’re just hidden).
• Use Option 2 if you want the fit to strictly adhere to your variable’s non-negative nature, making the plot more accurate to your data’s constraints.

内容的提问来源于stack exchange，提问作者James