Technical Implementation of Setting Individual Axis Limits with facet_wrap and scales="free"

Problem Background and Data Preparation

In data modeling analysis, it is often necessary to simultaneously display scatter plots of predicted vs. actual values and predicted vs. residual values. This comparative analysis helps evaluate model fitting performance and residual distribution characteristics. The original data contains three key variables: actual values (act), predicted values (pred), and residuals (resid).

Using the melt function from the reshape2 package to reshape the data, integrating act and resid variables into the value column, with pred as the identifier variable:

library(reshape2)
plot <- melt(results, id.vars = "pred")

This data format transformation enables generating multiple related but differently scaled facet plots using a single ggplot call.

Basic Faceted Plot Implementation

Using ggplot2's facet_wrap function to create faceted plots is the standard approach for handling such multi-plot comparisons. The basic implementation code is as follows:

library(ggplot2)
p <- ggplot(plot, aes(x = pred, y = value)) + 
    geom_point(size = 2.5) + 
    theme_bw()
p <- p + facet_wrap(~variable, scales = "free")
print(p)

The scales="free" parameter allows each facet to have independent axis ranges, which is crucial for handling variables of different magnitudes (such as actual values and residuals). However, this freedom also introduces new challenges: when specific facets (like actual vs predicted) require consistent x and y axis ranges, simple global axis limit settings cannot meet the requirements.

Challenges and Misconceptions in Axis Control

At first glance, calculating global min-max values and setting axis limits seems like a reasonable solution:

min_xy <- min(min(plot$pred), min(plot$value))
max_xy <- max(max(plot$pred), max(plot$value))

p <- ggplot(plot, aes(x = pred, y = value)) + 
    geom_point(size = 2.5) + 
    theme_bw()
p <- p + facet_wrap(~variable, scales = "free")
p <- p + scale_x_continuous(limits = c(min_xy, max_xy))
p <- p + scale_y_continuous(limits = c(min_xy, max_xy))

The problem with this approach is that global axis limits affect all facets, including the residual plot. Since residuals typically have different numerical ranges, this forced uniformity causes the residual plot's axis range to be too large, with data points concentrated in the center of the graph, making it difficult to observe residual distribution characteristics.

Core Principles of the geom_blank Solution

A more elegant solution utilizes the geom_blank layer to extend the data range of specific facets. geom_blank does not draw any visible elements in the plot but affects axis range calculations. The specific implementation steps are as follows:

First, calculate the axis range required for the actual vs predicted facet:

range_act <- range(range(results$act), range(results$pred))

Then create a dummy data frame containing the extended range data:

dummy <- data.frame(pred = range_act, 
                    value = range_act,
                    variable = "act", 
                    stringsAsFactors = FALSE)

Finally, add the geom_blank layer during plotting:

ggplot(d, aes(x = pred, y = value)) +
    facet_wrap(~variable, scales = "free") +
    geom_point(size = 2.5) + 
    geom_blank(data = dummy) + 
    theme_bw()

The key to this method is that the dummy data specifies only variable="act", thus only affecting the axis range of the actual vs predicted facet without interfering with the residual plot's axis settings.

Technical Details and Parameter Optimization

Several important technical details need attention during implementation:

Application of the range function: The range function returns the minimum and maximum values of a vector, which is more concise and efficient than using min and max functions separately. The expression range(range(a), range(b)) correctly calculates the overall range of multiple vectors.

Dummy data construction: Ensure the column names and data types of the dummy data frame match the original data, particularly for factor-type variable columns that require the stringsAsFactors=FALSE parameter to maintain consistency.

Layer order: The position of the geom_blank layer does not affect its functionality, but for code clarity, it is recommended to place it after geom_point.

Alternative Solutions Analysis and Comparison

Another possible solution is using the scales="free_x" parameter:

p <- ggplot(plot, aes(x = pred, y = value)) + 
    geom_point(size = 2.5) + 
    theme_bw()
p <- p + facet_wrap(~variable, scales = "free_x")

This method only allows free scaling of the x-axis while keeping the y-axis uniform. In certain specific scenarios, this approach may be simpler and more effective, but it cannot meet complex requirements where different facets need different y-axis ranges.

Practical Application Extensions

The geom_blank method can be further extended to accommodate more complex facet control requirements. For example, when different axis ranges are needed for multiple facet groups, corresponding dummy data frames can be created:

dummy1 <- data.frame(pred = range1, value = range1, variable = "group1")
dummy2 <- data.frame(pred = range2, value = range2, variable = "group2")
# Add multiple geom_blank layers in the plot

This approach is particularly useful when fine control of axis ranges for multiple facet subsets is required, as mentioned in the reference article regarding setting different y-axis ranges for different facet groups.

Summary and Best Practices

The method of using geom_blank layers to control specific facet axis ranges offers the following advantages: concise and clear code with logical clarity; no interference with other facets' axis settings; maintenance of ggplot2 syntax consistency; easy extension and modification.

In practical applications, it is recommended to: carefully plan facet structure and axis requirements; use appropriate variable naming to improve code readability; conduct thorough testing and validation in complex scenarios. This method is not only suitable for modeling result visualization but can also be widely applied to various statistical analysis charts requiring fine control of facet axes.