What kind of plot is suitable to show the relationship between two numeric variables?

Aggregation and representing uncertainty#

More complex datasets will have multiple measurements for the same value of the x variable. The default behavior in seaborn is to aggregate the multiple measurements at each x value by plotting the mean and the 95% confidence interval around the mean:

fmri = sns.load_dataset("fmri")
sns.relplot(data=fmri, x="timepoint", y="signal", kind="line")

The confidence intervals are computed using bootstrapping, which can be time-intensive for larger datasets. It’s therefore possible to disable them:

sns.relplot(
    data=fmri, kind="line",
    x="timepoint", y="signal", errorbar=None,
)

Another good option, especially with larger data, is to represent the spread of the distribution at each timepoint by plotting the standard deviation instead of a confidence interval:

sns.relplot(
    data=fmri, kind="line",
    x="timepoint", y="signal", errorbar="sd",
)

To turn off aggregation altogether, set the estimator parameter to None This might produce a strange effect when the data have multiple observations at each point.

sns.relplot(
    data=fmri, kind="line",
    x="timepoint", y="signal",
    estimator=None,
)

Plotting subsets of data with semantic mappings#

The lineplot() function has the same flexibility as scatterplot(): it can show up to three additional variables by modifying the hue, size, and style of the plot elements. It does so using the same API as scatterplot(), meaning that we don’t need to stop and think about the parameters that control the look of lines vs. points in matplotlib.

Using semantics in lineplot() will also determine how the data get aggregated. For example, adding a hue semantic with two levels splits the plot into two lines and error bands, coloring each to indicate which subset of the data they correspond to.

sns.relplot(
    data=fmri, kind="line",
    x="timepoint", y="signal", hue="event",
)

Adding a style semantic to a line plot changes the pattern of dashes in the line by default:

sns.relplot(
    data=fmri, kind="line",
    x="timepoint", y="signal",
    hue="region", style="event",
)

But you can identify subsets by the markers used at each observation, either together with the dashes or instead of them:

sns.relplot(
    data=fmri, kind="line",
    x="timepoint", y="signal", hue="region", style="event",
    dashes=False, markers=True,
)

As with scatter plots, be cautious about making line plots using multiple semantics. While sometimes informative, they can also be difficult to parse and interpret. But even when you are only examining changes across one additional variable, it can be useful to alter both the color and style of the lines. This can make the plot more accessible when printed to black-and-white or viewed by someone with color blindness:

sns.relplot(
    data=fmri, kind="line",
    x="timepoint", y="signal", hue="event", style="event",
)

When you are working with repeated measures data (that is, you have units that were sampled multiple times), you can also plot each sampling unit separately without distinguishing them through semantics. This avoids cluttering the legend:

sns.relplot(
    data=fmri.query("event == 'stim'"), kind="line",
    x="timepoint", y="signal", hue="region",
    units="subject", estimator=None,
)

The default colormap and handling of the legend in lineplot() also depends on whether the hue semantic is categorical or numeric:

dots = sns.load_dataset("dots").query("align == 'dots'")
sns.relplot(
    data=dots, kind="line",
    x="time", y="firing_rate",
    hue="coherence", style="choice",
)

It may happen that, even though the hue variable is numeric, it is poorly represented by a linear color scale. That’s the case here, where the levels of the hue variable are logarithmically scaled. You can provide specific color values for each line by passing a list or dictionary:

palette = sns.cubehelix_palette(light=.8, n_colors=6)
sns.relplot(
    data=dots, kind="line",
    x="time", y="firing_rate",
    hue="coherence", style="choice", palette=palette,
)

Or you can alter how the colormap is normalized:

from matplotlib.colors import LogNorm
palette = sns.cubehelix_palette(light=.7, n_colors=6)
sns.relplot(
    data=dots.query("coherence > 0"), kind="line",
    x="time", y="firing_rate",
    hue="coherence", style="choice",
    hue_norm=LogNorm(),
)

The third semantic, size, changes the width of the lines:

sns.relplot(
    data=dots, kind="line",
    x="time", y="firing_rate",
    size="coherence", style="choice",
)

While the size variable will typically be numeric, it’s also possible to map a categorical variable with the width of the lines. Be cautious when doing so, because it will be difficult to distinguish much more than “thick” vs “thin” lines. However, dashes can be hard to perceive when lines have high-frequency variability, so using different widths may be more effective in that case:

sns.relplot(
    data=dots, kind="line",
    x="time", y="firing_rate",
    hue="coherence", size="choice", palette=palette,
)

Controlling sorting and orientation#

Because lineplot() assumes that you are most often trying to draw y as a function of x, the default behavior is to sort the data by the x values before plotting. However, this can be disabled:

healthexp = sns.load_dataset("healthexp").sort_values("Year")
sns.relplot(
    data=healthexp, kind="line",
    x="Spending_USD", y="Life_Expectancy", hue="Country",
    sort=False
)

It’s also possible to sort (and aggregate) along the y axis:

sns.relplot(
    data=fmri, kind="line",
     x="signal", y="timepoint", hue="event",
    orient="y",
)