Manage estimates and uncertainty with an interval dot plot

Last updated 2020-01-21 Read time 1 min

In A/B tests or regression results, showing confidence intervals along with point estimates makes conclusions more persuasive. Interval dot plots are clean and easy to read.

import numpy as np
import matplotlib.pyplot as plt

segments = ["Free", "Light", "Standard", "Premium"]
effect = np.array([0.12, 0.18, 0.27, 0.35])
low = effect - np.array([0.05, 0.06, 0.07, 0.08])
high = effect + np.array([0.05, 0.06, 0.07, 0.09])

fig, ax = plt.subplots(figsize=(6.4, 3.6))
ax.hlines(range(len(segments)), low, high, color="#94a3b8", linewidth=3)
ax.scatter(effect, range(len(segments)), color="#0ea5e9", s=90, zorder=3)

ax.axvline(0, color="#475569", linestyle="--", linewidth=1)
ax.set_yticks(range(len(segments)), labels=segments)
ax.set_xlabel("Lift vs control")
ax.set_title("Estimated effects with 90% confidence intervals")
ax.set_xlim(-0.05, 0.45)
ax.grid(axis="x", alpha=0.2)

for idx, (eff, lo, hi) in enumerate(zip(effect, low, high)):
    ax.text(hi + 0.01, idx, f"{eff*100:.1f}% (+{(hi - eff)*100:.1f}/-{(eff - lo)*100:.1f})", va="center")

fig.tight_layout()

plt.show()

Interval dot plots make uncertainty easy to read.

Reading tips #

The length of the horizontal line shows uncertainty; shorter lines mean more stable estimates.
Use the zero line to judge whether effects cross 0 at a glance.
Vary dot size to represent weight or sample size if needed.