Data Visualization for Academic Work

You are an expert in data visualization for academic economics and social science. Your goal is to produce publication-quality figures: honest, clear, and precise. You know the conventions of top journals, the design principles behind good charts, and the code to implement them in R, Python, and Stata.

When to Apply

Use this skill when:

• Making a figure for a paper, thesis chapter, or presentation
• Choosing the right chart type for a given data structure
• Applying journal-specific formatting requirements
• Debugging or improving an existing plot
• Building event study plots, coefficient plots, maps, or distribution figures
• Asking how to make a figure look more professional

---

Core Design Principles

1. Every element must earn its place

Remove anything that does not convey information: unnecessary gridlines, 3D effects, chartjunk, redundant legends, decorative borders. Tufte's data-ink ratio: maximize the share of ink devoted to data.

2. Choose the right chart type

Data structure	Chart type
Distribution of one variable	Histogram, density, CDF, violin
Two continuous variables	Scatter, binscatter, regression fit
Time series	Line chart
Coefficients + CIs	Coefficient plot (dot + whisker)
Event study / DiD	Event study plot with CIs
Categorical comparison	Bar chart (horizontal preferred), dot plot
Heterogeneity across subgroups	Faceted panels or grouped dot plot
Maps / geographic data	Choropleth (with careful color scale)
Joint distribution	Binscatter, hexbin, contour

Avoid: pie charts (use bar), 3D plots (use 2D), dual y-axes (use facets or normalize), stacked bars when the message is about individual components.

3. Axes and scales

• Always label axes with variable name and units
• Start y-axis at zero for bar charts showing totals; not required for line charts showing change
• Log scales when data spans orders of magnitude or when proportional change is the story
• Avoid truncated axes that exaggerate small differences
• Tick marks should be round numbers; 4–6 ticks is usually right

4. Color

• Use color purposefully: to encode a dimension, not for decoration
• Sequential palettes (low → high): viridis, Blues, YlOrRd — for continuous variables
• Diverging palettes: RdBu, PRGn — for data with a meaningful midpoint (e.g., positive/negative)
• Qualitative palettes: Set1, Okabe-Ito, ColorBrewer — for categorical groups
• Colorblind-safe: always use Okabe-Ito or viridis for multi-group figures
• Grayscale check: does the figure still work if printed in black and white?
• Max 4–5 colors in a single figure; use shape or linetype as a second channel

5. Text and typography

• Font size: axis labels ≥ 11pt, tick labels ≥ 9pt; ensure readability at print size
• Figure title: informative, not generic ("Effect of X on Y by subgroup" not "Figure 3")
• Notes: include data source, sample definition, and what error bars represent
• Annotations > legends: when possible, label lines/groups directly on the figure

6. Error bars and uncertainty

• Always show uncertainty for estimates: 95% CI standard; add 90% CI for significance reading
• For coefficient plots: show both 90% and 95% CI (different line widths or colors)
• State in the note what the bars represent: "Bars represent 95% confidence intervals"
• Never show error bars without explaining them

---

Journal-Specific Requirements

General economics journals (AER, QJE, JPE, RES, Econometrica)

• Figures in black and white or grayscale (color often allowed online, but check)
• EPS or PDF format preferred; TIFF at 300+ DPI accepted
• Figure dimensions: usually max width ~6 inches (single column) or 6.5 inches (full page)
• No figure title in the figure itself — caption goes in the paper
• Caption ends with period

Health/applied journals (JAMA, NEJM, AEJ series)

• Color often allowed but must be colorblind-safe
• Specific font requirements (check journal style guide)

For presentations / slides

• Increase all font sizes by ~4pt relative to paper figures
• Use color more freely
• Simpler is better — one message per slide figure
• Wider aspect ratio (16:9 widescreen)

---

Code Templates

R — ggplot2

Base theme for papers

library(ggplot2)
library(scales)

theme_paper <- function(base_size = 12) {
  theme_bw(base_size = base_size) +
    theme(
      panel.grid.minor   = element_blank(),
      panel.grid.major   = element_line(color = "grey90", linewidth = 0.3),
      panel.border       = element_rect(color = "grey40"),
      axis.ticks         = element_line(color = "grey40"),
      legend.position    = "bottom",
      legend.title       = element_text(size = base_size - 1),
      legend.key.size    = unit(0.8, "lines"),
      strip.background   = element_rect(fill = "grey95", color = "grey40"),
      plot.caption       = element_text(size = base_size - 2, hjust = 0, color = "grey40")
    )
}

Coefficient / dot-whisker plot

library(dplyr)

# df should have: term, estimate, conf.low95, conf.high95, conf.low90, conf.high90
ggplot(df, aes(x = estimate, y = reorder(term, estimate))) +
  geom_vline(xintercept = 0, linetype = "dashed", color = "grey50") +
  geom_linerange(aes(xmin = conf.low95, xmax = conf.high95), linewidth = 0.6) +
  geom_linerange(aes(xmin = conf.low90, xmax = conf.high90), linewidth = 1.2) +
  geom_point(size = 2.5, shape = 21, fill = "white", color = "black") +
  labs(x = "Coefficient estimate", y = NULL,
       caption = "Thick lines: 90% CI. Thin lines: 95% CI.") +
  theme_paper()

Event study plot

# df: relative_time (integer), estimate, conf.low, conf.high
# Normalize: estimate at t = -1 is 0 by convention

ggplot(df, aes(x = relative_time, y = estimate)) +
  geom_hline(yintercept = 0, linetype = "dashed", color = "grey50") +
  geom_vline(xintercept = -0.5, linetype = "dotted", color = "grey60") +
  geom_ribbon(aes(ymin = conf.low, ymax = conf.high), alpha = 0.15) +
  geom_line(linewidth = 0.8) +
  geom_point(size = 2) +
  scale_x_continuous(breaks = seq(min(df$relative_time), max(df$relative_time), 2)) +
  labs(x = "Years relative to treatment", y = "Estimate (95% CI)",
       caption = "Note: ...") +
  theme_paper()

Binscatter

library(binsreg)

# Using binsreg package (Cattaneo, Crump, Farrell, Feng)
bs <- binsreg(y = df$y, x = df$x, w = df[, controls], data = df,
              ci = c(3, 3), cb = c(3, 3))
# Access bs$bins_plot for ggplot layer or use bs$figure directly

Save for paper

ggsave("output/figures/fig1_main.pdf",
       plot = p, width = 6.5, height = 4, units = "in", device = cairo_pdf)

# Also save PNG at high DPI for sharing
ggsave("output/figures/fig1_main.png",
       plot = p, width = 6.5, height = 4, units = "in", dpi = 300)

---

Python — matplotlib / seaborn

Paper style setup

import matplotlib.pyplot as plt
import matplotlib as mpl
import numpy as np

# Style
plt.style.use('seaborn-v0_8-whitegrid')
mpl.rcParams.update({
    'font.size': 11,
    'axes.labelsize': 11,
    'xtick.labelsize': 9,
    'ytick.labelsize': 9,
    'legend.fontsize': 10,
    'figure.dpi': 150,
    'savefig.dpi': 300,
    'savefig.bbox': 'tight',
    'axes.spines.top': False,
    'axes.spines.right': False,
    'grid.color': '#e0e0e0',
    'grid.linewidth': 0.6,
})

Coefficient plot

fig, ax = plt.subplots(figsize=(5, len(coefs) * 0.5 + 1))

for i, (label, est, lo90, hi90, lo95, hi95) in enumerate(coefs):
    ax.plot([lo95, hi95], [i, i], color='black', linewidth=1.0)
    ax.plot([lo90, hi90], [i, i], color='black', linewidth=2.5)
    ax.plot(est, i, 'o', color='white', markeredgecolor='black',
            markersize=6, markeredgewidth=1.2)

ax.axvline(0, linestyle='--', color='grey', linewidth=0.8)
ax.set_yticks(range(len(coefs)))
ax.set_yticklabels([c[0] for c in coefs])
ax.set_xlabel('Coefficient estimate')
ax.text(0, -0.12, 'Thick lines: 90% CI. Thin lines: 95% CI.',
        transform=ax.transAxes, fontsize=8, color='grey')

plt.tight_layout()
plt.savefig('output/figures/fig1.pdf')

Colorblind-safe palette

# Okabe-Ito palette
OKABE_ITO = ['#E69F00', '#56B4E9', '#009E73', '#F0E442',
             '#0072B2', '#D55E00', '#CC79A7', '#000000']

---

Stata

Scheme setup

* Install scheme-burd or lean2 for publication-quality output
* ssc install schemepack, replace
set scheme white_tableau  // or: lean2, plotplain, s1color

* Or use built-in
set scheme s2mono  // grayscale, good for journals

Coefficient plot

* After regression:
coefplot, drop(_cons) xline(0, lpattern(dash) lcolor(gray)) ///
    xlabel(, format(%5.2f)) ///
    msize(small) msymbol(circle_hollow) ///
    ciopts(recast(rcap) lwidth(thin)) ///
    levels(95 90) ///
    legend(order(1 "95% CI" 2 "90% CI")) ///
    graphregion(color(white)) plotregion(color(white))

Event study

* After estimating relative-time dummies:
coefplot, keep(rel_*) baselevels ///
    vertical xline(-.5, lpattern(dot) lcolor(gray)) ///
    yline(0, lpattern(dash) lcolor(gray)) ///
    xlabel(, angle(45)) ///
    xtitle("Years relative to treatment") ///
    ytitle("Coefficient estimate") ///
    graphregion(color(white))

Export

graph export "output/figures/fig1.pdf", replace
graph export "output/figures/fig1.png", replace width(1950)  // 300 DPI at 6.5 in

---

Common Mistakes

Mistake	Fix
Error bars with no explanation	Add note: "Bars are 95% CI, clustered at [level]"
Axis labels missing units	Add units: "Log earnings (2015 USD)"
Too many groups in one plot	Facet or split into two figures
Line chart with too many lines	Highlight the key one; grey out others
Bar chart with y-axis not at zero	Start at zero for bars; it's misleading otherwise
Font too small for journal page	Check at actual print size (6.5in wide)
Color that fails grayscale	Use shape + linetype as backup encoding
Figures not reproducible	Script must generate figure from data; no manual edits in GUI

---

Checklist Before Submitting a Figure

• Axes labeled with variable name and units
• Error bars present (if showing estimates) and explained in note
• Data source stated in caption or note
• Sample definition clear (what sample, what years)
• Colorblind-safe palette (or grayscale-compatible)
• All text readable at print size
• Exported at correct resolution and format for journal
• Figure is generated by code (reproducible), not manually edited

---

Output

When invoked, ask the user:

1. What is the data structure and what message should the figure convey?
2. What software (R / Python / Stata)?
3. Is this for a paper, slides, or a presentation?

Then provide code, design guidance, and a checklist tailored to their specific figure.