AEM 6850
Empirical Methods for Applied Economists
Prof. Ariel Ortiz-Bobea
12 · Publication graphics
Tuesday, October 6, 2026
Figure 1, ready for Nature
- Question: did farmland values rise more in counties where home values rose more?
- a Farmland value per acre (Census of Agriculture) and b median value of an owner-occupied home (1940 census; 2018–2022 American Community Survey), each as its 2022 value divided by its 1940 value. c One against the other, 2,985 counties.
- Real terms: each year’s dollars divided by the consumer price index. 183 mm wide, Nature’s two-column width.
Figure 1 with R’s default settings
The same data, with pdf(), par(mfrow = c(1, 3)), map(fill = TRUE), cut(x, 6) and plot(), with nothing else set. The code is at the start of Part 2.
- A 7 × 7 in page; no journal prints that size, and two thirds of it is blank
- Fonts not embedded: the file holds no copy of the typeface
- Six classes of equal width: one outlier stretches them, and 3,052 of 3,053 counties get the same red
- No key, and black county borders that turn the East solid
- The x label is a variable name,
d$house_ratio; the y label does not show
One Nature column, drawn in the R Console
R Console: writes cars-nature.pdf
# cars comes with R: speed (mph) and
# stopping distance (ft) of 50 cars
cairo_pdf("cars-nature.pdf",
width = 89 / 25.4, height = 70 / 25.4,
pointsize = 7, family = "Helvetica")
par(mar = c(3, 3.2, 1.6, 0.6), mgp = c(1.8, 0.5, 0),
cex.axis = 6 / 7, # 6-pt tick labels
lwd = 0.5 / 0.75, # 0.5-pt box; axes 0.75 pt
las = 1)
plot(cars$speed, cars$dist, pch = 16, cex = 0.5,
xlab = "Speed (mph)",
ylab = "Stopping distance (ft)")
mtext("a", side = 3, line = 0.4, adj = 0,
font = 2, cex = 8 / 7) # 8-pt bold letter
dev.off()
The PDF, 89 × 70 mm.
- The width, font, text sizes and line weight come from the table on the previous slide.
Categorical, sequential and diverging colour scales
- Deuteranopia: red–green colour blindness. The eye lacks the cones that sense green, so reds and greens look alike.
“About 1 in 12 men have color vision deficiency”, most often red and green (National Eye Institute). Okabe-Ito colours: Wong (Nature Methods, 2011), the paper Nature’s guide cites. Simulation: Machado, Oliveira and Fernandes (IEEE TVCG, 2009), full severity.
The default code for Figure 1
code/figure1_default.R (part); needs data_clean/ from code/2_clean_data.R
library(maps)
d <- readRDS("data_clean/counties.rds")
poly <- map("county", plot = FALSE, namesonly = TRUE)
fips <- sprintf("%05d",
county.fips$fips[match(poly, county.fips$polyname)])
pdf("figures/default/figure1.pdf")
par(mfrow = c(1, 3))
map("county", fill = TRUE,
col = heat.colors(6)[cut(d$land_ratio[match(fips, d$fips)], 6)])
title("Farmland")
map("county", fill = TRUE,
col = heat.colors(6)[cut(d$house_ratio[match(fips, d$fips)], 6)])
title("Houses")
plot(d$house_ratio, d$land_ratio)
dev.off()
dhas one row per county.land_ratioandhouse_ratioare the 2022 value divided by the 1940 value.- Download the project (a zip file), then its data (Cornell NetID login): replace the project’s
datafolder with the one in the data zip.
The target: a picture of Figure 1 for Nature
target/figure1.png: the figure from “Figure 1, ready for Nature” as a picture at 300 dpi (pixels per inch): 2,159 × 730 pixels, 183 mm wide in print.- It stands in for a published figure whose look you want to copy. We drew it before class with the project’s finished script (Part 3).
- The project’s
match/folder holds three things: this picture, a copy of the default code, and the data. - Claude Code starts in
match/, with no list of Nature’s rules and none of the code that drew the target. - The prompt says “Use only the files in this folder”, and Claude Code asks before it reads anything outside it.
What Claude Code drew
- A test run before class: 335 seconds, 12 rounds of drawing and comparing.
- Its own pixel-by-pixel comparison found the two almost the same.
Claude Code’s figure misses two of Nature’s rules
| Text | Nature’s rule (Part 1) | Claude Code’s code (test run) | Size in print |
|---|---|---|---|
| Panel and axis titles | at most 7 pt | cex = 1.05 and 1.04 at pointsize = 7 |
7.35 and 7.28 pt |
| Panel letters | 8-pt bold | font = 2, cex = 1.05 |
7 × 1.05 = 7.35 pt |
- A picture shows how a figure looks, without the numbers behind it. Claude Code set the sizes to match the widths of words in pixels: at 300 dpi, half a point is about two pixels.
- Matching a picture gets close, and will rarely be exact.
- Meeting a journal’s rules exactly takes the rules themselves, which Part 3 gives the agent.
The checking script on three versions of Figure 1
| Figure | Width | Fonts embedded | Text size | Panel letters | Other five checks |
|---|---|---|---|---|---|
Default code |
FAIL, 177.8 mm | FAIL, 0 of 2 | FAIL, up to 9.2 pt | FAIL, none | PASS |
 Claude Code (Part 2) |
PASS | PASS | FAIL, up to 7.3 pt | FAIL, 7.3 pt | PASS |
Target |
PASS | PASS | PASS, 6 to 7 pt | PASS, 8 pt | PASS |
Other five checks: height, font, text on page, line weight, vector only.
Terminal, in the project folder: one run per PDF
Rscript tools/check_fig.R figures/default/figure1.pdf natureThe Nature and AER versions
- AER: three files,
Figure1a.pdftoFigure1c.pdf, each 115 mm wide, text 8 and 9 pt, “Panel A.” in each title. - All of these are our house rules. One file per panel follows the file names in the style guide’s example.
- Nature: one file, 183 × 62 mm, text 6 and 7 pt, letters a, b, c in 8-pt bold (the slide “Figure 1, ready for Nature”).