Data Analysis for PolicyMaking in Georgia
Module 5 · Building Your Policy Finding › 5.3 Writing up a finding
Subsection 5.3

Writing up a finding

~3 min

Reading

A data finding written for a policy audience is different from a finding written for an academic journal. It needs to be understandable to a legislator, an advocacy director, or a journalist—not just to another statistician. But plain language cannot come at the cost of accuracy. The skill is precision in plain language.

Plain language principles

Lead with the result, not the method. Start with what you found, not how you found it. Save the methodology for the note section.

Use one number that matters. Do not list five metrics when one tells the story. Pick the most meaningful comparison and lead with it. Other numbers go in the supporting evidence.

Name the comparison explicitly. "Rejection rates were higher" is not a finding—higher than what? "Mail ballot rejection rates in majority-Black counties (2.1%) were 2.3× the rate in majority-white counties (0.9%)" is a finding.

Hedging appropriately

Hedging is not weakness—it is intellectual honesty. The right hedge communicates exactly what your data does and does not show. Phrases like "consistent with," "suggests," and "is associated with" are appropriate when causation cannot be established. Phrases like "proves" and "demonstrates that discrimination occurred" are almost never appropriate for observational analysis.

After Callais, the most important hedge is: "This finding documents an effects disparity. Whether this disparity was produced by intentional racial discrimination would require additional evidence including examination of the decision-making record."

What the data does NOT show

Every finding section should include at least one sentence about what the data cannot tell you. This is not a weakness—it shows you understand your own analysis. Examples:

  • "This analysis does not account for variations in voter mail experience, which may correlate with both rejection rates and race."
  • "These estimates rely on ecological inference and cannot be verified at the individual level."
  • "This comparison does not control for differences in election administration staffing across counties."

Example of a well-written finding

Well-written (for a Policy Data Brief): "In the 2020 Georgia general election, counties where Black voters made up more than 40% of active registrants showed mail ballot rejection rates of 2.1%, compared to 0.9% in counties where Black voters made up less than 20% of active registrants—a difference of 2.3×. The most common rejection reason in high-Black-share counties was signature mismatch (44% of rejections). This analysis is based on county-level aggregates and cannot confirm this pattern holds within individual precincts or that the disparity results from any specific policy decision."

Writing Up a Finding — Weak vs. Strong Language Weak (do not write this) "Black voters had higher rejection rates. This shows the system is discriminating against them. Georgia should fix this." Problems: ✗ No specific number ✗ No comparison stated ✗ "Discriminating" infers intent from effects ✗ Recommendation is vague ("fix this") ✗ No methodology or limitation noted Strong (aim for this) "In the 2020 GA general election, counties with >40% Black active registrants had 2.1% mail ballot rejection — 2.3× the 0.9% rate in <20% Black counties. Sig. mismatch was 44% of rejections. This analysis is county-level aggregate; intent would require additional evidence." Why it works: ✓ Specific number (2.3×) ✓ Comparison stated explicitly ✓ Hedged correctly (effects ≠ intent)
Diagram 5.3 · Weak vs. strong finding language. A well-written finding includes a specific number, an explicit comparison, and a clear hedge about what the data does not show.