TEPI Explorer

CSC 596 · Project 2

Tucson Equity Priority Index

A visual analytics system for TEPI and Census data, 1990 to 2020–24

Aditya Retnanto
October 19, 2026

Five questions

TEPI ranks Pima County's 718 block groups for equity priority. Do similar block groups sit together? Were its high-priority areas served over time? Is it tied to geography? How do district lines treat them? And would it survive privacy protection?

T1 · SimilarityDo block groups with similar profiles sit near each other?Similarity view (UMAP / PCA) linked to the map · Dashboard
T2 · TemporalIf TEPI had existed since 1990, were high-priority areas targeted?TEPI recomputed for 1990–2024, economic indicators, time curves and gap chart · Dashboard
T3 · GeospatialIs there a geographical relationship in TEPI values?Block groups as a network of neighbours, global and local Moran's I, by decade · Neighbours
T4 · RedistrictingDid redistricting move or gather high-priority neighbourhoods?Legislative lines of 2001, 2011, 2021 as a network; who moved, by tier · Redistricting
T5 · PrivacyWould TEPI survive census-style differential privacy?Noise on the counts, then hot spots and PCA re-checked · Privacy

Data

TEPI City of Tucson, pilot index, July 2024
- 718 block groups
- 12 indicators in 2 domains (Economic, Personal)
- Percentile ranks, 5 tiers
1990 – 2000 Decennial census long form
- 1990 (STF3A, Census FTP) and 2000 (SF3, Census API)
- No health insurance or disability question
2009–13 – 2020–24 American Community Survey, 5-year estimates (Census API)
- 12 releases, with margins of error
- TEPI matches the 2019–23 release for 9 of its 12 indicators
Boundaries Census TIGERweb and boundary files
- Block groups of 1990, 2000, 2010 and 2020
- State legislative districts of 2001, 2011 and 2021

Everything before 2016–20 is moved onto today's 2020 block groups by area, so it is an estimate. All of it is one table: tepi_timeseries.csv, 718 block groups × 14 releases.

2010 2020 ⓘ

ACS release The 12 indicators rebuilt from the raw ACS against TEPI's published values. 9 match the 2019–23 release; unemployment, children and seniors match no release. A block group "matches" within 0.15 points (TEPI rounds to 2 decimals).

Our ACS value vs TEPI's published value, by indicator

ⓘ

Release Select into Landmarks

Map: TEPI by block group

ⓘ

Time curves

ⓘ Each curve Shared space A − B gap

Similarity

ⓘ Perplexity Iter. Neighbours Min dist

⚙ Model

Definitions TEPI's Standard

Time-curve units Time-curve distance

Indicators

Global Moran's I

Do neighbouring areas have similar values?

Compare each area with the average of its neighbours, and sum over the map. I = Σ z_i · lag_i ÷ Σ z_i²

Click a square to flip it.

Variable Significance Decade Show Select into

Block groups and their neighbours

ⓘ

State legislative districts drawn in Colour by Change from previous election Descriptive only ⓘ

Block groups and district lines

ⓘ

Mayoral vote by TEPI tier, 2011 → 2023

ⓘ

Each district: residents by TEPI tier, and the mayoral vote

ⓘ

TEPI distribution in the selected district

ⓘ

Data preprocessing

Rebuild TEPI's 12 indicators from raw Census tables, with TEPI's own definitions
- Found by testing alternatives: some divide household counts by total population; people of color excludes Hispanic white residents; no insurance counts the uninsured under 65 only
- Check: 9 of 12 come from the 2019–23 release: 8 match TEPI exactly, poverty in 78% of block groups (r = 0.999)
Fill the 3 TEPI-only indicators (unemployment, children, seniors)
- They match no ACS release (likely Esri estimates): TEPI's published values are used for every ACS release
- 1990 and 2000 use that census's own values
Remap releases older than 2016–20 onto today's 2020 block groups
- From 1990, 2000 and 2010 block groups by area: a spatial join of the old and new boundaries
- Check: at the seam (2015–19 vs 2016–20) rates still correlate 0.80–0.90 across block groups
Handle what is missing or different
- 1990 and 2000 have no insurance or disability question: TEPI is recomputed without them
- Income in 2024 dollars (CPI-U); margins of error carried through for the ACS
Recompute TEPI for all 14 releases
- Percentile ranks, averaged up TEPI's tree, cut into 5 tiers
- Check: the 2019–23 recompute matches the published TEPI (ρ = 1.00, 98.5% same tier)

Output: one table, tepi_timeseries.csv (718 block groups × 14 releases), plus a network of neighbouring block groups for Moran's I.

Methods

What-if scoring What-if
- Re-scores a block group with TEPI's own formula after its inputs change
- Change any of the 12 indicators and the reference release
Time curves · gap chart T2 · Dashboard
- One point per release, placed by MDS so similar profiles sit close, joined in time order (Bach et al. 2016)
- Change the groups, indicators, units, distance and time span
UMAP · t-SNE · PCA T1, T5 · Dashboard, Privacy
- Places block groups with similar indicator profiles close together
- Change the method, neighbours, minimum distance, perplexity and indicators
Moran's I T3, T5 · Neighbours, Privacy
- Tests whether neighbouring block groups have similar values; the local version finds hot and cold spots (Anselin 1995)
- Change the variable, significance level and release
District network T4 · Redistricting
- Block groups as nodes, districts as Voronoi regions, animated over the 2001, 2011 and 2021 lines; who moved, by tier
- Change the cycle, hover a bar to find its block groups
Differential privacy (Laplace) T5 · Privacy
- Noise on every count, TEPI recomputed, Moran's I and PCA re-run
- Change the privacy budget ε and the PCA inputs
Change the district lines (2001, 2011, 2021) and the plan shown

Neighbours = block groups sharing a boundary line (1,928 links). Selections (Groups A and B) are shared by every view on the Dashboard and Neighbours pages.

Marks and channels

One colour rule across the system: luminance (light → dark blue) is always TEPI tier; hue is always a category (group, class, district, indicator); dotted is always an interpolated release.

Page	View	Marks	Channels
1 · Dashboard	Map	Areas (block groups); pins	Luminance (5 blues) = TEPI tier · outline hue = Group A (orange) / B (black, dashed) · position = geography
	Time curves	Points (releases) + lines (time order)	Position = MDS of group profiles (distance = change) · luminance = time (light → dark) · hue = County / A / B · dotted = interpolated release
	A − B gap	Lines + points	x = release year · y = gap in points (A − B) · hue = indicator · dotted = interpolated release · 0 line = no gap
	Similarity	Points; bars (PCA loadings)	Position = UMAP / t-SNE / PCA (near = similar profile) · luminance = TEPI tier · outline = group · bar length = loading, hue = sign
2 · Neighbours	Hot-spot map	Areas, or nodes + links	Hue = local Moran class (red High–High, blue Low–Low, orange / light blue outliers, grey not significant) · node size = population
	Moran scatterplot	Points + regression line	x = block group (z) · y = neighbours' mean (z) · hue = class · slope = global Moran's I
	Hot spots by decade	Stacked bars	Length = number of block groups · hue = class · label = share of hot spots kept
3 · Redistricting	District network	Points + links + Voronoi areas (pieces)	Core vs margin: fill = times moved (grey, light / dark orange) · deltas: dashed line = old border · solid line = new border · area hue = new district · fill strength = piece moved / stayed · orange outline = selected piece · dot size = moved · dot luminance = TEPI tier
	Who moved?	Stacked bars	Length = share (moved, stayed) or count (per moved piece, old → new) · luminance = TEPI tier · highlighted row = selected piece
4 · Privacy	Moran's I, neighbours kept	Points (runs), tick, line	x = ε (ordinal) · y = statistic · tick = median · dashed = without noise
	PCA under noise	Points + rings	Position = noise-free PCA · luminance = TEPI tier · ring size = average move under noise

How TEPI turns 12 census indicators into a priority tier

Hover a box to trace what feeds into it and what it feeds. Line width = that input's share of the final average. Box values (P) are for the what-if block group below.

What-if: set a block group's indicator values

Values are percentages of residents or households. Each one is ranked against the real block groups, exactly as TEPI does, and the diagram updates.

Start from Rank against

Construction steps

Select indicators. 12 block-group percentages (Esri Business Analyst / ACS). Chosen from a review of existing vulnerability indices, limited to demographic variables, and revised by a community survey. Every one is read as "higher % = more vulnerable".
Percentile-rank each indicator across all block groups:
P = (rank − 1) / (n − 1) × 100
Chosen to reduce outlier influence and keep rank, not magnitude. Ties share the lowest rank, so 0% gets P = 0.
Sub-index = mean of its inputs' percentiles, then re-ranked. Five sub-indices; Age (children, seniors) is nested inside Physical.
Theme = mean of its sub-index percentiles: Economic (3) and Personal (2), then re-ranked.
TEPI = mean of the Economic and Personal percentiles, then re-ranked. The mean is compensatory: a high value on one input can offset a low one.
Decision heuristic: the TEPI percentile is cut into quintiles, giving five tiers of about 143 block groups each. "High" priority = the top 20%.

Equal weights at each level do not give equal weights overall. Because the tree is uneven, low English proficiency and education each carry 12.5% of the final average, while children and seniors carry 4.2%.

Build the ML pipeline

1. Variables

2. Missing values

Fill with median Drop block group

3. Scaling

4. k-means clusters:

ML clusters on the first two PCA components

Color by ML cluster TEPI tier Drag to lasso-select (Shift to add) · click a point for details · click empty space to clear

Expert tiers vs. ML clusters

Block groups per combination. Click a cell, a row or a column to select it; Shift-click to add more.

Why differential privacy?

Published tables can be turned back into people
- The Census Bureau rebuilt person records (block, sex, age, race, ethnicity) from 6.2 billion statistics in the 2010 Census tables
- Exactly right for 46% of the population; linked to commercial data, about 52 million identities confirmed
The response: noise on purpose
- Since the 2020 Census, every published count carries calibrated random noise (differential privacy, the TopDown algorithm)
- The ACS does not use it today; the decennial census does
TEPI is exposed
- Built from the smallest published geography, block groups
- About small groups: the unemployed, households without a vehicle, often only 20–25 people in a block group
Our question: if TEPI's inputs were protected this way, would its answers survive?
- Hot spots on the neighbour network (T3) and the similarity view (PCA, T1)

Source: M. Hawes, U.S. Census Bureau (2020), database reconstruction experiment on the 2010 Census.

How differential privacy works

The guarantee
- Any output is almost as likely with or without any one person:
  P[M(D) ∈ S] ≤ e^ε · P[M(D′) ∈ S]
- Smaller ε = stronger privacy = more noise; budgets add up over everything published
The Laplace mechanism
- One person changes a count by at most 1
- So add noise η ~ Laplace(0, 1/ε): typical size √2/ε people, whatever the count
- Round back to whole counts, 0 ≤ numerator ≤ denominator
In our system
- Noise on every numerator and denominator behind TEPI's 12 inputs (2019–23), TEPI recomputed, Moran's I and PCA re-run; 20 runs per ε
- Per-count ε: far stricter than the Census, whose 2020 redistricting data split a total ε = 19.61 (17.14 persons + 2.47 housing units) over all tables and levels, with discrete Gaussian noise

The same noise, different counts

Privacy budget ε PCA inputs ⓘ

PCA under noise: how far do block groups move at the chosen ε?

ⓘ

Is the clustering robust? Global Moran's I under noise

ⓘ

Is the dimensionality reduction robust? PCA neighbours kept

ⓘ

AI-assisted development

Claude Code wrote the code: 20 Python scripts and 17 D3 modules. I steered it: of my 199 messages over five days, 103 asked for a change; the rest said go ahead. Half of those changes (51) were about display or interaction.

My 103 change requests, by kind a message can count twice

Display and visual design34which view, which encoding, what to cut

Analysis direction, domain ideas31the questions and methods

Scope and narrative24what to keep, drop, merge

Interaction20selection, linking, replay, navigation

Correcting AI errors or misreadings15wrong output or a misunderstood request

What the AI brought

The data pipeline: Census API, 1990 and 2000 long form, remapping onto 2020 block groups
Reverse-engineering TEPI's definitions by scoring candidates; TEPI reproduced (ρ = 1.00)
Fast builds of methods I named: ReCom, ALARM, Laplace noise, Moran's I
Measures that make claims testable: geography check, neighbours kept, gap chart

What I brought

The questions: redistricting, differential privacy, polygons as a network, every DR method, core vs margin
The design: which view answers which question, encodings (no TEPI blues for districts, PCA first)
The interactions: shift-click, two groups, finding replays, slide sorter, clickable pieces
Interpretation: Voting Rights Act, the 2008 recession, overlapping releases

Blind baseline. A fresh session given only the rubric and the CSV proposed a broad, textbook plan (hot spots, weight sliders, a random forest; 12 views, untested hypotheses) and none of the questions above. It did catch one point I had only noted: compare ACS releases that do not overlap.

AI reflection: one view, six rounds

Example: the redistricting view

AskedLegislative lines of 2001, 2011, 2021 over TEPI: who moved?
ProducedMaps coloured by TEPI; a force graph that collapsed into a ball
Evaluated"Don't use TEPI colours"; "you don't get it: drag a node"
ProducedVoronoi districts over the network, with islands; bars I could not read
Refined (mine)A redistricting is a delta: old × new lines make clickable pieces
Refined (mine)All three plans: a stable core and the margins that move

Mistakes I caught

Time curves built per block group, not per group (I read the paper)
A map slider running backwards; district shapes not aligned between years
A Moran's I axis zoomed so tight that noise looked large
District hues in TEPI blue; tooltips off the screen; a slow hot-spot page
Unverified facts: a privacy budget quoted from memory; Laplace called the Census's mechanism

How I checked

Every number against something independent: published TEPI, the official canvass, the seam between block-group vintages
Every view by using it on a question: if I could not read the answer, it was redone
Claims sourced or cut: the ε of 19.61 checked against the Census Bureau's release

Lessons. The AI raised the floor, not the ceiling. It was fast and good at checking numbers, but it optimised what existed: it never proposed the questions, and its first visual design was usually cluttered or ambiguous. The design work, which view, which encoding, which interaction, stayed mine. What helped: naming the method or paper, asking it to prove claims against data, and a fresh agent for critique.

Tucson has been redistricted

State legislative lines were redrawn after every census: 2001, 2011 and 2021. Bold: each redistricting's districts; dashed: the lines they replaced.

State legislative districts, 2001 · 2011 · 2021, over TEPI of each era

Districts as a network: who stays, who moves

ⓘ

Who moved?

ⓘ