Willis Danielson

Willis Danielson

Demography, urbanization, the environment.

Environmental modeling · Time series

Forecasting Brighton snowpack

When it comes to snow accumulation, the 2025–26 winter was among the worst for the Wasatch Front. I wanted to see whether I could make a reasonable snow water equivalent (SWE) forecast for such an unconventional season. The model followed early accumulation fairly well, then missed the onset of melt in mid-March.

Model and evaluation

The model used rolling 60-day windows of SWE observations, along with sine and cosine features to encode where each date falls in the hydrological year. I used a multilayer perceptron with two hidden layers of 128 and 64 neurons, ReLU activations, and 15% dropout between the layers. The training data came from the Brighton SNOTEL site, going back to the start of the 1988 hydrological year. I removed leap days to give each year the same length and a consistent day-of-season index.

To estimate uncertainty, I used Monte Carlo dropout during inference. I kept dropout active across 300 forward passes using the same input. This produced a distribution of possible outcomes for each forecast step, which I summarized with the median, 10th percentile, and 90th percentile. The forecast was recursive. Each day’s median prediction fed back into the model so the next prediction had a complete 60-day window.

The first forecast was completed on December 22, 2025. Subsequent runs were roughly weekly through April 5, 2026. I reran the full sequence on June 23, 2026, so every forecast could be evaluated against the same observed season.

Past seasons that resembled this one typically continued accumulating snow into late March or beyond. With its limited feature set, the model had no way to foresee an early melt. It predicted below-average SWE that remained closer to the historical median and seasonal trend than the observations did. My read is that early accumulation was below average but within the range of comparable years. The unusually early melt was what disrupted forecast accuracy.

The model also struggled to generalize. It performed poorly when I retrained it on site-specific observations from the SNOTEL station near Snowbird, and changes to the architecture did not help. My next step is to try an ensemble of methods with a wider set of environmental features and observations from other sites.

First frame of the Brighton snow water equivalent forecast series
Brighton SNOTEL · Forecasts and observations, 2025–26Open GIF
Browse all 15 interactive forecast runs

Environmental inequality · Regression

Street trees and income in New York City

Who has access to the benefits of urban green space? I approached that question through the trees that line New York City’s streets. Drawing on the 2015 NYC Street Tree Census, I used the count of living street trees in each census tract as a proxy for neighborhood green space and examined its relationship with median household income.

In the adjusted model, median income was no longer a statistically significant predictor. The shares of White residents and owner-occupied housing were positively associated with tree counts. These associations point to the importance of considering race, housing tenure, and other neighborhood characteristics when studying urban environmental inequalities. The model cannot establish causation.

Why I used a negative binomial model

The simple ordinary least squares (OLS) fit in the chart suggests a positive association between income and street-tree counts. Because the outcome is a count and the spread of the residuals varies, I estimated a negative binomial regression with robust standard errors and controls for several tract-level sociodemographic characteristics.

Street tree count by tract median household incomeHover over a point to inspect a census tract. The line shows the unadjusted OLS fit.
Negative binomial model results
Model diagnostics
Dep. Variable: num_tree No. Observations: 2027
Model: GLM Df Residuals: 2018
Model Family: NegativeBinomial Df Model: 8
Link Function: Log Scale: 1.0000
Method: IRLS Log-Likelihood: -13375.
Date: Wed, 30 Jul 2025 Deviance: 623.18
Time: 14:39:35 Pearson chi2: 464.
No. Iterations: 30 Pseudo R-squ. (CS): 0.2132
Covariance Type: HC0
Coefficient estimates
coef std err z P>|z| [0.025 0.975]
Intercept 4.8999 0.054 90.698 0.000 4.794 5.006
C(medIncCat)[T.1.0] -0.0093 0.028 -0.327 0.744 -0.065 0.047
C(medIncCat)[T.2.0] 0.0416 0.035 1.196 0.232 -0.027 0.110
C(edu)[T.1.0] 0.0694 0.045 1.542 0.123 -0.019 0.158
C(edu)[T.2.0] 0.0793 0.049 1.621 0.105 -0.017 0.175
pctWhite 0.0015 0.000 3.957 0.000 0.001 0.002
tract_pop 1.833e-05 1.93e-05 0.951 0.342 -1.95e-05 5.61e-05
shape_area 8.152e-08 1.12e-08 7.258 0.000 5.95e-08 1.04e-07
pct_own_occ 0.0060 0.001 8.451 0.000 0.005 0.007

Housing · Demography

Residential building permits across the U.S.

Using data from the U.S. Census Bureau’s Building Permits Survey and linearly interpolated population estimates, I plotted new residential building permits per 100,000 residents for all 50 states and the District of Columbia from 2004 to 2019.

I’m struck by how broadly permitting began to decline before the housing crisis came to a head in 2008. North Dakota also stands out: its permitting trend looks remarkably similar to the growth in U.S. oil drilling during the oil boom.

Residential building permits per 100,000 residentsHover over a line for state and year. Use the legend to toggle regions.

Additional work

SOGIE vocabulary for AAC users

Specialized fields and subcultures often have large, interconnected vocabularies. It can be hard for newcomers to know which terms are foundational or how ideas relate. I wanted to explore that problem through network analysis, focusing on vocabulary for sexual orientation, gender identity, and expression (SOGIE) in augmentative and alternative communication (AAC).

Starting with LGBTQIA-related terms collected from several advocacy organizations, I constructed a graph using Wikipedia hyperlinks. Each node represents a term, and an edge connects two terms if either Wikipedia article links to the other. The result is a rough map of how these concepts are connected.

Explore the interactive network

Original node-and-edge graph of LGBTQIA vocabulary
Links between LGBTQIA vocabulary terms