An official website of the United States government
The conditional average treatment effect (CATE) is the best measure of individual causal effects given baseline covariates. However, the CATE only captures the (conditional) average, and can overlook risks and tail events, which are important to treatment choice. In aggregate analyses, this is usually addressed by measuring the distributional treatment effect (DTE), such as differences in quantiles or tail expectations between treatment groups. Hypothetically, one can similarly fit conditional quantile regressions in each treatment group and take their difference, but this would not be robust to misspecification or provide agnostic best-in-class predictions. We provide a new robust and model-agnostic methodology for learning the conditional DTE (CDTE) for a class of problems that includes conditional quantile treatment effects, conditional super-quantile treatment effects, and conditional treatment effects on coherent risk measures given by f-divergences. Our method is based on constructing a special pseudo-outcome and regressing it on covariates using any regression learner. Our method is model-agnostic in that it can provide the best projection of CDTE onto the regression model class. Our method is robust in that even if we learn these nuisances nonparametrically at very slow rates, we can still learn CDTEs at rates that depend on the class complexity and even conduct inferences on linear projections of CDTEs. We investigate the behavior of our proposal in simulations, as well as in a case study of 401(k) eligibility effects on wealth.
more »
« less
This content will become publicly available on May 12, 2027
Multi-quantile regression for extreme precipitation downscaling
Deep super-resolution networks for precipitation downscaling achieve strong bulk skill yet systematically under-predict the heavy-tail events that drive flood risk. We demonstrate that the primary obstacle is the loss function, not the data: under intensity-weighted MAE, real and synthetic labels at the same input are simply averaged, meaning data augmentation shifts the predicted mean rather than the conditional distribution. We resolve this with Q-SRDRN, a multi-quantile super-resolution network trained with pinball loss at tau in 0.50, 0.95, 0.99, 0.999. Two CNN-specific design choices make this practical: IncrementBound enforces monotonicity while preserving each quantile channel's gradient identity, and separate per-quantile output heads provide independent filter banks for bulk and tail detection. Under this design, data augmentation via cVAE becomes complementary: the median head absorbs synthetic patterns without contaminating upper quantiles. Empirically, on Florida (convective/tropical-cyclone dominated), the un-augmented Q-SRDRN P999 head detects 1,598 of 2,111 events at 200 mm/day versus 88 for the deterministic baseline--an 18x detection-rate gain (4.2% to 75.7%)--with 63% lower KL divergence and 3.9% lower RMSE. Adding cVAE-generated samples lifts the P50 channel from 14 to 1,038 hits at 200 mm/day. On California (atmospheric-river dominated), the architecture reaches near-perfect detection (P999 SEDI >= 0.996 through 300 mm/day). On Texas, the baseline catches only 2 of 10,720 events at 200 mm/day while the P999 head catches 8,776 (81.9%). While the cVAE does not transfer across regions, multi-quantile regression captures extremes wherever the large-scale signal is strong, while augmentation rescues the median where it is not.
more »
« less
- Award ID(s):
- 2331908
- PAR ID:
- 10703596
- Publisher / Repository:
- arXiv preprint: https://doi.org/10.48550/arXiv.2605.12762
- Date Published:
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Bardanis, M. (Ed.)It is unlikely to predict the distribution of soil suction in the field deterministically. It is well established that there are various sources of uncertainty in the measurement of matric suction, and the suction measurements in the field are even more critical because of the heterogeneities in the field conditions. Hence it becomes necessary to probabilistically characterize the suction in the field for enhanced reliability. The objective of this study was to conduct a probabilistic analysis of measured soil suction of two different test landfill covers, compacted clay cover (CC) and engineered turf cover (ETC), under similar meteorological events. The size of the two test landfill covers was 3 m × 3 m (10 ft. × 10 ft.) and 1.2 m (4ft.) in depth. The covers were constructed by excavating the existing subgrade, placing 6-mil plastic sheets, and backfilling the excavated soil, followed by layered compaction. Then the covers were instrumented identically with soil water potential sensors up to specified depths. One of the covers acted as the CC, and the other cover was ETC. In ETC, engineered turf was laid over the compacted soil. The engineered turf consisted of a structured LLDPE geomembrane overlain by synthetic turf (polyethylene fibers tufted through a double layer of woven polypropylene geotextiles). The sensors were connected to an automated data logging system and the collected data were probabilistically analyzed using the R program. There were significant inconsistencies in the descriptive statistical parameters of the measured soil suction at both covers under the same climatic conditions. Soil suction measured in the field ranged between almost 12 to 44 kPa in ETC, while it was in the range of almost 1 to 2020 kPa in the CC. The histogram and quantile-quantile (Q-Q) plot showed the data to be non-normally distributed in the field. A heavy-tailed leptokurtic (Kurtosis=13) distribution of suction was observed in the ETC with substantial outliers. In contrast, the suction distribution in CC was observed skewed to the right containing a thinner tail indicating an almost platykurtic distribution. The distribution of suction in the field under engineered turf was observed to be reasonably consistent with time compared to bare soil under the same meteorological events. The results obtained from this study revealed the engineered turf system to be an effective barrier to inducing changes in soil suction against climatic events.more » « less
-
Abstract. Observations of recent mass loss rates of the West Antarctic Ice Sheet (WAIS) raise concerns about its stability since a collapse would increase global sea levels by several meters. Future projections of these mass loss trends are often estimated using numerical ice sheet models, and recent studies have highlighted the need for models to be benchmarked against present-day observed mass change rates. Here, we present an improved initialization method that optimizes local agreement not only with observations of ice thickness and surface velocity but also with satellite-based estimates of mass change rates. This is achieved by a combination of tuned thermal forcing under the floating ice shelves and friction under the ice sheet. Starting from this improved present-day state, we generate an ensemble of future simulations of Antarctic mass change by varying model physical choices and parameter values while fixing the climate forcing at present-day values. The dynamical response shows slow grounding-line retreat over several centuries, followed by a phase of rapid mass loss over about 200 years with a consistent rate of ∼3 mm GMSL yr−1 (global mean sea level). We find that, for all ensemble members, the Thwaites Glacier and Pine Island Glacier collapse. Our results imply that present-day ocean thermal forcing, if held constant over multiple centuries, may be sufficient to deglaciate large parts of the WAIS, raising global mean sea level by at least a meter.more » « less
-
Abstract Global warming contributes to changes in local temperature and precipitation distributions. These shifts alter key statistical properties such as the mean, variance, and frequency of severe climate events, making accurate predictions more complex. In this study, we analyze temperature and precipitation data from the National Oceanic and Atmospheric Administration’s (NOAA) Twentieth Century Reanalysis Project, covering the period from 1950 to 2015 across the continental United States. We employ the expected shortfall (ES) regression method to examine trends in temperature and precipitation relative to time and El Niño–Southern Oscillation (ENSO). Unlike quantile regression, which focuses on specific conditional quantiles such as the median, ES provides insights into averages beyond a certain quantile, offering a more comprehensive description of the tails. Our findings reveal notable differences between ES and mean and quantile regression results, highlighting distinct spatial and temporal trends on interannual-to-decadal time scales. Specifically, ES regression captures significant temperature increases over the decadal time scale, especially in the upper tails, across the southern and central United States, which supports the disappearance of “global warming hole” over the recent decades. Additionally, ENSO events, particularly El Niño, are found to intensify cooling effects in the Gulf Coast region during the winter, particularly for upper tails. The findings demonstrate the usefulness of expected shortfall regression in estimating temperature and precipitation distributions and characterizing changes in the tails. Significance StatementThis study employs expected shortfall (ES) regression to analyze the tails of temperature and precipitation distributions across the United States, focusing on how these tail behaviors vary over time, especially in relation to El Niño–Southern Oscillation (ENSO). ES regression provides insights into the events beyond high quantile levels, such as intense heat waves and heavy rainfall, which are critical for understanding the risks associated with climate change. The findings underscore the importance of incorporating tail behavior when analyzing climate variability, which has direct implications for policy decisions and societal adaptation strategies in the face of global warming.more » « less
-
Abstract. We report on fall speed measurements of raindrops in light-to-heavyrain events from two climatically different regimes (Greeley,Colorado, and Huntsville, Alabama) using the high-resolution(50 µm) Meteorological Particle Spectrometer (MPS) anda third-generation (170 µm resolution) 2-D videodisdrometer (2DVD). To mitigate wind effects, especially for thesmall drops, both instruments were installed within a 2∕3-scaleDouble Fence Intercomparison Reference (DFIR) enclosure. Two casesinvolved light-to-moderate wind speeds/gusts while the third casewas a tornadic supercell and several squall lines that passed overthe site with high wind speeds/gusts. As a proxy for turbulentintensity, maximum wind speeds from 10 m height at theinstrumented site recorded every 3 s were differenced withthe 5 min average wind speeds and then squared. The fall speedsvs. size from 0.1 to 2 and >0.7 mm were derived from theMPS and the 2DVD, respectively. Consistency of fall speeds from thetwo instruments in the overlap region (0.7–2 mm) gaveconfidence in the data quality and processing methodologies. Ourresults indicate that under low turbulence, the mean fall speedsagree well with fits to the terminal velocity measured in thelaboratory by Gunn and Kinzer from 100 µm up toprecipitation sizes. The histograms of fall speeds for 0.5, 0.7, 1and 1.5 mm sizes were examined in detail under the sameconditions. The histogram shapes for the 1 and 1.5 mm sizeswere symmetric and in good agreement between the two instrumentswith no evidence of skewness or of sub- or super-terminal fallspeeds. The histograms of the smaller 0.5 and 0.7 mm dropsfrom MPS, while generally symmetric, showed that occasionaloccurrences of sub- and super-terminal fall speeds could not beruled out. In the supercell case, the very strong gusts andinferred high turbulence intensity caused a significant broadeningof the fall speed distributions with negative skewness (for drops of1.3, 2 and 3 mm). The mean fall speeds were also found todecrease nearly linearly with increasing turbulent intensityattaining values about 25–30 % less than the terminalvelocity of Gunn–Kinzer, i.e., sub-terminal fall speeds.more » « less
