Fixed Weather Windows Strip Hidden Bias from Grapevine Phenology Models

Observers visit plants on irregular dates, so the true moment a vine flowers or its berries ripen is known only to fall somewhere between the last visit on which the stage was absent and the first visit on which it was seen. A new methodological study argues that this interval censoring, combined with a subtle statistical trap in how weather variables are constructed, can make warming appear to delay grapevine development when it in fact accelerates it. The work, published in the journal Discover Data, presents a reproducible survival-analysis workflow designed to eliminate that bias, and demonstrates it on citizen-science records of European grapevine collected across the United States.

The central problem the authors identify is endogeneity. Phenological studies routinely summarize antecedent weather as cumulative heat, such as growing degree days, accumulated up to the observed event date. Because heat accumulates non-negatively through the season, a later event date mechanically produces a larger cumulative total, regardless of any real biological relationship between temperature and timing. When such a to-event covariate is then fed into a regression or survival model, the covariate is partly a function of the outcome itself, violating the exogeneity assumption that regression inference relies on. The consequence can be severe: the apparent direction of a temperature effect can be reversed, so that a genuine acceleration under warming shows up in the analysis as a delay.

To quantify how large this distortion can be, the researchers ran a simulation study with one thousand replications. In each replication they generated fifty phenological event times from an accelerated failure time model with a known true effect: a time ratio of 0.85, meaning that a one standard deviation increase in pre-season heat accelerates the event by fifteen percent. They then fit models using either the true exogenous pre-season heat or the endogenous cumulative heat accumulated to the event itself. The exogenous approach recovered the truth almost exactly, with a mean estimated time ratio of 0.851 and minimal error. The endogenous approach produced a mean estimate of 1.237, a forty-five percent upward bias that pointed in the wrong direction entirely, indicating a spurious delay rather than the true acceleration.

The remedy the authors propose is conceptually simple: summarize weather only within fixed pre-season windows that close before the earliest plausible detection. For flowering they used day of year 1 to 120, and for ripening day 1 to 180, requiring at least seventy percent data coverage within each window. Because these windows end before typical events occur, the covariates cannot be mechanically correlated with the outcome. The trade-off is that the window must be chosen carefully: if it closes too late, events can fall inside it and reintroduce partial endogeneity, while a very short window discards potentially relevant antecedent conditions. The paper treats window length as a design choice to be tested, and its sensitivity analyses show the consequences of moving it.

The empirical demonstration rests on the complete USA National Phenology Network holding for cultivated Vitis vinifera, retrieved in a single archive query spanning 2012 to 2025. The query returned 41,274 status records on just seven individual plants at five sites, contributed by sixteen observers, underscoring how sparsely commercial crop species are represented in a network dominated by wild and ornamental plants. Restricting to the two target phenophases, Open flowers and Ripe fruits, left 8,274 records, which collapse to fifty plant-site-year survival units carrying thirty flowering and thirty-seven ripening events. Two of the five sites were each visited only once and contribute no analytical power to the fixed-window models, which therefore rest on three sites in California and New York.

The framework converts the Status and Intensity protocol’s explicit no/yes sequences into interval-censored survival data, treating each plant-site-year as one observation with lower and upper bounds around the true transition date. Years in which the stage was never recorded are handled as right-censored rather than discarded, which matters because first-date regression drops exactly those uninformative-looking years that carry real information about late or absent events. Accelerated failure time models with Weibull and log-logistic distributions express covariate effects as time ratios: values below one indicate acceleration and values above one indicate delay. The pipeline, built in Python with the lifelines library, covers data ingestion, interval construction, quality control, covariate design, model fitting and reporting, and is openly archived with pinned software versions.

Applied to the grapevine data, the corrected models show a clear asymmetry between the two phenophases. Warmer pre-season conditions robustly advanced ripening: the Weibull fit gave a time ratio of 0.864 with a confidence interval from 0.813 to 0.917, meaning each additional standard deviation of pre-season heat, roughly 150 growing degree days in these data, multiplies the median ripening date by 0.86 and advances it by about twenty-nine days. The primary log-logistic specification gave a more modest ratio of 0.962, illustrating how sensitive the magnitude is to distributional assumptions even when the direction is consistent. Flowering, by contrast, showed only weak and statistically uncertain responses, with confidence intervals spanning unity in every specification. Precipitation effects could not be separated from no effect at this sample size, and the single marginally significant precipitation estimate disappeared when cumulative rainfall was replaced by a climatic water balance of precipitation minus reference evapotranspiration.

The diagnostic models using to-event covariates told the opposite story, and that is the point. Aggregating cumulative heat up to the detection date produced time ratios above one for both phenophases, apparently indicating later timing with greater warmth, which the simulation shows to be a mechanical artifact rather than biology. The authors are careful to frame the grapevine application as a methodological demonstration rather than substantive ecological inference. Cultivar is identifiable for only one of the seven plants, forty-three of the fifty analytical units lack cultivar information, most vines are irrigated, and the sample spans wine, table and ornamental grapes with heterogeneous management. Sensitivity analyses across base temperatures from five to twelve degrees Celsius, alternative window lengths, stricter BBCH-equivalent endpoints and a homogeneous single-site subset left every inferential conclusion intact while quantifying how much each choice moves the point estimates.

The study also probes the assumptions that survival analysis requires. Censoring was tested for informativeness, interval widths were characterized, with a median of four days, and hierarchical clustering among sites, individuals and years was assessed through intraclass correlations, revealing moderate site-level clustering for ripening that would roughly double the standard errors under a cluster-aware fit without changing any qualitative conclusion. From these checks the authors distill explicit conditions under which the framework transfers to other species and networks: visit frequency must be high relative to the between-year spread of event timing, right censoring must be moderate enough that parametric medians do not diverge from non-parametric estimates, sample size must support the specification, and the fixed window must close before the earliest detection. Where those conditions hold, the approach offers citizen-science phenology networks a way to extract climate signals that are honest about observation uncertainty and immune to one of the field’s most insidious statistical traps.

Subject of Research: A statistical framework for unbiased interval-censored survival modeling of grapevine phenology using fixed pre-season weather windows

Article Title: Fixed pre-season weather windows eliminate endogeneity bias in interval-censored survival models of grapevine phenology

Article References: Behnamian, S., & Fogh, F. (2026). Fixed pre-season weather windows eliminate endogeneity bias in interval-censored survival models of grapevine phenology. Discover Data, 4(1), Article 18. https://doi.org/10.1007/s44248-026-00118-x

Image Credits: AI Generated

DOI: 10.1007/s44248-026-00118-x

Keywords: grapevine, phenology, interval censoring, survival analysis, endogeneity, growing degree days, USA-NPN, NASA POWER, climate change, accelerated failure time models, citizen science, viticulture