📄 earth_science

The impact of data missingness mechanisms and degradation levels on rainfall agroclimatic data analyses

This study utilizes Monte Carlo simulations to demonstrate that missing not at random (MNAR) mechanisms and higher missingness volumes (up to 20%) significantly bias agroclimatic rainfall indices, with Maximum Consecutive Dry Days (CDD) proving most sensitive and dry regions least affected compared to wetter areas.

Original authors: Tinetariro Wesley Chikati, Moven Manjowe, Tapiwa Michelle Masawi

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tinetariro Wesley Chikati, Moven Manjowe, Tapiwa Michelle Masawi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to bake the perfect cake, but you only have a recipe that is missing a few pages. If you guess the missing ingredients, you might end up with a delicious treat, or you might serve a brick. This is the daily reality for scientists who study weather and farming. They rely on "rainfall data"—a long, continuous diary of how much rain falls every day—to predict when crops will grow, when droughts might hit, and how to feed millions of people. But in the real world, these weather diaries are often torn. Stations break, batteries die, or storms knock out equipment, leaving big holes in the record. The big question is: does it matter how the pages go missing? If the missing days are just random accidents, is that different from missing days that happen specifically during the heaviest storms? If we try to fill in those gaps with math, will our cake still taste right, or will the whole recipe collapse?

This study, conducted by researchers from Zimbabwe, dives into this messy problem by treating rainfall data like a game of "fill in the blanks." They took a perfect, complete record of rain and deliberately tore holes in it to see what happens. They tested three different ways data can go missing: "Missing Completely at Random" (like a random page falling out of a book), "Missing at Random" (where the missing data depends on values that were observed, but not on the missing values themselves), and "Missing Not at Random" (the worst case, where the missing pages are the ones with the most important rain, like the heavy storms). They also tested how bad it gets when they remove 5%, 10%, or 20% of the data.

The researchers didn't just guess; they ran a massive computer simulation (a "Monte Carlo" experiment) over 2,000 times to see how these missing pieces messed up five specific weather measurements: the longest consecutive wet days (days with at least 3mm of rain), the longest consecutive dry days (days with less than 3mm of rain), the total rain for the season, when the rainy season starts, and when it ends. They found that the way data disappears matters a huge amount. When the missing data was "Not at Random" (hiding the heavy rain), the results were the most distorted, causing the biggest errors. Surprisingly, the measurements for when the season starts and stops were tough cookies; they stayed mostly accurate even when data was missing. However, the measurement for the longest consecutive dry days was a total disaster. In dry areas, even a small amount of missing data made the calculation of dry spells go haywire, sometimes throwing off the results by nearly 100%.

The study also revealed a funny twist: wetter regions with lots of rain suffered bigger errors in their total rainfall numbers when data went missing, simply because there was more "stuff" to lose. But in the driest parts of the country, the total rain numbers stayed surprisingly steady because there wasn't much rain to begin with. However, those dry regions paid a different price: their ability to track dry spells completely fell apart. The researchers concluded that while we can't always stop weather stations from breaking, we need to be very careful about how we fill in the gaps. If we assume missing data is random when it's actually hiding the storms, our predictions for farmers could be dangerously wrong, potentially leading to poor planning and reduced harvests. The study suggests that keeping our weather instruments running is the best recipe of all, because no amount of math can perfectly fix a torn diary.

Technical Summary: The Impact of Data Missingness Mechanisms and Degradation Levels on Rainfall Agroclimatic Data Analyses

Problem Statement
Reliable weather data is the foundation for drought monitoring, hydrological modeling, and agricultural decision-making, particularly in agriculture-dependent economies like Zimbabwe. While ground station data is often treated as error-free, it frequently suffers from spatial gaps and temporal discontinuities due to transmission failures, instrument malfunctions, and human error. A critical gap exists in the systematic evaluation of how different data degradation structures—specifically Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR)—and varying volumes of missingness (5%, 10%, 20%) affect the accuracy of derived agroclimatic indices. Many existing studies and gap-filling methodologies implicitly assume randomness in missing data, potentially misrepresenting reality and introducing significant uncertainty into climate decision support systems.

Methodology
To address this, the authors conducted a controlled Monte Carlo Simulation (MCS) using daily rainfall data from 1981 to 2025 derived from the CHIRPS-V2 satellite dataset, validated against ground station coordinates in Zimbabwe. The study focused on seven stations, each representing a unique Agro-Ecological Zone (AEZ) ranging from high-rainfall Region I to arid Region Vb.

The methodology involved:

Data Degradation: Artificially degrading the complete time series by introducing missing data according to three mechanisms (MCAR, MAR, MNAR) at three volume levels (5%, 10%, 20%).
Simulation Scale: The simulation ran 2,000 iterations per scenario to ensure statistical stability, utilizing a pseudo-random number generator seeded with the value 42 to ensure reproducibility.
Target Metrics: Five key seasonal agroclimatic indices were calculated for both wet and dry seasons:
- Maximum Consecutive Dry Days (CDD)
- Maximum Consecutive Wet Days (CWD)
- Seasonal Total Rainfall
- Season Onset
- Season Cessation
Error Assessment: The study compared degraded time series against the complete reference using robust error metrics suitable for non-normally distributed rainfall data: Median Bias (MdB), Normalised Median Absolute Deviation (NMAD), and Spearman's rank correlation coefficient ( $\rho$ ).

Key Results
The study yielded several distinct findings regarding the sensitivity of agroclimatic indices to data degradation:

Mechanism Sensitivity: The MNAR mechanism consistently caused the most severe bias across all indices, while MAR resulted in the least severe bias. For instance, under 20% MNAR, Seasonal Total Rainfall showed a median bias of -244.3 mm yr⁻¹, whereas MAR showed significantly lower deviations.
Volume Scaling: Uncertainty and bias in the indices increased disproportionately as the volume of missingness scaled upward. A threshold effect was observed where deviations became severe at 20% missingness.
Index Resilience:
- Most Resilient: Season Onset and Season Cessation were the most robust to missingness. Onset maintained an NMAD of 0% across all volumes, and Cessation peaked at only 4% under the worst-case MNAR scenario.
- Most Sensitive: CDD was the most sensitive metric. In the dry season, CDD variance became catastrophic, with NMAD reaching up to 92% across all scenarios. This indicates that missing data during dry periods severely destabilizes the calculation of consecutive dry spells.
Spatial Variability:
- Wet vs. Dry Regions: Wetter regions (e.g., Chipinge, Region I) experienced the highest absolute bias in Seasonal Total Rainfall due to the larger volume of rainfall at risk of being lost. Conversely, drier regions (e.g., Beitbridge, Region Vb) showed minimal bias in total rainfall totals.
- Continuity Metrics: Paradoxically, drier regions suffered more significant degradation in continuity metrics (like CDD and temporal correlation). In arid stations, even 5% missingness caused a collapse in the temporal integrity of CDD calculations, as gaps made it impossible to distinguish between dry and wet spells.

Significance and Claims
The paper claims that the magnitude of errors introduced by data missingness is highly sensitive to the local climate baseline and the specific missingness mechanism, rather than just the volume of missing data. The authors assert that:

MNAR is the most critical threat: Assuming randomness (MCAR/MAR) when data is actually MNAR leads to the most severe underestimation of seasonal totals and mischaracterization of drought/wet spells.
Context matters: While low baseline rainfall acts as a "natural ceiling" limiting the bias in total rainfall sums, it simultaneously makes continuity metrics like CDD extremely fragile.
Operational Implications: The findings suggest that current data completeness thresholds (e.g., the 10% missing data limit for the PICSA program) may need re-evaluation, as even low volumes of missingness in specific contexts (like dry seasons in arid regions) can render continuity indices unreliable.

The study concludes that maintaining continuous operation of observation instruments is paramount. It emphasizes that while imputation methods are useful, the underlying pattern of missingness remains the primary driver of error, and the results should be viewed as an illustrative quantification of these risks rather than a definitive characterization of all possible scenarios due to the computational limits of finite simulations.

Technical Summary: The Impact of Data Missingness Mechanisms and Degradation Levels on Rainfall Agroclimatic Data Analyses

More like this