📊 statistics

Three inference traps in exploratory space–time analysis: evidence from population ageing in Türkiye

This paper identifies and provides diagnostics for three common inference traps in exploratory space–time analysis—specifically regarding the misinterpretation of local statistics, the conflation of temporal trends with spatial clustering in hot spot analysis, and the misattribution of class-position effects to spatial contagion—demonstrating through Türkiye's population ageing data that observed spatial patterns stem from shared regional forcing rather than neighbor-to-neighbor diffusion.

Original authors: Kerem Yavuz ARSLANLI

Published 2026-10-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kerem Yavuz ARSLANLI

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 looking at a map of a country and seeing how a specific condition, like the proportion of older people, spreads across the landscape over time. Scientists have long used special tools to draw these maps, hoping to spot where the condition is gathering in tight groups, where it is suddenly appearing in new places, or whether it is jumping from one town to its neighbor like a contagion. These tools are designed to separate random noise from real patterns, helping researchers understand if a trend is driven by shared regional forces, such as local economies or migration, or if it is actually spreading from person to place to place. However, just as a camera can distort an image if the lens is dirty or the settings are wrong, these statistical maps can sometimes show a picture that looks convincing but is actually misleading. The way the data is processed can create the illusion of a pattern where none exists, or hide a real one behind a wall of technical confusion.

A recent study by Kerem Yavuz Arslanli at Istanbul Technical University takes a close look at three specific ways these maps can trick researchers. The study uses the real-world example of population aging across the districts of Türkiye between 2007 and 2025. Over this period, the country saw a rapid shift where the population grew older, but this shift happened at different speeds in different regions, with some areas aging quickly while others remained younger. The researcher applied three common methods used to analyze this kind of data to see if the conclusions drawn from them were actually correct. What he found was that three very common practices in the field were leading to false conclusions about how aging was spreading, and he offered a simple way to fix each one.

The first trap involves comparing two different types of maps that are often used side-by-side. One map highlights areas where high values cluster together, while the other highlights areas where low values cluster together. Researchers often look at these two maps and assume that if they show different spots as important, it means the two methods are seeing different things. Arslanli showed that this disagreement is usually an illusion caused by how the computer calculates the results. When both maps are calculated using the exact same method for testing their reliability, they actually point to the exact same locations. The only difference is in the labels they give to a few unusual spots. In the case of Türkiye, once the calculation method was standardized, the two maps agreed on nearly every single district, proving that the apparent conflict was just a technical artifact, not a real difference in the data.

The second trap occurs when researchers try to track how hot spots of a condition emerge over time. This is often done by looking at a "space-time cube," which stacks maps of different years on top of each other to see how the pattern moves. The problem arises when the condition being studied is changing rapidly across the entire country at the same time. In Türkiye, the number of older people increased steadily everywhere. When the researchers ran the standard analysis on the raw numbers, the computer flagged almost half of the districts as "new" or "consecutive" hot spots simply because the national average had risen. It looked like aging was exploding in new places, but it was actually just the whole country getting older at once. When the researchers adjusted the data to remove this national trend and looked only at which districts were aging faster than their neighbors, the picture changed completely. The number of "new" hot spots dropped dramatically, revealing that the real story was not a sudden explosion of aging, but a steady intensification of existing clusters in the north and a genuine weakening of youthfulness in the southeast.

The third and perhaps most subtle trap involves a method used to see if neighbors influence each other. This technique checks if a district that is surrounded by older neighbors is more likely to age faster itself, which would suggest a contagious spread of aging. The standard test suggested that neighbors did indeed influence each other, with older neighbors making a district age faster. However, Arslanli realized that this result was misleading because it did not account for where a district stood within its own age group. A district that is already quite old within its own category and has older neighbors is simply closer to the next level of aging, so it is statistically more likely to move up. When the analysis was corrected to account for a district's own starting age, the influence of the neighbors flipped. Instead of speeding up aging, having older neighbors was actually associated with a slightly slower rate of aging. This suggests that aging in Türkiye is not spreading from neighbor to neighbor like a virus, but is instead driven by shared regional factors that affect entire areas at once, causing them to age together rather than infecting one another.

The study concludes that while population aging in Türkiye is indeed strongly clustered and getting more intense, the evidence does not support the idea that it is spreading contagiously between districts. The patterns seen on the maps are real, but the interpretation of how they formed was often wrong due to these three technical pitfalls. By fixing the way the data is tested, the way trends are removed, and the way neighbor effects are measured, the study provides a clearer, more accurate picture of the demographic shifts reshaping the country. It serves as a reminder that in the world of data analysis, the tools we use to see the world can sometimes distort what we see, and that checking the settings of those tools is just as important as the data itself.

Technical Summary: Three Inference Traps in Exploratory Space–Time Analysis

Problem Statement
Exploratory local statistics, including Local Indicators of Spatial Association (LISA), Getis–Ord $G_i^*$ , Emerging Hot Spot Analysis (EHSA), and Spatial Markov chains, are standard tools in regional analysis. However, the paper argues that three common practices in combining these methods can lead to misleading inferences regarding spatial clustering, the emergence of hot spots, and the diffusion of processes. Specifically, researchers often misinterpret discrepancies between statistics, attribute temporal trends to spatial emergence, and conflate spatial sorting with contagion effects. The paper uses the diffusion of population ageing across Türkiye's districts (2007–2025) as a test case to document these traps and provide diagnostics for each.

Methodology
The study utilizes an annual, register-based panel of Türkiye's districts (harmonized to 848 consistent units) from 2007 to 2025. The primary indicator is the ageing index ( $AI = 100 \times P_{65+} / P_{0-14}$ ). The analysis employs three distinct methodological investigations:

Trap 1 (Inference Regimes): The paper compares $G_i^*$ and Local Moran's $I$ ( $LISA$ ) under three regimes: (i) conventional mixed regimes (analytical $G_i^*$ vs. permutation $LISA$ ), (ii) mixed regimes with one-sided analytical $G_i^*$ , and (iii) a common regime using 99,999 conditional permutations for both statistics. Theoretical propositions are derived to show that under conditional permutation with identical neighbor sets and folded pseudo-p-values, the statistics must yield identical significance rankings.
Trap 2 (Trending Variables in EHSA): Emerging Hot Spot Analysis is applied to the raw ageing index and to an index standardized within each year (removing the common national trend). The study compares the resulting classifications against a static 2025 $G_i^*$ map to determine if "emerging" categories reflect spatial clustering or merely the national trend.
Trap 3 (Spatial Markov Chains): A spatial Markov chain is used to test for spatial independence in transitions between quintiles of the ageing index. To distinguish between diffusion (neighbour effects) and spatial sorting (position within a class), the study employs a diagnostic logit model and growth regression. These models condition the probability of upward transitions on both the unit's own level and its spatial lag, controlling for fixed effects.

Robustness checks are performed across alternative spatial weights (queen contiguity, k-nearest neighbors, distance bands), sub-periods, population thresholds, and the exclusion of earthquake-affected provinces.

Key Contributions and Results

Trap 1: The Illusion of Disagreement.
The paper demonstrates that apparent disagreements between $G_i^*$ and $LISA$ maps often stem from mixing inference regimes (analytical approximations vs. permutation) and test sidedness (two-sided vs. one-sided/folded). Under a common regime of conditional permutation with folded pseudo-p-values, the two statistics flag identical locations; all 969 pseudo-p-values coincide. The only remaining difference is in the labeling of spatial outliers: $LISA$ identifies them based on Moran scatterplot quadrants, while $G_i^*$ labels them based on the level of the neighborhood including the focal unit. The study concludes that differences in significance are artifacts of implementation, not statistical properties.
Trap 2: Trend vs. Spatial Clustering in EHSA.
Applying EHSA to a variable with a strong common trend (the raw ageing index) produces "consecutive" and "new" hot spots that largely record the national trend rather than the emergence of spatial clusters. The share of significant hot-spot bins rises from 0% to 45% over the period, with low agreement ( $\kappa = 0.25$ ) with static clustering maps. When the index is standardized within each year to remove the common trend, the share of hot spots stabilizes at ~19%, and agreement with static maps improves significantly ( $\kappa = 0.54$ ). The corrected analysis reveals that while the relative youthfulness of the south-east is genuinely weakening, the "emergence" of hot spots in the raw data was an artifact of the national ageing trajectory.
Trap 3: Spatial Sorting vs. Contagion.
The spatial Markov chain decisively rejects spatial independence ( $LR = 269.6, p < 0.001$ ), a result often interpreted as evidence of diffusion. However, the paper shows this significance reflects the spatial sorting of units within discrete classes. Units with older neighbors tend to lie higher within their own discrete class (closer to the transition threshold). Once the unit's own level is controlled for in a logit model, the neighbor effect reverses: units surrounded by older districts actually experience slower upward transitions. This suggests the significant test result is driven by the spatial autocorrelation of the variable's level, not by contagion accelerating the process.

Significance and Claims
The paper claims that the evidence for population ageing in Türkiye is consistent with strong spatial clustering and intensification, driven by shared regional forcing (fertility regimes, migration fields, labor markets) rather than contagion between neighboring districts. The study does not claim to offer new causal mechanisms for ageing but rather provides a set of inferential diagnostics to prevent misinterpretation of standard exploratory tools.

The author asserts that:

Comparisons of local statistics must fix the inference regime and test sidedness; under conditional permutation, $G_i^*$ and $LISA$ are functionally equivalent for flagging locations.
EHSA applied to trending variables requires standardization within time periods to distinguish spatial emergence from temporal trends.
Significant spatial Markov tests do not automatically imply diffusion; a within-class diagnostic is necessary to rule out the effect of a unit's position within its discrete class.

The paper concludes that while district ageing in Türkiye is strongly clustered, the evidence supports convergence within regions and shared regional drivers rather than inter-district contagion. The analysis serves as a corrective to descriptive mapping by adding rigorous inferential checks.

More like this