Arbitrary modeling choices are inevitable in scientific studies. Yet, few empirical studies in conservation science report the effects these arbitrary choices have on estimated results. I explored the effects of subjective modeling choices in the context of counterfactual impact evaluations. Over 5000 candidate models based on reasonable changes in the choice of statistical matching algorithms (e.g., genetic and nearest distance mahalanobis matching), the parametrization of these algorithms (e.g., number of matches), and the inclusion of specific covariates (e.g., distance to nearest city, slope, or rainfall) were valid for studying the effect of Virunga National Park in Democratic Republic of the Congo on changes in tree cover loss and carbon storage over time. I randomly picked 2000 of the 5000 candidate models to determine how much and which subjective modeling choices affected the results the most. All valid models indicated that tree cover loss decreased and carbon storage increased in Virunga National Park from 2000 to 2019. Nonetheless, the order of magnitude of the estimates varied by a factor of 3 (from −4.78 to −13.12 percentage points decrease in tree cover loss and from 20 to 46 t Ce/ha for carbon storage). My results highlight that modeling choices, notably the choice of the matching algorithm, can have significant effects on point estimates and suggest that more structured robustness checks are a key step toward more credible findings in conservation science.

中文翻译：

---

主观建模选择和保护科学影响评估的稳健性

在科学研究中，任意的建模选择是不可避免的。然而，很少有保护科学的实证研究报告这些任意选择对估计结果的影响。我在反事实影响评估的背景下探索了主观建模选择的影响。超过 5000 个候选模型基于统计匹配算法的选择（例如遗传和最近距离马氏匹配）、这些算法的参数化（例如匹配数）以及特定协变量的包含（例如到最近的城市、斜坡或降雨）对于研究刚果民主共和国维龙加国家公园对树木覆盖损失和碳储存随时间变化的影响是有效的。我从 5000 个候选模型中随机挑选了 2000 个，以确定对结果影响最大的主观建模选择和影响。所有有效模型都表明，从 2000 年到 2019 年，维龙加国家公园的树木覆盖损失减少，碳储量增加。 尽管如此，估计的数量级相差 3 倍（树木覆盖减少从 -4.78 到 -13.12 个百分点）损失和 20 至 46 吨 Ce/ha 用于碳储存）。我的结果强调，建模选择，尤其是匹配算法的选择，可以对点估计产生重大影响，并表明更结构化的稳健性检查是在保护科学中获得更可信发现的关键一步。所有有效模型都表明，从 2000 年到 2019 年，维龙加国家公园的树木覆盖损失减少，碳储量增加。 尽管如此，估计的数量级相差 3 倍（树木覆盖减少从 -4.78 到 -13.12 个百分点）损失和 20 至 46 吨 Ce/ha 用于碳储存）。我的结果强调，建模选择，尤其是匹配算法的选择，可以对点估计产生重大影响，并表明更结构化的稳健性检查是在保护科学中获得更可信发现的关键一步。所有有效模型都表明，从 2000 年到 2019 年，维龙加国家公园的树木覆盖损失减少，碳储量增加。 尽管如此，估计的数量级相差 3 倍（树木覆盖减少从 -4.78 到 -13.12 个百分点）损失和 20 至 46 吨 Ce/ha 用于碳储存）。我的结果强调，建模选择，尤其是匹配算法的选择，可以对点估计产生重大影响，并表明更结构化的稳健性检查是在保护科学中获得更可信发现的关键一步。