Bias in random forest variable importance measures: illustrations, sources and a solution

Towards more accurate classification of risk of arrest among offenders on community supervision: An application of machine learning versus logistic regression

BACKGROUND

Although there is general consensus about the behavioural, clinical and sociodemographic variables that are risk factors for reoffending, optimal statistical modelling of these variables is less clear. Machine learning techniques offer an approach that may provide greater accuracy than traditional methods.

AIM

To compare the performance of advanced machine learning techniques (classification trees and random forests) to logistic regression in classifying correlates of rearrest among adult probationers and parolees in the United States.

METHOD

Data were from the subgroup of people on probation or parole who had taken part in the National Survey on Drug Use and Health for the years 2015-2019. We compared the performance of logistic regression, classification trees and random forests, using receiver operating characteristic curves, to examine the correlates of arrest within the past 12 months.

RESULTS

We found that machine learning techniques, specifically random forests, possessed significantly greater accuracy than logistic regression in classifying correlates of arrest.

CONCLUSIONS

Our findings suggest the potential for enhanced risk classification. The next step would be to develop applications for criminal justice and clinical practice to inform better support and management strategies for former offenders in the community.

Significant increase in natural disturbance impacts on European forests since 1950

Over the last decades, the natural disturbance is increasingly putting pressure on European forests. Shifts in disturbance regimes may compromise forest functioning and the continuous provisioning of ecosystem services to society, including their climate change mitigation potential. Although forests are central to many European policies, we lack the long-term empirical data needed for thoroughly understanding disturbance dynamics, modeling them, and developing adaptive management strategies. Here, we present a unique database of >170,000 records of ground-based natural disturbance observations in European forests from 1950 to 2019. Reported data confirm a significant increase in forest disturbance in 34 European countries, causing on an average of 43.8 million m³ of disturbed timber volume per year over the 70-year study period. This value is likely a conservative estimate due to under-reporting, especially of small-scale disturbances. We used machine learning techniques for assessing the magnitude of unreported disturbances, which are estimated to be between 8.6 and 18.3 million m³ /year. In the last 20 years, disturbances on average accounted for 16% of the mean annual harvest in Europe. Wind was the most important disturbance agent over the study period (46% of total damage), followed by fire (24%) and bark beetles (17%). Bark beetle disturbance doubled its share of the total damage in the last 20 years. Forest disturbances can profoundly impact ecosystem services (e.g., climate change mitigation), affect regional forest resource provisioning and consequently disrupt long-term management planning objectives and timber markets. We conclude that adaptation to changing disturbance regimes must be placed at the core of the European forest management and policy debate. Furthermore, a coherent and homogeneous monitoring system of natural disturbances is urgently needed in Europe, to better observe and respond to the ongoing changes in forest disturbance regimes.

Spatiotemporal influences of climate and humans on muskox range dynamics over multiple millennia

Processes leading to range contractions and population declines of Arctic megafauna during the late Pleistocene and early Holocene are uncertain, with intense debate on the roles of human hunting, climatic change, and their synergy. Obstacles to a resolution have included an overreliance on correlative rather than process-explicit approaches for inferring drivers of distributional and demographic change. Here, we disentangle the ecological mechanisms and threats that were integral in the decline and extinction of the muskox (Ovibos moschatus) in Eurasia and in its expansion in North America using process-explicit macroecological models. The approach integrates modern and fossil occurrence records, ancient DNA, spatiotemporal reconstructions of past climatic change, species-specific population ecology, and the growth and spread of anatomically modern humans. We show that accurately reconstructing inferences of past demographic changes for muskox over the last 21,000 years require high dispersal abilities, large maximum densities, and a small Allee effect. Analyses of validated process-explicit projections indicate that climatic change was the primary driver of muskox distribution shifts and demographic changes across its previously extensive (circumpolar) range, with populations responding negatively to rapid warming events. Regional analyses show that the range collapse and extinction of the muskox in Europe (~13,000 years ago) was likely caused by humans operating in synergy with climatic warming. In Canada and Greenland, climatic change and human activities probably combined to drive recent population sizes. The impact of past climatic change on the range and extinction dynamics of muskox during the Pleistocene-Holocene transition signals a vulnerability of this species to future increased warming. By better establishing the ecological processes that shaped the distribution of the muskox through space and time, we show that process-explicit macroecological models have important applications for the future conservation and management of this iconic species in a warming Arctic.

Predicting disease activity in patients with multiple sclerosis: An explainable machine‐learning approach in the Mavenclad trials

Multiple sclerosis (MS) is among the most common autoimmune disabling neurological conditions of young adults and affects more than 2.3 million people worldwide. Predicting future disease activity in patients with MS based on their pathophysiology and current treatment is pivotal to orientate future treatment. In this respect, we used machine learning to predict disease activity status in patients with MS and identify the most predictive covariates of this activity. The analysis is conducted on a pooled population of 1935 patients enrolled in three cladribine tablets clinical trials with different outcomes: relapsing–remitting MS (from CLARITY and CLARITY‐Extension trials) and patients experiencing a first demyelinating event (from the ORACLE‐MS trial). We applied gradient‐boosting (from XgBoost library) and Shapley Additive Explanations (SHAP) methods to identify patients' covariates that predict disease activity 3 and 6 months before their clinical observation, including patient baseline characteristics, longitudinal magnetic resonance imaging readouts, and neurological and laboratory measures. The most predictive covariates for early identification of disease activity in patients were found to be treatment duration, higher number of new combined unique active lesion count, higher number of new T1 hypointense black holes, and higher age‐related MS severity score. The outcome of this analysis improves our understanding of the mechanism of onset of disease activity in patients with MS by allowing their early identification in clinical settings and prompting preventive measures, therapeutic interventions, or more frequent patient monitoring.

Information-based summary statistics for spatial genetic structure inference

The measurement of biodiversity at all levels of organisation is an essential first step to understand the ecological and evolutionary processes that drive spatial patterns of biodiversity. Ecologists have explored the use of a large range of different summary statistics and have come to the view that information-based summary statistics, and in particular so-called Hill numbers, are a useful tool to measure biodiversity. Population geneticists, on the other hand, have largely focused on summary statistics based on heterozygosity and allelic richness measures. However, recent studies proposed the adoption of information-based summary statistics in population genetics studies. Here, we performed a comprehensive assessment of the power of this family of summary statistics to inform about genetic diversity spatial patterns and we compared it with that of traditional population genetics approaches, namely measures based on allelic richness and heterozygosity. To give an unbiased evaluation, we used three machine learning methods to test the performance of different sets of summary statistics to discriminate between spatial scenarios. We defined three distinct sets, (a) one based on allelic richness measures which included the Jaccard index, (b) a set based on heterozygosity that included F_ST , and (c) a set based on Hill numbers derived from Shannon entropy, which included the recently proposed Shannon differentiation, ΔD. The results showed that the latter performed as well or, under some specific spatial scenarios, even better than the traditional population genetics measures. Interestingly, we found that a rarely or never used genetic differentiation measure based on allelic richness, Jaccard dissimilarity (J), showed the highest discriminatory power to discriminate among spatial scenarios, followed by Shannon differentiation ΔD. We concluded, therefore, that information-based measures as well as Jaccard dissimilarity, represent excellent additions to the population genetics toolkit.