Search

Numerous environmental drivers (e.g., temperature, salinity) have been proposed as controls to recruitment. However, these relationships often become discredited over time as further data become available. It remains unclear if these environment-recruitment relationships are spurious results of examining many variables or if environmental drivers' influences on recruitment change over time as other variables exert control, e.g., with shifting climate. We propose that it is of value to re-examine discredited environment-recruitment relationships using the poor-recruitment paradigm (Gross et al. 2022, Fish. Res. 252, 106329). This approach examines whether extreme environmental conditions are associated with poor recruitment, the idea being that non-extreme conditions are uninformative in terms of predicting recruitment. This allows one to detect patterns of poor recruitment regardless of an environment-recruitment relationship's perceived (lack of) significance. We apply the poor-recruitment paradigm approach to various stocks of Atlantic cod (Gadus morhua) in the North Atlantic for both historical environment-recruitment data from a meta-analysis (Myers 1998, Rev. Fish Biol. Fisher. 8, 285) and for recent environment-recruitment data from the 2023 stock assessments for stocks in the Gulf of Maine and Georges Bank in the northwest Atlantic Ocean. From re-examining nine historical data sets, we find poor recruitment can be predicted from previously discredited or uninformative environmental variables involving temperature, salinity and zooplankton abundance. For recent cod data, we examine three stocks each with six environmental variables measured during two seasons (spring and autumn). We find that temperature continues to have predictive value for poor recruitment, but zooplankton abundance is less informative and does not always follow the poor-recruitment paradigm. These findings suggest that (1) these predictors continue to be important over time for cod recruitment and (2) that for the recent regime, temperature may be a better predictor of poor recruitment than zooplankton abundance. [ABSTRACT FROM AUTHOR]

Recently, researchers have argued that the impressive performance of Natural Language Inference (NLI) models is highly due to the spurious correlations existing in training data, which makes models vulnerable and poorly generalized. Some work has made preliminary debiased attempts by developing data-driven interventions or model-level debiased learning. Despite the progress, existing debiased methods either suffered from the high cost of data annotation processing, or required elaborate design to identify biased factors. By conducting detailed investigations and data analysis, we argue that label information can provide meaningful guidance to identify these spurious correlations in training data, which has not been paid enough attention. Thus, we design a novel Label-aware Debiased Causal Reasoning Network (LDCRN). Specifically, according to the data analysis, we first build a causal graph to describe causal relations and spurious correlations in NLI. Then, we employ an NLI model (e.g., RoBERTa) to calculate total causal effect of input sentences to labels. Meanwhile, we design a novel label-aware biased module to model spurious correlations and calculate their causal effect in a fine-grained manner. The debiasing process is realized by subtracting this causal effect from total causal effect. Finally, extensive experiments over two well-known NLI datasets and multiple human-annotated challenging test sets are conducted to prove the superiority of LDCRN. Moreover, we have developed novel challenging test sets based on MultiNLI to facilitate the community.

Deep neural networks may be susceptible to learning spurious correlations that hold on average but not in atypical test samples. As with the recent emergence of vision transformer (ViT) models, it remains unexplored how spurious correlations are manifested in such architectures. In this paper, we systematically investigate the robustness of different transformer architectures to spurious correlations on three challenging benchmark datasets. Our study reveals that for transformers, larger models and more pre-training data significantly improve robustness to spurious correlations. Key to their success is the ability to generalize better from the examples where spurious correlations do not hold. Further, we perform extensive ablations and experiments to understand the role of the self-attention mechanism in providing robustness under spuriously correlated environments. We hope that our work will inspire future research on further understanding the robustness of ViT models to spurious correlations. [ABSTRACT FROM AUTHOR]

Patterns in datasets, while key to successful predictions, can also be misleading. In this study, a feature filtering workflow based on interpretable dimensionality reduction techniques was developed to diagnose the spurious correlations in high-entropy alloy dataset with phase structure and hardness labels. It is found that the presumed linear relationship between valence electron concentration (VEC) and hardness in high-entropy alloys is spurious, which highly influenced by the constitute elements of alloy system and determined by properties related to atomic radius. In addition, electron work function (w) and cohesive energy (Ec) have similar relationships with hardness, which indicated that these electronic features should be excluded for hardness prediction. Hence, this process can serve as a preliminary step in feature selection to mitigate the influence of non-causal features on prediction.

Purpose: Criticisms of the acute to chronic workload ratio (ACWR) have been that the mathematical coupling inherent in the traditional calculation of the ACWR results in a spurious correlation. The purposes of this commentary are (1) to examine how mathematical coupling causes spurious correlations and (2) to use a case study from actual monitoring data to determine how mathematical coupling affects the ACWR. Methods: Training and competition workload (TL) data were obtained from international-level open-skill (basketball) and closed-skill (weightlifting) athletes before their respective qualifying tournaments for the 2016 Olympic Games. Correlations between acute TL, chronic TL, and the ACWR as coupledAincoupled variations were examined. These variables were also compared using both rolling averages and exponentially weighted moving averages to account for any potential benefits of one calculation method over another. Results: Although there were some significant differences between coupled and uncoupled chronic TL and ACWR data, the effect sizes of these differences were almost all trivial (g = 0.04-0.21). Correlations ranged from r = .55 to .76, .17 to .53, and .88 to .99 for acute to chronic TL, acute to uncoupled chronic TL, and ACWR to uncoupled ACWR, respectively. Conclusions: There may be low risk of mathematical coupling causing spurious correlations in the TL-injury-risk relationship. Varying levels of correlation seem to exist naturally between acute and chronic TL variables regardless of coupling. The trivial to small effect sizes and large to nearly perfect correlations between coupled and uncoupled AWCRs also imply that mathematical coupling may have little effect on either calculation method, if practitioners choose to apply the ACWR for TL monitoring purposes. [ABSTRACT FROM AUTHOR]

Artificial intelligence (AI) has been successful at solving numerous problems in machine perception. In radiology, AI systems are rapidly evolving and show progress in guiding treatment decisions, diagnosing, localizing disease on medical images, and improving radiologists' efficiency. A critical component to deploying AI in radiology is to gain confidence in a developed system's efficacy and safety. The current gold standard approach is to conduct an analytical validation of performance on a generalization dataset from one or more institutions, followed by a clinical validation study of the system's efficacy during deployment. Clinical validation studies are time-consuming, and best practices dictate limited re-use of analytical validation data, so it is ideal to know ahead of time if a system is likely to fail analytical or clinical validation. In this paper, we describe a series of sanity tests to identify when a system performs well on development data for the wrong reasons. We illustrate the sanity tests' value by designing a deep learning system to classify pancreatic cancer seen in computed tomography scans.

Background: Significant progress has been made in the practice of conducting causal analysis using network models. Despite this progress, there is limited evidence that hospital risk managers are using these analytical models. Objective: This article introduces the causal network, its related concepts, and methods of analysis. The article demonstrates how hospital risk managers can use existing regression software to construct a causal network and identify root causes of an adverse event. Methods: Causal networks depict cause and effect in a set of variables. In this context, causes are strong correlations that meet 3 additional criteria: (1) causes occur prior to effects, (2) there is an articulated mechanism for how causes lead to effects, and (3) the association between cause and effect is not spurious, meaning the association persists even after other variables are statistically controlled for (a method of analysis called counterfactual). A causal network can be constructed through repeated use of least absolute shrinkage and selection operator (LASSO) regression. In the proposed regressions, the response variable is any variable in the data. The independent variables are variables that occur prior to the response variable. By design, the statistically significant coefficients in the time-constrained LASSO regressions identify "direct" causes of the response variables. When direct causes of all variables are identified, then the entire network model, including root causes, has been specified. In the final step, the parameters of the network model (ie, strength of causal associations) are estimated by fitting the network structure to the available data. We demonstrate these concepts through fitting a network model to simulated data for causes of excessive boarding in emergency departments. Results: The network (involving 12 causes, over 4 periods, and 1 sentinel event) was accurately recovered from the simulated case reports. The recovered network did not differ from the original network used to simulate the data in any of the 156 possible links. The recovered network allowed the identification of root and direct causes. It showed that hospital occupancy rate, and not emergency department efficiency, was root cause of excessive emergency department boarding. Discussion: Causal networks can provide insights into root, and direct, causes of an adverse event. These models provide empirical tests of causes of adverse events. We encourage the use of these methods by hospital risk managers. [ABSTRACT FROM AUTHOR]

The covariance matrix estimated from the ensemble data assimilation always suffers from filter collapse because of the spurious correlations induced by the finite ensemble size. The localization technique is applied to ameliorate this issue, which has been suggested to be effective. In this paper, an adaptive scheme for Schur product covariance localization is proposed, which is easy and efficient to implement in the ensemble data assimilation frameworks. A Gaussian-shaped taper function is selected as the localization taper function for the Schur product in the adaptive localization scheme, and the localization radius is obtained adaptively through a certain criterion of correlations with the background ensembles. An idealized Lorenz96 model with an ensemble Kalman filter is firstly examined, showing that the adaptive localization scheme helps to significantly reduce the spurious correlations in the small ensemble with low computational cost and provides accurate covariances that are similar to those derived from a much larger ensemble. The investigations of adaptive localization radius reveal that the optimal radius is model-parameter-dependent, vertical-level-dependent and nearly flow-dependent with weather scenarios in a realistic model; for example, the radius of model parameter zonal wind is generally larger than that of temperature. The adaptivity of the localization scheme is also illustrated in the ensemble framework and shows that the adaptive scheme has a positive effect on the assimilated analysis as the well-tuned localization.

Much time and effort, as well as funding, is being devoted to Genome Wide Association Studies (GWAS) for identifying genetic causes of variation (single nucleotide polymorphisms or SNPs) in human cognitive abilities and educational attainments (CA and EA). After years of finding only very weak associations, usually failing to replicate, attention has turned to aggregates of otherwise non-significant SNPs (called polygenic scores, or PGS) and some associations with traits are now being reported. Here we show how, in the context of CA and EA as approximation measures, spurious correlations in GWAS/PGS can arise in a number of ways, particularly from genetic population structure. We review recent studies suggesting that attempts to control for such confounds have been quite inadequate, and also criticize the underlying statistical assumptions and genetic model. [ABSTRACT FROM AUTHOR]

Deep learning techniques have dominated the literature on aspect-based sentiment analysis (ABSA), achieving state-of-the-art performance. However, deep models generally suffer from spurious correlations between input features and output labels, which significantly hurts the robustness and generalization capability. In this paper, we propose to reduce spurious correlations for ABSA, via a novel C ontrastive V ariational I nformation B ottleneck framework (called CVIB). The proposed CVIB framework is composed of an original network and a self-pruned network, and these two networks are optimized simultaneously via contrastive learning. Concretely, we employ the Variational Information Bottleneck (VIB) principle to learn an informative and compressed network (self-pruned network) from the original network, which discards the superfluous patterns or spurious correlations between input features and prediction labels. Then, self-pruning contrastive learning is devised to pull together semantically similar positive pairs and push away dissimilar pairs, where the representations of the anchor learned by the original and self-pruned networks respectively are regarded as a positive pair while the representations of two different sentences within a mini-batch are treated as a negative pair. To verify the effectiveness of our CVIB method, we conduct extensive experiments on five benchmark ABSA datasets. The experimental results show that our approach achieves better performance than the strong competitors in terms of overall prediction performance, robustness, and generalization. [ABSTRACT FROM AUTHOR]

Multivariate analyses constitute an integral part of today's marine integrated ecosystem assessments (IEAs). Principal component analysis (PCA) is one of the most common of these techniques, and the method has been used repeatedly to summarize the dynamics of marine ecosystems. There seems to be little recognition of the potential pitfalls associated with performing PCA on time-series that are autocorrelated and/or non-stationary. We investigate how the descriptive performance of PCAs may be affected by the structure of the underlying timeseries and question whether such analyses can provide useful summaries of ecosystem trajectories. For this purpose, we reanalyse four datasets from the Barents, Norwegian, Baltic, and North Seas. We compare the results with those obtained from simulated datasets that share similar trend and autocorrelation properties, but in which the variables are unrelated. We show that most of the patterns revealed by the PCA can emerge from random time-series and that the fraction of the variance that cannot be accounted for by random processes is minimal. The Norwegian Sea dataset is a pathological case in which the variance explained by the first two components only exceeds what would be expected from randomly simulated time-series by 2%. We conclude that outputs from explorative multivariate analyses provide very little insight into ecosystem status, trajectories and functioning. IEA groups need to be equipped with methods that can provide better insight into how marine ecosystems function, the drivers of their changes and their possible future trajectories. [ABSTRACT FROM AUTHOR]

Ecoinformatics, as defined in this review, is the use of preexisting data sets to address questions in ecology. We provide the first review of ecoinformatics methods in agricultural entomology. Ecoinformatics methods have been used to address the full range of questions studied by agricultural entomologists, enabled by the special opportunities associated with data sets, nearly all of which have been observational, that are larger and more diverse and that embrace larger spatial and temporal scales than most experimental studies do. We argue that ecoinformatics research methods and traditional, experimental research methods have strengths and weaknesses that are largely complementary. We address the important interpretational challenges associated with observational data sets, highlight common pitfalls, and propose some best practices for researchers using these methods. Ecoinformatics methods hold great promise as a vehicle for capitalizing on the explosion of data emanating from farmers, researchers, and the public, as novel sampling and sensing techniques are developed and digital data sharing becomes more widespread. [ABSTRACT FROM AUTHOR]

In 2010, the American College of Cardiology Foundation and American Heart Association could not recommend brachial artery percentage flow-mediated dilation (FMD%) for risk assessment of coronary artery disease (CAD) in asymptomatic adults. We aimed to scrutinise past and recently published findings regarding FMD% in this same context of clinical utility and conclude that (1) the question of whether brachial FMD% is a suitable substitute for coronary vasodilation is addressed by method agreement statistics rather than the correlation coefficients that have been reported in past studies. Also, the much-repeated view that brachial FMD% and coronary vasodilation are 'closely related' is not entirely justified, even before the influence of baseline lumen diameters on this relationship is accounted for; (2) along with the specialist training and the considerable time (≥1 h) that is required for the FMD% protocol, the error in individual measurements and population reference ranges is too large for clinical decisions to be robust on individual patients; (3) many interventions that are proposed to change FMD% also change baseline artery diameter, which can bias estimates of any intervention effects on the flow-mediated response per se, and (4) the FMD% index generates spurious correlations between shear rate, artery diameter and endothelial function, which may help to explain the apparent paradoxes of FMD% being higher in obese people and lower in athletes. In conclusion, the clinical relevance of brachial artery flow-mediated dilation is unclear at present. The dependence of the chosen index, FMD%, on initial artery size has contributed to this lack of clarity. [ABSTRACT FROM AUTHOR]

Direct gradient analyses in spatial genetics provide unique opportunities to describe the inherent complexity of genetic variation in wildlife species and are the object of many methodological developments. However, multicollinearity among explanatory variables is a systemic issue in multivariate regression analyses and is likely to cause serious difficulties in properly interpreting results of direct gradient analyses, with the risk of erroneous conclusions, misdirected research and inefficient or counterproductive conservation measures. Using simulated data sets along with linear and logistic regressions on distance matrices, we illustrate how commonality analysis ( CA), a detailed variance-partitioning procedure that was recently introduced in the field of ecology, can be used to deal with nonindependence among spatial predictors. By decomposing model fit indices into unique and common (or shared) variance components, CA allows identifying the location and magnitude of multicollinearity, revealing spurious correlations and thus thoroughly improving the interpretation of multivariate regressions. Despite a few inherent limitations, especially in the case of resistance model optimization, this review highlights the great potential of CA to account for complex multicollinearity patterns in spatial genetics and identifies future applications and lines of research. We strongly urge spatial geneticists to systematically investigate commonalities when performing direct gradient analyses. [ABSTRACT FROM AUTHOR]

Spurious correlations occur when two independent time series are found to be correlated according to the typical statistical procedure for testing the null hypothesis of zero correlation in the population. Using a Monte Carlo analysis, this study examines the spurious correlation phenomenon for two independent stationary AR(1) processes and it finds that if an alternative testing procedure is applied, spurious behaviour is eliminated using the variance of the sample correlation coefficient of these two series, suggested by Bartlett (1935). [ABSTRACT FROM AUTHOR]

This paper addresses three intractable difficulties associated with the statistical analysis of compositional data, such as percentages or ppm. These are: (1) that such data do not follow multivariate normal distributions thus rendering inappropriate, standard parametric statistical tests and estimation procedures, (2) the covariance/correlation coefficients between specific pairs of components are determined in whole or in part by the presence or absence of other components, and, (3) the negative bias property. That is, at least one covariance and therefore at least one correlation, must be negative, hence the remaining correlations are prevented from ranging freely between −1 and +1. It follows that correlation coefficients formed from compositional data are not only not absolute, but also frequently spurious. Standard multivariate procedures based on them are unreliable, and intrinsic associations between components inferred from strong positive correlations in particular, are potentially false. In a recent 2009 paper, it was reported that 59 surface sediment samples from 7 regions in the Polish exclusive economic zone had been chemically analyzed for 16 elements. Enrichment factors together with crude correlation coefficients between selected elements were presented. All these quantities were computed from the initial raw compositional data resulting from the chemical analyses In this paper, a statistical procedure is presented which is distinctly different to the enrichment factor computations based on the same raw compositional data. The procedure generates a log-ratio measure of the abundance of each element in each of the seven regions, thus enabling comparisons of relative levels of pollution between the regions. Although the two techniques are quite unrelated, it is shown that in general, extremely high or low measures of the relative abundances in the regions are associated with correspondingly high or low values of the enrichment factors in the same regions that were reported in the 2009 paper. That is, the statistical analysis confirms the results of the enrichment factor data in the identification of the most to the least polluted regions. In an additional analysis, the residue term was excluded from each sediment sample by rescaling the 16 element concentrations to sum to 100%, thus forming 59 residue-free sub-compositions. Crude correlation coefficients were computed for pairs of elements of this sub-compositional data. These revealed that certain correlations based on the initial raw data that were reported in the 2009 paper for the same pairs of elements, were not only inconsistent, but sometimes also contradictory. Such contradictions imply that intrinsic geochemical element associations inferred in that paper from such correlations were false. [ABSTRACT FROM AUTHOR]

Using a small ensemble size in the ensemble Kalman filter methodology is efficient for updating numerical reservoir models but can result in poor updates following spurious correlations between observations and model variables. The most common approach for reducing the effect of spurious correlations on model updates is multiplication of the estimated covariance by a tapering function that eliminates all correlations beyond a prespecified distance. Distance-dependent tapering is not always appropriate, however. In this paper, we describe efficient methods for discriminating between the real and the spurious correlations in the Kalman gain matrix by using the bootstrap method to assess the confidence level of each element from the Kalman gain matrix. The new method is tested on a small linear problem, and on a water flooding reservoir history matching problem. For the water flooding example, a small ensemble size of 30 was used to compute the Kalman gain in both the screened EnKF and standard EnKF methods. The new method resulted in significantly smaller root mean squared errors of the estimated model parameters and greater variability in the final updated ensemble. [ABSTRACT FROM AUTHOR]

The estimation of correlation coefficients has become an easy task since comfortable computer software is available to the end that in numerous publications whole batteries of correlation coefficient matrices are being presented. However, the effects of confounding variables are not always recognized and not searched for. Hence false conclusions are drawn from otherwise important scientific studies. This study offers a tool which allows researcher to estimate in advance those correlation coefficients which are promising candidates of being biased by a third variable. This is time sparing because only targeted coefficients need then to be recalculated. Furthermore, the tabulated data are given which allow to roughly and quickly estimate partial correlation coefficients. [ABSTRACT FROM AUTHOR]

Normalization of trace element contents of sediment through division by the content of an immobile element – usually aluminum – is common practice in marine geochemical studies. This might appear to be a simple way of correcting for dilution by sedimentary phases barren of a particular trace element – for instance carbonates or silica – and for comparison with its content in a standard clay or shale. The purpose of this paper is to revive the awareness of the pitfalls of this practice by giving numerical examples of the often-unexpected results of normalization. Taking a statistical perspective, it is first shown that uncorrelated variables acquire spurious correlations when normalized. But normalization can also increase, decrease, change sign of, or even blur the correlations between unmodified variables. Next, a number of realistic scenarios are worked out to show that the correlations between normalized element contents still suffer from the closure effect. Only in a few simple cases it is possible to extract, from the normalized data, realistic estimates of trace element contents in distinct sedimentary phases, such as organic matter. This is ascertained from data on copper contents in sediment from the Black Sea and Arabian Sea. When the coefficient of variation (i.e. standard deviation divided by the mean) of the aluminum data is relatively low and much lower than for the original values of the trace elements, comparisons of correlations can equally well or even better be made on the basis of the unmodified values. Then, the comparison of trace element data with their values for standard shales is discussed. The inherent problem is that the composition of the commonly used standard shale and, consequently, the reference values of normalized elements are not necessarily representative of the local or regional sediments in the study area. Also, this method falls short of identifying the processes responsible for enrichment or diminution in trace element contents. Finally, consideration is given to some aspects of a proper use of Al normalization [Copyright &y& Elsevier]

Studies on the effect of non- g ability residuals have often employed double adjustment for general cognitive ability (g), as they have calculated the ability residuals adjusting for g and then calculated the effect of the non- g residuals while adjusting for g. The present simulations demonstrate that the double adjustments may result in spurious negative associations between the non- g residual on one cognitive ability, e.g. verbal ability, and variables with a positive association with another ability, e.g. SAT math and math ability. In analyses of data from the 1997 National Longitudinal Survey of Youth (NLSY97), the negative associations between non- g residuals on verbal and math ability and aptitude test scores on the other ability vanished when not double adjusting for g. This indicates that the observed negative associations may be spurious and not due to differential investment of time and effort in one ability at the expense of the other ability, as suggested in the literature. Researchers of the effects of specific abilities are recommended to validate their findings and interpretations with analyses not double adjusting for g. • Negative associations involving non- g ability residuals have been identified. • These associations may be spurious and due to double adjustment for g. • Claims that associations are due to differential investment can be questioned. • Effects of specific abilities should be verified by not double adjusting for g. [ABSTRACT FROM AUTHOR]

The covariance matrix estimated from the ensemble data assimilation always suffers from filter collapse because of the spurious correlations induced by the finite ensemble size. The localization technique is applied to ameliorate this issue, which has been suggested to be effective. In this paper, an adaptive scheme for Schur product covariance localization is proposed, which is easy and efficient to implement in the ensemble data assimilation frameworks. A Gaussian-shaped taper function is selected as the localization taper function for the Schur product in the adaptive localization scheme, and the localization radius is obtained adaptively through a certain criterion of correlations with the background ensembles. An idealized Lorenz96 model with an ensemble Kalman filter is firstly examined, showing that the adaptive localization scheme helps to significantly reduce the spurious correlations in the small ensemble with low computational cost and provides accurate covariances that are similar to those derived from a much larger ensemble. The investigations of adaptive localization radius reveal that the optimal radius is model-parameter-dependent, vertical-level-dependent and nearly flow-dependent with weather scenarios in a realistic model; for example, the radius of model parameter zonal wind is generally larger than that of temperature. The adaptivity of the localization scheme is also illustrated in the ensemble framework and shows that the adaptive scheme has a positive effect on the assimilated analysis as the well-tuned localization. [ABSTRACT FROM AUTHOR]

In this paper, we propose an efficient and practical implementation of the ensemble Kalman filter (EnKF) via the distribution-free Ledoit and Wolf (LW) covariance matrix estimator. Initially, we develop a tractable implementation of the LW estimator in high-dimensional probability spaces such as those found in the context of operational data assimilation. We employ this well-conditioned, full-rank covariance matrix estimator to approximate background error covariances and to mitigate the impact of spurious correlations during assimilation steps. The proposed formulation can be coupled within the EnKF framework to derive a matrix-free implementation via an iterative Woodbury formula. Experimental tests are performed by using an Atmospheric General Circulation Model. The numerical results are compared with those of the EnKF based on the Rao–Blackwell Ledoit, and Wolf covariance matrix estimator (EnKF-RBLW) wherein Gaussian assumptions are a must on prior members. The outcomes reveal that the use of the proposed filter can mitigate the impact of spurious correlations during assimilation stages, and even more, the proposed method can improve on the results of the EnKF-RBLW as a consequence of the Gaussian relaxation on prior errors. [ABSTRACT FROM AUTHOR]

The Cancer Genome Atlas database offers the possibility of analyzing genome-wide expression RNA-Seq cancer data using paired counts, that is, studies where expression data are collected in pairs of normal and cancer cells, by taking samples from the same individual. Correlation of gene expression profiles is the most common analysis to study co-expression groups, which is used to find biological interpretation of -omics big data. The aim of the paper is threefold: firstly we show for the first time, the presence of a "regulation-correlation bias" in RNA-Seq paired expression data, that is an artifactual link between the expression status (up- or down-regulation) of a gene pair and the sign of the corresponding correlation coefficient. Secondly, we provide a statistical model able to theoretically explain the reasons for the presence of such a bias. Thirdly, we present a bias-removal algorithm, called SEaCorAl, able to effectively reduce bias effects and improve the biological significance of correlation analysis. Validation of the SEaCorAl algorithm is performed by showing a significant increase in the ability to detect biologically meaningful associations of positive correlations and a significant increase of the modularity of the resulting unbiased correlation network., (Copyright © 2021 The Author(s). Published by Elsevier Ltd.. All rights reserved.)