Biased Estimates Research Articles

Poverty mapping in the age of machine learning

Recent years have witnessed considerable methodological advances in poverty mapping, much of which has focused on the application of modern machine-learning approaches to remotely-sensed data. Poverty maps produced with these methods generally share a common validation procedure, which assesses model performance by comparing sub-national poverty estimates with survey-based, direct estimates. While unbiased, direct estimates can be imprecise measures of true poverty rates, meaning that it is unclear whether these validation procedures are informative of actual model performance. In this paper, we use a rich dataset from Mexico to provide a more rigorous assessment of the modern approach to poverty mapping by evaluating its performance against a credible ground truth. We find that the modern method under-performs relative to benchmark traditional methods, largely because of the limited predictive capacity of remotely-sensed covariates. For a given covariate set, we also find that machine learning produces more biased poverty estimates than the traditional procedures, particularly for the poorest geographic areas.

Latent Vector Autoregressive Modeling: A Stepwise Estimation Approach

Researchers often study dynamic processes of latent variables in everyday life, such as the interplay of positive and negative affect over time. An intuitive approach is to first estimate the measurement model of the latent variables, then compute factor scores, and finally use these factor scores as observed scores in vector autoregressive modeling. However, this approach neglects the uncertainty in the factor scores, leading to biased parameter estimates and threatening the validity of conclusions about the dynamic process. We propose Three-Step Latent Vector Autoregression that adheres to this stepwise procedure while correcting for the factor scores’ uncertainty. Stepwise approaches offer various advantages, for example the ability to visualize and inspect the factor scores. A simulation study demonstrates that the method performs well in obtaining correct parameter estimates of a dynamic process. We also provide an empirical example and scripts for implementation in the open-source software R using the lavaan package.

Robust and time-resolved estimation of cardiac sympathetic and parasympathetic indices.

The time-resolved analysis of heart rate (HR) and heart rate variability (HRV) is crucial for the evaluation of the dynamic changes of autonomic activity under different clinical and behavioral conditions. Standard HRV analysis is performed in the frequency domain because the sympathetic activations tend to increase low-frequency HRV oscillations, while the parasympathetic ones increase high-frequency HRV oscillations. However, a strict separation of HRV in frequency bands may cause biased estimations, especially in the low frequency range. To overcome this limitation, we propose a robust estimator that combines HR and HRV dynamics, based on the correlation of the Poincaré plot descriptors of interbeat intervals from the electrocardiogram. To validate our method, we used electrocardiograms gathered from open databases where standardized paradigms were applied to elicit changes in autonomic activity. Our proposal outperforms the standard spectral approach for the estimation of low- and high-frequency fluctuations in HRV, and its performance is comparable to newer methods. Our method constitutes a valuable, robust, time-resolved, and cost-effective tool for a better understanding of autonomic activity through HR and HRV in healthy state and potentially for pathological conditions.

Improved Quantification of MicroPET/CT Imaging Using CT-derived Scaling Factors.

Combined micro-PET/CT scanners are widely employed to investigate models of brain disorders in rodents using PET-based coregistration. We examined if CT-based coregistration could improve estimates of brain dimensions and consequently estimates of nondisplaceable binding potential (BPND) in rodent PET studies. PET and CT scans were acquired on 5 female and 5 male CD-1 mice with 3-[18F]fluoro-5-(2-pyridinylethynyl)benzonitrile ([18F]FPEB), a radiotracer for the metabotropic glutamate receptor subtype 5 (mGluR5). In the proposed PET/CT (PTCT) approach, the tracer-specific standard volume was dimension-customized to each animal using the scaling factors from CT-to-standard CT coregistration to simplify PET-to-standard PET coregistration (i.e., 3 CT- and 6 PET-derived parameters). For comparison, conventional PET-based coregistration was performed with 9 (PT9) or 12 (PT12) parameters. PET frames were transferred to the standard space by the three approaches (PTCT, PT9, and PT12) to obtain regional time-activity curves (TACs) and BPND in 14 standard volumes of interest (VOIs). Lastly, CT images of the animals were transferred to the standard space by CT-based parameters from PTCT and with the scaling factors replaced with those from PET-based PT9 to evaluate agreement of the skull to the standard CT. The PET-based approaches showed various degrees of underestimations of scaling factors in the posterior-anterior-direction compared to PTCT, which resulted in negatively proportional overestimation of radioactivity in the cerebellum (reference region) up to 20%, and proportional, more prominent underestimation of BPND in target regions down to -50%. The skulls of individual animals agreed with the standard skull for scaling factors from PTCT but not for the scaling factors from PT9, which suggested inaccuracy of the latter. The results indicated that conventional PET-based coregistration approaches could yield biased estimates of BPND in proportion to errors of brain dimensions when applied to tracers for which the cerebellum serves as reference region. The proposed PTCT provides evidence of a quantitative improvement over PET-based approaches for brain studies using micro-PET/CT scanners.

Anisotropic elasticity measurements of the retina using optical coherence elastography

Anisotropic elasticity measurements of the retina are essential for retinal disease diagnosis and function assessment. Optical coherence elastography (OCE) is a high-resolution imaging technique for mapping the elasticity distribution of tissues. However, previous OCE measurements quantified the tissue elasticity in a single direction, resulting in a biased estimation of the elastic properties. In this study, we propose an OCE method with acoustic radiation force (ARF) excitation to map the retinal anisotropic elasticity in the depth and lateral directions. The axial elasticity was analyzed using the natural frequency of free vibration, and the lateral elasticity was quantified using the elastic wave velocity. After evaluating the feasibility of the OCE method on the phantoms, the anisotropic elasticity of ex vivo porcine retinas was mapped. The results show that the OCE method with ARF excitation can assess the elasticity in orthogonal directions and provide a comprehensive understanding of the elasticity of the anisotropic tissues.

Assortative mating across nine psychiatric disorders is consistent and persistent over cultures and generations.

Emerging evidence has shown that assortative mating (AM) is a key factor that shapes the landscape of complex human traits. It can increase the overall prevalence of disorders, influence occurrences of comorbidities, and bias estimation of genetic architectures. However, there is lack of large-scale studies to examine the cultural differences and the generational trends of AM for psychiatric disorders. Here, using national registry datasets, we conduct the largest scale of AM analyses on nine psychiatric disorders, with up to 1.4 million mated cases and 6 million matched controls. We performed meta-analyses on AM estimates from Taiwan, Denmark, and Sweden, to examine the potential impact of cultural differences. Generational changes for people born after 1930s were investigated as well. We found that AM of psychiatric disorders are consistent across nations and persistent over generations, with a small proportion of disorders showing generational changes of AM. Our results provide additional insight into the mechanisms of AM across psychiatric disorders and have evident implications on the estimation of the genetic architectures of psychiatric disorders.

Retaliatory Tariffs of China during the Trade War: A Firm-level Overview

Abstract In this paper, we offer a comprehensive survey of the latest research regarding the pass-through of tariffs during the trade war between China and the United States, with an emphasis on the investigations conducted from the Chinese perspective. While one might anticipate symmetric effects on both the United States and China, these nations exhibit substantial disparities in terms of trade structure and policy. We underscore two salient concerns that have been largely neglected in extant literature: (1) the exemption of processing imports from retaliatory tariffs in China, which may engender biased estimations of pass-through at the product level, and (2) the heterogeneous distribution of retaliatory tariffs across product classifications, firm ownership structures, and geographical regions, necessitating more granular estimations at the firm level. Furthermore, we provide a detailed description of firm import tariffs, emphasizing the crucial role of tailored policies in mitigating the repercussions of the trade conflict.

Nonlinear and threshold effects of the built environment, road vehicles and air pollution on urban vitality

The impact of factors such as the built environment, road vehicles, and air quality on urban vitality attracts increasing interest in urban planning and design research. However, tacit assumptions of linear relationships between these factors have been embedded in most studies, leading to biased estimations of their effects on urban vitality. This study addresses the gap by using machine learning models and SHAP (SHapley Additive exPlanations) to investigate the non-linear and threshold effects of the built environment, road vehicles and air pollution on urban vitality, using Manhattan as a study case. Urban vitality was represented by pedestrian presence in 29,540 street-view images. Results showed that Extreme Gradient Boosting outperformed Ordinary Least Squares, Random Forest, and Gradient Boosting Decision Trees in urban vitality estimation. It reveals that while the built environment variables explained a significant portion (77.5 %) of the variance in urban vitality, road vehicles (such as bicycles, buses, cars and motorbikes) and ozone concentrations accounted for 15.18 % and 1.46 %, respectively. The built environment and road vehicle factors exhibit positive nonlinear relationships with urban vitality. Meanwhile, ozone concentration demonstrated a negative threshold effect on urban vitality with a threshold at 27.5 ppb. This study advances our understanding of the threshold effect mechanism of the factors on urban vitality, offering insights into fostering sustainable urban environment.

Importance of accounting for imperfect detection of plants in the estimation of population growth rates

Detection of plant individuals is imperfect. Not accounting for this issue can result in biased estimates of demographic parameters as important as population growth rates. In mobile organisms, a common practice is to explicitly account for detection probability during the estimation of most demographic parameters, but no study in plant populations has examined the consequences of ignoring imperfect detectability on the estimation of population growth rates. The lack of accounting for detection probability occurs because plant demographers have frequently assumed that detection is perfect, and because there is a scarcity of studies that formally compare the performance of estimation methods that incorporate detection probabilities with respect to methods that ignore detectabilities. Based on field data of five plant species and data simulations, we compared the performance of three methods that estimate population growth rates, two that do not estimate detection probabilities (direct counts of individuals and the minimum‐number‐alive method) and the other that explicitly accounts for detection probabilities (temporal symmetry models). Our aims were 1) to estimate detection probabilities, and 2) to evaluate the performance of these three methods by calculating bias, accuracy, and precision in their estimates of population growth rates. Our five plant species had imperfect detection. Estimates of population growth rates that explicitly incorporate detectabilities had better performance (less biased estimates, with higher accuracy and precision) than those obtained with the two methods that do not calculate detection probabilities. In these latter methods, bias increases as detection probability decreases. Our findings highlight the importance of using robust analytical methods that account for detection probability of plants during the estimation of critical demographic parameters such as population growth rates. In this way, estimates of plant population parameters will reliably indicate their actual status and quantitative trends.

Rethinking the residual approach: Leveraging machine learning to operationalize cognitive resilience in Alzheimer's disease.

Cognitive resilience describes the phenomenon of individuals evading cognitive decline despite prominent Alzheimer's disease neuropathology. Operationalization and measurement of this latent construct is non-trivial as it cannot be directly observed. The residual approach has been widely applied to estimate CR, where the degree of resilience is estimated through a linear model's residuals. We demonstrate that this approach makes specific, uncontrollable assumptions and likely leads to biased and erroneous resilience estimates. We propose an alternative strategy which overcomes the standard approach's limitations using machine learning principles. Our proposed approach makes fewer assumptions about the data and construct to be measured and achieves better estimation accuracy on simulated ground-truth data.

Response is increased using postal rather than electronic questionnaires – new results from an updated Cochrane Systematic Review

BackgroundA decade ago paper questionnaires were more common in epidemiology than those administered online, but increasing Internet access may have changed this. Researchers planning to use a self-administered questionnaire should know whether response rates to questionnaires administered electronically differ to those of questionnaires administered by post. We analysed trials included in a recently updated Cochrane Review to answer this question.MethodsWe exported data of randomised controlled trials included in three comparisons in the Cochrane Review that had evaluated hypotheses relevant to our research objective and imported them into Stata for a series of meta-analyses not conducted in the Cochrane review. We pooled odds ratios for response using random effects meta-analyses. We explored causes of heterogeneity among study results using subgroups. We assessed evidence for reporting bias using Harbord’s modified test for small-study effects.ResultsTwenty-seven trials (66,118 participants) evaluated the effect on response of an electronic questionnaire compared with postal. Results were heterogeneous (I-squared = 98%). There was evidence for biased (greater) effect estimates in studies at high risk of bias; A synthesis of studies at low risk of bias indicates that response was increased (OR = 1.43; 95% CI 1.08–1.89) using postal questionnaires. Ten trials (39,523 participants) evaluated the effect of providing a choice of mode (postal or electronic) compared to an electronic questionnaire only. Response was increased with a choice of mode (OR = 1.63; 95% CI 1.18–2.26). Eight trials (20,909 participants) evaluated the effect of a choice of mode (electronic or postal) compared to a postal questionnaire only. There was no evidence for an effect on response of a choice of mode compared with postal only (OR = 0.94; 95% CI 0.86–1.02).ConclusionsPostal questionnaires should be used in preference to, or offered in addition to, electronic modes.

Development and evaluation of hydraulic geometry relationships for large Brazilian river basins

ABSTRACT River cross-section characteristics are not available for many regions of the world, and they are generally expressed as hydraulic geometry (HG) relationships. These relationships usually represent average characteristics of a region, which can generate biased estimates in smaller basins. Therefore, to improve the representation of river channel geomorphological characteristics, the objective of this study was to use a large database of river cross-section geometries to refine HG relationships for five river basins in Brazil. HG relationships were developed for the entire basins and for hydraulically homogeneous regions, obtained through a geographical convenience method. The relationship between the coefficients of the geometric relationships and morphoclimatic characteristics was assessed, showing a trend between precipitation and slope and the coefficients of the depth relationships. The estimated bankfull width and depth values were also compared to observed data and global datasets, showing an improvement in estimates with the more refined HG relationships.

Multiple imputation Procedure for Record Linkage and Causal Inference to Estimate the Effects of home-delivered meals

To assess the impact of home-delivered meals programs offered by Meals on Wheels (MOW) a dataset that includes enrollment into MOW and healthcare utilization is required. Because of confidentiality restrictions this data is dispersed over MOW programs internal systems and Medicare claims. In the absence of unique identifying identifiers, MOW beneficiaries are linked to Medicare beneficiaries using probabilistic linking algorithm. However, this procedure may erroneously link individuals across the two datasets, which may result in biased treatment effect estimates and suboptimal interval estimates. We propose a two-stage multiple imputation framework to estimate causal effects when the covariates and outcome information are stored in Medicare claims and the treatment assignment is MOW program data. In the first stage, we create multiple datasets in which MOW beneficiaries are linked to Medicare beneficiaries using a Bayesian record linkage technique. In the second stage, we match Medicare beneficiaries that were not enrolled in MOW programs to those that did. Using these matches we multiply imputed for each MOW beneficiary their unobserved healthcare utilization if they did not receive MOW. This procedure propagates the errors in the linking process and the matching process, and can be used to estimate effects of interventions in other linked datasets when there are no unique identifiers.

Identifying hospitalization episodes of care among people with and without HIV in British Columbia, Canada

BackgroundHospitalizations are a resource-intensive form of healthcare use, particularly for persons with chronic conditions such as those with HIV. Interhospital transfers typically appear as separate records in Canadian databases; misclassifying transfers as independent hospitalizations can bias key metrics such as readmission rates. We examined approaches of combining sequential, related records into hospitalization episodes of care (HEoCs) among persons with and without HIV (PWH; PWoH) in British Columbia (BC), Canada. MethodsBC hospitalization records (1992 to 2020) were sourced from the Comparative Outcomes and Service Utilization Trends (COAST) study, a data linkage that includes samples of PWH and PWoH. We constructed 8 HEoC definitions that varied by the: a) time gap between records, and b) transfer indication. Comparisons were informed by the proportion of multi-record HEoCs (mHEoCs; episodes with multiple hospitalization records) generated, and feasibility given data quality. ResultsWe analyzed 98,553 hospitalization records from 13,498 PWH, and 1,874,507 hospitalization records from 385,011 PWoH. Across the definitions, the proportion of mHEoCs varied from 2.46% to 5.27% for PWH and 2.73% to 4.18% for PWoH. Definitions requiring no transfer indication yielded the highest proportion of mHEoCs, whereas those requiring two-way agreement of hospital identifiers yielded the lowest proportion of mHEoCs. Patterns were comparable among PWH and PWoH. A pragmatic approach to defining HEoCs can be a reasonable option for general purposes – requiring at least one populated hospital identifier field, and ≤ 1 day gap between hospitalizations. ConclusionsVarious approaches can be employed to combine sequential, related hospitalization records into HEoCs to help provide less biased estimates of hospitalization-related metrics.

Identifiability and Estimation for Potential-Outcome Means with Misclassified Outcomes

Potential outcomes play a fundamental and important role in many causal inference problems. If the potential-outcome means are identifiable, a series of causal effect measures, including the risk difference, the risk ratio, and the treatment benefit rate, among others, can also be identified. However, current identification and estimation methods for these means often implicitly assume that the collected data for analysis are measured precisely. In many fields such as medicine and economics, the collected variables may be subject to measurement errors, such as medical diagnostic results and individual wage data. Misclassification, as a non-classic measurement error, can lead to severely biased estimates in causal inference. In this paper, we leverage a combined sample to study the identifiability of potential-outcome means corresponding to different treatment levers under a plausible misclassification assumption for the outcome, allowing the misclassification probability to depend on not only the true outcome but also the covariates. Furthermore, we propose the multiply-robust and semiparametric efficient estimators for the means, consistent even under partial misspecification of the observed data law, based on the semiparametric theory framework. The simulation studies and real data analysis demonstrate the satisfactory performance of the proposed method.

Allometric equations for biomass and carbon pool estimation in short rotation Pinus radiata stands of the Western Cape, South Africa

AbstractThe quantification of tree growth and carbon storage over time is an important task for sustainable forest management and carbon sequestration projects. For the South African short-rotation Pinus radiata (D. Don) forests, this knowledge is lacking. We developed allometric equations and compared the estimated weights to previously published biomass studies and we used Dirichlet Regression (DR) modelling to ensure additivity of the component proportions. The biomass components and their contribution to carbon storage depend strongly on forest structure and mean tree size but also on-site conditions and tree architecture. Our first two hypotheses were that the (1) best model for stemwood (SW), bark and total mass will include the combined variable DBH2H and (2) that the DR will yield statistically similar estimates for all components when compared to the best models. Our third hypothesis was that allometric equations developed for sites with high resource availability (e.g. wet, fertile sites) will yield biased estimates when extrapolated to sites with lower levels of resource availability (drier and/or infertile sites). The results indicated that DBH2H was the best variable to describe SW, bark and total mass and the DR yield similar estimates for all component proportions when compared to the best models. There were strong similarities in the SW and total mass of independent test sites in comparison to the SW and total mass of this study but greater variability in the bark, needle and branch mass. This can be associated to site and seasonal differences as well as variability in tree architecture brought about by different silvicultural operations on individual sites. Previously developed equations by other authors for sites with high resource availability overpredicted the SW and total mass of the models developed in this study. Our set of additive component equations performed well even when applied to sites of similar productivity over a climate gradient. The presented new equations bridge the gap in knowledge where allometric equations for short rotation Radiata pine stands are lacking.

Joint modeling of mixed skewed longitudinal responses using convolution of normal and log-normal distributions: a Bayesian approach

This paper investigates the joint modeling of mixed ordinal and continuous longitudinal responses using a random effects model and applying a conditional approach. For the ordinal responses, a latent variable model with a logistic distribution is employed. To address skewness in the data, the model incorporates normal and log-normal convolution (NLNC) for both the error term and the random effects in the longitudinal model. Parameter estimation is carried out within a Bayesian framework using Gibbs sampling. The performance of the proposed model is evaluated through simulation studies, comparing it to joint models of mixed ordinal and continuous longitudinal responses assuming skew-normal or normal distributions. The results indicate that joint models using skew-normal or normal distributions can lead to biased parameter estimates, whereas the NLNC joint model performs better overall. Additionally, the proposed method is applied to analyze data from the British Household Panel Survey (BHPS), using life satisfaction and annual income as the correlated ordinal and continuous longitudinal responses, respectively, with annual income showing significant skewness. The results demonstrate that the proposed model provides the best fit for capturing the substantial skewness in the data.

Random forest analysis and lasso regression outperform traditional methods in identifying missing data auxiliary variables when the MAR mechanism is nonlinear (p.s. Stop using Little's MCAR test).

The selection of auxiliary variables is an important first step in appropriately implementing missing data methods such as full information maximum likelihood (FIML) estimation or multiple imputation. However, practical guidelines and statistical tests for selecting useful auxiliary variables are somewhat lacking, leading to potentially biased estimates. We propose the use of random forest analysis and lasso regression as alternative methods to select auxiliary variables, particularly in situations in which the missing data pattern is nonlinear or otherwise complex (i.e., interactive relationships between variables and missingness). Monte Carlo simulations demonstrate the effectiveness of random forest analysis and lasso regression compared to traditional methods (t-tests, Little's MCAR test, logistic regressions), in terms of both selecting auxiliary variables and the performance of said auxiliary variables when incorporated in an analysis with missing data. Both techniques outperformed traditional methods, providing a promising direction for improvement of practical methods for handling missing data in statistical analyses.

Iterative UKF under generalized maximum correntropy criterion for intermittent observation systems with complex non-Gaussian noise

The traditional unscented Kalman filters (UKFs) under the maximum correntropy criterion provide a powerful tool for nonlinear state estimation with heavy-tailed non-Gaussian noise. Nevertheless, the above-mentioned filters may yield biased estimates because the Gaussian kernel function can only handle certain types of non-Gaussian noise. Additionally, the use of statistical linearization methods can result in approximation errors when solving linear observation equations, while the system may also experience observation data loss. Therefore, a new iterative UKF with intermittent observations under the generalized maximum correntropy criterion is proposed for systems with complex non-Gaussian noise, called GMCC-IO-IUKF. Firstly, the connection between the UKF with and without intermittent observations is established by designing a coefficient matrix including intermittent observation variables, so as to derive the UKF with intermittent observations under the maximum correntropy criterion. Secondly, for the measurement update of GMCC-IO-IUKF, a nonlinear regression augmented model that can deal with both prediction and observation errors is established using the coefficient matrix and the nonlinear function. To better adapt to different types of non-Gaussian noise, the generalized Gaussian kernel function is substituted for the traditional Gaussian kernel function. Theoretically, GMCC-IO-IUKF can achieve better estimation performance by directly employing the nonlinear function and the latest iteration value. Finally, a classical target tracking model is used to evaluate the estimation performance and feasibility of our proposed GMCC-IO-IUKF algorithm. It appears from the experiment results that our proposed GMCC-IO-IUKF can not only promote estimation precision but also handle complex non-Gaussian noise flexibly.

Fl-IRT-ing with Psychometrics to Improve NLP Bias Measurement

To prevent ordinary people from being harmed by natural language processing (NLP) technology, finding ways to measure the extent to which a language model is biased (e.g., regarding gender) has become an active area of research. One popular class of NLP bias measures are bias benchmark datasets—collections of test items that are meant to assess a language model’s preference for stereotypical versus non-stereotypical language. In this paper, we argue that such bias benchmarks should be assessed with models from the psychometric framework of item response theory (IRT). Specifically, we tie an introduction to basic IRT concepts and models with a discussion of how they could be relevant to the evaluation, interpretation and improvement of bias benchmark datasets. Regarding evaluation, IRT provides us with methodological tools for assessing the quality of both individual test items (e.g., the extent to which an item can differentiate highly biased from less biased language models) as well as benchmarks as a whole (e.g., the extent to which the benchmark allows us to assess not only severe but also subtle levels of model bias). Through such diagnostic tools, the quality of benchmark datasets could be improved, for example by deleting or reworking poorly performing items. Finally, in regards to interpretation, we argue that IRT models’ estimates for language model bias are conceptually superior to traditional accuracy-based evaluation metrics, as the former take into account more information than just whether or not a language model provided a biased response.