Established Statistical Methods Research Articles

Variable ranking based on the estimated degree of separation for two distributions of data by the length of the receiver operating characteristic curve

Variable responses are fundamental for all experiments, and they can consist of information-rich, redundant, and low signal intensities. A dataset can consist of a collection of variable responses over multiple classes or groups. Usually some of the variables are removed in a dataset that contain very little information. Sometimes all the variables are used in the data analysis phase. It is common practice to discriminate between two distributions of data; however, there is no formal algorithm to arrive at a degree of separation (DS) between two distributions of data. The DS is defined herein as the average of the sum of the areas from the probability density functions (PDFs) of A and B that contain a≥percentage of A and/or B. Thus, DS90 is the average of the sum of the PDF areas of A and B that contain ≥90% of A and/or B. To arrive at a DS value, two synthesized PDFs or very large experimental datasets are required. Experimentally it is common practice to generate relatively small datasets. Therefore, the challenge was to find a statistical parameter that can be used on small datasets to estimate and highly correlate with the DS90 parameter. Established statistical methods include the overlap area of the two data distribution profiles, Welch’s t-test, Kolmogorov–Smirnov (K–S) test, Mann–Whitney–Wilcoxon test, and the area under the receiver operating characteristics (ROC) curve (AUC). The area between the ROC curve and diagonal (ACD) and the length of the ROC curve (LROC) are introduced. The established, ACD, and LROC methods were correlated to the DS90 when applied on many pairs of synthesized PDFs. The LROC method provided the best linear correlation with, and estimation of, the DS90. The estimated DS90 from the LROC (DS90–LROC) is applied to a database, as an example, of three Italian wines consisting of thirteen variable responses for variable ranking consideration. An important highlight of the DS90–LROC method is utilizing the LROC curve methodology to test all variables one-at-a-time with all pairs of classes in a dataset.

Integrative multi-omics module network inference with Lemon-Tree.

Module network inference is an established statistical method to reconstruct co-expression modules and their upstream regulatory programs from integrated multi-omics datasets measuring the activity levels of various cellular components across different individuals, experimental conditions or time points of a dynamic process. We have developed Lemon-Tree, an open-source, platform-independent, modular, extensible software package implementing state-of-the-art ensemble methods for module network inference. We benchmarked Lemon-Tree using large-scale tumor datasets and showed that Lemon-Tree algorithms compare favorably with state-of-the-art module network inference software. We also analyzed a large dataset of somatic copy-number alterations and gene expression levels measured in glioblastoma samples from The Cancer Genome Atlas and found that Lemon-Tree correctly identifies known glioblastoma oncogenes and tumor suppressors as master regulators in the inferred module network. Novel candidate driver genes predicted by Lemon-Tree were validated using tumor pathway and survival analyses. Lemon-Tree is available from http://lemon-tree.googlecode.com under the GNU General Public License version 2.0.

Evapotranspiration Variability and Its Association with Vegetation Dynamics in the Nile Basin, 2002–2011

Evapotranspiration (ET) is a vital component in land-atmosphere interactions. In drylands, over 90% of annual rainfall evaporates. The Nile Basin in Africa is about 42% dryland in a region experiencing rapid population growth and development. The relationship of ET with climate, vegetation and land cover in the basin during 2002–2011 is analyzed using thermal-based Simplified Surface Energy Balance Operational (SSEBop) ET, Normalized Difference Vegetation Index (NDVI)-based MODIS Terrestrial (MOD16) ET, MODIS-derived NDVI as a proxy for vegetation productivity and rainfall from Tropical Rainfall Measuring Mission (TRMM). Interannual variability and trends are analyzed using established statistical methods. Analysis based on thermal-based ET revealed that >50% of the study area exhibited negative ET anomalies for 7 years (2009, driest), while >60% exhibited positive ET anomalies for 3 years (2007, wettest). NDVI-based monthly ET correlated strongly (r > 0.77) with vegetation than thermal-based ET (0.52 < r < 0.73) at p < 0.001. Climate-zone averaged thermal-based ET anomalies positively correlated (r = 0.6, p < 0.05) with rainfall in 4 of the 9 investigated climate zones. Thermal-based and NDVI-based ET estimates revealed minor discrepancies over rainfed croplands (60 mm/yr higher for thermal-based ET), but a significant divergence over wetlands (440 mm/yr higher for thermal-based ET). Only 5% of the study area exhibited statistically significant trends in ET.

Survival analysis on data streams: Analyzing temporal events in dynamically changing environments

Abstract In this paper, we introduce a method for survival analysis on data streams. Survival analysis (also known as event history analysis) is an established statistical method for the study of temporal “events” or, more specifically, questions regarding the temporal distribution of the occurrence of events and their dependence on covariates of the data sources. To make this method applicable in the setting of data streams, we propose an adaptive variant of a model that is closely related to the well-known Cox proportional hazard model. Adopting a sliding window approach, our method continuously updates its parameters based on the event data in the current time window. As a proof of concept, we present two case studies in which our method is used for different types of spatio-temporal data analysis, namely, the analysis of earthquake data and Twitter data. In an attempt to explain the frequency of events by the spatial location of the data source, both studies use the location as covariates of the sources.

English

  It may be intuitive to assume that the quality of issue corrective actions (iCA) for customers with mobile terminal (MT) issues can be improved by collecting samples of the faulty products in question for verifying the issues reported from the field. This study was created to test this hypothesis using three different established statistical analysis methods: binary logistic regression, Kruskal-Wallis statistical tests and Mood’s Median Test. The methods are used because the data collected from in-house database tool is categorical. In this study, these methods were employed to test if the presence or absence of corrective actions process (CAP) samples had any effect on the perceived quality of issue corrective actions (P-QoiCA) or any effect on the perceived quality of issue resolution time (P-QoiRT) for service to the customers. This study also examined the effect of the quality of the collected samples (QoSa) on the absolute issue resolution time (iRT). The study checked if the frequency of requests for samples differs between MT products of different software platforms (SW_P). Additionally the paper investigated how the sample collection turnaround time (SC-TAT) differed in different sales areas (SA). The main findings were that the collected samples had no measurable effect upon the P-QoiCA or on P-QoiRT.  Also it has been shown that QoSa and the SC-TAT had no effect on the absolute time (iRT) to resolve customers’ issues. In addition to the aforementioned findings, the study noted as an aside that the Moods’ median test found significant differences in iRT between different software platforms (SW_ P).   Key words: Corrective actions process (CAP)-samples, non-parametric tests, log-normal distribution, logistic regression, perceived quality.

IBBiG: iterative binary bi-clustering of gene sets

Motivation: Meta-analysis of genomics data seeks to identify genes associated with a biological phenotype across multiple datasets; however, merging data from different platforms by their features (genes) is challenging. Meta-analysis using functionally or biologically characterized gene sets simplifies data integration is biologically intuitive and is seen as having great potential, but is an emerging field with few established statistical methods.Results: We transform gene expression profiles into binary gene set profiles by discretizing results of gene set enrichment analyses and apply a new iterative bi-clustering algorithm (iBBiG) to identify groups of gene sets that are coordinately associated with groups of phenotypes across multiple studies. iBBiG is optimized for meta-analysis of large numbers of diverse genomics data that may have unmatched samples. It does not require prior knowledge of the number or size of clusters. When applied to simulated data, it outperforms commonly used clustering methods, discovers overlapping clusters of diverse sizes and is robust in the presence of noise. We apply it to meta-analysis of breast cancer studies, where iBBiG extracted novel gene set—phenotype association that predicted tumor metastases within tumor subtypes.Availability: Implemented in the Bioconductor package iBBiGContact: aedin@jimmy.harvard.edu

The Amino Acid-Polyamine-Organocation Superfamily

The amino acid-polyamine-organocation (APC) superfamily has been shown to include five recognized families, four of which are specific for amino acids and their derivatives. Recent high-resolution X-ray crystallographic data have shown that four additional transporter families (BCCT, TC No. 2.A.15; SSS, 2.A.21; NSS, 2.A.22; and NCS1, 2.A.39), transporting a wide range of solutes, exhibit sufficiently similar folds to suggest a common evolutionary origin. We have used established statistical methods, based on sequence similarity, to show that these families are, in fact, members of the APC superfamily. We also identify two additional families (NCS2, 2.A.40; SulP, 2.A.53) as being members of this superfamily. Repeat sequences, each having five transmembrane α-helical segments and arising via ancient intragenic duplications, are demonstrated for all of these families, further strengthening the conclusion of homology. The APC superfamily appears to be the second largest superfamily of secondary carriers, the largest being the major facilitator superfamily (MFS). Although the topology of the members of the APC superfamily differs from that of the MFS, both families appear to have arisen from a common ancestral 2 TMS hairpin structure that underwent intragenic triplication followed by loss of a TMS in the APC family, to give the repeat units that are characteristic of these two superfamilies.

Phase 0 Clinical Chemoprevention Trial of the Akt Inhibitor SR13668

SR13668, an orally active Akt pathway inhibitor, has demonstrated cancer chemopreventive potential in preclinical studies. To accelerate the clinical development of this promising agent, we designed and conducted the first-ever phase 0 chemoprevention trial to evaluate and compare the effects of food and formulation on SR13668 bioavailability. Healthy adult volunteers were randomly assigned to receive a single, 38-mg oral dose of SR13668 in one of five different formulations, with or without food. On the basis of existing animal data, SR13668 in a PEG400/Labrasol oral solution was defined as the reference formulation. Blood samples were obtained pre- and post-agent administration for pharmacokinetic analyses. Area under the plasma concentration-time curve (AUC(0-∞)) was defined as the primary endpoint. Data were analyzed and compared using established statistical methods for phase 0 trials with a limited sample size. Participants (n = 20) were rapidly accrued over a 5-month period. Complete pharmacokinetic data were available for 18 randomized participants. AUC(0-∞) values were highest in the fed state (range = 122-439 ng/mL × hours) and were statistically significantly different across formulations (P = 0.007), with Solutol HS15 providing the highest bioavailability. SR13668 time to peak plasma concentration (3 hours; range, 2-6 hours) and half-life were (11.2 ± 3.1 hours) were not formulation-dependent. Using a novel, highly efficient study design, we rapidly identified a lead formulation of SR13668 for further clinical testing. Our findings support application of the phase 0 trial paradigm to accelerate chemoprevention agent development.

Abstract CN02-03: Phase 0 chemoprevention trial of SR13668, a novel AKT inhibitor, in normal volunteers

Abstract Purpose: SR13668, an orally active AKT pathway inhibitor, has demonstrated cancer chemopreventive potential in preclinical studies. To accelerate the clinical development of this promising agent, we designed and conducted the first-ever phase 0 chemoprevention trial to evaluate and compare the effects of food and formulation on SR13668 bioavailability. Patients and Methods: Healthy adult volunteers were randomly assigned to receive a single 38-mg oral dose of SR13668 in one of five different formulations, with or without food. Based on existing animal data, SR13668 in a PEG400/Labrasol@ oral solution (formulation 1) was defined as the reference formulation. Four self-emulsifying solid dispersion test formulations in capsules were also investigated: Solutol@ HS 15 (formulation 2), Solutol@ HS 15/Vitamin E TGPS (50/50 w/w %) (formulation 3), Vitamin E TGPS (formulation 4), and Myrj 53 (formulation 5). SR163668 bioavailability was compared using a two-stage study design. In stage I, participants were randomized to receive the formulation 1 with or without a high-fat meal after an overnight fast. In Stage II, participants were randomly assigned to receive the test formulations, using the preferred dietary condition identified in stage I. Blood samples were obtained pre- and post-agent administration for pharmacokinetic analyses. Area under the plasma concentration-time curve (AUC) was defined as the primary endpoint. Data were analyzed and compared using established statistical methods for phase 0 trials with a limited sample size. Results: Participants (N=20) were rapidly accrued over a 5-month period. Complete pharmacokinetic data were available for 18 randomized participants. The time to peak plasma concentration and half-life values for SR13668 were 3 hours (range, 2 — 6 hours) and 11.2 ± 3.1 hours, respectively, and were not dependent on the formulation. In stage I, the oral bioavailability of SR13668 in formulation 1 was greater under the fed versus fasting dietary state (p = 0.05, Wilcoxon rank-sum test). Therefore, the fed state was selected as the preferred dietary condition for stage II. In stage II, the bioavailability of SR13668 for each of the test formulations were compared with the bioavailability of formulation 1 in the fed state. While not significantly different from formulation 1, formulation 2 AUC0-24h was significantly different (p = 0.004, ANOVA) from formulation 3, 4, and 5 (AUC0-24h, formulation 1- 222 ± 53 ng/ml·hr; AUC0-24h, formulation 2 — 302 ± 78 ng/ml·hr; AUC-177±39ng/ml·hr; AUC-144±29 0-24h, formulation30-24h, formulation 4 ng/ml·hr; AUC0-24h, formulation 5- 98 ± 21 ng/ml·hr). Formulation 1 AUC0-24h was almost 40% lower than the value achieved with formulation 2, and not significantly different from formulation 3, 4, and 5. Thus, formulation 2 provided the highest bioavailability of SR13668. Conclusions: Using a novel, highly efficient study design, we rapidly identified a lead formulation (Solutol@ HS 15) of SR13668 for further clinical testing. Our findings support application of the phase 0 trial paradigm to accelerate chemoprevention agent development. Supported by NIH grants N01-CN-35000, CA15083-34C3 and UL1 RR024150. Citation Information: Cancer Prev Res 2010;3(12 Suppl):CN02-03.

Darwinian and demographic forces affecting human protein coding genes.

Past demographic changes can produce distortions in patterns of genetic variation that can mimic the appearance of natural selection unless the demographic effects are explicitly removed. Here we fit a detailed model of human demography that incorporates divergence, migration, admixture, and changes in population size to directly sequenced data from 13,400 protein coding genes from 20 European-American and 19 African-American individuals. Based on this demographic model, we use several new and established statistical methods for identifying genes with extreme patterns of polymorphism likely to be caused by Darwinian selection, providing the first genome-wide analysis of allele frequency distributions in humans based on directly sequenced data. The tests are based on observations of excesses of high frequency-derived alleles, excesses of low frequency-derived alleles, and excesses of differences in allele frequencies between populations. We detect numerous new genes with strong evidence of selection, including a number of genes related to psychiatric and other diseases. We also show that microRNA controlled genes evolve under extremely high constraints and are more likely to undergo negative selection than other genes. Furthermore, we show that genes involved in muscle development have been subject to positive selection during recent human history. In accordance with previous studies, we find evidence for negative selection against mutations in genes associated with Mendelian disease and positive selection acting on genes associated with several complex diseases.

Modeling the effect of time-dependent exposure on intensive care unit mortality

To illustrate modern survival models with focus on the temporal dynamics of intensive care data. A typical situation is given in which time-dependent exposures and competing events are present. We briefly review the following established statistical methods: logistic regression, regression models for event-specific hazards and the subdistribution hazard. These approaches are compared by showing advantages as well as disadvantages. All methods are applied to real data from a study of day-by-day ICU surveillance. Standard logistic regression ignores the time-dependent nature of the data and is only a crude approach. Cumulative hazards and probability plots add important information and provide a deep insight into the temporal dynamics. This paper might help to encourage researchers working in hospital epidemiology to apply adequate statistical models to complex medical questions.

Extracting Protein-Protein Interactions from MEDLINE using the Hidden Vector State model

A major challenge in text mining for biomedicine is automatically extracting protein-protein interactions from the vast amount of biomedical literature. We have constructed an information extraction system based on the Hidden Vector State (HVS) model for protein-protein interactions. The HVS model can be trained using only lightly annotated data whilst simultaneously retaining sufficient ability to capture the hierarchical structure. When applied in extracting protein-protein interactions, we found that it performed better than other established statistical methods and achieved 61.5% in F-score with balanced recall and precision values. Moreover, the statistical nature of the pure data-driven HVS model makes it intrinsically robust and it can be easily adapted to other domains.

Region-specific network plasticity in simulated and living cortical networks: comparison of the center of activity trajectory (CAT) with other statistics

Electrically interfaced cortical networks cultured in vitro can be used as a model for studying the network mechanisms of learning and memory. Lasting changes in functional connectivity have been difficult to detect with extracellular multi-electrode arrays using standard firing rate statistics. We used both simulated and living networks to compare the ability of various statistics to quantify functional plasticity at the network level. Using a simulated integrate-and-fire neural network, we compared five established statistical methods to one of our own design, called center of activity trajectory (CAT). CAT, which depicts dynamics of the location-weighted average of spatiotemporal patterns of action potentials across the physical space of the neuronal circuitry, was the most sensitive statistic for detecting tetanus-induced plasticity in both simulated and living networks. By reducing the dimensionality of multi-unit data while still including spatial information, CAT allows efficient real-time computation of spatiotemporal activity patterns. Thus, CAT will be useful for studies in vivo or in vitro in which the locations of recording sites on multi-electrode probes are important.

Understanding mechanism of spatial ecological phenomena: a preface to the special feature on “Spatial statistics”

Understanding of the mechanism underlying the spatial distribution of animal and plant populations is one of the central themes in ecology. Ecological patterns tend to have scale dependency, which is problematic for a comparative approach, but could itself indicate the underlying ecology governed by biological and physical phenomena operating on characteristic spatial scales (O’Neill 1989; Powell 1989; Levin 1992; Tuda 1993). Theoretical studies such as host–parasitoid models by Hassell et al. (1991) and Comins et al. (1992) have explored population behavior in spatially explicit lattice space. Similar phenomena can also be found in other fields of science (e.g., Kaneko 1984). There has been a gap between the advanced theory and understanding of crude spatial data in spatial ecology. Recent advances in spatial statistics, however, shed light on empirical spatial ecology (e.g., Dale et al. 2002; Perry et al. 2002). Direct observation of marked organisms (e.g., Tuda and Shima 2002) and bottle experiments (Huffaker 1958; recent reviews, Bonsall and Hassell 2005; Cadotte et al. 2005; Jessup et al. 2005; Holyoak and Lawler 2005; Sabelis et al. 2005; Tuda and Shimada 2005) may be the most basic and simple approach to capturing spatial mechanisms in ecology. Nevertheless, spatial statistics proposed or applied here can bring greater understanding of mechanisms when larger sets of data with a wider range of spatial scales are available. In practice, information we obtain may be limited to a single snapshot of multiple transitional states or quantitative attributes of a single state on a map. On the contrary, richer information may be available as timeserial data of spatial patterns more like an animation. Here, I have invited papers tackling ecological spatial dynamics: Schlicht and Iwasa (2007) aim to handle both snapshots and animations of forest disturbance patterns. The other two papers (Yamanaka et al. 2007; Kim et al. 2007) elucidate spatial correlations from animations of quantitative spatio-temporal data for multiple transitional states, making full use of established spatial statistical methods. Schlicht and Iwasa (2007) proposed two statistical methods and tested their relevancy by fitting them to three hypothetical spatial patterns, generated by simulating models that employed different numbers of states (two, three, and continuous states respectively). The first method that detects directionality of regeneration waves succeeded in distinguishing two-state and three-state hypothetical data, whereas the second method, which detects spatial-scale dependency in the variance of fractions of states, distinguished the three-state and continuous-state hypothetical data. The kind of data most suitable for the two methods is the temporal dynamics of spatial patterns for the first method and single snapshots of spatial patterns for the second method. The authors applied these methods to observed forest disturbance patterns and detected proper models. This study not only provides forest ecologists with the most parsimonious descriptive models that may narrow down possible ecological mechanisms underlying observed spatial patterns, but can also be extended to analyses of other spatial distributions of plants and animals. Yamanaka et al. (2007) tested two hypotheses, resource concentration and reaction diffusion, on the mechanism inducing spatial distribution of the ragweed beetle. The nonparametric autoand cross-correlation functions, developed by Bjornstad and his colleagues, were applied to the 3-year seasonal spatial dynamics of a resource (the ragweed) and its herbivorous predator (the ragweed beetle). The resource concentration hypothesis was only weakly supported. The nonparametric crosscorrelation with time lags between the weed and the beetle, on the other hand, supported trophic interaction between the two at a certain spatial lag, thereby leading Spatial statistics in ecology

The association between environmental events and self‐injurious behaviour in Cornelia de Lange syndrome

There has been limited empirical research into the environmental causes of self-injury in Cornelia de Lange syndrome. The present study examined the variability of self-injurious behaviour in Cornelia de Lange syndrome across environmental setting events. Additionally, the association between setting events and more specific environmental events was examined. A descriptive analysis of observational data on eight children with Cornelia de Lange syndrome aged between 4 and 14 years was carried out. The association between self-injurious behaviour and four environmental setting events and between specific environmental events and setting events was examined using established statistical methods for observational data. Seven out of eight of the participants showed at least one form of self-injurious behaviour that was associated with a particular setting event. The study also demonstrated that the relationship between setting events and environmental events is extremely variable across individuals. Self-injurious behaviour in some individuals with Cornelia de Lange syndrome is associated with environmental events although the precise nature of the association warrants clarification. Using broad setting events as a methodological tool in isolation provides some insight into the role of specific environmental factors in maintaining self-injurious behaviour but the integrity of setting events must be established.

The interaction effects of binary mixtures of benzene and toluene on the developing heart of medaka ( Oryzias latipes)

The United States Environmental Protection Agency (USEPA) has pursued the estimation of risk of adverse health effects from exposure to chemical mixtures since the early 1980s. Methods used to calculate risk estimates of mixtures were often based on single chemical information that required assumptions of dose-addition or response-addition and did not consider possible changes in response due to interaction effects among chemicals. Full factorial designs for laboratory studies can produce interactions information, but these are expensive to perform and may not provide the information needed to evaluate specific environmentally relevant mixtures. In this research, groups of Japanese medaka ( Oryzias latipes) embryos were exposed to binary mixtures of benzene and toluene as well as to each of these chemicals alone. Endpoint specific dose–response models were built for the hydrocarbon mixture under an assumption of dose-additivity, using the single chemical dose–response information on benzene and toluene. The endpoints included heart rate, heart rate progression, and lethality. Results included a synergistic response for heart rate at 72 h of development, and either additivity or antagonism for all other endpoints at 96 h of development. This work uses an established statistical method to evaluate the toxicity of an environmentally relevant mixture to ascertain whether interaction effects are occurring, thus providing additional information on toxicity.

Comparison of equations describing band broadening in high-performance liquid chromatography

Several models are well established that describe band broadening in gas and liquid chromatography, including those due to Van Deemter and Knox. Comparison of competing models is complicated if raw data are noisy or if the equations to be fitted to data contain many adjustable parameters. This paper describes a comparison of fitting the Van Deemter, Knox and other equations to low noise data gathered during the separation of propyl- and methylparaben by HPLC. Equations are compared using established statistical methods, including analysis of residuals, inference of parameter estimates and Akaikes Information Criterion for model identification. This work indicates that equations that account for non-linear band broadening at elevated mobile phase velocities are more successful at describing the relationship between height equivalent to a theoretical plate, H, and the velocity of the mobile phase, u.

Empirical and substantive models, the Bayesian paradigm, and meta-analysis in functional brain imaging

Functional neuroimaging research is currently rediscovering and adapting established statistical methods for its use, including design of experiments, the general linear model, contrasts, random field theory, longitudinal models, Fourier analysis, and general signal and image processing methods. This brief paper gives an example of comparative performance of five different statistical models applied to the same set of data generated in an fMRI study of motor cortex. These methods include a two-sample t-statistic, a Kolmogorov-Smirnov statistic, a principal component/canonical variates approach, a pruned feed-forward artificial neural network with one hidden layer, and a frequency domain regression convolution model allowing for spatially varying hemodynamic responses. Produced by essentially empirical statistical models, there appear to be more similarities than differences in these spatial activation patterns, yet all lack explicit incorporation of substantive prior information. Distinctions are drawn between exploratory models for hypothesis generation and confirmatory models for hypothesis testing. In addition, the Bayesian paradigm helps to combine empirical and substantive models, and meta-analysis provides a rational means by which to combine information over a range of similar results affected minimally by publication bias.

Prediction of elderly seating comfort from physical condition, chair type and gender

Many persons over 60 years of age have unique problems sitting on chairs. These problems which are caused by decreased mobility, strength or disease suggest that the chairs should be selected carefully with some scientific basis that incorporates the varying demands of the elderly community. Identification of the proper chair for a particular individual, especially if it has to be low cost and nonmotorized, is a difficult job for institutions, elderly themselves, their families and the furniture industry. Keeping in view these problems a prediction model is developed using established statistical methods to predict comfort of sitting of a particular individual sitting in a particular chair. The study is based on experimental data collected on 18 male and female subjects over the age of 64. The equation uses chair type, body fat %, i.e. above average fat etc., body size, i.e. lean etc., and gender as inputs. The equation was tested on new female and male elderly subjects sitting on two new chairs. The model holds out excellent on the validation despite a low coefficient of determination value obtained for the equation. The paper not only presents a comfort prediction model but also shows a new direction for elderly seating research by employing statistical methods and elderly variables which do not appear in the published literature on elderly seating in the manner presented in this paper. This model will have wide applications especially for practioners (e.g., human factors engineers, occupational therapists, product developers) who need such information in a hurry. This approach also has heuristic value for the researcher in providing guidelines for the testing of other critical variables that may impact on the use of a chair or of other pieces of equipment.

Improving model selection by nonconvergent methods

Many techniques for model selection in the field of neural networks correspond to well established statistical methods. For example, architecture modifications based on test variables calculated after convergence of the training process can be viewed as part of a hypothesis testing search, and the use of complexity penalty terms is essentially a type of regularization or biased regression. The method of “stopped” or “cross-validation” training, on the other hand, in which an oversized network is trained until the error on a further validation set of examples deteriorates, then training is stopped, is a true innovation since model selection doesn't require convergence of the training process. Here, the training process is used to perform a directed search of the parameter space for a model which doesn't overfit the data and thus demonstrates superior generalization performance. In this paper we show that this performance can be significantly enhanced by expanding the “nonconvergent method” of stopped training to include dynamic topology modifications (dynamic weight pruning) and modified complexity penalty term methods in which the weighting of the penalty term is adjusted during the training process. On an extensive sequence of simulation examples we demonstrate the general superiority of the “extended” nonconvergent methods compared to classical penalty term methods, simple stopped training, and methods which only vary the number of hidden units.