Using Bayesian methods for the estimation of uncertainty in complex statistical models

The research objectives of this thesis were to contribute to Bayesian statistical methodology by contributing to risk assessment statistical methodology, and to spatial and spatio-temporal methodology, by modelling error structures using complex hierarchical models. Specifically, I hoped to consider two applied areas, and use these applications as a springboard for developing new statistical methods as well as undertaking analyses which might give answers to particular applied questions. Thus, this thesis considers a series of models, firstly in the context of risk assessments for recycled water, and secondly in the context of water usage by crops. The research objective was to model error structures using hierarchical models in two problems, namely risk assessment analyses for wastewater, and secondly, in a four dimensional dataset, assessing differences between cropping systems over time and over three spatial dimensions. The aim was to use the simplicity and insight afforded by Bayesian networks to develop appropriate models for risk scenarios, and again to use Bayesian hierarchical models to explore the necessarily complex modelling of four dimensional agricultural data. The specific objectives of the research were to develop a method for the calculation of credible intervals for the point estimates of Bayesian networks; to develop a model structure to incorporate all the experimental uncertainty associated with various constants thereby allowing the calculation of more credible credible intervals for a risk assessment; to model a single day’s data from the agricultural dataset which satisfactorily captured the complexities of the data; to build a model for several days’ data, in order to consider how the full data might be modelled; and finally to build a model for the full four dimensional dataset and to consider the timevarying nature of the contrast of interest, having satisfactorily accounted for possible spatial and temporal autocorrelations. This work forms five papers, two of which have been published, with two submitted, and the final paper still in draft. The first two objectives were met by recasting the risk assessments as directed, acyclic graphs (DAGs). In the first case, we elicited uncertainty for the conditional probabilities needed by the Bayesian net, incorporated these into a corresponding DAG, and used Markov chain Monte Carlo (MCMC) to find credible intervals, for all the scenarios and outcomes of interest. In the second case, we incorporated the experimental data underlying the risk assessment constants into the DAG, and also treated some of that data as needing to be modelled as an ‘errors-invariables’ problem [Fuller, 1987]. This illustrated a simple method for the incorporation of experimental error into risk assessments. In considering one day of the three-dimensional agricultural data, it became clear that geostatistical models or conditional autoregressive (CAR) models over the three dimensions were not the best way to approach the data. Instead CAR models are used with neighbours only in the same depth layer. This gave flexibility to the model, allowing both the spatially structured and non-structured variances to differ at all depths. We call this model the CAR layered model. Given the experimental design, the fixed part of the model could have been modelled as a set of means by treatment and by depth, but doing so allows little insight into how the treatment effects vary with depth. Hence, a number of essentially non-parametric approaches were taken to see the effects of depth on treatment, with the model of choice incorporating an errors-in-variables approach for depth in addition to a non-parametric smooth. The statistical contribution here was the introduction of the CAR layered model, the applied contribution the analysis of moisture over depth and estimation of the contrast of interest together with its credible intervals. These models were fitted using WinBUGS [Lunn et al., 2000]. The work in the fifth paper deals with the fact that with large datasets, the use of WinBUGS becomes more problematic because of its highly correlated term by term updating. In this work, we introduce a Gibbs sampler with block updating for the CAR layered model. The Gibbs sampler was implemented by Chris Strickland using pyMCMC [Strickland, 2010]. This framework is then used to consider five days data, and we show that moisture in the soil for all the various treatments reaches levels particular to each treatment at a depth of 200 cm and thereafter stays constant, albeit with increasing variances with depth. In an analysis across three spatial dimensions and across time, there are many interactions of time and the spatial dimensions to be considered. Hence, we chose to use a daily model and to repeat the analysis at all time points, effectively creating an interaction model of time by the daily model. Such an approach allows great flexibility. However, this approach does not allow insight into the way in which the parameter of interest varies over time. Hence, a two-stage approach was also used, with estimates from the first-stage being analysed as a set of time series. We see this spatio-temporal interaction model as being a useful approach to data measured across three spatial dimensions and time, since it does not assume additivity of the random spatial or temporal effects.

Identifying parent perspectives to guide active travel intervention with small populations : case study in Denver, Colorado

This paper investigates a strategy for guiding school-based active travel intervention. School-based active travel programs address the travel behaviors and perceptions of small target populations (i.e., at individual schools) so they can encourage people to walk or bike. Thus, planners need to know as much as possible about the behaviors and perceptions of their target populations. However, existing strategies for modeling travel behavior and segmenting audiences typically work with larger populations and may not capture the attitudinal diversity of smaller groups. This case study used Q technique to identify salient travel-related attitude types among parents at an elementary school in Denver, Colorado; 161 parents presented their perspectives about school travel by rank-ordering 36 statements from strongly disagree to strongly agree in a normalized distribution, single centered around no opinion. Thirty-nine respondents' cases were selected for case-wise cluster analysis in SPSS according to criteria that made them most likely to walk: proximity to school, grade, and bus service. Analysis revealed five core perspectives that were then correlated with the larger respondent pool: optimistic walkers, fair-weather walkers, drivers of necessity, determined drivers, and fence sitters. Core perspectives are presented—characterized by parents' opinions, personal characteristics, and reported travel behaviors—and recommendations are made for possible intervention approaches. The study concludes that Q technique provides a fine-grained assessment of travel behavior for small populations, which would benefit small-scale behavioral interventions

Body composition and its relationship to metabolic risk factors in Asian children

Obesity is a major public health problem in both developed and developing countries. The body mass index (BMI) is the most common index used to define obesity. The universal application of the same BMI classification across different ethnic groups is being challenged due to the inability of the index to differentiate fat mass (FM) and fat�]free mass (FFM) and the recognized ethnic differences in body composition. A better understanding of the body composition of Asian children from different backgrounds would help to better understand the obesity�]related health risks of people in this region. Moreover, the limitations of the BMI underscore the necessity to use where possible, more accurate measures of body fat assessment in research and clinical settings in addition to BMI, particularly in relation to the monitoring of prevention and treatment efforts. The aim of the first study was to determine the ethnic difference in the relationship between BMI and percent body fat (%BF) in pre�]pubertal Asian children from China, Lebanon, Malaysia, the Philippines, and Thailand. A total of 1039 children aged 8�]10 y were recruited using a non�]random purposive sampling approach aiming to encompass a wide BMI range from the five countries. Percent body fat (%BF) was determined using the deuterium dilution technique to quantify total body water (TBW) and subsequently derive proportions of FM and FFM. The study highlighted the sex and ethnic differences between BMI and %BF in Asian children from different countries. Girls had approximately 4.0% higher %BF compared with boys at a given BMI. Filipino boys tended to have a lower %BF than their Chinese, Lebanese, Malay and Thai counterparts at the same age and BMI level (corrected mean %BF was 25.7�}0.8%, 27.4�}0.4%, 27.1�}0.6%, 27.7�}0.5%, 28.1�}0.5% for Filipino, Chinese, Lebanese, Malay and Thai boys, respectively), although they differed significantly from Thai and Malay boys. Thai girls had approximately 2.0% higher %BF values than Chinese, Lebanese, Filipino and Malay counterparts (however no significant difference was seen among the four ethnic groups) at a given BMI (corrected mean %BF was 31.1�}0.5%, 28.6�}0.4%, 29.2�}0.6%, 29.5�}0.6%, 29.5�}0.5% for Thai, Chinese, Lebanese, Malay and Filipino girls, respectively). However, the ethnic difference in BMI�]%BF relationship varied by BMI. Compared with Caucasians, Asian children had a BMI 3�]6 units lower for a given %BF. More than one third of obese Asian children in the study were not identified using the WHO classification and more than half were not identified using the International Obesity Task Force (IOTF) classification. However, use of the Chinese classification increased the sensitivity by 19.7%, 18.1%, 2.3%, 2.3%, and 11.3% for Chinese, Lebanese, Malay, Filipino and Thai girls, respectively. A further aim of the first study was to determine the ethnic difference in body fat distribution in pre�]pubertal Asian children from China, Lebanon, Malaysia, and Thailand. The skin fold thicknesses, height, weight, waist circumference (WC) and total adiposity (as determined by deuterium dilution technique) of 922 children from the four countries was assessed. Chinese boys and girls had a similar trunk�]to�]extremity skin fold thickness ratio to Thai counterparts and both groups had higher ratios than the Malays and Lebanese at a given total FM. At a given BMI, both Chinese and Thai boys and girls had a higher WC than Malays and Lebanese (corrected mean WC was 68.1�}0.2 cm, 67.8�}0.3 cm, 65.8�}0.4 cm, 64.1�}0.3 cm for Chinese, Thai, Lebanese and Malay boys, respectively; 64.2�}0.2 cm, 65.0�}0.3 cm, 62.9�}0.4 cm, 60.6�}0.3 cm for Chinese, Thai, Lebanese and Malay girls, respectively). Chinese boys and girls had lower trunk fat adjusted subscapular/suprailiac skinfold ratio compared with Lebanese and Malay counterparts. The second study aimed to develop and cross�]validate bioelectrical impedance analysis (BIA) prediction equations of TBW and FFM for Asian pre�]pubertal children from China, Lebanon, Malaysia, the Philippines, and Thailand. Data on height, weight, age, gender, resistance and reactance measured by BIA were collected from 948 Asian children (492 boys and 456 girls) aged 8�]10 y from the five countries. The deuterium dilution technique was used as the criterion method for the estimation of TBW and FFM. The BIA equations were developed from the validation group (630 children randomly selected from the total sample) using stepwise multiple regression analysis and cross�]validated in a separate group (318 children) using the Bland�]Altman approach. Age, gender and ethnicity influenced the relationship between the resistance index (RI = height2/resistance), TBW and FFM. The BIA prediction equation for the estimation of TBW was: TBW (kg) = 0.231�~Height2 (cm)/resistance (ƒ¶) + 0.066�~Height (cm) + 0.188�~Weight (kg) + 0.128�~Age (yr) + 0.500�~Sex (male=1, female=0) . 0.316�~Ethnicity (Thai ethnicity=1, others=0) �] 4.574, and for the estimation of FFM: FFM (kg) = 0.299�~Height2 (cm)/resistance (ƒ¶) + 0.086�~Height (cm) + 0.245�~Weight (kg) + 0.260�~Age (yr) + 0.901�~Sex (male=1, female=0) �] 0.415�~Ethnicity (Thai ethnicity=1, others=0) �] 6.952. The R2 was 88.0% (root mean square error, RSME = 1.3 kg), 88.3% (RSME = 1.7 kg) for TBW and FFM equation, respectively. No significant difference between measured and predicted TBW and between measured and predicted FFM for the whole cross�]validation sample was found (bias = �]0.1�}1.4 kg, pure error = 1.4�}2.0 kg for TBW and bias = �]0.2�}1.9 kg, pure error = 1.8�}2.6 kg for FFM). However, the prediction equation for estimation of TBW/FFM tended to overestimate TBW/FFM at lower levels while underestimate at higher levels of TBW/FFM. Accuracy of the general equation for TBW and FFM compared favorably with both BMI�]specific and ethnic�]specific equations. There were significant differences between predicted TBW and FFM from external BIA equations derived from Caucasian populations and measured values in Asian children. There were three specific aims of the third study. The first was to explore the relationship between obesity and metabolic syndrome and abnormalities in Chinese children. A total of 608 boys and 800 girls aged 6�]12 y were recruited from four cities in China. Three definitions of pediatric metabolic syndrome and abnormalities were used, including the International Diabetes Federation (IDF) and National Cholesterol Education Program (NCEP) definition for adults modified by Cook et al. and de Ferranti et al. The prevalence of metabolic syndrome varied with different definitions, was highest using the de Ferranti definition (5.4%, 24.6% and 42.0%, respectively for normal�]weight, overweight and obese children), followed by the Cook definition (1.5%, 8.1%, and 25.1%, respectively), and the IDF definition (0.5%, 1.8% and 8.3%, respectively). Overweight and obese children had a higher risk of developing the metabolic syndrome compared to normal�]weight children (odds ratio varied with different definitions from 3.958 to 6.866 for overweight children, and 12.640�]26.007 for obese children). Overweight and obesity also increased the risk of developing metabolic abnormalities. Central obesity and high triglycerides (TG) were the most common while hyperglycemia was the least frequent in Chinese children regardless of different definitions. The second purpose was to determine the best obesity index for the prediction of cardiovascular (CV) risk factor clustering across a 2�]y follow�]up among BMI, %BF, WC and waist�]to�]height ratio (WHtR) in Chinese children. Height, weight, WC, %BF as determined by BIA, blood pressure, TG, high�]density lipoprotein cholesterol (HDL�]C), and fasting glucose were collected at baseline and 2 years later in 292 boys and 277 girls aged 8�]10 y. The results showed the percentage of children who remained overweight/obese defined on the basis of BMI, WC, WHtR and %BF was 89.7%, 93.5%, 84.5%, and 80.4%, respectively after 2 years. Obesity indices at baseline significantly correlated with TG, HDL�]C, and blood pressure at both baseline and 2 years later with a similar strength of correlations. BMI at baseline explained the greatest variance of later blood pressure. WC at baseline explained the greatest variance of later HDL�]C and glucose, while WHtR at baseline was the main predictor of later TG. Receiver�]operating characteristic (ROC) analysis explored the ability of the four indices to identify the later presence of CV risk. The overweight/obese children defined on the basis of BMI, WC, WHtR or %BF were more likely to develop CV risk 2 years later with relative risk (RR) scores of 3.670, 3.762, 2.767, and 2.804, respectively. The final purpose of the third study was to develop age�] and gender�]specific percentiles of WC and WHtR and cut�]off points of WC and WHtR for the prediction of CV risk in Chinese children. Smoothed percentile curves of WC and WHtR were produced in 2830 boys and 2699 girls aged 6�]12 y randomly selected from southern and northern China using the LMS method. The optimal age�] and gender�]specific thresholds of WC and WHtR for the prediction of cardiovascular risk factors clustering were derived in a sub�]sample (n=1845) by ROC analysis. Age�] and gender�]specific WC and WHtR percentiles were constructed. The WC thresholds were at the 90th and 84th percentiles for Chinese boys and girls, respectively, with sensitivity and specificity ranging from 67.2% to 83.3%. The WHtR thresholds were at the 91st and 94th percentiles for Chinese boys and girls, respectively, with sensitivity and specificity ranging from 78.6% to 88.9%. The cut�]offs of both WC and WHtR were age�] and gender�]dependent. In conclusion, the current thesis quantifies the ethnic differences in the BMI�]%BF relationship and body fat distribution between Asian children from different origins and confirms the necessity to consider ethnic differences in body composition when developing BMI and other obesity index criteria for obesity in Asian children. Moreover, ethnicity is also important in BIA prediction equations. In addition, WC and WHtR percentiles and thresholds for the prediction of CV risk in Chinese children differ from other populations. Although there was no advantage of WC or WHtR over BMI or %BF in the prediction of CV risk, obese children had a higher risk of developing the metabolic syndrome and abnormalities than normal�]weight children regardless of the obesity index used.

Ergodic capacity of the exponentially correlated slotted amplify and forward relay channel

In this paper we analyze the performance degradation of slotted amplify-and-forward protocol in wireless environments with high node density where the number of relays grows asymptotically large. Channel gains between source-destination pairs in such networks can no longer be independent. We analyze the degradation of performance in such wireless environments where channel gains are exponentially correlated by looking at the capacity per channel use. Theoretical results for eigenvalue distribution and the capacity are derived and compared with the simulation results. Both analytical and simulated results show that the capacity given by the asymptotic mutual information decreases with the network density.

Sequential Fusion Using Correlated Decisions for Controlled Verification Errors

Fusion techniques have received considerable attention for achieving lower error rates with biometrics. A fused classifier architecture based on sequential integration of multi-instance and multi-sample fusion schemes allows controlled trade-off between false alarms and false rejects. Expressions for each type of error for the fused system have previously been derived for the case of statistically independent classifier decisions. It is shown in this paper that the performance of this architecture can be improved by modelling the correlation between classifier decisions. Correlation modelling also enables better tuning of fusion model parameters, ‘N’, the number of classifiers and ‘M’, the number of attempts/samples, and facilitates the determination of error bounds for false rejects and false accepts for each specific user. Error trade-off performance of the architecture is evaluated using HMM based speaker verification on utterances of individual digits. Results show that performance is improved for the case of favourable correlated decisions. The architecture investigated here is directly applicable to speaker verification from spoken digit strings such as credit card numbers in telephone or voice over internet protocol based applications. It is also applicable to other biometric modalities such as finger prints and handwriting samples.

Sequential fusion of decisions from adaptive and random samples for controlled verification of errors

Fusion techniques have received considerable attention for achieving performance improvement with biometrics. While a multi-sample fusion architecture reduces false rejects, it also increases false accepts. This impact on performance also depends on the nature of subsequent attempts, i.e., random or adaptive. Expressions for error rates are presented and experimentally evaluated in this work by considering the multi-sample fusion architecture for text-dependent speaker verification using HMM based digit dependent speaker models. Analysis incorporating correlation modeling demonstrates that the use of adaptive samples improves overall fusion performance compared to randomly repeated samples. For a text dependent speaker verification system using digit strings, sequential decision fusion of seven instances with three random samples is shown to reduce the overall error of the verification system by 26% which can be further reduced by 6% for adaptive samples. This analysis novel in its treatment of random and adaptive multiple presentations within a sequential fused decision architecture, is also applicable to other biometric modalities such as finger prints and handwriting samples.

Severity of driver injury and vehicle damage in traffic crashes at intersections: A Bayesian hierarchical analysis

Most crash severity studies ignored severity correlations between driver-vehicle units involved in the same crashes. Models without accounting for these within-crash correlations will result in biased estimates in the factor effects. This study developed a Bayesian hierarchical binomial logistic model to identify the significant factors affecting the severity level of driver injury and vehicle damage in traffic crashes at signalized intersections. Crash data in Singapore were employed to calibrate the model. Model fitness assessment and comparison using Intra-class Correlation Coefficient (ICC) and Deviance Information Criterion (DIC) ensured the suitability of introducing the crash-level random effects. Crashes occurring in peak time, in good street lighting condition, involving pedestrian injuries are associated with a lower severity, while those in night time, at T/Y type intersections, on right-most lane, and installed with red light camera have larger odds of being severe. Moreover, heavy vehicles have a better resistance on severe crash, while crashes involving two-wheel vehicles, young or aged drivers, and the involvement of offending party are more likely to result in severe injuries.

Modeling random effect and excess zeros in road traffic accident prediction

Poisson distribution has often been used for count like accident data. Negative Binomial (NB) distribution has been adopted in the count data to take care of the over-dispersion problem. However, Poisson and NB distributions are incapable of taking into account some unobserved heterogeneities due to spatial and temporal effects of accident data. To overcome this problem, Random Effect models have been developed. Again another challenge with existing traffic accident prediction models is the distribution of excess zero accident observations in some accident data. Although Zero-Inflated Poisson (ZIP) model is capable of handling the dual-state system in accident data with excess zero observations, it does not accommodate the within-location correlation and between-location correlation heterogeneities which are the basic motivations for the need of the Random Effect models. This paper proposes an effective way of fitting ZIP model with location specific random effects and for model calibration and assessment the Bayesian analysis is recommended.

Measuring voluntary dietary change in response to exercise : a focus on dietary fat

Traditional treatments for weight management have focussed on prescribed dietary restriction or regular exercise, or a combination of both. However recidivism for such prescribed treatments remains high, particularly among the overweight and obese. The aim of this thesis was to investigate voluntary dietary changes in the presence of prescribed mixed-mode exercise, conducted over 16 weeks. With the implementation of a single lifestyle change (exercise) it was postulated that the onerous burden of concomitant dietary and exercise compliance would be reduced, leading to voluntary lifestyle changes in such areas as diet. In addition, the failure of exercise as a single weight loss treatment has been reported to be due to compensatory energy intakes, although much of the evidence is from acute exercise studies, necessitating investigation of compensatory intakes during a long-term exercise intervention. Following 16 weeks of moderate intensity exercise, 30 overweight and obese (BMI≥25.00 kg.m-2) men and women showed small but statistically significant decreases in mean dietary fat intakes, without compensatory increases in other macronutrient or total energy intakes. Indeed total energy intakes were significantly lower for men and women following the exercise intervention, due to the decreases in dietary fat intakes. There was a risk that acceptance of the statistical validity of the small changes to dietary fat intakes may have constituted a Type 1 error, with false rejection of the Null hypothesis. Oro-sensory perceptions to changes in fat loads were therefore investigated to determine whether the measured dietary fat changes were detectable by the human palate. The ability to detect small changes in dietary fat provides sensory feedback for self-initiated dietary changes, but lean and overweight participants were unable to distinguish changes to fat loads of similar magnitudes to that measured in the exercise intervention study. Accuracy of the dietary measurement instrument was improved with the effects of random error (day-to-day variability) minimised with the use of a statistically validated 8-day, multiple-pass, 24 hour dietary recall instrument. However systematic error (underreporting) may have masked the magnitude of dietary change, particularly the reduction in dietary fat intakes. A purported biomarker (plasma Apolipoprotein A-IV) (apoA-IV) was subsequently investigated, to monitor systematic error in self-reported dietary intakes. Changes in plasma apoA-IV concentrations were directly correlated with increased and decreased changes to dietary fat intakes, suggesting that this objective marker may be a useful tool to improve the accuracy of dietary measurement in overweight and obese populations, who are susceptible to dietary underreporting.

Document clustering evaluation : Divergence from a random baseline

Divergence from a random baseline is a technique for the evaluation of document clustering. It ensures cluster quality measures are performing work that prevents ineffective clusterings from giving high scores to clusterings that provide no useful result. These concepts are defined and analysed using intrinsic and extrinsic approaches to the evaluation of document cluster quality. This includes the classical clusters to categories approach and a novel approach that uses ad hoc information retrieval. The divergence from a random baseline approach is able to differentiate ineffective clusterings encountered in the INEX XML Mining track. It also appears to perform a normalisation similar to the Normalised Mutual Information (NMI) measure but it can be applied to any measure of cluster quality. When it is applied to the intrinsic measure of distortion as measured by RMSE, subtraction from a random baseline provides a clear optimum that is not apparent otherwise. This approach can be applied to any clustering evaluation. This paper describes its use in the context of document clustering evaluation.

Jackknife for bias reduction in predictive regressions

One of the fundamental econometric models in finance is predictive regression. The standard least squares method produces biased coefficient estimates when the regressor is persistent and its innovations are correlated with those of the dependent variable. This article proposes a general and convenient method based on the jackknife technique to tackle the estimation problem. The proposed method reduces the bias for both single- and multiple-regressor models and for both short- and long-horizon regressions. The effectiveness of the proposed method is demonstrated by simulations. An empirical application to equity premium prediction using the dividend yield and the short rate highlights the differences between the results by the standard approach and those by the bias-reduced estimator. The significant predictive variables under the ordinary least squares become insignificant after adjusting for the finite-sample bias. These discrepancies suggest that bias reduction in predictive regressions is important in practical applications.

Effect of Australia's walk to work day campaign on adults' active commuting and physical activity behavior

Purpose. To determine whether Australia's Walk to Work Day media campaign resulted in behavioural change among targeted groups. Methods. Pre- and postcampaign telephone surveys of a cohort of adults aged 18 to 65 years (n = 1100, 55% response rate) were randomly sampled from Australian major melropolitan areas. Tests for dependent samples were applied (McNemax chi(2) or paired t-test). Results. Among participants who did not usually actively commute to work was a significant decrease in car only use an increase in walking combined with public transport. Among those who were employed was a significant increase in total time walking (+16 min/wk; t [780] = 2.04, p < .05) and in other moderate physical activity (+120 min/wk; t [1087] = 4.76, p < .005), resulting in a significant decrease in the proportion who were inactive (chi(2) (1) = 6.1, p < .05). Conclusion. Although nonexperimental, the Walk to Work Day initiative elicited short-term changes in targeted behaviors among target groups. Reinforcement by integrating worksite health promotion strategies may be required for sustained effects.