Adaptive Random Effects/Coefficients Modeling

Adaptive fractional polynomial modeling of general correlated outcomes is formulated to address nonlinearity in means, variances/dispersions, and correlations. Means and variances/dispersions are modeled using generalized linear models in fixed effects/coefficients. Correlations are modeled using random effects/coefficients. Nonlinearity is addressed using power transforms of primary (untransformed) predictors. Parameter estimation is based on extended linear mixed modeling generalizing both generalized estimating equations and linear mixed modeling. Models are evaluated using likelihood cross-validation (LCV) scores and are generated adaptively using a heuristic search controlled by LCV scores. Cases covered include linear, Poisson, logistic, exponential, and discrete regression of correlated continuous, count/rate, dichotomous, positive continuous, and discrete numeric outcomes treated as normally, Poisson, Bernoulli, exponentially, and discrete numerically distributed, respectively. Example analyses are also generated for these five cases to compare adaptive random effects/coefficients modeling of correlated outcomes to previously developed adaptive modeling based on directly specified covariance structures. Adaptive random effects/coefficients modeling substantially outperforms direct covariance modeling in the linear, exponential, and discrete regression example analyses. It generates equivalent results in the logistic regression example analyses and it is substantially outperformed in the Poisson regression case. Random effects/coefficients modeling of correlated outcomes can provide substantial improvements in model selection compared to directly specified covariance modeling. However, directly specified covariance modeling can generate competitive or substantially better results in some cases while usually requiring less computation time.

An Adaptive Approach for Hazard Regression Modeling

Regression models for survival time data involve estimation of the hazard rate as a function of predictor variables and associated slope parameters. An adaptive approach is formulated for such hazard regression modeling. The hazard rate is modeled using fractional polynomials, that is, linear combinations of products of power transforms of time together with other available predictors. These fractional polynomial models are restricted to generating positive-valued hazard rates and decreasing survival times. Exponentially distributed survival times are a special case. Parameters are estimated using maximum likelihood estimation allowing for right censored survival times. Models are evaluated and compared using likelihood cross-validation (LCV) scores. LCV scores and tolerance parameters are used to control an adaptive search through alternative fractional polynomial hazard rate models to identify effective models for the underlying survival time data. These methods are demonstrated using two different survival time data sets including survival times for lung cancer patients and for multiple myeloma patients. For the lung cancer data, the hazard rate depends distinctly on time. However, controlling for cell type provides a distinct improvement while the hazard rate depends only on cell type and no longer on time. Furthermore, Cox regression is unable to identify a cell type effect. For the multiple myeloma data, the hazard rate also depends distinctly on time. Moreover, consideration of hemoglobin at diagnosis provides a distinct improvement, the hazard rate still depends distinctly on time, and hemoglobin distinctly moderates the effect of time on the hazard rate. These results indicate that adaptive hazard rate modeling can provide unique insights into survival time data.

Adaptive Conditional Hazard Regression Modeling of Multiple Event Times

Recurrent event time data and more general multiple event time data are commonly analyzed using extensions of Cox regression, or proportional hazards regression, as used with single event time data. These methods treat covariates, either time-invariant or time-varying, as having multiplicative effects while general dependence on time is left un-estimated. An adaptive approach is formulated for analyzing multiple event time data. Conditional hazard rates are modeled in terms of dependence on both time and covariates using fractional polynomials restricted so that the conditional hazard rates are positive-valued and so that excess time probability functions (generalizing survival functions for single event times) are decreasing. Maximum likelihood is used to estimate parameters adjusting for right censored event times. Likelihood cross-validation (LCV) scores are used to compare models. Adaptive searches through alternate conditional hazard rate models are controlled by LCV scores combined with tolerance parameters. These searches identify effective models for the underlying multiple event time data. Conditional hazard regression is demonstrated using data on times between tumor recurrence for bladder cancer patients. Analyses of theory-based models for these data using extensions of Cox regression provide conflicting results on effects to treatment group and the initial number of tumors. On the other hand, fractional polynomial analyses of these theory-based models provide consistent results identifying significant effects to treatment group and initial number of tumors using both model-based and robust empirical tests. Adaptive analyses further identify distinct moderation by group of the effect of tumor order and an additive effect to group after controlling for nonlinear effects to initial number of tumors and tumor order. Results of example analyses indicate that adaptive conditional hazard rate modeling can generate useful insights into multiple event time data.

Adaptive Conditional Hazard Regression Modeling of Multiple Event Times

Recurrent event time data and more general multiple event time data are commonly analyzed using extensions of Cox regression, or proportional hazards regression, as used with single event time data. These methods treat covariates, either time-invariant or time-varying, as having multiplicative effects while general dependence on time is left un-estimated. An adaptive approach is formulated for analyzing multiple event time data. Conditional hazard rates are modeled in terms of dependence on both time and covariates using fractional polynomials restricted so that the conditional hazard rates are positive-valued and so that excess time probability functions (generalizing survival functions for single event times) are decreasing. Maximum likelihood is used to estimate parameters adjusting for right censored event times. Likelihood cross-validation (LCV) scores are used to compare models. Adaptive searches through alternate conditional hazard rate models are controlled by LCV scores combined with tolerance parameters. These searches identify effective models for the underlying multiple event time data. Conditional hazard regression is demonstrated using data on times between tumor recurrence for bladder cancer patients. Analyses of theory-based models for these data using extensions of Cox regression provide conflicting results on effects to treatment group and the initial number of tumors. On the other hand, fractional polynomial analyses of these theory-based models provide consistent results identifying significant effects to treatment group and initial number of tumors using both model-based and robust empirical tests. Adaptive analyses further identify distinct moderation by group of the effect of tumor order and an additive effect to group after controlling for nonlinear effects to initial number of tumors and tumor order. Results of example analyses indicate that adaptive conditional hazard rate modeling can generate useful insights into multiple event time data.

Modeling Individual Patient Count/Rate Data over Time with Applications to Cancer Pain Flares and Cancer Pain Medication Usage

The purpose of this article is to investigate approaches for modeling individual patient count/rate data over time accounting for temporal correlation and non-constant dispersions while requiring reasonable amounts of time to search over alternative models for those data. This research addresses formulations for two approaches for extending generalized estimating equations (GEE) modeling. These approaches use a likelihood-like function based on the multivariate normal density. The first approach augments standard GEE equations to include equations for estimation of dispersion parameters. The second approach is based on estimating equations determined by partial derivatives of the likelihood-like function with respect to all model parameters and so extends linear mixed modeling. Three correlation structures are considered including independent, exchangeable, and spatial autoregressive of order 1 correlations. The likelihood-like function is used to formulate a likelihood-like cross-validation (LCV) score for use in evaluating models. Example analyses are presented using these two modeling approaches applied to three data sets of counts/rates over time for individual cancer patients including pain flares per day, as needed pain medications taken per day, and around the clock pain medications taken per day per dose. Means and dispersions are modeled as possibly nonlinear functions of time using adaptive regression modeling methods to search through alternative models compared using LCV scores. The results of these analyses demonstrate that extended linear mixed modeling is preferable for modeling individual patient count/rate data over time, because in example analyses, it either generates better LCV scores or more parsimonious models and requires substantially less time.

Regression Modeling of Individual-Patient Correlated Discrete Outcomes with Applications to Cancer Pain Ratings

Purpose: To formulate and demonstrate methods for regression modeling of probabilities and dispersions for individual-patient longitudinal outcomes taking on discrete numeric values. Methods: Three alternatives for modeling of outcome probabilities are considered. Multinomial probabilities are based on different intercepts and slopes for probabilities of different outcome values. Ordinal probabilities are based on different intercepts and the same slope for probabilities of different outcome values. Censored Poisson probabilities are based on the same intercept and slope for probabilities of different outcome values. Parameters are estimated with extended linear mixed modeling maximizing a likelihood-like function based on the multivariate normal density that accounts for within-patient correlation. Formulas are provided for gradient vectors and Hessian matrices for estimating model parameters. The likelihood-like function is also used to compute cross-validation scores for alternative models and to control an adaptive modeling process for identifying possibly nonlinear functional relationships in predictors for probabilities and dispersions. Example analyses are provided of daily pain ratings for a cancer patient over a period of 97 days. Results: The censored Poisson approach is preferable for modeling these data, and presumably other data sets of this kind, because it generates a competitive model with fewer parameters in less time than the other two approaches. The generated probabilities for this model are distinctly nonlinear in time while the dispersions are distinctly nonconstant over time, demonstrating the need for adaptive modeling of such data. The analyses also address the dependence of these daily pain ratings on time and the daily numbers of pain flares. Probabilities and dispersions change differently over time for different numbers of pain flares. Conclusions: Adaptive modeling of daily pain ratings for individual cancer patients is an effective way to identify nonlinear relationships in time as well as in other predictors such as the number of pain flares.

Adaptive Fractional Polynomial Modeling

Regression analyses reported in the applied research literature commonly assume that relationships are linear in predictors without assessing this assumption. Fractional polynomials provide a general approach for addressing nonlinearity through power transforms of predictors using real valued powers. An adaptive approach for generating fractional polynomial models is presented based on heuristic search through alternative power transforms of predictors guided by k-fold likelihood cross-validation (LCV) scores and controlled by tolerance parameters indicating how much a reduction in the LCV score can be tolerated at given stages of the search. The search optionally can generate geometric combinations, that is, products of power transforms of multiple predictors, thereby supporting nonlinear moderation analyses. Positive valued continuous outcomes can be power transformed as well as predictors. These methods are demonstrated using data from a study of family management for mothers of children with chronic physical conditions. The example analyses demonstrate that power transformation of a predictor may be required to identify that a relationship holds between that predictor and an outcome (dependent or response) variable. Consideration of geometric combinations can identify moderation effects not identifiable using linear relationships or power transforms of interactions. Power transformation of positive valued continuous outcomes along with their primary predictors can resolve model assumption problems.

A Reassessment of Birth Defects for Children of Participants of the Air Force Health Study

The Air Force Health Study (AFHS), also called the Ranch Hand Study, investigated the impact of exposure to dioxin the toxic contaminant in Agent Orange on health, survival, and reproductive outcomes of male Air Force Vietnam War veterans. It was concluded that available reproductive outcome data did not provide support for an adverse association with paternal dioxin exposure. A more extensive set of AFHS data was used to reassess this conclusion, restricting to the case of birth defects in children fathered after the start of the first Vietnam War tour. Analyses started by repeating published analyses, followed by assessing decisions made in those analyses, for example, of excluding participants with dioxin levels below the detectable limit, using a threshold of 10 parts per trillion for a high dioxin level, and not adjusting for multiple conceptions/children of the same participant. Using data for all participants with measured dioxin levels, both veterans who served in Operation Ranch Hand and other non-Ranch Hand veterans, and after accounting for correlation within children of the same participant, the occurrence for children fathered after the start of the first tour of a major defect, a non-major defect, and multiple defects depended significantly on participants having a high dioxin level. These conclusions were not changed by consideration of covariates. In contrast to prior published analyses, the more extensive AFHS data provided support for an adverse effect of paternal dioxin exposure on birth defects. However, the study had many limitations that could have affected the conclusions.

Adaptive Classification Methods for Predicting Transitions in the Nursing Workforce

Earlier analyses of transitions from licensed practical nurse (LPN) to registered nurse (RN) in the North Carolina (NC) nursing workforce in terms of 11 categorical predictors were limited by not considering parsimonious classifications based on these predictors and by substantial amounts of missing data. To address these issues, we formulated adaptive classification methods. Secondary analyses of data collected by the NC State Board of Nursing were also conducted to demonstrate adaptive classification methods by modeling the occurrence of LPN-to-RN transitions in the NC nursing workforce from 2001-2013. These methods combine levels (values) for one or more categorical predictors into parsimonious classifications. Missing values for a predictor are treated as one level for that predictor so that the complete data can be used in the analyses;the missing level is imputed by combining it with other levels of a predictor. An adaptive nested classification generated the best model for predicting an LPN-to-RN transition based on three predictors in order of importance: year of first LPN licensure, work setting at transition, and age at first LPN licensure. These results demonstrate that adaptive classification can identify effective and parsimonious classifications for predicting dichotomous outcomes such as the occurrence of an LPN-to-RN transition.

Adaptive Regression for Nonlinear Interrupted Time Series Analyses with Application to Birth Defects in Children of Vietnam War Veterans

The purpose of this article is to provide an overview of adaptive regression modeling and demonstrate its use in conducting nonlinear analyses of interrupted time series (ITS) data. Adaptive regression modeling is based on heuristic search over alternative models for data controlled by likelihood-cross validation (LCV) scores with larger scores indicating better models. Extended linear mixed models are used for correlated data like ITS data. Power transforms of predictor variables are used to account for nonlinearity. The use of adaptive regression modeling for assessing ITS effects is demonstrated using data on annual proportions of major birth defects in children fathered by male Air Force veterans of the Vietnam War over a 59-year period. The interruption for this ITS is conception after versus before the start of a participant’s first tour in the Vietnam War. Whether the ITS effect is related to dioxin exposure is also addressed. Dioxin is a highly toxic contaminant of the herbicide Agent Orange used in the Vietnam War. The core findings of the reported analyses are that a substantial adverse ITS interruption effect is identified and that this adverse effect can reasonably be attributed to participants having a high dioxin exposure level. Moreover, these results indicate that adaptive regression modeling can identify nonlinear ITS effects in general situations that can lead to consequential insights into nonlinear relationships over time, possibly varying with other available predictors.

An Analysis of Specific Categories of Birth Defects and Developmental Disabilities for Children of Participants of the Air Force Health Study

Background: The Air Force Health Study collected reproductive outcomes for live-born children of male Air Force veterans of the Vietnam War. Methods: Dioxin values for participants were obtained from blood samples. Analyses were conducted of occurrence of 16 specific categories of birth defects and developmental disabilities. Children were categorized as conceived before and after the start of participants’ Vietnam War service. Children conceived before the start of Vietnam War service were treated as being conceived when their fathers had unquantifiable dioxin values. Children conceived after the start of Vietnam War service for participants with missing dioxin values were excluded from primary analyses, but were used to assess the impact of their exclusion on conclusions. Correlation between values for specific categories for multiple children fathered by the same participant was accounted for. The dose-response relationship was treated as a step function increasing for dioxin values larger than adaptively identified individual thresholds changing with the specific category. Results: For 15 of 16 specific categories, the probability of occurrence increased substantially for a sufficiently high dioxin level above identified thresholds. Exclusion of children due to missing dioxin likely did not affect these results. Conclusions: Results supported the conclusion of substantial adverse effects on a wide variety of specific categories of birth defects and developmental disabilities due to sufficiently high exposures to dioxin, a toxic contaminant of Agent Orange used for herbicide spraying in the Vietnam War. Results may hold more generally, but might also have been affected by a variety of limitations.