Mixture models, robustness, and the weighted likelihood methodology

Problems associated with the analysis of data from a mixture of distributions include the presence of outliers in the sample, the fact that a component may not be well represented in the data, and the problem of biases that occur when the model is slightly misspecified. We study the performance of weighted likelihood in this context. The method produces estimates with low bias and mean squared error, and it is useful in that it unearths data substructures in the form of multiple roots. This in turn indicates multiple potential mixture model fits due to the presence of more components than originally specified in the model. To compute the weighted likelihood estimates, we use as starting values the method of moment estimates computed on bootstrap subsamples drawn from the data. We address a number of important practical issues involving bootstrap sample size selection, the role of starting values, and the behavior of the roots. The algorithm used to compute the weighted likelihood estimates is competitive with EM, and it is similar to EM when the components are not well separated. Moreover, we propose a new statistical stopping rule for the termination of the algorithm. An example and a small simulation study illustrate the above points.     

Confidence bands for low-dose risk estimation with quantal response data

We study the use of simultaneous confidence bands for low-dose risk estimation with quantal response data, and derive methods for estimating simultaneous upper confidence limits on predicted extra risk under a multistage model. By inverting the upper bands on extra risk, we obtain simultaneous lower bounds on the benchmark dose (BMD). Monte Carlo evaluations explore characteristics of the simultaneous limits under this setting, and a suite of actual data sets are used to compare existing methods for placing lower limits on the BMD.     

Radiation-related cancer risks at low doses among atomic bomb survivors

To clarify the information in the Radiation Effects Research Foundation data regarding cancer risks of low radiation doses, we focus on survivors with doses less than 0.5 Sv. For reasons indicated, we also restrict attention mainly to survivors within 3, 000 m of the hypocenter of the bombs. Analysis is of solid cancer incidence from 1958-1994, involving 7,000 cancer cases among 50,000 survivors in that dose and distance range. The results provide useful risk estimates for doses as low as 0.05-0.1 Sv, which are not overestimated by linear risk estimates computed from the wider dose ranges 0-2 Sv or 0-4 Sv. There is a statistically significant risk in the range 0-0.1 Sv, and an upper confidence limit on any possible threshold is computed as 0.06 Sv. It is indicated that modification of the neutron dose estimates currently under consideration would not markedly change the conclusions.     

Point and interval estimation of baseline risk and treatment effect based on logistic model for observational studies

In observational studies with dichotomous outcome of a population, researchers usually report treatment effect alone, although both baseline risk and treatment effect are needed to evaluate the significance of the treatment effect to the population. In this article, we study point and interval estimates including confidence region of baseline risk and treatment effect based on logistic model, where baseline risk is the risk of outcome of the population under control treatment while treatment effect is measured by the risk difference between outcomes of the population under active versus control treatments. Using approximate normal distribution of the maximum‐likelihood (ML) estimate of the model parameters, we obtain an approximate joint distribution of the ML estimate of the baseline risk and the treatment effect. Using the approximate joint distribution, we obtain point estimate and confidence region of the baseline risk and the treatment effect as well as point estimate and confidence interval of the treatment effect when the ML estimate of the baseline risk falls into specified range. These interval estimates reflect nonnormality of the joint distribution of the ML estimate of the baseline risk and the treatment effect. The method can be easily implemented by using any software that generates normal distribution. The method can also be used to obtain point and interval estimates of baseline risk and any other measure of treatment effect such as risk ratio and the number needed to treat. The method can also be extended from logistic model to other models such as log‐linear model.     

A mixture model for occupational exposure mean testing with a limit of detection

Information from detectable exposure measurements randomly sampled from a left-truncated log-normal distribution may be used to evaluate the distribution of nondetectable values that fall below an analytic limit of detection. If the proportion of nondetects is larger than expected under log normality, alternative models to account for these unobserved data should be considered. We discuss one such model that incorporates a mixture of true zero exposures and a log-normal distribution with possible left censoring, previously considered in a different context by Moulton and Halsey (1995, Biometrics 51, 1570-1578). A particular relationship is demonstrated between maximum likelihood parameter estimates based on this mixture model and those assuming either left-truncated or left-censored data. These results emphasize the need for caution when choosing a model to fit data involving nondetectable values. A one-sided likelihood ratio test for comparing mean exposure under the mixture model to an occupational exposure limit is then developed and evaluated via simulations. An example demonstrates the potential impact of specifying an incorrect model for the nondetectable values.     

Mixture models for continuous data in dose-response studies when some animals are unaffected by treatment.

A mixture model is described for dose-response studies where measurements on a continuous variable suggest that some animals are not affected by treatment. The model combines a logistic regression on dose for the probability an animal will "respond" to treatment with a linear regression on dose for the mean of the responders. Maximum likelihood estimation via the EM algorithm is described and likelihood ratio tests are used to distinguish between the full model and meaningful reduced-parameter versions. Use of the model is illustrated with three real-data examples.     

Inference of genome-wide mutation rates and distributions of mutation effects for fitness traits: a simulation study.

The properties and limitations of maximum likelihood (ML) inference of genome-wide mutation rates (U) and parameters of distributions of mutation effects are investigated. Mutation parameters are estimated from simulated experiments in which mutations randomly accumulate in inbred lines. ML produces more accurate estimates than the procedure of Bateman and Mukai and is more robust if the data do not conform to the model assumed. Unbiased ML estimates of the mutation effects distribution parameters can be obtained if a value for U can be assumed, but if U is estimated simultaneously with the distribution parameters, likelihood may increase monotonically as a function of U. If the distribution of mutation effects is leptokurtic, the number of mutation events per line is large, or if genotypic values are poorly estimated, only a lower limit for U, an upper limit for the mean mutation effect, and a lower limit for the kurtosis of the distribution can be given. It is argued that such lower (upper) limits are appropriate minima (maxima). Estimates of the mean mutational effect are unbiased but may convey little about the properties of the distribution if it is leptokurtic.     

More Accurate Dose-Response Estimation Using Monte-Carlo Uncertainty Analysis: The Data Cube Approach

The traditional q₁ ^* methodology for constructing upper confidence limits (UCLs) for the low-dose slopes of quantal dose-response functions has two limitations: (i) it is based on an asymptotic statistical result that has been shown via Monte Carlo simulation not to hold in practice for small, real bioassay experiments (Portier and Hoel, 1983); and (ii) it assumes that the multistage model (which represents cumulative hazard as a polynomial function of dose) is correct. This paper presents an uncertainty analysis approach for fitting dose-response functions to data that does not require specific parametric assumptions or depend on asymptotic results. It has the advantage that the resulting estimates of the dose-response function (and uncertainties about it) no longer depend on the validity of an assumed parametric family nor on the accuracy of the asymptotic approximation. The method derives posterior densities for the true response rates in the dose groups, rather than deriving posterior densities for model parameters, as in other Bayesian approaches (Sielken, 1991), or resampling the observed data points, as in the bootstrap and other resampling methods. It does so by conditioning constrained maximum-entropy priors on the observed data. Monte Carlo sampling of the posterior (constrained, conditioned) probability distributions generate values of response probabilities that might be observed if the experiment were repeated with very large sample sizes. A dose-response curve is fit to each such simulated dataset. If no parametric model has been specified, then a generalized representation (e.g., a power-series or orthonormal polynomial expansion) of the unknown dose-response function is fit to each simulated dataset using “model-free” methods. The simulation-based frequency distribution of all the dose-response curves fit to the simulated datasets yields a posterior distribution function for the low-dose slope of the dose-response curve. An upper confidence limit on the low-dose slope is obtained directly from this posterior distribution. This “Data Cube” procedure is illustrated with a real dataset for benzene, and is seen to produce more policy-relevant insights than does the traditional q₁ ^* methodology. For example, it shows how far apart are the 90%, 95%, and 99% limits and reveals how uncertainty about total and incremental risk vary with dose level (typically being dominated at low doses by uncertainty about the response of the control group, and being dominated at high doses by sampling variability). Strengths and limitations of the Data Cube approach are summarized, and potential decision-analytic applications to making better informed risk management decisions are briefly discussed.     

High dimensional mediation analysis with latent variables

We propose a model for high dimensional mediation analysis that includes latent variables. We describe our model in the context of an epidemiologic study for incident breast cancer with one exposure and a large number of biomarkers (i.e., potential mediators). We assume that the exposure directly influences a group of latent, or unmeasured, factors which are associated with both the outcome and a subset of the biomarkers. The biomarkers associated with the latent factors linking the exposure to the outcome are considered “mediators.” We derive the likelihood for this model and develop an expectation‐maximization algorithm to maximize an L1‐penalized version of this likelihood to limit the number of factors and associated biomarkers. We show that the resulting estimates are consistent and that the estimates of the nonzero parameters have an asymptotically normal distribution. In simulations, procedures based on this new model can have significantly higher power for detecting the mediating biomarkers compared with the simpler approaches. We apply our method to a study that evaluates the relationship between body mass index, 481 metabolic measurements, and estrogen‐receptor positive breast cancer.     

Application of iterative robust model‐based optimal experimental design for the calibration of biocatalytic models

The aim of model calibration is to estimate unique parameter values from available experimental data, here applied to a biocatalytic process. The traditional approach of first gathering data followed by performing a model calibration is inefficient, since the information gathered during experimentation is not actively used to optimize the experimental design. By applying an iterative robust model‐based optimal experimental design, the limited amount of data collected is used to design additional informative experiments. The algorithm is used here to calibrate the initial reaction rate of an ω‐transaminase catalyzed reaction in a more accurate way. The parameter confidence region estimated from the Fisher Information Matrix is compared with the likelihood confidence region, which is not only more accurate but also a computationally more expensive method. As a result, an important deviation between both approaches is found, confirming that linearization methods should be applied with care for nonlinear models. © 2017 American Institute of Chemical Engineers Biotechnol. Prog., 33:1278–1293, 2017     

Interval mapping of quantitative trait loci for time-to-event data with the proportional hazards mixture cure model

Interval mapping using normal mixture models has been an important tool for analyzing quantitative traits in experimental organisms. When the primary phenotype is time-to-event, it is natural to use survival models such as Cox's proportional hazards model instead of normal mixtures to model the phenotype distribution. An extra challenge for modeling time-to-event data is that the underlying population may consist of susceptible and nonsusceptible subjects. In this article, we propose a semiparametric proportional hazards mixture cure model which allows missing covariates. We discuss applications to quantitative trait loci (QTL) mapping when the primary trait is time-to-event from a population of mixed susceptibility. This model can be used to characterize QTL effects on both susceptibility and time-to-event distribution, and to estimate QTL location. The model can naturally incorporate covariate effects of other risk factors. Maximum likelihood estimates for the parameters in the model as well as their corresponding variance estimates can be obtained numerically using an EM-type algorithm. The proposed methods are assessed by simulations under practical settings and illustrated using a real data set containing survival times of mice after infection with Listeria monocytogenes. An extension to multiple intervals is also discussed.     

Estimating the immunity coverage required to prevent epidemics in a community of households

An estimation of the immunity coverage needed to prevent future outbreaks of an infectious disease is considered for a community of households. Data on outbreak size in a sample of households from one epidemic are used to derive maximum likelihood estimates and confidence bounds for parameters of a stochastic model for disease transmission in a community of households. These parameter estimates induce estimates and confidence bounds for the basic reproduction number and the critical immunity coverage, which are the parameters of main interest when aiming at preventing major outbreaks in the future. The case when individuals are homogeneous, apart from the size of their household, is considered in detail. The generalization to the case with variable infectivity, susceptibility and/or mixing behaviour is discussed more briefly. The methods are illustrated with an application to data on influenza in Tecumseh, Michigan.     

A Bayesian approach to finite mixture models in bioassay via data augmentation and Gibbs sampling and its application to-insecticide resistance

After continued treatment with an insecticide, within the population of the susceptible insects, resistant strains will occur. It is important to know whether there are any resistant strains, what the proportions are, and what the median lethal doses are for the insecticide. Lwin and Martin (1989, Biometrics 45, 721-732) propose a probit mixture model and use the EM algorithm to obtain the maximum likelihood estimates for the parameters. This approach has difficulties in estimating the confidence intervals and in testing the number of components. We propose a Bayesian approach to obtaining the credible intervals for the location and scale of the tolerances in each component and for the mixture proportions by using data augmentation and Gibbs sampler. We use Bayes factor for model selection and determining the number of components. We illustrate the method with data published in Lwin and Martin (1989).     

A frailty mixture cure model with application to hospital admission data [corrected

Mixture cure models have been utilized to analyze survival data with possible cure. This paper considers the inclusion of frailty into the mixture cure model to model recurrent event data with a cure fraction. An attractive feature of the proposed model is the allowance for heterogeneity in risk among those individuals experiencing the event of interest in addition to the incorporation of a cured component. Maximum likelihood estimates can be obtained using the Expectation Maximization algorithm and standard errors are calculated from the Bootstrap method. The model is applied to hospital readmission data among colorectal cancer patients.     

An augmented probit model for missing predictable covariates in quantal bioassay with small sample size

Summary Quantal bioassay experiments relate the amount or potency of some compound; for example, poison, antibody, or drug to a binary outcome such as death or infection in animals. For infectious diseases, probit regression is commonly used for inference and a key measure of potency is given by the ID_P , the amount that results in P% of the animals being infected. In some experiments, a validation set may be used where both direct and proxy measures of the dose are available on a subset of animals with the proxy being available on all. The proxy variable can be viewed as a messy reflection of the direct variable, leading to an errors‐in‐variables problem. We develop a model for the validation set and use a constrained seemingly unrelated regression (SUR) model to obtain the distribution of the direct measure conditional on the proxy. We use the conditional distribution to derive a pseudo‐likelihood based on probit regression and use the parametric bootstrap for statistical inference. We re‐evaluate an old experiment in 21 monkeys where neutralizing antibodies (nABs) to HIV were measured using an old (proxy) assay in all monkeys and with a new (direct) assay in a validation set of 11 who had sufficient stored plasma. Using our methods, we obtain an estimate of the ID₁ for the new assay, an important target for HIV vaccine candidates. In simulations, we compare the pseudo‐likelihood estimates with regression calibration and a full joint likelihood approach.     

SMR analysis of historical follow-up studies with missing death certificates

The evaluation of epidemiological follow-up studies is frequently based on a comparison of the number O of deaths observed in the cohort from a specified cause with the expected number E calculated from person years in the cohort and mortality rates from a reference population. The ratio SMR = 100 x O/E is called the standardized mortality ratio (SMR). While person years can easily be calculated from the cohort and reference rates are generally available from the national statistical offices or the World Health Organization (WHO), problems can arise with the accessibility of the causes of death of the deceased study participants. However, the information that a person has died may be available, e.g., from population registers. In this paper, a statistical model for this situation is developed to derive a maximum likelihood (ML) estimator for the true (but unknown) number O* of deaths from a specified cause, which uses the known number O of deaths from this cause and the proportion p of all known causes of death among all decreased participants. It is shown that the standardized mortality ratio SMR* based on this estimated number is just SMR* = SMR/p. Easily computable confidence limits can be obtained by dividing the usual confidence limits of the SMR by the opposite limit of the proportion p. However, the confidence level alpha has to be adjusted appropriately.     

A Monte Carlo maximum likelihood method for estimating uncertainty arising from shared errors in exposures in epidemiological studies of nuclear workers

Errors in the estimation of exposures or doses are a major source of uncertainty in epidemiological studies of cancer among nuclear workers. This paper presents a Monte Carlo maximum likelihood method that can be used for estimating a confidence interval that reflects both statistical sampling error and uncertainty in the measurement of exposures. The method is illustrated by application to an analysis of all cancer (excluding leukemia) mortality in a study of nuclear workers at the Oak Ridge National Laboratory (ORNL). Monte Carlo methods were used to generate 10,000 data sets with a simulated corrected dose estimate for each member of the cohort based on the estimated distribution of errors in doses. A Cox proportional hazards model was applied to each of these simulated data sets. A partial likelihood, averaged over all of the simulations, was generated; the central risk estimate and confidence interval were estimated from this partial likelihood. The conventional unsimulated analysis of the ORNL study yielded an excess relative risk (ERR) of 5.38 per Sv (90% confidence interval 0.54-12.58). The Monte Carlo maximum likelihood method yielded a slightly lower ERR (4.82 per Sv) and wider confidence interval (0.41-13.31).     

A Mixture‐of‐Genotypes Model for the Distribution of Thermostable Phenol Sulfotransferase Activity

A statistical method for parametric density estimation based upon a mixture‐of‐genotypes model is developed for the thermostable phenol sulfotransferase (SULT1A1) activity which has a putative role in modifying risk for colon and prostate cancer/polyps. The EM algorithm for the general mixture model is modified to accommodate the genetic constraints and is used to estimate genotype frequencies from the distribution of the SULT1A1 phenotype. A parametric bootstrap likelihood ratio test is considered as a testing method for the number of mixing components. The size and power of the test is then investigated and compared with the conventional chi‐squared test. The relative risk associated with genotypes defined by this model is also investigated through the generalized linear model. This analysis revealed that a genotype with the highest mean value of SULT1A1 activity has greater impact on cancer risk than others. This result suggests that the phenotype with a higher SULT1A1 activity might be important in studying the association between the cancer risk and SULT1A1 activity. (© 2004 WILEY‐VCH Verlag GmbH & Co. KGaA, Weinheim)