Full random effects models (FREM): A practical usage guide.

The full random-effects model (FREM) is an innovative and relatively novel covariate modeling technique. It differs from other covariate modeling approaches in that it treats covariates as observations and captures their impact on model parameters using their covariances. These unique characteristics mean that FREM is insensitive to correlations between covariates and implicitly handles missing covariate data. In practice, this implies that covariates are less likely to be excluded from the modeling scope in light of the observed data. FREM has been shown to be a useful modeling method for small datasets, but its pre-specification properties make it a very compelling modeling choice for late-stage phases of drug development. The present tutorial aims to explain what FREM models are and how they can be used in practice.

Using forest plots to interpret covariate effects in pharmacometric models.

The inclusion of covariates in pharmacometric models is important due to their ability to explain variability in drug exposure and response. Clear communication of the impact of covariates is needed to support informed decision making in clinical practice and in drug development. However, effectively conveying these effects to key stakeholders and decision makers can be challenging. Forest plots have been proposed to meet these communication needs. However, forest plots for the illustration of covariate effects in pharmacometrics are complex combinations of model predictions, uncertainty estimates, tabulated results, and reference lines and intervals. The purpose of this tutorial is to outline the aspects that influence the interpretation of forest plots, recommend best practices, and offer specific guidance for a clear and transparent communication of covariate effects.

Properties of the full random-effect modeling approach with missing covariate data.

During drug development, a key step is the identification of relevant covariates predicting between-subject variations in drug response. The full random effects model (FREM) is one of the full-covariate approaches used to identify relevant covariates in nonlinear mixed effects models. Here we explore the ability of FREM to handle missing (both missing completely at random (MCAR) and missing at random (MAR)) covariate data and compare it to the full fixed-effects model (FFEM) approach, applied either with complete case analysis or mean imputation. A global health dataset (20 421 children) was used to develop a FREM describing the changes of height for age Z-score (HAZ) over time. Simulated datasets (n = 1000) were generated with variable rates of missing (MCAR) covariate data (0%-90%) and different proportions of missing (MAR) data condition on either observed covariates or predicted HAZ. The three methods were used to re-estimate model and compared in terms of bias and precision which showed that FREM had only minor increases in bias and minor loss of precision at increasing percentages of missing (MCAR) covariate data and performed similarly in the MAR scenarios. Conversely, the FFEM approaches either collapsed at ≥ $$ \ge $$ 70% of missing (MCAR) covariate data (FFEM complete case analysis) or had large bias increases and loss of precision (FFEM with mean imputation). Our results suggest that FREM is an appropriate approach to covariate modeling for datasets with missing (MCAR and MAR) covariate data, such as in global health studies.

Efficient and relevant stepwise covariate model building for pharmacometrics.

Covariate modeling is an important opportunity for pharmacometrics to influence decision making in drug development. The stepwise covariate model (SCM) building procedure is the most common method for covariate model development. Despite its advantages, the traditional SCM method is known to have long runtimes and the suboptimal ability to select relevant covariates, especially in more complex phase III settings. In this work, two alternative approaches are presented: SCM+, which introduces the "adaptive scope reduction" and changes to general estimation settings, and "stage-wise filtering," which groups covariates into categories based on their importance (mechanistic, structural, and exploratory). The three methods (SCM, SCM+, and SCM+ with stage-wise filtering) are applied to data from a simulated phase III population pharmacokinetic study and are compared in terms of efficiency and relevance. The two SCM+ methods were considerably more efficient than the traditional SCM: the number of function evaluations was reduced by 70% for SCM+ and by 76% for SCM+ with stage-wise filtering compared to SCM; the corresponding number of executed models was reduced by 44% for SCM+ and 70% for SCM+ with stage-wise filtering. In addition, among the three methods, SCM+ with stage-wise filtering selected the highest number of relevant covariates. Given the improved efficiency and ability to select relevant covariates shown in this work, the use of SCM+ and stage-wise filtering can greatly increase the efficiency of covariate modeling in drug development, which will ultimately facilitate more timely support for decision making.

Population pharmacokinetic analysis of tepotinib, an oral MET kinase inhibitor, including data from the VISION study.

PURPOSE: Tepotinib is a highly selective, potent, mesenchymal-epithelial transition factor (MET) inhibitor, approved for the treatment of non-small cell lung cancer (NSCLC) harboring MET exon 14 skipping. Objectives of this population pharmacokinetic (PK) analysis were to evaluate the dose-exposure relationship of tepotinib and its major circulating metabolite, MSC2571109A, and to identify the intrinsic/extrinsic factors that are predictive of PK variability. METHODS: Data were included from 12 studies in patients with cancer and in healthy participants. A sequential modeling approach was used to analyze the parent and metabolite data, including covariate analyses. Potential associations between observed covariates and PK parameters were illustrated using bootstrap analysis-based forest plots. RESULTS: A two-compartment model with sequential zero- and first-order absorption, and a first-order elimination from the central compartment, best described the plasma PK of tepotinib in humans across the dose range of 30-1400 mg. The bioavailability of tepotinib was shown to be dose dependent, although bioavailability decreased primarily at doses above the therapeutic dose of 500 mg. The intrinsic factors of race, age, sex, body weight, mild/moderate hepatic impairment and mild/moderate renal impairment, along with the extrinsic factors of opioid analgesic and gefitinib intake, had no relevant effect on tepotinib PK. Tepotinib has a long effective half-life of ~ 32 h. CONCLUSIONS: Tepotinib shows dose proportionality up to at least the therapeutic dose, and time-independent clearance with a profile appropriate for once-daily dosing. None of the covariates identified had a clinically meaningful effect on tepotinib exposure or required dose adjustments.

A quantitative approach to the choice of number of samples for percentile estimation in bootstrap and visual predictive check analyses.

Understanding the uncertainty in parameter estimates or in derived secondary variables is important in all data analysis activities. In pharmacometrics, this is often done based on the standard errors from the variance-covariance matrix of the estimates. Confidence intervals derived in this way are by definition symmetrical, which may lead to implausible outcomes, and will require translation to generate uncertainties in derived variables. An often-used alternative is numerical percentile estimation by, for example, nonparametric bootstraps to circumvent these issues. Visual predictive checks (VPCs), which is a commonly used model diagnostic tool in pharmacometric analyses, also rely on the estimation of percentiles through numerical approaches. Given the cost in terms of run times and processing times for these methods, it is important to consider the trade-off between the number of bootstrap samples or simulated data sets in the VPCs, to the increase in precision related to a large number of bootstrap samples or simulated data sets. The objective with this tutorial is to provide a quantitative framework for assessing the precision in estimated percentile limits in bootstrap and visual predictive checks analyses to facilitate an informed choice of confidence interval width, number of bootstrap samples/simulated data sets, and required level of precision.

An introduction to the full random effects model.

The full random-effects model (FREM) is a method for determining covariate effects in mixed-effects models. Covariates are modeled as random variables, described by mean and variance. The method captures the covariate effects in estimated covariances between individual parameters and covariates. This approach is robust against issues that may cause reduced performance in methods based on estimating fixed effects (e.g., correlated covariates where the effects cannot be simultaneously identified in fixed-effects methods). FREM covariate parameterization and transformation of covariate data records can be used to alter the covariate-parameter relation. Four relations (linear, log-linear, exponential, and power) were implemented and shown to provide estimates equivalent to their fixed-effects counterparts. Comparisons between FREM and mathematically equivalent full fixed-effects models (FFEMs) were performed in original and simulated data, in the presence and absence of non-normally distributed and highly correlated covariates. These comparisons show that both FREM and FFEM perform well in the examined cases, with a slightly better estimation accuracy of parameter interindividual variability (IIV) in FREM. In addition, FREM offers the unique advantage of letting a single estimation simultaneously provide covariate effect coefficient estimates and IIV estimates for any subset of the examined covariates, including the effect of each covariate in isolation. Such subsets can be used to apply the model across data sources with different sets of available covariates, or to communicate covariate effects in a way that is not conditional on other covariates.

Conditional distribution modeling as an alternative method for covariates simulation: Comparison with joint multivariate normal and bootstrap techniques.

Clinical trial simulation (CTS) is a valuable tool in drug development. To obtain realistic scenarios, the subjects included in the CTS must be representative of the target population. Common ways of generating virtual subjects are based upon bootstrap (BS) procedures or multivariate normal distributions (MVNDs). Here, we investigated the performance of an alternative method based on conditional distributions (CDs). Covariate data from a hypertension drug development program were used. The methods were evaluated based on the original data set (internal evaluation) and on their ability to reproduce an older, unobserved population (extrapolation). Similar results were obtained in the internal evaluation for summary statistics, yet BS was able to preserve the correlation structure of the empirical distribution, which was not adequately reproduced by MVND; CD was in between BS and MVND. BS does not allow to extrapolate to an unobserved population. When the data set used to inform the extrapolation was well approximated by an MVND, the results from CD and MVND were comparable. However, improved extrapolation performance was observed for CD when deviations from normality assumptions occurred. If CTS is used to simulate within the observed distribution, BS is the preferred method. When extrapolating to new populations, a parametric method like CD/MVND is needed. In case the empirical multivariate distribution is characterized by linearly related covariates and unimodal marginal distributions, MVND can be used because of the simpler statistical framework and well-established use; however, if uncertainty about the MVND assumptions exists, CD will increase the confidence in the simulations compared to MVND.

A Physiologically-Based Pharmacokinetic Framework for Prediction of Drug Exposure in Malnourished Children.

Malnutrition in children is a global health problem, particularly in developing countries. The effects of an insufficient supply of nutrients on body composition and physiological functions may have implications for drug disposition and ultimately affect the clinical outcome in this vulnerable population. Physiologically-based pharmacokinetic (PBPK) modeling can be used to predict the effect of malnutrition as it links physiological changes to pharmacokinetic (PK) consequences. However, the absence of detailed information on body composition and the limited availability of controlled clinical trials in malnourished children complicates the establishment and evaluation of a generic PBPK model in this population. In this manuscript we describe the creation of physiologically-based bridge to a malnourished pediatric population, by combining information on (a) the differences in body composition between healthy and malnourished adults and (b) the differences in physiology between healthy adults and children. Model performance was confirmed using clinical reference data. This study presents a physiologically-based translational framework for prediction of drug disposition in malnourished children. The model is readily applicable for dose recommendation strategies to address the urgent medicinal needs of this vulnerable population.

A PK/PD Analysis of Circulating Biomarkers and Their Relationship to Tumor Response in Atezolizumab-Treated non-small Cell Lung Cancer Patients.

To assess circulating biomarkers as predictors of antitumor response to atezolizumab (anti-programmed death-ligand 1 (PD-L1), Tecentriq) serum pharmacokinetic (PK) and 95 plasma biomarkers were analyzed in 88 patients with relapsed/refractory non-small cell lung cancer (NSCLC) receiving atezolizumab i.v. q3w (10-20 mg/kg) in the PCD4989g phase I clinical trial. Following exploratory analyses, two plasma biomarkers were chosen for further study and correlation with change in tumor size (the sum of the longest diameter) was assessed in a pharmacokinetic/pharmacodynamic (PK/PD) tumor modeling framework. When longitudinal kinetics of biomarkers and tumor size were modeled, tumor shrinkage was found to significantly correlate with area under the curve (AUC), baseline factors (metastatic sites, liver metastases, and smoking status), and relative change in interleukin (IL)-18 level from baseline at day 21 (RCFBIL -18,d21 ). Although AUC was a major predictor of tumor shrinkage, the effect was estimated to dissipate with an average half-life of 80 days, whereas RCFBIL -18,d21 seemed relevant to the duration of the response.

Dosing and Switching Strategies for Paliperidone Palmitate 3-Month Formulation in Patients with Schizophrenia Based on Population Pharmacokinetic Modeling and Simulation, and Clinical Trial Data.

Paliperidone palmitate 3-month formulation (PP3M), a long-acting injectable atypical antipsychotic, was recently approved in the US and Europe for the treatment of schizophrenia in adult patients who have already been treated with paliperidone palmitate 1-month formulation (PP1M) for ≥4 months. This article reviews the pharmacokinetic rationale for the approved dosing regimens for PP3M, dosing windows, management of missed doses and treatment discontinuation, switching to other formulations, and dosing in special populations. Approved PP3M dosing regimens are based on the comparisons of simulations with predefined dosing regimens using paliperidone palmitate and oral paliperidone extended release (ER) population pharmacokinetic models (one-compartment model with two saturable absorption processes for PP3M; one-compartment model with parallel zero- and first-order absorption for PP1M; two-compartment model with sequential zero- and first-order absorption for ER) versus clinical trial data. Covariates were obtained by resampling subject covariates from the pharmacokinetics database for PP1M and PP3M. Simulation scenarios with varying doses and covariate values were generated. The population median and 90% prediction interval of the simulated concentration-time profiles were plotted for simulation outcomes evaluation. Simulations described in this paper provide (a) simulated plasma exposures for switching from PP1M to PP3M, (b) support for a once-every-3-months injection cycle, (c) information on dosing windows and managing missed doses of PP3M, (d) important guidance on PP3M dosing in special patient populations, and (e) key PP3M pharmacokinetic exposure metrics based on the population pharmacokinetic PP3M model. Population pharmacokinetics provided practical guidance to establish dosing regimens for PP3M.

Population Pharmacokinetics of a Novel Once-Every 3 Months Intramuscular Formulation of Paliperidone Palmitate in Patients with Schizophrenia.

OBJECTIVES: Our objective was to characterize the population pharmacokinetics of paliperidone after intramuscular administration of its long-acting 3-month formulation palmitate ester at various doses and at different injection sites (deltoid and gluteal muscles). METHODS: This retrospective analysis included pooled data from 651 subjects from one phase I study (single injection of the 3-month formulation) and one phase III study (multiple injections of both 1- and 3-month formulations). A total of 8990 pharmacokinetic samples with valid concentration time points were available for this analysis. Nonlinear mixed-effects modelling of the pooled data was conducted using NONMEM software. Knowledge from a previously developed 1-month formulation model was used as a starting point to build the 3-month formulation model. RESULTS: The final model describing the plasma concentrations after administration of the 3-month formulation was a one-compartment model with first-order elimination and two saturable absorption processes (rapid and slow). The apparent volume of distribution estimated for the 3-month formulation was not the same as for the previously modelled 1-month formulation. Apparent clearance (CL), apparent volume of distribution (V), and fraction of the absorbed dose (F3) were estimated to be 3.84 l/h, 1960 L, and 20.9 %. For slow absorption, the maximum absorption rate constant (k a1 max), amount of paliperidone at the absorption site when half of the maximum absorption rate was achieved (k amt1 50), and Hill factor (γ) were estimated to be 90.4 µg/h, 120 mg, and 1.44, respectively. For rapid absorption, the maximum absorption rate constant (k a3 max) and amount of paliperidone at the absorption site when half of the maximum absorption rate was achieved (k amt3 50) were estimated to be 164 µg/h and 21.4 mg, respectively. CONCLUSION: The final model with two saturable absorption processes provided a good description of the pharmacokinetic characteristics of paliperidone after intramuscular administration of its long-acting 3-month formulation palmitate ester. In addition to the structural covariates (creatinine clearance on CL, body mass index on V, and injection volume on both absorption rates), injection site and sex were identified as covariates on k a max of the slow absorption process (k a1 max). Clinical trial registration numbers: NCT01559272, NCT01529515, and NCT01515423.

Population pharmacokinetics of tanezumab in phase 3 clinical trials for osteoarthritis pain.

AIMS: The aims were to 1) develop the pharmacokinetics model to describe and predict observed tanezumab concentrations over time, 2) test possible covariate parameter relationships that could influence clearance and distribution and 3) assess the impact of fixed dosing vs. a dosing regimen adjusted by body weight. METHODS: Individual concentration-time data were determined from 1608 patients in four phase 3 studies conducted to assess efficacy and safety of intravenous tanezumab. Patients received two or three intravenous doses (2.5, 5 or 10 mg) every 8 weeks. Blood samples for assessment of tanezumab PK were collected at baseline, 1 h post-dose and at weeks 4, 8, 16 and 24 (or early termination) in all studies. Blood samples were collected at week 32 in two studies. Plasma samples were analyzed using a sensitive, specific, validated enzyme-linked immunosorbent assay. RESULTS: A two compartment model with parallel linear and non-linear elimination processes adequately described the data. Population estimates for clearance (CL), central volume (V1 ), peripheral volume (V2 ), inter-compartmental clearance, maximum elimination capacity (VM) and concentration at half-maximum elimination capacity were 0.135 l day(-1) , 2.71 l, 1.98 l, 0.371 l day(-1) , 8.03 µg day(-1) and 27.7 ng ml(-1) , respectively. Inter-individual variability (IIV) was included on CL, V1 , V2 and VM. A mixture model accounted for the distribution of residual error. While gender, dose and creatinine clearance were significant covariates, only body weight as a covariate of CL, V1 and V2 significantly reduced IIV. CONCLUSIONS: The small increase in variability associated with fixed dosing is consistent with other monoclonal antibodies and does not change risk : benefit.

A Bayesian decision support tool for efficient dose individualization of warfarin in adults and children.

BACKGROUND: Warfarin is the most widely prescribed anticoagulant for the prevention and treatment of thromboembolic events. Although highly effective, the use of warfarin is limited by a narrow therapeutic range combined with a more than ten-fold difference in the dose required for adequate anticoagulation in adults. An optimal dose that leads to a favourable balance between the wanted antithrombotic effect and the risk of bleeding as measured by the prothrombin time International Normalised Ratio (INR) must be found for each patient. A model describing the time-course of the INR response can be used to aid dose selection before starting therapy (a priori dose prediction) and after therapy has been initiated (a posteriori dose revision). RESULTS: In this paper we describe a warfarin decision support tool. It was transferred from a population PKPD-model for warfarin developed in NONMEM to a platform independent tool written in Java. The tool proved capable of solving a system of differential equations that represent the pharmacokinetics and pharmacodynamics of warfarin with a performance comparable to NONMEM. To estimate an a priori dose the user enters information on body weight, age, baseline and target INR, and optionally CYP2C9 and VKORC1 genotype. By adding information about previous doses and INR observations, the tool will suggest a new dose a posteriori through Bayesian forecasting. Results are displayed as the predicted dose per day and per week, and graphically as the predicted INR curve. The tool can also be used to predict INR following any given dose regimen, e.g. a fixed or an individualized loading-dose regimen. CONCLUSIONS: We believe that this type of mechanism-based decision support tool could be useful for initiating and maintaining warfarin therapy in the clinic. It will ensure more consistent dose adjustment practices between prescribers, and provide efficient and truly individualized warfarin dosing in both children and adults.

Characterizing variability in warfarin dose requirements in children using modelling and simulation.

AIMS: Although genetic, clinical and demographic factors have been shown to explain approximately half of the inter-individual variability in warfarin dose requirement in adults, less is known about causes of dose variability in children. This study aimed to identify and quantify major genetic, clinical and demographic sources of warfarin dose variability in children using modelling and simulation. METHODS: Clinical, demographic and genetic data from 163 children with a median age of 6.3 years (range 0.06-18.9 years), covering over 183 years of warfarin therapy and 6445 INR observations were used to update and optimize a published adult pharmacometric warfarin model for use in children. RESULTS: Genotype effects in children were found to be comparable with what has been reported for adults, with CYP2C9 explaining up to a four-fold difference in dose (CYP2C9 *1/*1 vs. *3/*3) and VKORC1 explaining up to a two-fold difference in dose (VKORC1 G/G vs. A/A), respectively. The relationship between bodyweight and warfarin dose was non-linear, with a three-fold difference in dose for a four-fold difference in bodyweight. In addition, age, baseline and target INR, and time since initiation of therapy, but not CYP4F2 genotype, had a significant impact on typical warfarin dose requirements in children. CONCLUSIONS: The updated model provides quantitative estimates of major clinical, demographic and genetic factors impacting on warfarin dose variability in children. With this new knowledge more individualized dosing regimens can be developed and prospectively evaluated in the pursuit of improving both efficacy and safety of warfarin therapy in children.

Warfarin dose prediction in children using pharmacometric bridging--comparison with published pharmacogenetic dosing algorithms.

PURPOSE: Numerous studies have investigated causes of warfarin dose variability in adults, whereas studies in children are limited both in numbers and size. Mechanism-based population modelling provides an opportunity to condense and propagate prior knowledge from one population to another. The main objectives with this study were to evaluate the predictive performance of a theoretically bridged adult warfarin model in children, and to compare accuracy in dose prediction relative to published warfarin algorithms for children. METHOD: An adult population pharmacokinetic/pharmacodynamic (PK/PD) model for warfarin, with CYP2C9 and VKORC1 genotype, age and target international normalized ratio (INR) as dose predictors, was bridged to children using allometric scaling methods. Its predictive properties were evaluated in an external data set of children 0-18 years old, including comparison of dose prediction accuracy with three pharmacogenetics-based algorithms for children. RESULTS: Overall, the bridged model predicted INR response well in 64 warfarin-treated Swedish children (median age 4.3 years), but with a tendency to overpredict INR in children ≤2 years old. The bridged model predicted 20 of 49 children (41 %) within ± 20 % of actual maintenance dose (median age 7.2 years). In comparison, the published dosing algorithms predicted 33-41 % of the children within ±20 % of actual dose. Dose optimization with the bridged model based on up to three individual INR observations increased the proportion within ±20 % of actual dose to 70 %. CONCLUSION: A mechanism-based population model developed on adult data provides a promising first step towards more individualized warfarin therapy in children.

A fast method for testing covariates in population PK/PD Models.

The development of covariate models within the population modeling program like NONMEM is generally a time-consuming and non-trivial task. In this study, a fast procedure to approximate the change in objective function values of covariate-parameter models is presented and evaluated. The proposed method is a first-order conditional estimation (FOCE)-based linear approximation of the influence of covariates on the model predictions. Simulated and real datasets were used to compare this method with the conventional nonlinear mixed effect model using both first-order (FO) and FOCE approximations. The methods were mainly assessed in terms of difference in objective function values (ΔOFV) between base and covariate models. The FOCE linearization was superior to the FO linearization and showed a high degree of concordance with corresponding nonlinear models in ΔOFV. The linear and nonlinear FOCE models provided similar coefficient estimates and identified the same covariate-parameter relations as statistically significant or non-significant for the real and simulated datasets. The time required to fit tesaglitazar and docetaxel datasets with 4 and 15 parameter-covariate relations using the linearization method was 5.1 and 0.5 min compared with 152 and 34 h, respectively, with the nonlinear models. The FOCE linearization method allows for a fast estimation of covariate-parameter relations models with good concordance with the nonlinear models. This allows a more efficient model building and may allow the utilization of model building techniques that would otherwise be too time-consuming.

Modeling disease progression in acute stroke using clinical assessment scales.

This article demonstrates techniques for describing and predicting disease progression in acute stroke by modeling scores measured using clinical assessment scales, accommodating dropout as an additional source of information. Scores assessed using the National Institutes of Health Stroke Scale and the Barthel Index in acute stroke patients were used to model the time course of disease progression. Simultaneous continuous and probabilistic models for describing the nature and magnitude of score changes were developed, and used to model the trajectory of disease progression using scale scores. The models described the observed data well, and exhibited good simulation properties. Applications include longitudinal analysis of stroke scale data, clinical trial simulation, and prognostic forecasting. Based upon experience in other areas, it is likely that application of this modeling methodology will enable reductions in the number of patients needed to carry out clinical studies of treatments for acute stroke.

The use of clinical irrelevance criteria in covariate model building with application to dofetilide pharmacokinetic data.

To characterise the pharmacokinetics of dofetilide in patients and to identify clinically relevant parameter-covariate relationships. To investigate three different modelling strategies in covariate model building using dofetilide as an example: (1) using statistical criteria only or in combination with clinical irrelevance criteria for covariate selection, (2) applying covariate effects on total clearance or separately on non-renal and renal clearances and (3) using separate data sets for covariate selection and parameter estimation. Pooled concentration-time data (1,445 patients, 10,133 observations) from phase III clinical trials was used. A population pharmacokinetic model was developed using NONMEM. Stepwise covariate model building was applied to identify important covariates using the strategies described above. Inclusion and exclusion of covariates using clinical irrelevance was based on reduction in interindividual variability and changes in parameters at the extremes of the covariate distribution. Parametric separation of the elimination pathways was accomplished using creatinine clearance as an indicator of renal function. The pooled data was split in three parts which were used for covariate selection, parameter estimation and evaluation of predictive performance. Parameter estimations were done using the first-order (FO) and the first-order conditional estimation (FOCE) methods. A one-compartment model with first order absorption adequately described the data. Using clinical irrelevance criteria resulted in models containing less parameter-covariate relationships with a minor loss in predictive power. A larger number of covariates were found significant when the elimination was divided into a renal part and a non-renal part, but no gain in predictive power could be seen with this data set. The FO and FOCE estimation methods gave almost identical final covariate model structures with similar predictive performance. Clinical irrelevance criteria may be valuable for practical reasons since stricter inclusion/exclusion criteria shortens the run times of the covariate model building procedure and because only the covariates important for the predictive performance are included in the model.

Comparison of proportional and differential odds models for mixed-effects analysis of categorical data.

In this work a model for analyzing categorical data is presented; the differential odds model. Unlike the commonly used proportional odds model, this model does not assume that a covariate affects all categories equally on the log odds scale. The differential odds model was compared to the proportional odds model, by assessing statistical significance and improvement of predictive performance when applying the differential odds model to data previously analyzed using the proportional odds model. Three clinical studies; 3-category T-cell receptor density data, 5-category diarrhea data and 6-category sedation data, were re-analyzed with the differential odds model. As expected, no improvements were seen with T-cell receptor density and diarrhea data. However, for the more complex measurement sedation, the differential odds model provided both statistical improvements and improvements in simulation properties. The estimated actual critical value was for all data lower than the nominal value, using the number of added parameters as the degree of freedom, i.e. the differential odds model is statistically indicated to a less extent than expected. The differential odds model had the desired property of not being indicated when not necessary, but it may provide improvements when the data does not represent a categorization of continuous data.