Dichotomization Of Continuous Variables Research Articles

MO703: Nonlinear Association of Fluid Overload to Technique Failure in Peritoneal Dialysis? Application of a Cubic Spline Model

Abstract BACKGROUND AND AIMS Severe volume overload in peritoneal dialysis (PD) patients is associated with an elevated risk of technique failure [including death and transfer to hemodialysis (HD)] compared to moderate volume overload, euvolemia and volume depletion (Vrtovsnik et al., 2021 Clin Kidney J). Severe volume overload is often defined by a single cut-off value. However, such dichotomization of continuous variables in analytical models has shortcomings, especially when a nonlinear association is likely to be present. The aim of the present analysis was to investigate whether the association of volume overload to technique failure was nonlinear in nature. METHOD The IPOD-PD study enrolled incident patients, on either continuous ambulatory PD or automated PD. Besides regular documentation of demographic, medical and laboratory data, body composition data from measurements employing the Body Composition Monitor (BCM, Fresenius Medical Care, Bad Homburg, Germany) were recorded. In contrast to the previously published analysis (Vrtovsnik et al., 2021 Clin Kidney J), a competing risk analysis (kidney transplantation as competing risk) with cubic spline analysis was applied to take the continuous nature of the parameter fluid overload into account. Of 1054 incident PD patients from the IPOD-PD patient cohort (Van Biesen et al., 2019 CJASN), after excluding those who terminated the study before month 6, 719 patients were included in the present analysis. The ‘starting point’ of the follow-up was month 6 on PD, and technique failure (composite endpoint of change to HD and death) in the next 12 months was recorded as the outcome. The association of relative fluid overload with technique failure was estimated using penalized cubic splines regression, with kidney transplantation as competing risk. The model was adjusted for same variables as in previous study, including age, percentage of PD patients in the center, and number of incident PD patients in the last 2 years in the center, which were derived from backward selection based on Akaike's information criteria. RESULTS The patients included in the present analysis (n = 719) were in average 58.4 ± 14.9 years old, and 57% were male. Mean relative fluid overload was 9.7% at baseline, and 8.0% at month 6. With the means of cubic spline analysis, the course of subdistributional hazard (SDH) over the range of relative fluid overload as measured in our patient cohort was illustrated. Whereas SDH was rather constant in the range of fluid overload representing euvolemia, it increased strongly at more elevated relative fluid overload levels, which could accordingly be considered as overhydration (Fig. 1). Of covariates included in the model, increasing age was significantly associated with an elevated risk of technique failure and being treated in a center with a high versus low proportion of PD patients (≥ versus < 56.5% of the patients on PD) was significantly associated with a lower risk of technique failure. CONCLUSION An analytical approach considering the continuous nature of the parameter fluid overload allows to visualize a differentiated nonlinear pattern of associations to technique failure.

Determining Amyloid-β Positivity Using 18F-AZD4694 PET Imaging.

Amyloid-β deposition into plaques is a pathologic hallmark of Alzheimer disease appearing years before the onset of symptoms. Although cerebral amyloid-β deposition occurs on a continuum, dichotomization into positive and negative groups has advantages for diagnosis, clinical management, and population enrichment for clinical trials. 18F-AZD4694 (also known as 18F-NAV4694) is an amyloid-β imaging ligand with high affinity for amyloid-β plaques. Despite being used in multiple academic centers, no studies have assessed a quantitative cutoff for amyloid-β positivity using 18F-AZD4694 PET. Methods: We assessed 176 individuals [young adults (n = 22), cognitively unimpaired elderly (n = 89), and cognitively impaired (n = 65)] who underwent amyloid-β PET with 18F-AZD4694, lumbar puncture, structural MRI, and genotyping for APOEε418F-AZD4694 values were normalized using the cerebellar gray matter as a reference region. We compared 5 methods for deriving a quantitative threshold for 18F-AZD4694 PET positivity: comparison with young-control SUV ratios (SUVRs), receiver-operating-characteristic (ROC) curves based on clinical classification of cognitively unimpaired elderly versus Alzheimer disease dementia, ROC curves based on visual Aβ-positive/Aβ-negative classification, gaussian mixture modeling, and comparison with cerebrospinal fluid measures of amyloid-β, specifically the Aβ42/Aβ40 ratio. Results: We observed good convergence among the 4 methods: ROC curves based on visual classification (optimal cut point, 1.55 SUVR), ROC curves based on clinical classification (optimal cut point, 1.56 SUVR) gaussian mixture modeling (optimal cut point, 1.55 SUVR), and comparison with cerebrospinal fluid measures of amyloid-β (optimal cut point, 1.51 SUVR). Means and 2 SDs from young controls resulted in a lower threshold (1.33 SUVR) that did not agree with the other methods and labeled most elderly individuals as Aβ-positive. Conclusion: Good convergence was obtained among several methods for determining an optimal cutoff for 18F-AZD4694 PET positivity. Despite conceptual and analytic idiosyncrasies linked with dichotomization of continuous variables, an 18F-AZD4694 threshold of 1.55 SUVR had reliable discriminative accuracy. Although clinical use of amyloid PET is currently by visual inspection of scans, quantitative thresholds may be helpful to arbitrate disagreement among raters or in borderline cases.

Low Perioperative Serum Prealbumin Predicts Early Recurrence After Curative Pulmonary Resection for Non‐Small‐Cell Lung Cancer

To the Editor We read with great interest the article by Kawai and Ota [1], recently published in World Journal of Surgery. They evaluated the relationship between the perioperative nutritional status and postoperative early recurrence of lung cancer, claiming a significant risk effect of low perioperative serum prealbumin. The serum level of prealbumin was used as an indicator of nutritional status, and the authors split their study sample into two groups: patients with preoperative prealbumin\23 mg/dl and postoperative prealbumin\15 mg/dl were included in the ‘‘low-prealbumin’’ group. The dichotomization of continuous variables is a common approach in clinical research because of its perceived advantages. It exemplifies the presentation and the interpretation of the results, because the two groups represent ‘‘low’’ and ‘‘high’’ levels of serum prealbumin, and so the features observed in the ‘‘low’’ group are indirectly linked by the reader to the effect of a decrease of serum prealbumin (and vice versa). This approach also exemplifies the statistical analysis, leading to comparisons between the two groups; in the simplest case, such comparisons are usually performed by Student’s t-test and the v test, which are familiar to most researchers. However, dichotomizing continuous variables leads to several problems that often are underestimated [2, 3]. For example, the variation in outcome between groups may be underestimated because the variability may be subsumed within each group. The possible non-linearity relationship between serum prealbumin and outcome could be hidden by the above-mentioned binary split. It also defines an unrealistic scenario, because subjects belonging to the same group are considered similar even if their difference in serum prealbumin could suggest otherwise. In the same way, subjects close to but on the opposite side of the cutpoint are described as having a different outcome rather than a similar one. It causes loss of information and a consequent decrease in statistical power. In their study, Kawai and Ota defined cutoff values of serum prealbumin to achieve more practical criteria for defining impaired nutritional status. We agree that a cutoff may be easy to use by clinicians, but these authors’ choice was based on their results, and it was completely arbitrary. Nevertheless, there is no reason to suppose that a underlying dichotomy in serum albumin exists; it would be like a threshold that changes the risk status of patients crossing it. In their study, Kawai and Ota provide useful information about the role of serum prealbumin as a prognostic marker, but we think that serum prealbumin—and age too—should have been included in the multivariable analysis as continuous variables and the identification of possible cutoffs should be left to further studies with appropriate mathematical methods.

Pick up the Pieces

Pick up the Pieces

Challenges in Nonlinear Structural Equation Modeling

Abstract. Challenges in evaluating nonlinear effects in multiple regression analyses include reliability, validity, multicollinearity, and dichotomization of continuous variables. While reliability and validity issues are solved by employing nonlinear structural equation modeling, multicollinearity remains a problem which may even be aggravated when using latent variable approaches. Further challenges of nonlinear latent analyses comprise the distribution of latent product terms, a problem especially relevant for approaches using maximum likelihood estimation methods based on multivariate normally distributed variables, and unbiased estimates of nonlinear effects under multicollinearity. The only methods that explicitly take the nonnormality of nonlinear latent models into account are latent moderated structural equations (LMS) and quasi-maximum likelihood (QML). In a small simulation study both methods yielded unbiased parameter estimates and correct estimates of standard errors for inferential statistics. The advantages and limitations of nonlinear structural equation modeling are discussed.

Is Self-Confidence a Bias Factor in Higher-Order Catastrophe Models? An Exploratory Analysis—A Critique

The current paper criticizes the concept, research methodology, data analyses, and validity of the conclusions made in Hardy, Woodman, and Carrington’s (2004) article published in this journal. In their repeated-measures analysis of data from the performances of 7 golfers, they did not examine changes in performance scores on successive holes. Instead, Hardy et al. used several ANOVA models to examine how performance varied with respect to somatic and cognitive anxiety level and self-confidence interaction. By doing so, their findings produced effects which we argue to be conceptually and empirically limited. We also address problems associated with dichotomization of continuous variables, measurement errors when splitting data, eradication of random significant effects, cell sizes in segmental quadrant analysis, and correlation between somatic and cognitive anxiety. We believe these difficulties prevent any reliable conclusions and/or generalizations from being made.

What you see may not be what you get: a brief, nontechnical introduction to overfitting in regression-type models.

Statistical models, such as linear or logistic regression or survival analysis, are frequently used as a means to answer scientific questions in psychosomatic research. Many who use these techniques, however, apparently fail to appreciate fully the problem of overfitting, ie, capitalizing on the idiosyncrasies of the sample at hand. Overfitted models will fail to replicate in future samples, thus creating considerable uncertainty about the scientific merit of the finding. The present article is a nontechnical discussion of the concept of overfitting and is intended to be accessible to readers with varying levels of statistical expertise. The notion of overfitting is presented in terms of asking too much from the available data. Given a certain number of observations in a data set, there is an upper limit to the complexity of the model that can be derived with any acceptable degree of uncertainty. Complexity arises as a function of the number of degrees of freedom expended (the number of predictors including complex terms such as interactions and nonlinear terms) against the same data set during any stage of the data analysis. Theoretical and empirical evidence--with a special focus on the results of computer simulation studies--is presented to demonstrate the practical consequences of overfitting with respect to scientific inference. Three common practices--automated variable selection, pretesting of candidate predictors, and dichotomization of continuous variables--are shown to pose a considerable risk for spurious findings in models. The dilemma between overfitting and exploring candidate confounders is also discussed. Alternative means of guarding against overfitting are discussed, including variable aggregation and the fixing of coefficients a priori. Techniques that account and correct for complexity, including shrinkage and penalization, also are introduced.

What You See May Not Be What You Get: A Brief, Nontechnical Introduction to Overfitting in Regression-Type Models

Objective: Statistical models, such as linear or logistic regression or survival analysis, are frequently used as a means to answer scientific questions in psychosomatic research. Many who use these techniques, however, apparently fail to appreciate fully the problem of overfitting, ie, capitalizing on the idiosyncrasies of the sample at hand. Overfitted models will fail to replicate in future samples, thus creating considerable uncertainty about the scientific merit of the finding. The present article is a nontechnical discussion of the concept of overfitting and is intended to be accessible to readers with varying levels of statistical expertise. The notion of overfitting is presented in terms of asking too much from the available data. Given a certain number of observations in a data set, there is an upper limit to the complexity of the model that can be derived with any acceptable degree of uncertainty. Complexity arises as a function of the number of degrees of freedom expended (the number of predictors including complex terms such as interactions and nonlinear terms) against the same data set during any stage of the data analysis. Theoretical and empirical evidence—with a special focus on the results of computer simulation studies—is presented to demonstrate the practical consequences of overfitting with respect to scientific inference. Three common practices—automated variable selection, pretesting of candidate predictors, and dichotomization of continuous variables—are shown to pose a considerable risk for spurious findings in models. The dilemma between overfitting and exploring candidate confounders is also discussed. Alternative means of guarding against overfitting are discussed, including variable aggregation and the fixing of coefficients a priori. Techniques that account and correct for complexity, including shrinkage and penalization, also are introduced.

L'analyse des variables indépendantes continues et catégorielles : alternatives à la dichotomisation

Summary : The analysis of continuous and categorical independent variables : Alternatives to dichotomization. Transforming continuous independent variables into categorial ones makes the statistical analyses simpler because after dichotomization, the effects of these variables can be examined via an analysis of variance (ANOVA) rather than a multiple regression analysis. However, this simplicity comes at a high price. When a continuous variable is dichotomized, one artificially introduces random error which decreases the statistical power of the inferential analyses. Assuming a normal distribution, the decrease in statistical power is equivalent to the exclusion of approximately 38 % of the participants. In this article, we present the problems associated with the dichotomization of continuous variables and we discuss various strategies that allow researchers to analyze experimental designs with continuous and categorical independent variables. Key words : continuous variables, quantitative variables, regression analysis, mean deviation form, dichotomization.

Dichotomization of continuous variables: The implications for meta-analysis.

In many studies included in meta-analyses, the independent variable measure, the dependent variable measure, or both, have been artificially dichotomized, attenuating the correlation from its true value and resulting in (a) a downward distortion in the mean correlation and (b) an upward distortion in the apparent real variation of correlations across studies. We present (a) exact corrections for this distortion for the case in which only one of the variables has been dichotomized and (b) methods for making approximate corrections when both variables have been artificially dichotomized

Dichotomization of continuous variables: The implications for meta-analysis.

Dichotomization of continuous variables: The implications for meta-analysis.