Reanalysis of the National Supported Work Experiment Using ODA

Ariel Linden & Paul R. Yarnold

Linden Consulting Group, LLC & Optimal Data Analysis, LLC

Data from the National Supported Work (NSW) randomized experiment have been used frequently over the past 30 years to demonstrate the implementation of various non-experimental methods for drawing causal inferences about treatment effects. In the present study we reanalyze the NSW data using ODA and compare results with those estimated using t-tests. Statistical results were largely consistent between methods, however ODA found 22.2% (2 of 9) preintervention characteristics to be imbalanced. Given that ODA avoids assumptions required of parametric methods, and is insensitive to skewed data and outliers, ODA should be considered the preferred approach when evaluating data from randomized experiments.

View journal article

Using ODA in the Evaluation of Randomized Controlled Trials: Application to Survival Outcomes

Ariel Linden & Paul R. Yarnold

Linden Consulting Group, LLC & Optimal Data Analysis, LLC

In a recent series of papers, ODA has been applied to observational data to draw causal inferences about treatment effects. Presently, ODA is applied to survival outcomes from a randomized controlled trial, with a reanalysis of a study by Linden and Butterworth [2014] that investigated (as a secondary outcome) the effect of a comprehensive hospital-based intervention in reducing mortality at 90 days for chronically ill patients. In the original analysis, differences in mortality rates between treatment and control groups were estimated using logistic regression and calculated as both risk differences and risk ratios, and a treatment effect was found in the subgroup of patients with chronic obstructive pulmonary disease (COPD), but not in the other subgroup of patients with congestive heart failure (CHF). In the present study, we reanalyze these results using both Cox regression, and weighted ODA, wherein the weight of every subject is their follow-up time (i.e., number of days of follow-up).

View journal article

Using ODA in the Evaluation of Randomized Controlled Trials

Ariel Linden & Paul R. Yarnold

Linden Consulting Group, LLC & Optimal Data Analysis, LLC

In a recent series of papers, ODA has been applied to observational data to draw causal inferences about treatment effects. In this article ODA is applied to data from a randomized controlled trial, with a reanalysis of a study by Linden and Butterworth [2014] that investigated the effect of a comprehensive hospital-based intervention in reducing readmissions for chronically ill patients. In the original analysis, negative binomial regression was used to evaluate readmission rates and emergency department visit rates at 30 and 90 days, and no treatment effects were found. However, ODA is a superior analytic approach because of its insensitivity to skewed data, model-free permutation tests to derive P values, identification of the threshold value which best discriminates intervention and control groups, use of a chance- and complexity-corrected indexes of classification accuracy, and cross-validation to assess generalizability of the findings.

View journal article

Obtaining LOO p in Analysis Involving Three or More Class Categories

Paul R. Yarnold

Optimal Data Analysis, LLC

For class variables with two categories, ODA and MegaODA software employ Fisher’s one-tailed exact test to assess p associated with LOO classification performance. For class variables having three or more categories, LOO p is not provided. This article discusses how to use ODA and MegaODA software to obtain LOO p in this situation.

View journal article

ANOVA with Three Between-Groups Factors vs. Novometric Analysis

Paul R. Yarnold

Optimal Data Analysis, LLC

ANOVA with three between-groups factors is used to compare an ordered attribute (score) between four independent categories of class variable “A”, three independent categories of class variable “B”, and two independent categories of class variable “C”. The novometric multiple regression analogue is demonstrated.

View journal article