Many-levelled continuation ratio models for frequency of alcohol and drug use data

Mark S. Chambers, Christopher DrovandiarXiv 2022 · DOI · arXiv

Abstract

Studies of alcohol and drug use are often interested in the number of days that people use the substance of interest over an interval, such as 28 days before a survey date. Although count models are often used for this purpose, they are not strictly appropriate for this type of data because the response variable is bounded above. Furthermore, if some peoples' substance use behaviors are characterized by various weekly patterns of use, summaries of substance days-of-use used over longer periods can exhibit multiple modes. These characteristics of substance days-of-use data are not easily fitted with conventional parametric model families. We propose a continuation ratio ordinal model for substance days-of-use data. Instead of grouping the set of possible response values into a small set of ordinal categories, each possible value is assigned its own category. This allows the exact numeric distribution implied by the predicted ordinal response to be recovered. We demonstrate the proposed model using survey data reporting days of alcohol use over 28-day intervals. We show the continuation ratio model is better able to capture the complexity in the drinking days dataset compared to binomial, hurdle-negative binomial and beta-binomial models.

Introduction

Modeling the frequency of substance use presents significant statistical challenges. Researchers typically measure use over a fixed interval (e.g., 28 days), resulting in data that is both bounded and often multi-modal. Standard count models, such as Poisson or Negative Binomial distributions, are theoretically inappropriate as they assume an infinite upper bound. Furthermore, human behavior often follows weekly routines, creating "spikes" at intervals of seven or four days. Traditional distributions tend to smooth over these irregularities, losing critical behavioral signals. This paper proposes a many-levelled continuation ratio model to address these issues by treating every possible day-count as a distinct ordinal category.

Methods

The proposed approach utilizes a continuation ratio ordinal model. Unlike typical ordinal models that use broad categories (e.g., low, medium, high), this model assigns a unique category to every integer value from 0 to 28. The model conceptualizes the outcome as a sequence of binary decisions: given a participant used a substance for at least n days, what are the odds they used it for at least n+1 days?

By estimating a separate threshold parameter for each day, the model can account for non-linear transitions, such as the difficulty of moving from 6 to 7 days of use compared to 5 to 6. The model was implemented in a Bayesian framework using the brms package in R and compared against Binomial, Beta-Binomial, and Hurdle-Negative Binomial alternatives.

Data and Analysis

The study analyzed data from the Australians’ Drug use: Adapting to Pandemic Threats (ADAPT) study, involving 452 participants who regularly used illicit drugs prior to the COVID-19 pandemic. Data was collected across four longitudinal waves between May 2020 and July 2021, focusing on self-reported alcohol consumption over the 28 days preceding each survey.

Distribution of reported drinking days showing periodic modes every four days.

As shown in Figure 1, the raw data exhibits clear periodic modes at 4, 8, 12, 16, 20, 24, and 28 days. These peaks likely reflect weekly drinking habits or survey design prompts.

Results

The continuation ratio (C-Ratio) model significantly outperformed alternative models across several metrics. In terms of fit, the Leave-One-Out Information Criterion (LOO-IC) for the C-Ratio model (5665) was substantially lower than that of the Beta-Binomial (6473) and standard Binomial (8571) models.

Comparison of Empirical Cumulative Distribution Functions (ECDF) across different models. The C-Ratio model (bottom right) most closely follows the observed data (black line).

Visual analysis via Empirical Cumulative Distribution Function (ECDF) plots (Figure 2) confirms that the C-Ratio model is uniquely capable of tracking the "staircase" effect of the observed data. While the Beta-Binomial model offers an improvement over the simple Binomial, it still fails to capture the local heterogeneity and periodic spikes characteristic of the dataset.

Inference using the C-Ratio model revealed that participants in isolation or quarantine had approximately 15% higher odds of more frequent drinking (OR 1.15). Additionally, a general decline in drinking frequency was observed as the pandemic progressed through subsequent survey waves.

Conclusion

By reframing "days of use" as a many-levelled ordinal problem, researchers can gain the flexibility needed to model complex, bounded behavioral data. This approach allows for the recovery of exact numeric distributions while accurately capturing multi-modal patterns that standard parametric models miss. This methodology is broadly applicable to any bounded frequency data, including health symptoms or habit tracking.

Summary updated

Abstract

Studies of alcohol and drug use are often interested in the number of days that people use the substance of interest over an interval, such as 28 days before a survey date. Although count models are often used for this purpose, they are not strictly appropriate for this type of data because the response variable is bounded above. Furthermore, if some peoples' substance use behaviors are characterized by various weekly patterns of use, summaries of substance days-of-use used over longer periods can exhibit multiple modes. These characteristics of substance days-of-use data are not easily fitted with conventional parametric model families. We propose a continuation ratio ordinal model for substance days-of-use data. Instead of grouping the set of possible response values into a small set of ordinal categories, each possible value is assigned its own category. This allows the exact numeric distribution implied by the predicted ordinal response to be recovered. We demonstrate the proposed model using survey data reporting days of alcohol use over 28-day intervals. We show the continuation ratio model is better able to capture the complexity in the drinking days dataset compared to binomial, hurdle-negative binomial and beta-binomial models.
BibTeX
@misc{chambers2022many,
  author = {Mark S. Chambers and Christopher Drovandi},
  title = {{Many-levelled continuation ratio models for frequency of alcohol and drug use data}},
  year = {2022},
  eprint = {2205.11942},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2205.11942},
}
Figure