Estimating Long-Term Drinking Patterns for People with Lifetime Alcohol Use Disorder

Abstract

Background. There is a lack of data on alcohol consumption over time. This study characterizes the long-term drinking patterns of people with lifetime alcohol use disorders who have engaged in treatment or informal care. Methods. We developed multinomial logit models using the National Epidemiologic Survey on Alcohol and Related Conditions (NESARC) to estimate short-term transition probabilities (TPs) among the 4 World Health Organization drinking risk levels (low, medium, high, and very high risk) and abstinence by age, sex, and race/ethnicity. We applied an optimization algorithm to convert 3-year TPs from NESARC to 1-year TPs, then used simulated annealing to calibrate TPs to a propensity-scored matched set of participants derived from a separate 16-year study of alcohol consumption. We validated the resulting long-term TPs using NESARC-III, a cross-sectional study conducted on a different cohort. Results. Across 24 demographic groups, the 1-year probability of remaining in the same state averaged 0.93, 0.81, 0.49, 0.51, and 0.63 for abstinent, low, medium, high, and very high-risk states, respectively. After calibration to the 16-year study data (N = 420), resulting TPs produced state distributions that hit the calibration target. We find that the abstinent or low-risk states are very stable, and the annual probability of leaving the very high-risk state increases by about 20 percentage points beyond 8 years. Limitations. TPs for some demographic groups had small cell sizes. The data used to calibrate long-term TPs are based on a geographically narrow study. Conclusions. This study is the first to characterize long-term drinking patterns by combining short-term representative data with long-term data on drinking behaviors. Current research is using these patterns to estimate the long-term cost effectiveness of alcohol treatment.

Keywords: alcohol use disorder; calibration; drinking patterns; transition probabilities.

Age Groups (at the time of the NESARC-III survey)

Gender

Race/ethnicity

Years between drinking risk state measurements

Figure 1.

Transitions between drinking states. This figure displays health states, defined as mutually exclusive states of drinking behavior, represented by the circles. The four drinking states are defined by World Health Organization (2000) risk levels depending on grams of alcohol consumed per day, and a fifth state represents abstinence. The grams of alcohol per day for each risk level are shown inside the circles. Arrows show that people can move between all states or remain in the same state. Individual trajectories are described based on yearly transition between all health states. F: female; M: male.

Figure 2.

Developing transition probabilities. This figure describes the sequential process by which the transition probabilities were developed, beginning with the estimation of the multinomial logit models (top left) and proceeding through calibration with long-term data (bottom). NESARC: National Epidemiologic Survey on Alcohol and Related Conditions.

Figure 3.

Simulated annealing results. This figure shows how the distance measure (sum of squared differences of state distributions) changed over time as the simulated annealing algorithm was used to calibrate the second set of transition probabilities. The full line representing the distance at a given iteration became more stable as the “temperature” of the system fell.

Figure 4.

Transition probabilities by set. This figure displays the average transition probabilities for each set: short term (0–3y) and long term (3–8y and 8y+). The transition probabilities for each of the 24 demographic groups in this study were weighted by their respective proportion of the study population (those with lifetime alcohol use disorder and treatment participation) in the National Epidemiologic Survey on Alcohol and Related Conditions-III. Vertical bars represent 95% credible intervals.

Figure 5.

Model validation. This figure shows the observed drinking state distribution in the National Epidemiologic Survey on Alcohol and Related Conditions-III (NESARC-III) and the predicted drinking state distribution from applying the transition probabilities to an earlier time point (N=1244). The means and 95% confidence intervals for the observed proportions of the NESARC-III sample in each drinking risk state are shown in gray, and the predicted proportions in each drinking state are shown in black with 95% credible intervals. NESARC-III weights were applied, and survey design was accounted for to calculate observed state distributions.

Figure 6.

Model validation by starting state. This figure shows the observed drinking state distribution in the National Epidemiologic Survey on Alcohol and Related Conditions-III (NESARC-III) and the predicted drinking state distribution from applying the transition probabilities to an earlier time point, stratified by starting state. The means and 95% confidence intervals for the observed proportions of the NESARC-III sample in each drinking risk state are shown in gray, and the predicted proportions in each drinking state are shown in black with 95% credible intervals. NESARC-III weights and survey design were applied to calculate observed state distributions.

Figure 7.

Example case: 25-year-old males. This figure shows drinking state distributions over time for males starting at age 25 of different races and starting drinking risk states. This figure represents the influence of starting state and race.

Figure 8.

Example case: 50-year-old female. This figure shows drinking state distributions over time for females starting at age 50 of different races and starting drinking risk states. This figure represents the influence of starting state and race.

Figure 9.

Example case: non-Hispanic white, very high–risk drinker. This figure shows drinking state distributions over time for non-Hispanic white males and females starting at the very high–risk drinking state for different starting ages and genders. This figure represents the influence of age and gender.