Format: Paper Presentation (English)
Date: May 31, 2019
Location: San Francisco, California, USA
Introduction: Individual participant data (IPD) meta-analysis has emerged as an important methodological approach to generating robust large-scale evidence in medical research as well as in the social and behavioral sciences. IPD meta-analysis can accommodate variations in study design to a much greater extent than traditional meta-analysis using summary statistics, as well as a wider array of outcome variables, particularly those that occur in prevention research.
In prevention research, count or frequency variables are common, such as the number of drinks consumed per week. Such outcomes are typically highly skewed with high frequencies of zeroes and more appropriately modeled using zero-altered count regression approaches. Another key challenge in meta-analysis is how to accommodate studies with heterogeneous designs, including differing numbers of treatment arms and longitudinal assessments. Typical options such as collapsing multiple treatment arms into a single arm, excluding treatment arms or studies, or conducting multiple univariate meta-analyses are not ideal due to loss of information and inefficient use of data.
We present a Bayesian multilevel modeling (BMLM) approach to highly skewed count outcomes with many zeroes in a one-step IPD meta-analysis that simultaneously accommodates between-study differences in the number of treatment groups and the number and timing of assessments when deriving an overall effect size estimate. A version of this modeling approach was used in a previous substantively-focused paper, which has been developed further for dissemination. We will provide an in-depth examination of the methodological issues encountered in our previous IPD meta-analysis, a rationale for the analysis approach we developed, and how prevention researchers can apply the same method to their own meta-analyses, using provided code in R.
Methods: We used a novel formulation of a Bayesian multilevel hurdle negative binomial model to calculate study-specific and overall treatment effects on zero-altered drinking outcomes across heterogenous studies. Illustrative data come from Project INTEGRATE, an IPD meta-analysis study of brief motivational interventions to reduce excessive alcohol use and related harm among college students.
Results: The one-step BMLM we developed was a feasible approach for producing study-specific and overall estimates of treatment effects with highly skewed count data. Sensitivity analyses supported the validity and robustness of the treatment effect estimates.
Conclusions: One-step IPD meta-analysis provides a feasible and flexible approach for combining IPD from heterogeneous studies that leverages all available information, while accommodating common differences in study design (e.g., varying treatment arms).
[ PDF slides ]
Format: Poster (English)
Date: June 24, 2014
Location: Bellevue, Washington, USA
Objective: For over two decades, brief motivational interventions (BMIs) have been implemented on college campuses to reduce heavy drinking and related negative consequences. Such interventions include in-person motivational interviews (MIs), which use an empathic and non-confrontational approach to increase motivation to change, often by providing personalized feedback (PF) designed to heighten awareness of patterns of use, norms, and related consequences, and non-in-person PF interventions delivered via mail, computer, or the web. Both narrative and meta-analytic reviews using aggregate data from published studies suggest at least short-term efficacy of BMIs, although overall effect sizes have been small.
Method: The present study was an individual participant-level data (IPD) meta-analysis of 17 randomized clinical trials evaluating BMIs. Unlike typical meta-analysis based on summary data, IPD meta-analysis allows for an analysis that correctly accommodates the sampling, sample characteristics, and distributions of the pooled data. In particular, highly skewed distributions with many zeroes that are typical for drinking measures need to be accounted for, and cannot be accommodated in meta-analyses of summary statistics. The data come from Project INTEGRATE, one of the largest IPD meta-analysis projects to date in the field of alcohol intervention research, representing 8,275 individuals each with 2 to 5 repeated measures up to 12 months post-baseline.
Results: We used Bayesian multilevel Hurdle Poisson and Gaussian models to estimate intervention effects on (1) drinks per week and (2) alcohol problems, respectively. Overall intervention effects were small and not statistically significant for drinks per week (Logit: OR=0.86, CI=0.71, 1.06; Count: RR=0.98, CI=0.95, 1.01) and alcohol problems (B=0.01, CI=-0.01, 0.04) at alpha = 0.05. We conducted post hoc comparisons of three BMI types (Individual MI with PF, PF only, and Group MI) vs. control. There were no statistically significant intervention effects for drinks per week. However, there was a small, statistically significant reduction in alcohol problems among participants who received an individual MI with PF (B=-0.06, CI=-0.11, 0.00). Short-term and long-term results were similar.
Conclusions: The results suggest a need for the continued development of more effective intervention strategies to reduce harmful drinking on college campuses. Supported by grants R01 AA019511 and T32 AA007455.
[ PDF poster ]
Format: Oral talk (English)
Date: May 20, 2014
Location: Storrs, Connecticut, USA
In this methodological illustration, we review challenges in individual participant-level data (IPD) meta-analysis with a focus on how to appropriately combine studies with varying numbers of treatments, a prevalent but under-addressed issue in meta-analysis. In the context of college drinking interventions, alcohol outcome data are particularly challenging due to highly-skewed distributions with many zeroes, a characteristic ignored in meta-analysis using summary statistics. We present a Bayesian multilevel modeling approach for combining two treatment arm and multiple arm trials in a distribution-appropriate IPD analysis. Illustrative data from Project INTEGRATE, an IPD study of brief motivational interventions designed to reduce excessive alcohol use and related harm among college students. The innovative analytical approach we present provides a practical method for estimating study-specific and overall treatment effects without the need to collapse intervention conditions within multi-intervention studies while accommodating non-normal distributions and other common characteristics of clinical trial data.
[ PDF slides ]
Format: Poster (Spanish)
Date: August 28, 2013
Location: Buenos Aires, Argentina
El abuso de alcohol tiene un gran coste para el individuo y la sociedad (Hingson et al., 2005; Perkins, 2002). La evaluación de los factores que predicen el uso de alcohol a través del tiempo es un foco importante de investigación, ya que el desarrollo de modelos teóricos puede dar información útil y hacer más efectivos los tratamientos. El análisis de datos del uso de alcohol es un reto debido a: (1) patrones de consumo crecientes y decrecientes durante la semana y (2) frecuencias altas de no consumo. El método más común es dividir los días de la semana en dos categorías (fines de semanas vs. días laborales), pero existe el riesgo de perder detalles importantes sobre días específicos. Además, los análisis típicos ignoran la alta tasa de no consumo, con la posible consecuencia de generar conclusiones incorrectas. Aquí se presenta un método alternativo que tiene en cuenta esas dos características importantes, usando los datos de un estudio de mujeres universitarias.
[ PDF poster ]
Format: Oral talk (English)
Date: May 21, 2013
Location: Storrs, Connecticut, USA
Daily drinking data often show highly skewed distributions that are bounded at zero as well as regular patterns across days of the week. Alcohol researchers have typically relied upon dummy variables for either weekend vs. weekday or for each day of the week. The present research evaluated the use of cyclical terms (i.e., sine and cosine regressors) in a zero-altered regression as a model for daily drinking data, in comparison to two dummy variable modeling strategies. Results showed that the cyclical model provided a more parsimonious approach than multiple dummy variables. The number of drinks when drinking had a rhythmic pattern that was reasonably approximated by cyclical terms, but either of the dummy variable approaches were a better model for the probability of any drinking. The combination of cyclical terms with zero-altered regression represents a feasible option for evaluating longitudinal drinking with high zero counts. However, drinking patterns are not perfectly sinusoidal, so care must be taken in evaluating the fit of models incorporating cyclical terms.
[ PDF slides ]
Last modified: 03 Jun 2019
dhuh@uw.edu
(206) 543-3692
(206) 543-1228