Empirical Investigation of Causal Inference in the Development of Smoking Behavior among Adolescents Population in the PATH Dataset

Advisor: Dr. Shu Xu, Dr. Raymond Niaura, Dr. Jennifer Cantrell

PRESENTED BY Zhihao Chen

Department of Biostatistics, NYU School of Global Public Health

05/10/2023

Outline

• Introduction
• Motivation and gaps in research
• Methods
• Study data, deign, and BN algorithms
• Results
• Univariate analysis
• Causal pathway leads to e-cig initiation
• Average Causal Effect
• Discussion
• Future research

2

Gaps in Previous Research

• Traditional General/Generalized Linear Models suffer from limitations including multicollinearity, non-linear relationship, increased computational cost, and low model interpretability.
• Limited guidance on selecting confounder variables

3

Causal Inference with Bayesian Network

• Causal inference allows researchers to infer “the response of an effect when a cause of the effect variable is changed.”
• There are several variable selection methods, of which we are interested in selecting variables along the causal pathway, with intuitive visualization of this pathway in the format of a probabilistic graphical model, naming the Bayesian Network (BN).

4

Bayesian Network

• A directed acyclic graph (DAG)
• Each node represents a variable
• The present of an edge between two nodes represents they are conditionally dependent
• The DAG of BN determines which nodes are conditionally (in)dependent
• BN is constructed by computing structural aspect of the graph, by adding, removing, reversing direction of the edges to optimize the Minimum Description Length (MDL) Score

5

Research Question: What is the average causal effect of Wave 2 e-cigarette initiation on Wave 3 past-30 days combustible cigarette use.

​

6

Data: Population Assessment of Tobacco and Health (PATH)

• This longitudinal cohort survey collected smoking behavior across multiple follow-up waves, with a baseline wave population of 45,971 adults and youth.
• Study sample: n = 7,340
• Wave 1 tobacco naïve Wave 2 cigarette naive population and Wave 3 new combustible cigarette users
• Imputed dataset from the Xu et al., (2022)

7

Measures

• Exposure (Wave 2): E-cig initiation
• Outcome (Wave 3): Past-30-day cigarette use
• Covariates (Wave 1):
• Age, Ethnicity, Smoking Susceptibility, Free tobacco product access, Exposure to second-hand smoking, Last time since self-related/environmental related mental health crisis

8

Bayesian Network (cont)

• Sum of Products Linear Equation (SopEQL) algorithm was used for structure learning and parameter learning
• Construct a causal network among variables by iteratively testing and selecting the most relevant variables for the network based on a scoring function.

[Violet: Put tech details in bullet points. E.g., software pacakge, algorithm, name of the scoring function et al. Remove the network on the right from this page because this is a part of your result (do not belong to the method section]

9

Results

• Univariate analysis [Violet: Need a description of univariate analysis results; e.g., xx.x% male …]
• [Add: causal pathway]
• An increase of 4.6% of prevalence risk was found
• E-cigarette initiation in Wave 2 youth population was associated with a higher risk of using combustible cigarettes in the past 30 days period in Wave 3, represented by an increase of 4.6% in the risk, should the sampled population did not frequently use e-product in Wave 2 as the counterfactual outcome.

10

11

Discussions and Conclusions

• E-cigarette and subsequent combustible cigarette
• Advantages of a BN approach
• Flexiable nonparamentric approach
• Able to handle complex interaction
• Visualization of data structure
• Infer causal effect directly from data

12

Conclusions

• Limitations
• The Bayesian Network is a method that is constantly being updated, thus there might be limitations from omitting algorithms that is at state-of-art
• Bias from data collection and the original PATH study
• Future research
• Policy application in tobacco reduction
• Adolescent intervention program on e-cigarette

13

References

• Soneji S. Errors in data input in meta-analysis on association between initial use of e-cigarettes and subsequent cigarette smoking among adolescents and young adults. JAMA Pediatrics. 2018;172(1):92. doi:10.1001/jamapediatrics.2017.4200
• BayesiaLab. BayesiaLab Overview. Bayesia. Accessed April 4, 2023. https://www.bayesia.com/articles/#!bayesialab-knowledge-hub/bayesialab-overview
• Xu S, Coffman DL, Liu B, Kollath-Cattano C, Pesko MF. Relationships Between E-cigarette Use and Subsequent Cigarette Initiation Among Adolescents in the PATH Study: an Entropy Balancing Propensity Score Analysis. Prev Sci. 2022. doi:10.1007/s11121-021-01326-4
• Conrady S, Jouffe L. Chapter 10: Causal Effect Identification and Estimation. In: Bayesian Networks and BayesiaLab: A Practical Introduction for Researchers. Franklin, TN: Bayesia USA; 2015. https://www.bayesia.com/articles/#!bayesialab-knowledge-hub/book. Accessed April 4, 2023.

​

​

​

​

14