A Viable Alternative When Propensity Scores Fail: Evaluation of Inverse Propensity Weighting and Sequential G-Estimation in a Two-Wave Mediation Model

Adjusting for Baseline Measurements of the Mediators and Outcome as a First Step Toward Eliminating Confounding Biases in Mediation Analysis

Mediation analysis prevails for researchers probing the etiological mechanisms through which treatment affects an outcome. A central challenge of mediation analysis is justifying sufficient baseline covariates that meet the causal assumption of no unmeasured confounding. But current practices routinely overlook this assumption. In this article, we suggest a relatively easy way to mitigate the risks of incorrect inferences resulting from unmeasured confounding: include pretreatment measurements of the mediator(s) and the outcome as baseline covariates. We explain why adjusting for pretreatment baseline measurements is a necessary first step toward eliminating confounding biases. We hope that such a practice can encourage explication, justification, and reflection of the causal assumptions underpinning mediation analysis toward improving the validity of causal inferences in psychology research.

How effective is reducing traffic speed for safer work zones? Methodology and a case study in Pennsylvania

Transportation agencies post and enforce reduced speed limits in work zones to ensure work zone safety, since traffic speed is found to be associated with work zone crash risks. However, prior findings on the relationship between speed and crash rate in work zones are inconsistent. This may be attributed to the methods of statistical associations between traffic speed and crash risks that do not necessarily discover true causal relations. In fact, work zone presence could lead to the reduction of actual traffic speed that influences crash risks, where it may also directly impose effects on crash risks as a result of work zone configurations. The actual traffic speed (not posted speed limit) is also known as a "mediator" where work zones can indirectly impact the crash risks. It is challenging to rigorously separate the causal effect of traffic speed on work zone crash risk from that directly caused by work zones. The underlying causal relation could help to determine what reduced post speed limit (with enforcement) is necessary to ensure work zone safety under the most desired "actual traffic speed". This study proposes to use the sequential g-estimation and the regression discontinuity design to estimate the controlled direct effect of traffic speed on work zone crashes. Two research gaps are identified and filled: inaccurate inferences of the effect of reduced speed limit in work zones as a result of ignoring (1) potential post-treatment bias since traffic speed is a mediator; and (2) potential confounding bias caused by unobservable roadway characteristics. The proposed methodology was applied to 4008 work zones in Pennsylvania from 2015 to 2017, and the results were validated through a series of robustness tests. The results indicate that the direct causal effect of the presence of work zones on crash risk is significantly positive when the traffic speed is relatively low (i.e., lower than 55 mph in this case study), while traffic speed has a positive causal effect on crash occurrences when the actual traffic speed is high (i.e., greater or equal to 55 mph). It suggests that strictly enforcing reduced posted speed limits in work zones is particularly effective when the actual traffic speed is greater than 55 mph. This is particularly true on roadways with high traffic volume (i.e., AADT > 20,000 vehicles per day), long, and daytime work zones (i.e., > 3000 m). On the other hand, the effect of enforcing reduced speed on work zone safety is unclear when the actual speed is already low. In this case, improving work zone configurations and driving behaviors may be more effective in reducing crash risks.

The Effect of Noninvariance on the Estimation of the Mediated Effect in the Two-Wave Mediation Model

The two-wave mediation model is the most suitable model for examining mediation effects in a randomized intervention and includes measures taken at pretest and posttest. When using self-report measures, the meaning of responses may change for the treatment group over the course of the intervention and result in noninvariance across groups at posttest, a phenomenon referred to as response shift. We investigate how the mediated effect would be impacted by noninvariance when using sum scores (i.e., assuming invariance). In a Monte Carlo simulation study, the magnitude and proportion of items that had noninvariant intercepts, the direction of noninvariance, number of items, effect size of the mediated effect and sample size were varied. Results showed increased Type I and Type II errors due to a biased estimate of the intervention effect on the mediator resulting from noninvariance. Thus, measurement noninvariance could lead to erroneous conclusions about the process underlying the intervention.

Assumptions Not Often Assessed or Satisfied in Published Mediation Analyses in Psychology and Psychiatry

Mediation analysis aims to investigate the mechanisms of action behind the effects of interventions or treatments. Given the history and common use of mediation in mental health research, we conducted this review to understand how mediation analysis is implemented in psychology and psychiatry and whether analyses adhere to, address, or justify the key underlying assumptions of their approaches. All articles (n = 206) were from top academic psychiatry or psychology journals in the PsycInfo database and were published in English from 2013 to 2018. Information extracted from each article related to study design, covariates adjusted for in the analysis, temporal ordering of variables, and the specific method used to perform the mediation analysis. In most studies, underlying assumptions were not adhered to. Only approximately 20% of articles had full temporal ordering of exposure, mediator, and outcome. Confounding of the exposure-mediator and/or mediator-outcome relationships was controlled for in fewer than half of the studies. In almost none of the articles were the underlying assumptions of their approaches discussed or causal mediation methods used. These results provide insights to how methodologists should aim to communicate methods, and motivation for more outreach to the research community on best practices for mediation analysis.

Clarifying the Implicit Assumptions of Two-Wave Mediation Models via the Latent Change Score Specification: An Evaluation of Model Fit Indices

Statistical mediation analysis is used to investigate mechanisms through which a randomized intervention causally affects an outcome variable. Mediation analysis is often carried out in a pretest-posttest control group design because it is a common choice for evaluating experimental manipulations in the behavioral and social sciences. There are four different two-wave (i.e., pretest-posttest) mediation models that can be estimated using either linear regression or a Latent Change Score (LCS) specification in Structural Equation Modeling: Analysis of Covariance, difference and residualized change scores, and a cross-sectional model. Linear regression modeling and the LCS specification of the two-wave mediation models provide identical mediated effect estimates but the two modeling approaches differ in their assumptions of model fit. Linear regression modeling assumes each of the four two-wave mediation models fit the data perfectly whereas the LCS specification allows researchers to evaluate the model constraints implied by the difference score, residualized change score, and cross-sectional models via model fit indices. Therefore, the purpose of this paper is to provide a conceptual and statistical comparison of two-wave mediation models. Models were compared on the assumptions they make about time-lags and cross-lagged effects as well as statistically using both standard measures of model fit (χ2, RMSEA, and CFI) and newly proposed T-size measures of model fit for the two-wave mediation models. Overall, the LCS specification makes clear the assumptions that are often implicitly made when fitting two-wave mediation models with regression. In a Monte Carlo simulation, the standard model fit indices and newly proposed T-size measures of model fit generally correctly identified the best fitting two-wave mediation model.

Causal Mediation Analysis in the Presence of Post-treatment Confounding Variables: A Monte Carlo Simulation Study

In many disciplines, mediating processes are usually investigated with randomized experiments and linear regression to determine if the treatment affects the outcome through a mediator. However, randomizing the treatment will not yield accurate causal direct and indirect estimates unless certain assumptions are satisfied since the mediator status is not randomized. This study describes methods to estimate causal direct and indirect effects and reports the results of a large Monte Carlo simulation study on the performance of the ordinary regression and modern causal mediation analysis methods, including a previously untested doubly robust sequential g-estimation method, when there are confounders of the mediator-to-outcome relation. Results show that failing to measure and incorporate potential post-treatment confounders in a mediation model leads to biased estimates, regardless of the analysis method used. Results emphasize the importance of measuring potential confounding variables and conducting sensitivity analysis.