A Brief Discussion on Time-Varying Confounding in the Evaluation of Programs, Practices, or Policies

hourglass image


At E4A, we’re constantly thinking about change: the relationship between changes – how health outcomes change in response to policy or practice changes – how to measure those changes, and how to account for other things that are changing at the same time (one type of confounding). This last point is crucial if we want to know whether a specific policy or practice change causes changes in health outcomes, or if those changes are the result of something else going on. It becomes even trickier to figure out what is driving changes in health outcome when the existence of or exposure to the policy or practice of interest occurs and reoccurs over time, with varying levels of intensity. Analyzing data in this type of scenario calls for a special set of statistical methods appropriate for evaluating the effects of time-varying conditions.

For instance, conditions such as receipt of mental health servicesaccess to greenspaceresidence in publicly subsidized housingsmoking status, exposure to varying levels of air pollution, and poverty can change over time and affect subsequent health status. We can describe these conditions in many ways, for example, based on whether individuals were ever exposed to the conditions; the duration of or most recent exposure; or the maximum intensity of the conditions, among many other possibilities. To estimate the health effects of exposure to a condition, we often need to consider the sequence or whole history of exposure, rather than the effect of exposure at a single point in time. For example, does sustained access to greenspace reduce cardiovascular risk? An important step in evaluating the health effects of time-varying changes is to specify how to characterize the pattern of exposure to the condition of interest. What feature of the condition (e.g., duration, frequency, intensity, etc.) is most relevant for our health outcome? Once the characterization is specified, confounding needs to be addressed.

Confounding is a major threat to the validity of all observational studies, and especially problematic for evaluating time-varying programs, policies, or interventions. Confounding occurs when other factors (confounders) affect the condition being evaluated and also affect the outcome of interest (for more on confounding, see our blog post and methods note on confounder versus instrumental variable designs). This “mixing” of effects distorts the causal effect of the condition on the outcome (the effect of the condition or intervention on the outcome is mixed with the effect of the confounder on the outcome). Just as conditions may vary over time, confounders may also vary over time.

Time-varying confounders are especially problematic when they are affected by prior conditions. When this is the case, adjusting for time-varying confounders is particularly challenging, because if we control for them using traditional regression, we remove part of the effect of interest, but if we don’t control for them, our estimates will be confounded. Although conventional regression methods are not appropriate when time-varying confounding occurs, several analytic methods can be used to address time-varying confounding. These methods support valid and actionable conclusions in studies with time-varying programs, policies, or interventions.

Example Research Questions

To better understand time-varying confounding and why it poses such a problem, consider the example of evaluating the effect of sustained public housing subsidies on children’s long-term cognitive outcomes. Housing subsidies may provide parents with stable housing that allows them to pursue additional education and improved work opportunities. Those enhanced work and income opportunities, attainable by virtue of stable housing may, in turn, affect both future eligibility for housing subsidies and children’s long-term health outcomes. To address such conundrums in time-varying settings, special methods are needed. We discuss these ideas and frequently asked questions about these methods in a separate Methods Note.

Possible Approaches

Several technical papers have been published in the last 35 years describing how to handle time-varying confounding appropriately. We summarize the main approaches (i.e., the so-called “G” methods) in our accompanying Methods Note. The uptake of methods for handling time-varying confounding has been slow, in part because the statistical tools can be complex and often require large sample sizes. In recent years, these methods have become easier to implement, and innovations have improved precision even for studies with moderate sample sizes.

Putting Evidence into Practice

E4A typically receives applications for experimental or quasi-experimental studies of interventions implemented at a single point in time because it is often more feasible to derive rigorous estimates of the effects of point-in-time interventions. Point-in-time studies provide essential evidence on the health effects of discrete interventions but may not reveal the effect of an intervention that is sustained or that varies in intensity over time. Yet evaluating such interventions might yield important evidence on policy impacts.

For instance, several E4A funded studies evaluate the causal effects of social safety net policies, such as the Women, Infants, and Children (WIC) supplement on child health and well-being. The WIC evaluation study funded by E4A originally leveraged instrument-based study designs to evaluate the health effects of receiving subsidies at a point-in-time. However, families' participation in WIC may decline over time due to other factors in families' lives. The influence of WIC support on health outcomes over time may differ for someone who has intermittent WIC support versus someone who has sustained WIC support. Similar questions about the effects of public housing subsidies, access to health care, and exposure to greenspace over the long-term are equally relevant. Evaluating these questions is methodologically challenging but overlooking them may substantially underestimate the benefits of these programs.

The slow adoption of methods for time-varying confounding is unfortunate because it makes research less relevant to changing population health. Exposure to many types of conditions would have little effect for people who received only short-term, small doses, but may have large and consequential effects after sustained exposure. For some conditions of special interest to health equity – for example, income – stability itself may be an important feature. The E4A-supported evaluation of the Stockton Economic Empowerment Demonstration examines the outcomes of income volatility itself, and it is important to understand how income variability influences health. Such analyses of long-term exposure or exposure variability will almost certainly require methods to address time-varying confounding.

Tools & Resources

The references below are good sources of information on methods for adjusting for time-varying confounding. The Methods Note for this blog also contains additional references and information on time-varying confounding. As always, we welcome your comments and feedback on these ideas!


  1. Clare, P. J., Dobbins, T. A., & Mattick, R. P. (2019). Causal models adjusting for time-varying confounding—a systematic review of the literature. International Journal of Epidemiology48(1), 254-265.
  2. Hernán MA, Robins JM (2020). G-methods for time-varying treatments. Causal Inference: What If (pp. 247-254). Charleston, SC: BN Publishing.
  3. Mansournia, M. A., Etminan, M., Danaei, G., Kaufman, J. S., & Collins, G. (2017). Handling time varying confounding in observational researchBMJ, 359.
  4. Vansteelandt, S., & Sjolander, A. (2016). Revisiting g-estimation of the effect of a time-varying exposure subject to time-varying confoundingEpidemiologic Methods5(1), 37-56. 
Blog posts

About the author(s)

Dakota Cintron, PhD, EdM, MS, is a postdoctoral scholar for the E4A Methods Laboratory. Dr. Cintron’s research focuses on the application, development, and assessment of quantitative methods in the social and behavioral sciences. 

Maria Glymour, ScD, MS, is an Associate Director of E4A that leads the E4A Methods Laboratory, as well as a social epidemiologist and Associate Professor in the Department of Epidemiology and Biostatistics at the University of California, San Francisco. She has dedicated much of her career to overcoming methodological problems encountered in observational epidemiology, in particular analyses of social determinants of health and dementia risk. 

Ellicott Matthay, PhD, is a social epidemiologist and postdoctoral scholar with E4A. She conducts methodological investigations to improve the way that research in her substantive areas is done, because she believes that improving the methodological rigor of applied studies is one of the most important steps to identifying effective prevention strategies.

Stay Connected