Background:

In observational studies on causality, including non-randomised therapeutic studies, the exposure and confounders are often considered constant over time. However, this assumption may not be valid, especially in case of time-varying confounding [Robins 1992, Robins 2000]. When exposure status affects confounder status, and this confounder again affects future exposure, conditioning on the confounder (as is done in a regular regression analysis), might induce bias rather than correct for confounding. To adequately estimate causal effects of e.g. interventions, initiation or stopping of medical therapy during follow-up should be taken into account, for example using time-dependent analyses. Several time-dependent analyses have been proposed, such as Cox proportional hazards analysis with time-dependent exposure, GEE, and inverse probability of treatment weighting (IPTW). In empirical data (observational or trials), however, these have not been compared throughout. Furthermore, theoretically the method could give different results, for example in case of time/varying confounding. However, to what extent time/varying confounding will affect studies over a wide range of clinical fields, is unclear. We therefore, aim to compare analyses for time-varying exposure for the association between breastfeeding and wheeze in two examples on observational studies (the Whistler cohort and the ALSPAC cohort), and compare the results to those from a randomised trial.

Objective:

To compare methods to assess time-varying exposure in observational studies on the effects of breastfeeding on asthma/eczema, and compare the associations found to those from a randomised controlled trial.

Research questions:

1. What is the association between breastfeeding and wheezing? (potential confounders: maternal smoking, being in a smoky room, daycare, maternal wheeze/eczema, age, weight and height, pets)

2. What is the association between breastfeeding and eczema? (potential confounders: maternal smoking, being in a smoky room, daycare, maternal wheeze/eczema, age, weight and height, pets)

Methods:

Several time-varying analyses will be applied and compared:

1) Exposure = ever vs. never breastfed (dichotomous measure);

2) Exposure = total duration of breastfeeding (i.e.,exposure is a continuous measure);

3) Exposure is time-dependent (i.e. breastfeeding at different ages of follow-up) and effects will be estimated using survival analysis with this time-dependent exposure. Confounders are considered constant over time (e.g. maternal smoking during pregnancy is taken as the constant value for the confounder "maternal smoking");

4) Exposure and confounders are both time-dependent. Effects are estimated using survival analysis.Follow-up is divided in discrete intervals (i.e., 1-year intervals). Each period is considered an (independent) observation. Covariates can differ at the baseline of each period.

5) Exposure and confounders are both time-dependent. Effects are estimated using Generalized Estimating Equations, thus taking dependency between observations into account.

6) Exposure and confounders are both time-dependent. Effects are estimated using marginal structural models using inverse probability of treatment weighting [Robins 2000]

7) Exposure and confounders are both time-dependent. Effects are estimated using G-estimation [Tilling 2002].