Gestational misalignment with fixed exposure windows: the potential dangers of zero-filling in distributed lag models

Distributed lag models (DLMs) are widely used in perinatal epidemiology to identify critical windows during pregnancy in which environmental exposures influence pregnancy/birth outcomes. A well-known complication of fitting DLMs in this context is that they require the same length of exposure history for every individual even though not all pregnancies are of the same duration. This misalignment often leads researchers to artificially extend exposure histories to a fixed length and fill post-birth weeks with zeroes (i.e. "zero-filling"). Despite its widespread use, the implications of zero-filling have not been formally evaluated.

We demonstrate conceptually that zero-filling induces a spurious association between gestational age and late-pregnancy exposures, thus introducing confounding that was otherwise not present in the observed data. We then conducted a simulation study and a real data application using air pollution and birth weight data from a Colorado-based cohort to compare zero-filling with alternative approaches to handle the misalignment between exposure window and gestation.

In our simulations, we found that zero-filling produced the largest bias, poorest coverage, and highest root mean squared error. In the analysis of the Colorado birth data, zero-filling produced implausibly strong associations. Adjusting for gestational age attenuated this bias. Alternative approaches of carrying forward the last pre-birth exposure value, using observed post-birth exposures, and, in some situations, truncation at 37 weeks eliminate this bias.

Zero-filling can cause bias in the estimated associations when using distributed lag models.