Table 2 Technical considerations required by IPD-MAs

Harmonization

Harmonization efforts often require intensive feedback from stakeholders and can therefore take an enormous amount of time—one study reported that their harmonization took nearly two years [70], and a report by Cochrane Group members suggested that researchers should not expect their first publication before 3 years [9].

Missing Data

What to do with missing data becomes more complex as studies are combined into one meta-analysis as the reasons for missingness (varying protocols for measurement; some studies may not capture the variable of interest; or, if studies all capture the variable of interest, the reason for its missingness may also vary across studies). The choice methods to account for these differences, then, require additional consideration, e.g., omission, imputation. And, if the choice for imputation is made, thoughtful decisions about imputation models and the order in which to conduct the next steps (one-step, two-step, Rubin’s rules) are required.

Pooling

There are trade-offs with both methods [71]: one-step meta-analysis methods require advanced statistical know-how, and two-step methods, though employing more widely-known methods (e.g. random-effects or inverse-variance fixed-effects), are a more arduous endeavor. Two-step methods were the more dominant across all medicinal fields in the past, but the one-step methods have been increasing some medical fields as knowledge of and software to assist have improved [72]. In fields like epidemiology, one-step methods have been said to be the dominant choice due to adjustment for covariates other than treatment.

Causal Methods

As this review showed, authors of studies included in this review focused their reporting on confounding control. However, there are additional forms of bias which can be present in longitudinal studies, whether purely observational or pooled with RCTs, such as time-varying confounding. Causal methods can be extremely beneficial to remove some forms of bias, e.g., unmeasured confounding, time-varying confounding. Some of these methods are common in the disciplines where they were created, e.g., Regression Discontinuity in Economics, but have been recently increasing in medicine. Implementing these novel methods across multiple studies with different study designs may require additional time and resources initially but may yield powerful results for the field of global health and population medicine.