Methods for addressing complex causal questions relating to health research in low- and middle-income countries

Identifying and understanding the causes of death and disease is key to improving health world-wide. But establishing and quantifying cause-effect relationships is difficult.
Suppose that a large number of HIV-positive patients are prescribed one of two drugs: a well-established drug, and a new competitor. If more patients on the new drug develop AIDS, can we conclude that it is inferior? What if the new drug was prescribed to more seriously ill patients?
Indeed, the decision as to which drug to prescribe is likely to be influenced by a patient‘s CD4 count, a measure of immunity and hence disease severity. But the drugs work precisely by increasing CD4 count. Thus CD4 count affects the choice of drug and vice versa. This so-called causal feedback makes drawing causal inferences even more difficult; but methods have been developed to address this.
This Fellowship aims to extend and apply these methods to address three important questions, each affected by causal feedback, in low- and middle-income country settings: (1) what is the causal effect of diarrhoea on stunted growth? (2) among HIV-positive individuals, what is the causal effect of antiretroviral therapy on hospitalisation? (3) what is the causal effect of maternal fertility on child mortality?

This proposal focuses on three situations in low- and middle-income countries in which epidemiologic data are used to address complex causal questions. The questions identified are complex as they face the problem of causal feedback. That is, interest lies in estimating the causal effect of a variable A on a variable B, but future values of A are also potentially affected by past values of B. For example, it is believed that high fertility leads to high child mortality and also that high child mortality can influence mothers‘ decisions to have more children. If the likely impact of family planning programmes on child mortality is to be estimated, then these two causal pathways - acting in opposing directions - must be teased apart.

Three methods (g-computation formula, inverse-probability-weighted-estimation of marginal structural models, g-estimation of structural nested models) have been developed to estimate causal effects in the presence of causal feedback, using longitudinal data. However, these methods suffer from some limitations.

The broad aim is to extend the three methods to address some of these limitations, to apply the extended methods in settings where, hitherto, they have not been applied, and in doing so, to make them more accessible to applied researchers.

Specifically, inverse-probability-weighted-estimation of MSMs will be extended to deal with repeated time-to-event outcomes, and g-estimation of SNMs will be extended to fit a logistic model for binary outcomes. Attrition can be dealt with in all three methods under the missing at random assumption, but extensions are required to deal with intermittent patterns of missingness. Sensitivity analyses, assessing the robustness of conclusions to violations of structural assumptions (such as missing at random and no unmeasured confounders), will be extended to longitudinal settings. Finally, existing software routines will be extended and new ones written in Stata and R.

These objectives are motivated by the need to address important complex causal questions in three settings, and the work will be illustrated using high-quality datasets from these settings.

The first concerns the causal effect of episodes of diarrhoea during early childhood on stunting at 24 months. Causal feedback is present since malnourishment (manifested in stunting) also leads to increased risk of diarrhoea.

The second concerns the effect of antiretroviral therapy on hospitalisation in HIV-positive individuals. Here, there is feedback between ART use and the important confounder, CD4 count.

The final setting is the one mentioned above, concerning the causal effect of fertility on child mortality.