Using the marginal structural model for finding the best moment to treat

Using the marginal structural model for finding the best moment to treat: Applied to a dataset consisting of dialysis patients treated with Erythropoiesis-Stimulating Agents

de Haas, J.J.

Jongbloed, G. (mentor)

Electrical Engineering, Mathematics and Computer Science

Delft Institute of Applied Mathematics

Probability, Risk and Statistics

2016-11-22

In this report we applied the marginal structural model and the naive Cox model on a simulated dataset, in order to find differences between these models. The marginal structural model is used to estimate causal effects in case of time-dependent confounding via inverse probability weights. The naive Cox model may produce a bias estimate of the causal effect, because it cannot correct for this time-dependent confounding. In the simulation study we saw that the estimated effect of treatment was different for both models. In some situations the models gave different conclusions, which could lead to a wrong decision in a real problem. We also applied the marginal structural model on `The Netherlands Cooperative Study on the Adequacy of Dialysis' (NECOSAD). Patients that participated on this study all have a chronic kidney disease. Because of kidney failure, these patients could suffer from anemia. In order to correct for this, they are treated with Erythropoiesis-Stimulating Agents (ESA). We compared treatment in dialysis patients with no/low ESA dose versus treatment with high ESA dose. We want to estimate the causal effect of treatment on survival. We used the marginal structural model, because Hb is a time-dependent confounder in this dataset. We saw an increased risk of mortality for the 'High ESA dose' patients in our analysis of 58%. This suggests that treatment with high ESA dose increases the risk of death. So treatment with high ESA dose is harmful according to our marginal structural model. At the moment, there is no strict protocol for doctors for when to start treatment. If treatment with high ESA dose is harmful it would be better to expose patients as short as possible to high doses. So we want to find the optimal regime to start treatment. This can be done by applying the dynamic marginal structural model. This model compares different regimes, such that we can find the regime which has the smallest causal effect on mortality. For this analysis we define different regimes of starting treatment. For instance, one regime could be `start treatment once the Hb level is below 10 g/dL'. With the dynamic marginal structural model we can compare these different regimes to find the best moment to treat. In our real data analysis we compared 6 regimes to find the best moment to treat based on Hb level. We applied the dynamic marginal structural Cox model and regime 3, initiate treatment with high ESA dose when Hb-level first drops below 11 g/dL, showed the best results.

inverse probability weights
marginal structural model
dynamic treatment regimes

http://resolver.tudelft.nl/uuid:06063c8a-1410-444c-bc6b-6b05e14af6f4

Student theses

master thesis

(c) 2016 de Haas, J.J.