Analyzing clinical health registries: improved software and mathematics of identifiability

We have developed new methods for estimating causal effects in survival analysis based on registry data. We have used techniques from stochastic differential equations, that are commonly used in finance, to reformulate the marginal structural models to situations with processes in continuous time. We have furthermore applied these techniques to develop methodology for identification and estimation of various causal effects i survival analysis. Furthermore, we have criticized the use of hazards as effect measures in event-history analysis in general. A consequence of this critique is that Cox-regression, a very popular method for estimating treatment effects in clinical studies, does not really give causal effects, not even for randomized controlled trials. Our critique does also have consequences for the interpretability of statistical hypotheses that are formulated with hazards. Several of the most commonly used hypothesis tests in survival analysis are problematic because of this. We have applied stability theory for differential equations to develop several alternative statistical methods that do not suffer from these problems.

Prosjektet har ført til: 1) Økt oppmerksomhet på at noen vanlige parametre som har har mye brukt som effektmål i kliniske studier ofte ikke har meningsfylte kausale fortolkninger. 2) Vi har introdusert en generell metode basert på differensialligninger for å estimere mange alternative mål som har en enklere kausal fortolkning i mange kyniske sammenhenger. 3) Vi har utviklet nye grafiske kriterier for å avgjøre hvorvidt en kausal effekt er identifiserbar fra de tilgjengelige observasjonene. 4) Vi har utviklet nye effektmål og tilhørende metoder for å estimere kausale effekter situasjoner, der man har konkurerende risikoer. 5) Vi har utviklet software i R-biblioteker for å utføre disse analysene.

A central theme in Kjetil Røysland's research at the Department of Biostatistics has been to extend modern causal reasoning to survival analysis. A key finding is that this is feasible if we let local characteristics for stochastic processes play the role of the conditional densities from the recursive factorisations of Bayesian networks. Local independence graphs relate to local characteristics as the directed acyclic graphs relate to the conditional densities in Bayesian networks. This gives a powerful generalisation of Causal Bayesian networks and enables us to apply many common ideas from causal inference directly to event history analysis. At the moment, applications of this methodology requires some seemingly unnecessary mathematical knowledge due to lack of generic identifiability conditions and adequate software. Nevertheless, it still provides an important opportunity to epidemiology and clinical research. The main objective of the current research proposal is to make causal inference based on local independence graphs more accessible to non-expert users. Our focus is two-folded, as we aim to both provide more generic identifiability conditions, and also to develop more mature software for performing actual data-analyses.