Mendelian Randomisation (MR) uses the special properties of germline genetic variation to strengthen causal inference regarding how modifiable exposures influence disease outcomes. Few papers were published in the first decade of its 20 year history, but since then there has been an exponential increase, which shows no sign of slowing down. The vast majority of current papers are, at best, uninformative, and at worst, nonsense.

In this talk I will restate the fundamental assumption of MR – that of gene-environment equivalence – and the conditions for causal effect identification and estimation. I will then group the threats under three headings:

Noodles: papers that are prima facie nonsense, generated from easily available two sample MR data and sub–ChatGPT text.
No Nulls: influential early MR findings suggested that some drug targets — such as HDL cholesterol level or C-reactive protein — were unlikely to be important. Null MR studies are now increasingly rarely seen.
Numb Skulls: the simplicity of the MR approach is being obscured by increasingly complex methods, the assumptions of which will be opaque to most readers and most authors of papers implementing them. The first of what should be a series of retractions of papers based on one such method has occurred
The common source of these threats will be discussed, approaches to mitigating them advanced, and some promising future directions introduced.