andrelaszlo on HackerNews asked how someone could draw a reasonable distinction between “direct” and “indirect” measurements in science. Below is how I answered. This is old hat to many folks and, needless to say, none of this is original to me.
[Edit: Sabine Hossenfelder has posted on this in the context of the BICEP2. I don’t agree with her that it’s reasonable to call this “direct” detection of inflation simply because we can’t think of a more direct way. I think a distinction in terms of Bayesian confidence in the causal chain is more useful.]
There’s a good philosophy of science argument to be made that there’s no precise and discrete distinction between direct and indirect measurement. In our model of the universe, there are always multiple physical steps that link the phenomena under investigation to our conscious perception. Therefore, any conclusions we draw from a perception are conditional on our confidence in the entire causal chain performing reliably (e.g. a gravitational wave induces a B-mode in the CMB, which propagates as a photon to our detectors, which heats up a transition-edge sensor, which increases the resistivity of the circuit, which flips a bit in the flash memory, which is read out to a monitor, which emits photons to our eye, which change the nerves firing in our brain). “Direct” measurements, then, are just ones that rely on a small number of reliable inferences, while “indirect” measurements rely on a large number of less reliable inferences.
Nonetheless, in practice there is a rather clear distinction which declares “direct” measurements to be those that take place locally (in space) using well-characterized equipment that we can (importantly) manipulate, and which is conditional only on physical laws which are very strongly established. All other measurements are called “indirect”, generally because they are observational (i.e. no manipulation of the experimental parameters), are conditional on tenuous ideas (i.e. naturalness arguments as indirect evidence for supersymmetry), and/or involve intermediary systems that are not well understood (e.g. galactic dynamics).
The classic example is dark matter detection. A detector built in your laboratory that produces clear evidence of a local interaction between the dark matter particle and the atoms composing the detector would be “direct detection”. Seeing an anomalous excess of gamma rays from the center of the galaxy whose energy and distribution is consistent with some theories which predict dark matter annihilation would be “indirect detection”. Likewise, pulsar slow downs are indirect measurements of gravitational waves, while gravitational wave detectors like the LISA proposal are direct.
Naturally, direct measurements have a much larger impact on your Bayesian credences than indirect ones. If someone says “I don’t trust that indirect measurement” they mean “one or more steps in the inference chain which connects the phenomena to our perceptions is unreliable”.