Separate frequentist inference, frequency probability, and estimation

This is a short post, but I think people need to separate frequentist inference, frequency probability, and estimation a bit better when describing statistical practices.

Some concepts in an oversimplified nutshell:

  • Frequentist inference: Inferring about an unknown parameter or process based on assumed or known frequentist properties. E.g., “If the generative parameter were zero, the [assumed, frequentist] probability of observing something at least as extreme as what I observed is low, so I will assume the generative parameter is not zero.” This inference is based on an assumed distribution resulting from repeated sampling of the same process, and from this assumed distribution, the frequency of occurrence of some effect (or more extreme) is small. One obtains data, employs an assumed distribution of events, and makes inferences about the relative frequency of occurrences.
  • Frequency probability: The computation of a probability from frequencies of events. This is required for frequentist inference, but is not limited to frequentist inference. Bayes theorem can use frequency-derived probabilities. Frequency-derived probabilites are sometimes used as data for a Bayesian model. Simulations of model-based inferential decisions use frequency probabilities — The frequentist probability of making Bayesian-informed decisions may be computed, and is irrelevant to frequentist inference per se. It is what it is — The probability of an event over repeated instances; this is useful for so many contexts outside of frequentist inference.
  • Estimation — Does not depend on frequentism or frequentist probabilities. I throw this in because I see people say “frequentism assumes a uniform prior” and that’s not really true. Maximum likelihood estimation may implicitly assume a uniform prior, but that is not relevant to frequentist inference. You could use penalized maximum likelihood (i.e., a non-uniform prior combined with a likelihood function) and still be frequentist; you’re a frequentist once your inferential logic depends on an assumed universe of possible repeated events.

You can be a Bayesian and accept the existence of frequency probabilities. But as a Bayesian, you do not construct frequency probabilities for use in inference, as inference comes from posterior uncertainty. Frequent_ists_ construct frequency probabilities with an assumed sample space of events, and use the implied frequency probability for inference. In fact, modern Bayesians can’t really reject the premise of frequency probabilities, given that posterior inferences are typically defined from the frequentist properties of an MCMC posterior sampler. What is p(theta > 0 | data)? You estimate it from the proportion of representative posterior MCMC samples that are above zero — It’s a frequency probability used to estimate a Bayesian quantity. There’s nothing inconsistent about this — Seriously, there isn’t. In the same way that a Bayesian can still flip a fair coin 100 times and see that 54% of the flips resulted in heads; that probability is not wrong or invalid, we just don’t use assumed sampling distributions of possible coin flips for inference about the fairness of a coin. The inference about fairness comes from posterior uncertainty given that 54/100 of the flips were heads, instead of “assuming the coin is fair, what is the probability of seeing 54/100 of coin flips?”. Frequency probabilities are not bound to frequentist inferences, but are required for frequentist inference.

You can be a frequentist and use non-uniform “priors” (in the sense of penalized maximum likelihood), because estimation is separate from frequentist inference. The frequentist inference would be employed the second you construct a hypothetical sampling distribution of possible events under some parameter given some statistical model (including a PML-estimated model) for use in inferring from some observed estimate. I.e., a frequentist may use PML to estimate a model, then construct a sampling distribution of PML-model derived estimates under some fixed parameter value, and compute the probability of a given estimate (or greater) from that sampling distribution. That’s still frequentist, even with a estimation penalty, simply because the inference is derived from the frequency of events under an assumed counter-factual.

TLDR; frequentist inference depends on frequentist probabilities and properties. Frequentist probabilities are not limited to frequentist inference. Bayesians may still use frequency-probabilities (as data, as priors [See Jaynes], in MCMC samplers, to understand frequentist properties of Bayesian estimators and decisions over repeated instances). Estimation is independent of frequentism, and frequentism is not defined by the use of implicit “uniform” priors (that is a property of the estimator, not the inferential logic).

Leave a Reply