Jump to main navigation

Tutorial 4.2 - Opposing Statistical Philosophies

27 Mar 2017

The two major opposing philosophies (frequentist and Bayesian) differ in how they consider population parameters and thus how interpretations are framed.

Frequentist (classical)

Variation in observed data allows the long-run frequency of different outcomes to be approximated (we can directly measure the likelihood of obtaining the data given a certain null hypothesis).

In the frequentist framework, population parameters (characteristics of the population) are considered fixed quantities (albeit unknown). That is, there is one true state (e.g. mean) per population. The observed data (and its characteristics - sample mean) however, represents just one possible outcome that could be collected from this true state. Multiple simultaneous data collection experiments would yield different outcomes. This notion that there is a single true population parameter and a distribution of possible sample outcomes is the basis for a framework that conveniently allows a logical and relatively simple algebraic link between data, hypotheses and probability.

Traditional (frequentist) statistics focus on determining the probability of obtaining the collected data, given a set of parameters (hypothesis). In probability notation, this can be written as: $$P(D|H)$$ where $D$ represents the collected data and $H$ represents the set of population parameters or hypothese(s).

The process of inference (hypothesis testing) is based on approximating long-run frequencies (probability) of repeated sampling from the stationary (non-varying) population(s). This approach to statistics provides relatively simple analytical methods to objectively evaluate research questions (albeit with rigid underlying assumptions) and thus gained enormous popularity across many disciplines.

The price for mathematical convenience is that under this philosophy, the associated statistical interpretations and conclusions are somewhat counter-intuitive and not quite aligned with the desired manner in which researchers would like to phrase research conclusions. Importantly, conclusions are strictly about the data, not the parameters or hypotheses.

Inference (probability and confidence intervals) is based on comparing the one observed outcome (parameter estimate) to all other outcomes that might be expected if the null hypothesis really was true. Moreover, as probabilities of point events always equal zero (e.g. the probability of obtaining a mean of 13.5646 is infinitely small), probability (the p-value) is calculated as the probability of obtaining data that is at least as extreme as that observed.

As a result of the somewhat obtuse association between hypotheses and the collected data, frequentist statistical outcomes (particularly p-values and confidence intervals) are often misused and misinterpreted. Of particular note;

  • A frequentist p-value is the probability of rejecting the null hypothesis, it is not a measure of the magnitude of an effect or the probability of a hypothesis being true.
  • Given that a statistical null hypothesis can never actually be true (e.g. a population slope is never going to be exactly zero), a p-value is really just a measure of whether the sample size is big enough to detect a non-zero effect.
  • A 95% confidence interval defines the proportion of repeated samples (95/100) of a particular spread of values that are likely to contain the true mean. It is not a range of values for which you are 95% confident the true mean lies between.


By contrast, the Bayesian framework considers the observed data to be fixed and a truth (a real property of a population) whilst the population parameters (and therefore also the hypotheses) are considered to have a distribution of possible values. Consequently, inferential statements can be made directly about the probability of hypotheses and parameters. Furthermore, outcomes depend on the observed data and not other more extreme (unobserved) data.

Bound up in this framework is the manner in which Bayesian philosophy treats knowledge and probability. Rather than being a long-run frequency of repeated outcomes (which never actually occur) as it is in frequentist statistics, probability is considered a degree of belief in an outcome. It naturally acknowledges that belief is an iterative process by which knowledge is continually refined on the basis of new information. In the purest sense, our updated (posterior) belief about a parameter or hypothesis is the weighted product of our prior belief in the parameter or hypothesis and the likelihood of obtaining the data given the parameter or hypothesis, such that our prior beliefs are reevaluated in light of the collected data.

The fundamental distinction between Bayesian and frequentist statistics is the opposing perspectives and interpretations of probability. Whereas, frequestists focus on the probability of the data given the (null) hypothesis ($P(D|H)$), Bayesians focus on the probability of the hypothesis given the data ($P(H|D)$).

Whilst the characteristics of a single sample (in particular, its variability) permit direct numerical insights (likelihood) into the characteristics of the data given a specific hypothesis (and hence frequentist probability), this is not the case for Bayesian hypotheses.

Nevertheless, further manipulations of conditional probability rules reveal a potential pathway to yield insights about a hypothesis given a collected sample (and thus Bayesian probability).

Recall the conditional probability law; $$P(A|B) = \frac{P(AB)}{P(B)} ~ \text{and equivalently, } P(B|A) = \frac{P(BA)}{PAB)}$$ These can be transformed to express in terms of $P(AB)$: $$P(AB) = P(A|B)\times P(B), \text{ and } P(BA) = P(B|A)\times P(A)$$ Since $P(AB)$ is the same as $P(BA)$, \begin{align*} P(A|B)\times P(B) &= P(B|A)\times P(A)\\ P(A|B) &= \frac{P(B|A)\times P(A)}{P(B)} \end{align*}

This probability statement (the general Bayes' rule) relates the conditional probability of outcome $A$ given $B$ to the conditional probability of $B$ given $A$. If we now substitute outcome $A$ for $H$ (hypothesis) and outcome $B$ for $D$ (data), it becomes clearer how Bayes' rule can be used to examine the probability of a parameter or hypothesis given a single sample of data. $$P(H|D) = \frac{P(D|H)\times P(H)}{P(D)}$$ where $P(H|D)$ is the posterior probability of the hypothesis (our beliefs about the hypothesis after inspiration from the newly collected data), $P(D|H)$ is the likelihood of the data, $P(H)$ is the prior probability of the hypothesis (our prior beliefs about the hypotheses before collecting the data) and $P(D)$ is the normalizing constant.

Bayesian statistics offer the following advantages;

  • Interpretive simplicity:
    • Since probability statements pertain to population parameters (or hypotheses), drawn conclusions are directly compatible with research or management questions.
  • Computationally robustness:
    • Design balance (unequal sample sizes) is not relevant
    • Multicollinearity is not relevant
    • There is no need to have expected counts greater than 5 in contingency analyses
  • Inferential flexibility:
    • The stationary joint posterior distribution reflects the relative credibility of all combinations of parameter values and from which we can explore any number and combination of inferences. For example, because this stationary joint posterior distribution never changes no matter what perspective (number and type of questions) it is examined from, we can explore all combinations of pairwise comparisons without requiring adjustments designed to protect against inflating type II errors. We simply derive samples of each new parameter (e.g. difference between two groups) from the existing parameter samples thereby obtaining the posterior distribution of these new parameters.
    • As we have the posterior density distributions for each parameter, we have inherent credible intervals for any parameter. That is we can state the probability that a parameter values lies within a particular range. We can also state the probability that two groups are different.
    • We get the covariances between parameters and therefore we can assess interactions in multiple regression

Despite all the merits of the Bayesian philosophy (and that its roots pre-date the frequentist movement), widespread adoption has been hindered by two substantial factors;

  1. Proponents have argued that the incorporation (indeed necessity) of prior beliefs introduces subjectivity into otherwise objective pursuits. Bayesians however, claim that considering so many other aspects of experimental design (choice of subjects and treatment levels etc) and indeed the research questions asked are discretionary, objectivity is really a fallacy. Rather than being a negative, true Bayesians consider the incorporation of priors to be a strength. Nevertheless, it is possible to define vague or non-informative priors
  2. Expressing probability as a function of the parameters and hypotheses rather than as a function of the data relies on applying an extension of conditional probability (called Bayes' rule) that is only tractable (solved with simple algebra) for the most trivial examples. Hence it is only through advances in modern computing power that Bayesian analyses have become feasible.
    For example, when only a discrete set of hypotheses are possible, then $P(D)$ essentially becomes the sum of all these possible scenarios (each $P(DH_i)$); $$P(D) = \sum{P(D|H)P(H)}$$ $$P(H|D) = \frac{P(D|H)\times P(H)}{\sum{P(D|H_i)P(H)}}$$ However, if there are potentially an infinite number of possible hypotheses (the typical case, at least in theory) this sum is replaced with an integral: $$P(H|D) = \frac{P(D|H)\times P(H)}{\int P(D|H_i)P(H)dH}$$

Whilst it is not always possible to integrate over all the possible scenarios, modern computing now permits more brute force solutions in which very large numbers of samples can be drawn from the parameter space which in turn can be used to recreate the exact posterior probability distribution.

Exponential family of distributions

The exponential distributions are a class of continuous distribution which can be characterized by two parameters. One of these parameters (the location parameter) is a function of the mean and the other (the dispersion parameter) is a function of the variance of the distribution. Note that recent developments have further extended generalized linear models to accommodate other non-exponential residual distributions.

End of instructions