The uniform-weights Bayes factor under weights the researcher states
Source:R/bf_rescaled.R
bf_rescaled.RdA range of shares is not a probability for the counts, so before
coding the researcher states how each theory spreads its weight over
the shares inside its range. bf_uniform_weights() uses the default,
which gives every share in a range the same weight. This function
takes the alternative the paper offers, so that a reader who would
have weighted the shares differently can see what those weights do to
the reported number.
Arguments
- y_W
Non-negative integer. Count (or summed integer weight) of evidence favorable to the working theory.
- y_R
Non-negative integer. Count (or summed integer weight) of evidence favorable to the rival.
- a1, b1
Positive numerics. The Beta pair squeezed onto the working theory's range, the shares above one half. Default
1and1, the uniform weights.- aR, bR
Positive numerics. The Beta pair squeezed onto the rival's range, the shares at or below one half. Default
1and1.
Details
The alternative starts from the Beta\((a, b)\) family, a family of distributions for a quantity between zero and one whose two parameters set where the weight piles up: with \(a > b\) toward one, with \(a < b\) toward zero, and with \(a = b\) symmetrically, concentrated in the middle when \(a = b\) is above one and pushed toward both ends when it is below one. Larger values of both parameters concentrate the weight more tightly.
Each theory's range is half as wide as the interval from zero to one,
so the researcher squeezes a Beta distribution onto it. If \(X\) has
the Beta\((a, b)\) distribution, then \(\theta = 1/2 + X/2\) lies
between one half and one, which is the working theory's range, and
\(\theta = X/2\) lies between zero and one half, which is the
rival's. Squeezing an interval to half its width doubles the density,
so the two weighting functions are
$$2 f_{a,b}(2\theta - 1) \quad\text{and}\quad 2 f_{a,b}(2\theta),$$
with \(f_{a,b}\) the Beta density. Setting both pairs to
\((1, 1)\) gives every share in each range the same weight and
returns bf_uniform_weights() exactly.
What the parameters mean substantively
The two pairs are a claim about what each theory predicts, and they move the reported number, so a researcher has to argue for them. At nine observations supporting the working theory against three supporting the rival, the uniform default gives 20.67. Beta(1/2, 1/2) on both sides piles weight at both ends of each range, including right next to one half where the two theories predict nearly the same share, so the counts separate the theories less and the value falls to 9.53. Beta(10, 10) on both sides concentrates each theory near the middle of its range, so the working theory predicts a share near three quarters and the rival a share near one quarter, the two predict more different things, and the value rises to 252.92.
The pairs need not match. A researcher who expects that under either
theory the search would still turn up evidence for the other puts both
densities next to one half, with a1 = 0.5, b1 = 1 for the working
theory and aR = 1, bR = 0.5 for the rival, and gets 7.35 at nine of
twelve. A researcher who expects each theory to leave almost no
evidence for the other pushes both densities to the far ends,
reversing those two pairs, and gets 28.08. The first value is below
the paper's threshold of 20 and the second is above it.
Weights on shares are not weights on observations
The weights here are weights on \(\theta\), the share of the
evidence supporting the working theory. They are a different object
from a weight on an observation, which acts as replication: a document
the researcher weights at ten enters as ten observations, and the
researcher passes that by summing the weights into y_W and y_R.
See also
bf_uniform_weights() for the uniform default;
bf_worst_case(), whose a1, b1 take the same rescaled pair for the
working theory's shares while the rival is evaluated at its most
favorable share.
Examples
# The uniform default, which is bf_uniform_weights(9, 3).
bf_rescaled(9, 3)
#> [1] 20.67196
# Beta(1/2, 1/2) on both sides: weight next to one half, where the
# two theories claim nearly the same share.
bf_rescaled(9, 3, 0.5, 0.5, 0.5, 0.5)
#> [1] 9.526137
# Beta(10, 10) on both sides: each theory concentrated in the middle
# of its range, so the two claims are far apart.
bf_rescaled(9, 3, 10, 10, 10, 10)
#> [1] 252.9172