Response bias measures the tendency to give one response over another. Unlike d’, the measure of sensitivity, there are multiple measures of response bias. For all response bias statistics in MacMillan and Creelman:
Suppose we have two observers, one with d’ of 1 (blue), the other with d’ of 2 (red). Both have \(\lambda\) = 1.5 (distance from mean of noise distribution). Do they have similar response bias?
Relative to the blue observer, red observer says “yes” more often! So, we would want to have a response bias measure reflect that. Need to take both signal and noise distributions into account. The three measures described below do that it different ways.
Criterion location, \(c\), is defined as the distance from the midpoint between the signal and noise distributions to the criterion.
\[ \begin{align} c = \lambda_{center} &= \lambda - \frac{1}{2}d' \\ &\text{replacing expressions for lambda and d' we can derive an expression in terms of H and FA: }\\ &=-z(FA) - \frac{1}{2}(z(H) - z(FA))\\ & = -z(FA)-\frac{1}{2}z(H) + \frac{1}{2} z(FA) = -\frac{1}{2}z(FA) - \frac{1}{2}z(H)\\ & = \frac{( z(H) + z(FA) )}{-2} \end{align} \]
Here is how c changes when d’ varies:
## Faces Normal.Yes Normal.No Hypno.Yes Hypno.No
## 1 Old 69 31 89 11
## 2 New 31 69 59 41
Calculating d’ and c:
dpNormal = qnorm((69/(69+31))) - qnorm((31/(31+69)))
dpHypno = qnorm((89/(89+11))) - qnorm((59/(59+41)))
cNormal = -0.5*(qnorm((69/(69+31))) + qnorm((31/(31+69))))
cHypno = -0.5*(qnorm((89/(89+11))) + qnorm((59/(59+41))))
(round(cbind(dpNormal, dpHypno),2))
## dpNormal dpHypno
## [1,] 0.99 1
(round(cbind(cNormal, cHypno),2))
## cNormal cHypno
## [1,] 0 -0.73
Hypnosis does not affect sensitivity, but it does lead to a more negative criterion c, i.e., higher false alarms compared to misses, i.e., more tendency to say ‘yes’
Now imagine that we train the observers to recognize faces and this results in an improvement in d’ from 1 to 2 (compare blue line in top and bottom figure) but no change in \(c\) (length of c line is the same).But doesn’t the second case look like the criterion is less biased? In the figure above, the criterion is to the left of the mean of the noise distribution, but it moved to the right of the noise distribution in the bottom figure.
Defined as ratio of c and d’: distance from criterion to the midpoint between the two distributions in units of d’
\[\begin{align} c' &= \frac{c}{d'} = \\ &=-\frac{1}{2} (\frac{z(H)+z(F)}{z(H)-z(F)}) \end{align}\]## [1] "pre-training where c = -0.73 and dprime = 1"
## [1] "cprime: "
## [1] -0.73
## [1] "post-training where c = -0.73 and dprime = 2"
## [1] "cprime: "
## [1] -0.365
Unlike c, c’ reflects a less extreme criterion post-training relative to pre-training.
The likelihood ratio at the position of the decision criterion, \(\lambda\), is another measure of response bias.
Likelihood ratio is the ratio of the probability density functions for signal and noise.
\[\begin{align} LR(x) = \Lambda(X) = \frac{P(X|\text{signal trial})}{P(X|\text{noise trial}}\\ \end{align}\]Note that the likelihood ratio is equal to the slope of the ROC curve. See the illustration below for a geometrical demonstration:
$$ Log-likelihood ratio \(ln(\beta)\) can be estimated using the equation with d’ and c.:
\[ ln(\beta) = d'c = -\frac{1}{2}(z(H) - z(FA)) * (z(H)+z(FA)) = -\frac{1}{2} ( z(H)^2 - z(FA)^2) \] or directly from density and distribution functions:
\[ ln(\beta) = \phi(z(H))/\phi(z(FA)) \]
Example: Hits = 0.915, False alarm: 0.265
H= 0.915
FA = 0.265
# method 1:
-0.5*(qnorm(H)^2-qnorm(FA)^2)
## [1] -0.7442759
# method 2:
log(dnorm(qnorm(H)) / dnorm(qnorm(FA)))
## [1] -0.7442759
Isobias curves trace curves connecting H and FA rates of constant response bias for each of the different response bias measures.
\[c = \lambda_{centre} = \lambda-d'/2 = \text{distance from midpoint between distribution means}\] \[ ln(\beta) = cd' \]
\[ c' = \frac{c}{d'}\]
Which ones are found empirically?
It depends on what the goal is.
(0.1 - -1)/(0.1 - (-0.01))
## [1] 10
A one-interval design comprising:
##
##
## After 10 min delay
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## Old 112 112 72 53 22 4 375
## New 7 38 50 117 101 62 375
##
##
## After 7 days delay
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## Old 49 94 75 60 75 22 375
## New 8 37 45 60 113 112 375
When given 6 response choices, the observer places 5 decision criteria separating each response choice.
##
##
## Step 1: 10 min delay proportions
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## Old 0.30 0.3 0.19 0.14 0.06 0.01
## New 0.02 0.1 0.13 0.31 0.27 0.17
##
##
## Step 2: 10 min delay cumulative sum of probabilities
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## Old 0.30 0.60 0.79 0.93 0.99 1
## New 0.02 0.12 0.25 0.56 0.83 1
##
##
## Step 3: 10 min delay z-transform
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## Old -0.524 0.253 0.806 1.476 2.326 Inf
## New -2.054 -1.175 -0.674 0.151 0.954 Inf
##
##
## Step 4: compute d-prime: z(Old) - z(New).. i.e., z(H) - z(FA)
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## 1.530 1.428 1.480 1.325 1.372 NaN
##
##
## Step 4: similarly can compute bias, c:-0.5*(z(Old) + z(New))
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## 1.2890 0.4610 -0.0660 -0.8135 -1.6400 -Inf
##
##
##
## 10 min delay results
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## Old -0.524 0.253 0.806 1.4760 2.326 Inf
## New -2.054 -1.175 -0.674 0.1510 0.954 Inf
## tenMinDelay_dPrime 1.530 1.428 1.480 1.3250 1.372 NaN
## tenMinDelay_c 1.289 0.461 -0.066 -0.8135 -1.640 -Inf
## [1] ""
##
##
## Do the same for 7 days
##
##
## 7 day results
## Old.3 Old.2 Old.1 New.1 New.2 New.3
## Old -1.126 -0.305 0.202 0.643 1.5550 Inf
## New -2.054 -1.175 -0.706 -0.253 0.5240 Inf
## sevenDayDelay_dPrime 0.928 0.870 0.908 0.896 1.0310 NaN
## sevenDayDelay_c 1.590 0.740 0.252 -0.195 -1.0395 -Inf
These data can be plotted in an ROC curve and a best-fit line can be found. Here, the slope is very close to 1, so we can fix the slope parameter to 1, fit the curve and extract d’ from the x or the y intercept.
## 10 min delay:
## (Intercept)
## 1.427
## 7 days delay:
## (Intercept)
## 0.9266
What if slopes are clearly not = 1??
Data from a memory experiment with low and high frequency words. The judgement is about whether the words are Old or New, and how certain the observer is about their answer.
## Low frequency
## Old.Sure Old.Maybe Uncertain.3 New.Maybe NewSure
## Old 61 15 15 5 4
## New 2 8 37 23 30
## High frequency
## Old.Sure Old.Maybe Uncertain.3 New.Maybe NewSure
## Old 37 25 18 11 9
## New 4 18 28 21 29
Repeat the same steps to get estimates of H and FA.
##
## Step 1: proportions
## Old.Sure Old.Maybe Uncertain.3 New.Maybe NewSure
## Old 0.61 0.15 0.15 0.05 0.04
## New 0.02 0.08 0.37 0.23 0.30
##
## Step 2: cumulative sum of probabilities
## Old.Sure Old.Maybe Uncertain.3 New.Maybe NewSure
## Old 0.61 0.76 0.91 0.96 1
## New 0.02 0.10 0.47 0.70 1
##
## Step 3: z-transform,d` and c calculation
## Old.Sure Old.Maybe Uncertain.3 New.Maybe NewSure
## Old 0.279 0.706 1.341 1.751 Inf
## New -2.054 -1.282 -0.075 0.524 Inf
## Old.Sure Old.Maybe Uncertain.3 New.Maybe NewSure
## Old 0.2790 0.706 1.341 1.7510 Inf
## New -2.0540 -1.282 -0.075 0.5240 Inf
## lowfreq.dPrime 2.3330 1.988 1.416 1.2270 NaN
## lowfreq.c 0.8875 0.288 -0.633 -1.1375 -Inf
Do the same for high frequency:
## Old.Sure Old.Maybe Uncertain.3 New.Maybe NewSure
## Old -0.3320 0.3050 0.842 1.341 Inf
## New -1.7510 -0.7720 0.000 0.553 Inf
## highfreq.dPrime 1.4190 1.0770 0.842 0.788 NaN
## highfreq.c 1.0415 0.2335 -0.421 -0.947 -Inf
Plotting the data reveals that the slopes are not equal, so the x- and y- intercepts are different. d’ is no longer a good measure.
Slope of ROC, \(s = d'_2/d'_1\)
\(\text{Equation for the unequal variance ROC curve: } z(H) = d'_2+\frac{d'_1}{d'_2}z(F)\)
\(A_z = \Phi(d_{YN}) = \Phi(d_a/\sqrt{2})\), Area under the ROC curve in proportion coordinates. Non-parametric measure of sensitivity.
Similar adjustments have to be made to replace c with an adjusted measure using the slope of the ROC (s)
\(c_1 = \frac{-1}{(1+s)} ( z(H) + z(F)\)
\(c_2 = \frac{-s}{(1+s)} ( z(H) + z(F)\)
\(c_a = \frac{-\sqrt{2}s}{\sqrt(1+s^2)(1+s)} ( z(H) + z(F)\)
\(c_e = \frac{-2s}{(1+s)^2} ( z(H) + z(F)\)
All measures are the same up to a multiplicative constant; when s = 1, they all = c
Likelihood ratio obtained from heights of density functions for H and FA. It is no longer equal to d’c. When variances are unequal, the two curves may intersect at 2 points. Using one single decision criterion line can lead to less than chance performance. The optimal decision involves two decision lines (bottom panel, below).
‘Yes’ responses to noise stimuli are random guesses, not from ‘actually thinking one has seen the stimulus’.
\[ \begin{align} H &= P(\text{"yes"}|S_2 \text{ or signal}) = q + u(1-q) \\ FA &= P(\text{"yes"}|S_1 \text{ or noise}) = u \\ &\text{replacing u with FA in the first line, we can get an expression for sensitivity, q}\\ q &= H - FA+FAq\\ q-FAq &= H-FA\\ q &=(H-FA)/(1-FA) \end{align} \]
Equation for the ROC curve: straight line, with (1-q) slope and q intercept. These are ‘non-regular’, because allow H to be >0 when FA = 0.
\[ \begin{align} H &=(1-q)FA + q \\ \end{align} \]
The figure below shows the decision space assumed by High threshold theory.
The main difference with the high threshold theory described above is that now, the threshold is “low” because it can be crossed by noise alone, i.e., the noise distribution extends beyond the threshold.
Observers can use one of two different strategies:
Here is the decision space for each strategy
Here is the predicted ROC curve:
It can be very difficult to obtain enough data to distinguish this prediction from that of signal detection theory. We will discuss this theory more in Kellen et al., 2016.
This theory underlies the use of proportion correct in yes/no tasks.
\[ p_{correct} = p_{\text{signal or }S_2}H + p_{\text{noise or }S_1}(1-F) \\ = p_{S_1}+ p_{S_2}H - p_{S_1}F \\ = \text{constant + weighted difference of H and F} \]
Sensitivity \(q = 2p_{correct} - 1\)
The ROC curve under this theory is as follows:
Yonelinas (1997) found ROC curves consistent with double threshold theory in an associative recognition experiment:
Bias measures for double-high-threshold theory:
Yes rate:
Error ratio:
The predicted shapes of iso-bias curves for the yes rate and error ratio look similar to those of c and c’..
Stems from the following ‘choice axiom’: > Odds of choosing one stimulus over another are not affected by other possible stimuli. Luce (1959)
This theory proposes another sensitivity measure, which is a nonlinear combination of H and FA:
\[ \alpha = \sqrt{\frac{H(1-F)}{(1-H)F}} \]
To express is as a difference of transformations of H and FA, we take logs on both sides, and we get:
\[ ln(\alpha) = \frac{1}{2} ln(\frac{H}{1-H})- \frac{1}{2}ln(\frac{F}{1-F}) \] note: \(ln(\alpha)\) is also called \(d_L\) in memory research.
Instead of the z-transformation, Choice Theory applies the log-odds transformation to H and F.
Calculating \[lg(\alpha)\] and d’ for the same pair of H and FA, we obtain:
F = 0.4
H = 0.8
(logAlpha = round(0.5*log(H/(1-H)) - 0.5*log(F/(1-F)),3))
## [1] 0.896
(dprime = qnorm(H)-qnorm(F))
## [1] 1.094968
When H vs FA are plotted in log-odds coordinates,the resulting ROC curve is a straight line.
In Linear coordinates, SDT and Choice Theory have very similar ROCs, except that Choice Theory cannot have asymmetric curves (i.e., only similar to the equal variance assumption of SDT ).
Criterion location: \(ln(b) = -\frac{1}{2} ln(\frac{H}{1-H}) + \frac{1}{2}ln(\frac{F}{1-F})\)
Relative criterion location: \(b' = ln(b)/2*ln(\alpha)\)
Likelihood ratio: \(\beta_L = \frac{H(1-H)}{F(1-F)}\)
To find a proportion given a log-odds value (logit) x, rearrange the equation for log-odds. \[ x = ln(p/(1-p))\\ e^x = p/(1-p)\\ e^x-pe^x = p\\ p = 1/(1-e^x) \] Using this knowledge, we can derive equations for H, F, and proportion correct (assuming unbiased observer and equal proportions of signal and noise) \[ H = \alpha/(\alpha + \beta)\\ F = 1/(1+\alpha\beta)\\ p_{correct} = \alpha/(\alpha+1)\\ \]
If only have 1 point on the ROC, various ‘nonparametric’ measures such as the area under the ROC curve (drawn with straight lines), is a combination of the above mentioned theories and is thus not better.
If obtain multiple points on the ROC, area under the ROC is a non-parametric measure of sensitivity.