Perceptual studies suggest that processing facial identity emphasizes upper-face information, whereas processing expressions of anger or happiness emphasizes the lower-face. The two goals of the present study were to determine (a) if the distributions of eye fixations reflect these upper/lower-face biases, and (b) whether this bias is task- or stimulus-driven. We presented a target face followed by a probe pair of morphed faces, neither of which was identical to the target. Subjects judged which of the pair was more similar to the target face while eye movements were recorded. In Experiment 1 the probe pair always differed from each other in both identity and expression on each trial. In one block subjects judged which probe face was more similar to the target face in identity, and in a second block subjects judged which probe face was more similar to the target face in expression. In Experiment 2 the two probe faces differed in either expression or identity, but not both. Subjects were not informed which dimension differed, but simply asked to judge which probe face was more similar to the target face. We found that subjects scanned the upper-face more than the lower-face during the identity task but the lower-face more than the upper-face during the expression task in Experiment 1 (task-driven effects), with significantly less variation in bias in Experiment 2 (stimulus-driven effects). We conclude that fixations correlate with regional variations of diagnostic information in different processing tasks, but that these reflect top-down task-driven guidance of information acquisition more than stimulus-driven effects.

^{1}

^{2}Analysis of these pilot data showed no significant difference in accuracy between judgments of identity and those of expression in the probes used, in either Experiment 1 or Experiment 2.

*y*

_{simfe}= 1) is read as “the probability that subject

*s,*for fixation

*i,*when given matrix

*m,*face

*f,*and expression

*e,*looks at the upper part of the face.” The variables

*b*

_{0},

*b*

_{1},

*b*

_{2}, and

*b*

_{3}are the fixed effects, i.e. the group average effects, induced by the experimental manipulations (

*dimension*

_{simfe}and

*experiment*

_{simfe}, which were given a treatment coding). The random effect term

*subject*

_{s}represents each subject's deviation from the group average preference for upper versus lower face fixation. The other random effects terms (

*matrix*

_{m},

*face*

_{f}, and

*expression*

_{e}) represent subject-invariant, between-item variation in the proportion of upper versus lower face fixations. Finally,

*ɛ*

_{simfe}is the residual term, representing remaining unexplained variance.

*l*

_{1}−

*l*

_{0}), where

*l*

_{0}and

*l*

_{1}denote the maximized log-likelihood of two models to be compared, one of which (

*l*

_{1}) comes from the model which has the predictors of interest; the other (

*l*

_{0}) comes from a model which differs only by not having these variables. This statistic has a null distribution approximating that of

*χ*

^{2}, with degrees of freedom obtained from the difference in the number of parameters, so a

*χ*

^{2}-test is used to assess whether a predictor contributes significantly to the model's fit. We also use the Wald statistic, calculated from the slope estimate and its standard error, to test whether the slope differs significantly from zero (see Agresti, 2002).

*SD*= 1.60). From the items, the face provided most subject-invariant variance (

*SD*= 0.11), followed by matrix and expression (both

*SD*= 0.05).

*χ*

^{2}(1) = 21.8,

*p*< 0.001] and dimension [

*χ*

^{2}(1) = 23.2,

*p*< 0.001]. We tested the effect of the originally hypothesized interaction between experiment and dimension, namely that there would be more fixations in the upper-face for the identity condition than for the expression condition for the task-driven experiment, and that the difference in upper-versus-lower fixations between dimensions would be less for the stimulus-driven experiment. The interaction was significant [

*χ*

^{2}(1) = 42.5,

*p*< 0.001]. The slope for the interaction term was significantly different to zero (slope = 0.56, 95% CI = 0.39–0.73, Wald

*z*= 6.5,

*p*< 0.001). Figures 5a and 5b present the mean fixation probabilities computed from the model.

*χ*

^{2}(1) = 2.91,

*p*= 0.09]. The effect was much larger for the task-driven experiment [

*χ*

^{2}(1) = 62.8,

*p*< 0.001]. For the task-driven experiment, the probability of a fixation on the upper-half of the face was significantly higher for identity processing than for expression processing (slope = 0.48, 95% CI = 0.36–0.59, Wald

*z*= 7.9,

*p*< 0.001), whereas there was a slight opposite trend for the stimulus-driven experiment (slope = −0.11, 95% CI = −0.23 to 0.02, Wald

*z*= −1.7,

*p*= 0.09).