Introduction to Experiment 2

Another potential contributor to the numerically small question effects of Experiment 1 was the reasonably high levels of memory strength available at test, as predicated on the “deep” levels-of-processing (LOP; Craik & Lockhart, 1972) achieved by the syllable counting task. Sufficiently high memory strength reduces the need to recruit higher-order strategies to accurately diagnose memory decisions (Jacoby et al., 1999; Koriat & Goldsmith, 1996). To investigate this mediating influence, memory strength was systematically manipulated in Experiment 2 by presenting different tasks at study – a pleasantness judgement task leading to “deep” LOP and a case judgement task leading to “shallow” LOP. This study manipulation was crossed with analogous manipulation of question emphasis as in the single dimension question condition in Experiment 1. To clarify, the case judgement task was expected to instil a lower level of overall memory strength than the syllable-counting task in

Experiment 1 (i.e. expected LOP: pleasantness > syllable counting > case

judgement), and therefore enhance the emphasis-driven shifts in criterion bias and decision accuracy. Such enhancement would further indicate that the question effect is rooted in higher-order strategies that are bear a greater influence on performance in situations of low overall memory strength and associated uncertainty (Benjamin, 2007; Kahneman, Slovic & Tversky, 1982).

2.3.1. Method

2.3.1.1. Participants. The sample comprised 29 native English speakers who reached the minimum performance threshold (20 female; mean age = 22.1; age range = 19 to 28) from a total of 31 recruited. Two participants were excluded for

86 poor performance based on the same threshold as in Experiment 1 (i.e. d′ < 0.1). Informed consent and participant compensation were identical to Experiment 1.

2.3.1.2. Stimuli. Word stimuli were randomly drawn from the same pool used in Experiment 1. Each of the four study-test blocks constituted 120 words (60 old and

60 new at each test phase), which were presented on PCs running MATLAB and Psychophysics Toolbox as before.

2.3.1.3. Procedure. After the presentation of on-screen instructions and a practice phase, participants completed four self-paced study-test blocks comprising blocked combinations of study LOP (shallow, deep) and test question emphasis (“old?”, “new?”; see Figure 2.2a). The two shallow LOP study phases involved a case

judgement task for 60 words presented serially in either uppercase or lowercase (30 of each case type) below the prompt “uppercase?” and with “yes” and “no” response options assigned to the keyboard. The two deep LOP study phases involved a pleasantness judgement task, wherein lowercase words appeared beneath a

“pleasant?” question with “yes” and “no” response options. A test phase immediately followed each study phase and comprised 60 old and 60 new words. Participants made old/new recognition decisions with test question emphasis varied analogously to the single dimension question condition in Experiment 1 i.e. with “old?” and “new?” questions presented in separate test blocks (with “yes” and “no” response options available throughout). As before, three-point confidence ratings were solicited 0.25s after each old/new decision. A 0.5 s fixation cross preceded each trial across all study and test blocks. Across participants, the study-test block order was pseudo- randomised such that only one level of a factor would change in any block transition. Hence, while it was possible for a participant to experience a shallow-“old?” to

87 shallow-“old” to deep-“new” block transition. This led to four orders in total to which participants were randomly assigned.

2.3.2. Results and Discussion

2.3.2.1. Equal variance sensitivity and bias. To confirm that the LOP manipulation affected memory strength as anticipated, a 2(LOP: shallow or deep) x 2(question emphasis: “old?” or “new?”) repeated measures ANOVA was conducted on EV estimates of d′. The ANOVA revealed an anticipated main effect of LOP, such that sensitivity was lower in the shallow LOP (M = 1.33, SD = 0.73) than the deep LOP condition (M= 2.70, SD = 0.52), F(1,28) = 109.14, p = .001, ηp2 = .80. There was

also an unexpected main effect of question emphasis such that d' was higher in the old (M= 2.10, SD = 0.55) than the newemphasis condition (M = 1.92, SD = 0.57),

F(1,28) = 5.84, p = .022, ηp2 = .17. There was no significant LOP x emphasis

interaction, F(1,28) = 1.06, p = .311, ηp2 = .04. The unexpected main effect of

emphasis on d' suggests that emphasis impacted on memory strength processes and this finding is elaborated on in ensuing sections.

The primary analysis on EV c was carried out using a 2(LOP) x 2(question emphasis) repeated measures ANOVA. A main effect of question emphasis was observed, with significantly higher c estimates in the old (M = 0.09, SD = 0.24) than the new emphasis condition (M = -0.03, SD = 0.32), F(1,28) = 6.65, p = .015, ηp2 =

.19 (see Figure 2.2b). Although there was no significant interaction effect, F(1,28) = 3.13, p = .088, ηp2 = .101, a numerical trend was observed, and planned pairwise

comparisons revealed that the disconfirmatory question bias achieved statistical significance in the shallow LOP condition but not in the deep LOP condition, t(28) =

88 3.05, p = .005, d = 0.62 and t(28) = .70, p = .487, d = 0.13 respectively. These

findings replicate the disconfirmatory emphasis bias observed in Experiment 1, and support our prediction of its enhancement under conditions of high overall memory strength.

From the same ANOVA on c, we found a main effect of LOP on criterion placement, such that c estimates were significantly lower in the deep (M = -0.11, SD = 0.29) than the shallow LOP conditions (M = 0.17, SD = 0.30), F(1,28) = 23.66, p = .001,

ηp2 = .46. Participants hence adopted a liberal “old” response strategy in the

condition of higher memory strength, and this finding is reminiscent of prior reports of criterion sensitivity to metacognitive inferences of overall memory strength (as

alluded to in the previous chapter, see Section 1.4.; Glanzer & Adams, 1985; Hirshman, 1995).

2.3.2.2. Unequal variance sensitivity and bias. As in Experiment 1, we probed the correspondence between the above equal variance SDT effects in equivalent

unequal variance analyses. A 2(LOP) x 2(question emphasis) repeated measures ANOVA on UEV d′ recovered a main effect of LOP, such that d' was significantly higher in the deep (M = 3.40, SD = 0.89) than the shallow LOP condition (M = 1.50,

SD = 0.85), F(1,28) = 88.89, p = 0.001, ηp2 = 0.76. A trend main effect of emphasis

was also observed, with sensitivity once again higher under old emphasis (M = 2.59,

SD = 0.67) compared to new emphasis (M = 2.32, SD = 0.89), F(1,28) = 3.16, p = 0.086, ηp2 = 0.10. The LOP by emphasis interaction was non-significant, F(1,28) =

1.62, p = 0.214, ηp2 = 0.06.

A 2(LOP) x 2(question emphasis) ANOVA on crel revealed a main effect of LOP,

89 conditions (M = -0.01, SD = 0.36), F(1,28) = 12.23, p = 0.002, ηp2 = 0.304. There

was neither a main effect of emphasis nor a significant interaction, F(1,28) < 1 and

F(1,28) = 2.73, p = 0.110, ηp2 = 0.089 respectively. However, planned pairwise

comparisons revealed a trend towards a significant disconfirmatory bias in the shallow LOP condition, t(28) = 1.87, p = .072, d = 0.36, with a higher crel under old

compared to new emphasis, but a clearly non-significant difference in crel in the deep

LOP condition, t(28) = 0.60, p = .556, d = 0.11 (see Table 2.1. for means). Overall, the effects of question emphasis on unequal variance estimates of d' and crel are

broadly consistent with those observed in the equal variance estimate analyses. The weakened effects of question emphasis on crel likely reflect the conservative nature

of the UEV model, in that it aims to account for added response variability associated with assessments of confidence.

2.3.2.3. Decision accuracy. As before, the accuracy of “old” and “new” decisions

under varying question emphases was analysed, separately for shallow and deep LOP conditions. A 2(decision type) x 2(question emphasis) repeated measures ANOVA for the shallow LOP condition revealed a significant main effect of decision type, F(1,28) = 9.77, p = .004, ηp2 = .26 (see Figure 2.2c). Participants’ “old”

decisions were characterised by greater overall accuracy (M = .76, SD = .10) than their “new” decisions (M = .71, SD = .09). Similar disparities in the response profiles of “old” and “new” decisions have been reported previously (e.g. Jaeger, Cox & Dobbins, 2012). No main effect of question emphasis was observed, F(1,28) = 1.46,

p = .237, ηp2 = .050. Crucially, the decision type x question emphasis interaction was

significant, F(1,28) = 9.04, p = .006, ηp2 = .244. This suggests that question

emphasis improved the accuracy of endorsing the emphasised decision i.e. “old?” improved the accuracy of “old” decisions and “new?” improved the accuracy of “new”

90 decisions (see Table 2.1. for means). Post-hoc comparisons revealed a significant difference in oldcorr across emphasis conditions and a nonsignificant difference for newcorr, t(28) = 2.23, p = .034, d = 0.41 and t(28) = 0.42, p = .679, d = 0.07

respectively. The greater influence of emphasis on oldcorr compared to newcorr parallels the decision accuracy analyses of Experiment 1 in suggesting that the strategies instilled by the question format manipulations might selectively improve the accuracy of “old” decision evaluations. This possibility is discussed in greater detail in the General Discussion in Chapter 7 (Section 7.3.).

A 2 x 2 ANOVA was also conducted for decision accuracy measures in the deep LOP condition, yielding a nonsignificant main effect of decision type and a

nonsignificant interaction, F(1,28) = 3.80, p = .061, ηp2 = .12 and F(1,28) = 1.00, p =

.326, ηp2 = .03 respectively. A significant main effect of question emphasis was

observed, with the question “old?” (M = .91, SD = .05) leading to greater overall decision accuracy than the question “new?” (M = .89, SD = .05), F(1,28) = 5.65, p = .025, ηp2 = .17 (see Figure 2.2d). Considered with the observed improvement in d′ for

“old?” questions in the same deep LOP condition, these findings raise the possibility that the question “old?” instilled a more rigorous source monitoring strategy that prioritised the recovery of recollected content (Johnson, Hashtroudi & Lindsay, 1993). In conditions of high overall memory strength (as in the deep LOP condition), this directly improved performance as a higher proportion of recollected old items were likely encountered at test. This contrasts the lack of a main effect of emphasis on decision accuracy or d′ in the shallow LOP condition, where old items were less likely to be associated with recollected content and hence the adoption of a

monitoring strategy that exclusively prioritised recollection exerted a less beneficial influence on performance. As such, these question format effects putatively impact

91 more directly on memory strength and related retrieval control processes, and hence reflect different strategic influences to the decision control processes of primary interest (see Chapter 7, Section 7.2. for further discussion).

To summarise, the results of Experiment 2 provided further evidence for question format serving as an implicit source of goal emphasis in the test environment. The trend effects observed in Experiment 1, namely the criterion bias reflecting a reduced likelihood of endorsing emphasised decisions coupled with improved accuracy of those decisions when made, were found to be enhanced in the shallow LOP condition as predicted. This is consistent with the heightened engagement of goal- driven strategies to counteract weakly diagnostic memory evidence and ensure accurate decision outcomes.

92

Figure 2.2. Design and results of Experiment 2. a) Design schematic showing study and test phase manipulations. b) Question emphasis effects on equal variance criterion (EV c), with separate lines denoting levels-of-processing (LOP) achieved at study. The remaining graphs show effects of question emphasis on decision accuracy measures in c) the shallow LOP condition and d) the deep LOP condition, with separate lines representing decision types. Error bars represent standard error of the mean.

93