Inhibition in the Processing of Garden-Path Sentences

The Hartman and Hasher (1991) garden-path sentence completion task has been used in several studies to assess the efficiency of the deletion function of inhibition (e.g., L. Hasher, R. Zacks, & C. P. May, 1999 ), with results suggesting that younger adults are efficient at suppressing once relevant but no longer appropriate information, whereas older adults generally are not (e.g., M. Hartman & L. Hasher, 1991 ; L. Hasher, M. B. Quig, & C. P. May, 1997 ; C. P. May & L. Hasher, 1998 ). An alternative interpretation of patterns of access to relevant and no-longer-relevant sentence endings focuses on the difficulty of selecting final words for sentence frames and on integration effects in implicit memory ( M. Hartman, 1995 ). This alternative is considered and found wanting on the basis of both new and old empirical data. On the basis of present data and related findings, it is concluded that the task does measure inhibitory efficiency.

Inhibition is a critical construct in many current theories of cognition, including integrative models of attention, memory, and language processing (e.g., Dagenbach & Carr, 1994 ; Gernsbacher & Faust, 1991 ; Hasher & Zacks, 1988 ) and models of life span development (e.g., Dempster, 1992 ; Harnishfeger, 1995 ; Hasher & Zacks, 1988 ; McDowd, Oseas-Kreger, & Filion, 1995 ) and abnormal functioning (e.g., Beech & Claridge, 1987 ; Enright & Beech, 1993 ; Shimamura, 1995 ). One perspective proposes that inhibition is responsible for suppressing irrelevant or marginally relevant information and for rejecting information that was once relevant but is no longer appropriate (e.g., Hasher, Zacks, & May, 1999 ). A number of tasks have been used in a converging operations approach to assess the efficiency with which no-longer-relevant information is deleted from consideration. These tasks include directed forgetting (e.g., Zacks, Radvansky, & Hasher, 1996 ), garden-path paragraph comprehension ( Hamm & Hasher, 1992 ), and memory for alternative endings of garden-path sentences (e.g., & Hasher, 1991 ; Hasher et al., 1997 ; May & Hasher, 1998 ).

The garden-path sentence task developed by Hartman and Hasher (1991) misleads people to first think of a highly probable word before providing them with a less probable word that completes the sentence and that they are to remember for a later memory test. An implicit memory test is then used to measure access to the two alternative completions of sentences ( Hartman & Hasher, 1991 ). For example, in the first phase of this task, participants generate endings for high-cloze sentence frames (approximate cloze values = .85; e.g., participants produce "lights" for the sentence frame, "Before you go to bed, turn off the __"). A target ending then appears and, for critical sentences, the target ending disconfirms the participant-generated ending (e.g., stove is the target ending for this example sentence frame). For filler sentences, the high-cloze ending that most participants generate is affirmed as the target ending. Participants are instructed to remember only the target ending for every sentence for a later memory test.

After a brief delay, accessibility of both the target endings (e.g., stove ) and the disconfirmed endings (e.g., lights ) from the critical sentences is tested through an indirect memory test. For this test, participants complete sentence frames that are missing their final word, but these sentence frames are only moderately predictive of their final words (approximate cloze value = .50). Three types of sentence frames are used to assess accessibility of the target and disconfirmed words from the first phase: (a) sentences that are moderately predictive of the target endings from Phase 1 (e.g., "She remodeled her kitchen and replaced the old __"; for "stove"); (b) sentences that are moderately predictive of the disconfirmed endings from Phase 1 (e.g., "The baby was fascinated by the bright __"; for "lights"); and (c) sentences that are moderately predictive of new, never-seen control endings (e.g., "The kitten sat peacefully on her owner's __"; for "lap"). ¹ Accessibility of the target and disconfirmed endings is assessed by comparing completion rates for those critical items with the completion rate for control items. Higher priming scores (e.g., target completions minus control completions) are interpreted as greater accessibility.

Previous work with this task indicates that patterns of priming differ across younger and older adults. For example, younger adults show reliable priming for the target items, but they fail to show above-baseline priming for the disconfirmed endings ( Hartman & Dusek, 1994 ; Hartman & Hasher, 1991 ; Hasher, Quig, & May, 1997 ). ² These data suggest that when participants are instructed to remember only the target items, younger adults successfully suppress the disconfirmed items and hence show priming only for targets. By contrast, across a number of studies, older adults show priming for disconfirmed endings as well as target endings ( Hartman & Dusek, 1994 ; Hartman & Hasher, 1991 ; Hasher et al., 1999 ; May & Hasher, 1998 ). ³ These data suggest that older adults, who have been shown elsewhere to suffer age-related inhibitory deficits (e.g., McDowd et al., 1995 ; Zacks & Hasher, 1994 ), are unable to suppress the disconfirmed items and thus show priming for these items. By an inhibitory view, then, the pattern of priming observed in the garden-path sentence completion task is a function of item relevance and participant age: Young adults suppress disconfirmed items and show priming for targets only; older adults are inefficient at deleting the disconfirmed items and thus show priming for both targets and disconfirmed items.

Recently, Hartman (1995) offered an alternative, noninhibitory interpretation for the sentence completion data. This integrative selection argument proposes that the pattern of priming found for younger and older adults in the implicit sentence completion task is determined by complex interactions among at least three factors: (a) the integration of the target item with its context, (b) the difficulty of selecting the target item, and (c) the age of the individual.

With respect to stimulus integration, the suggestion is that the more closely integrated (in terms of meaning) an item is with its context, the less priming will be seen for that item in the test phase. By this view, priming should not occur for expected items (i.e., the words generated by participants) because they are, by definition, easily integrated within the sentence context. Thus, inhibition is not the reason why disconfirmed items show no priming by younger adults; rather, this lack of priming is a result of the integration of those items with the context. By this view, priming should only occur for items that are difficult to integrate within the sentence context (in our study, for the unexpected, experimenter-provided items).

This argument is complicated, however, through the further claim that the degree of integration of an item with its context (and thus the relative lack of priming) depends critically on the difficulty of target selection. When target selection is easy, expected items will be integrated but unexpected items will not ( Hartman, 1995 ); thus, only unexpected items show priming. When target selection is difficult, expected items are not well integrated and the encoding of unexpected items is disrupted ⁴ ; thus, priming for expected items is robust and that for unexpected items is diminished. Furthermore, selection difficulty is determined not only by the specific task but also by the individual's age. Older adults are presumed to generally find selection of the target more difficult than younger adults, particularly in situations in which selection is based on a semantic criterion.

According to the integration—selection explanation, then, the pattern of priming typically observed for younger and older adults in the sentence completion task reflects the combined influence of integration, selection, and age factors. For younger adults, selection of the target is relatively easy, so the expected (disconfirmed) items are integrated and show no priming, and the unexpected (target) items are not integrated and show reliable priming. For older adults, selection of the target is more difficult than it is for younger adults. Consequently, older adults fail to integrate either the unexpected (target) items or the expected (disconfirmed) items and so show priming for both.

Hartman (1995) reported findings that are generally consistent with this integration—selection view. In particular, under presumed easy (i.e., perceptual) selection conditions, both younger and older adults showed priming for unexpected items only, regardless of the status of those items as targets or distractors (Experiments 1 and 2). ⁵ Under presumed difficult (i.e., semantic) selection conditions, younger adults showed reduced but significant priming for unexpected items and significant priming for expected items (again, regardless of item relevance; Experiment 3). Older adults, on the other hand, showed significant priming for expected items only (Experiment 4).

However, despite the general fit between Hartman's (1995) analysis of the Hartman and Hasher (1991) task and her findings, there are several reasons to suggest that the original (inhibitory) account of priming effects in this task may still be viable: (a) Hartman's account is complex and, in our view, somewhat ad hoc. (b) Hartman's experiments did not include what we consider to be a critical aspect of the original procedure; specifically, they did not require participants to generate and then reject the expected endings. (c) This is particularly important because the patterns of priming differ for read words versus generated words that are well integrated with their context at encoding. Specifically, integrated items that are read in a context (e.g., "To see your reflection, just look in a mirror ") do not show priming, but those that are generated (e.g., "To see your reflection just look in a m__") do (cf. Masson & MacLeod, 1992 , Experiments 8A and 7, respectively). (d) Hartman did not include a measure of selection difficulty. (e) Finally, it is difficult to reconcile Hartman's view with our results showing that priming patterns differ depending on whether testing occurs at an optimal or a nonoptimal time of day ( May & Hasher, 1998 ). Perhaps most striking is the finding that when tested at their optimal time of day, but not at other times, young adults show a below-baseline probability of generating the disconfirmed endings on the priming test. For all of these reasons, we think that it is premature to reject the original account of the Hartman and Hasher (1991) indirect memory test and that further study is justified.

Given that Hartman's (1995) new interpretation of the garden-path sentence completion task hinges on claims about variation in selection difficulty across different target selection tasks, we carried out two experiments in which we directly compared different selection tasks. We also recorded response times and error rates on the target selection tasks in an attempt to provide confirmation of presumed selection difficulty differences across tasks. Because it is the original Hartman and Hasher (1991) procedure that is at issue here, we duplicated that procedure in all respects except for the selection task. In particular, as was true in the Hartman and Hasher procedure, participants first generated an ending for each sentence frame; then they were shown both the expected and unexpected endings from which they selected the target according to the selection criterion designated by the experimental instructions. As was also true in the original procedure, the target, to-be-remembered word for critical items was the unexpected ending, and for filler items it was the expected ending.

Experiment 1 compared younger adults' performance on two perceptual selection tasks that were expected to differ in difficulty. In this study, half of the participants selected targets on the basis of letter case (upper- versus lowercase; easy selection condition), and half selected targets on the basis of font (upright versus italics; difficult selection condition). Experiment 2 compared a perceptual and a semantic selection task with younger and older adults. Participants in the perceptual task were cued on each trial to select either the word in italics or the word in upright font; those in the semantic task were cued on each trial to select either the expected item or the unexpected item. Note that our semantic task is directly derived from the procedure that Hartman (1995) used in Experiments 3 and 4, with the added component of a selection cue that instructed participants on a trial-by-trial basis whether to select the expected or the unexpected ending. By contrast, each participant in Hartman's study had a fixed selection criterion for all trials. Because of this difference, it seems likely that our semantic task was more difficult than hers. Similar to Hartman's claims about the relative difficulty of semantic versus perceptual selection tasks, we also expected that the semantic task in Experiment 2 would prove more difficult than the perceptual selection task. As a check on these intuitions about selection difficulty differences, we assessed selection time and error rates as measures of selection difficulty in both experiments.

In sum, the two experiments reported in this article examine the impact on priming effects of target selection tasks that were chosen (based on Hartman's [1995] arguments and our own intuitions) so as to include a range of selection difficulties. Consequently, Hartman's view would predict a variation in priming effects across selection tasks, especially in Experiment 2, which included older adults. To anticipate, the priming findings do not vary with selection task. They vary with age only. In the General Discussion section, additional arguments are detailed to support the inhibition account of priming effects in the Hartman and Hasher (1991) task.

In our first study, we tested younger adults in a modified version of the original Hartman and Hasher (1991) sentence completion paradigm, in which participants read high-cloze sentence frames and generated the expected endings for those sentences. After the expected ending was generated, two words appeared: a target and a distractor. Participants selected the to-be-remembered target item either on the basis of letter case (used as an easy selection condition) or font (used as a difficult selection condition). Selection difficulty was measured by the speed and accuracy of selection decisions. For critical sentences, the generated (expected) item was disconfirmed, and participants were to remember the new (unexpected) target item. For filler sentences, the generated, expected item was confirmed and the new, unexpected item served as the distractor.

Participants . Eighty-eight college students (age range = 18—22 years) from the University of Arizona participated in this study and received course credit for their involvement. Half of the participants were randomly assigned to the easy selection condition, and half were randomly assigned to the difficult selection condition.

Materials . Materials for the sentence completion task were those developed by Hartman and Hasher (1991) . Twenty-eight sentences with highly predictable endings (e.g., "Before you go to bed, turn off the __"; expected ending: lights ) were used in the learning phase of the experiment. For each of these sentences, a low-probability ending (e.g., stove for this example) was also available. Each low-probability ending created an unlikely but plausible sentence (e.g., "Before you go to bed, turn off the stove "). The high- and low-probability endings were relatively equal with respect to frequency of occurrence ( Kuçera & Francis, 1967 ), with mean frequencies of 116 and 190, respectively. The set of 28 critical sentences was divided into two subsets, each of which served as critical items in the learning phase for half of the participants in each selection difficulty condition. Thus, each participant viewed 14 critical sentences in the learning phase.

In addition to the critical sentences, 14 filler and 2 practice sentences with high-probability endings were also used in the learning phase. The filler and practice sentences were the same for all participants. The 14 fillers were intermixed with the 14 critical sentences, with the constraints that at least 2 filler sentences appeared at the beginning and end of each list of sentences and that no more than 3 critical sentences appeared consecutively.

For each of the 28 critical sentences developed for the learning phase, a pair of normatively moderate-cloze sentence frames (cloze value of approximately .50) was used in the test phase. One sentence in each pair was moderately predictive of the high-probability ending for the corresponding critical sentence, and the other was predictive of the critical low-probability ending. For example, for the critical sentence, "Before you go to bed, turn off the lights/stove, " the following sentences were included in Phase 2: "The baby was fascinated by the bright __" (moderately predictive of lights ), and "She remodeled the kitchen and replaced the old __" (moderately predictive of stove ). A total of 56 moderate-cloze sentences were used in the test phase, and each participant saw every moderate-cloze sentence. For each participant, 28 of the moderate-cloze sentences could be completed with the critical words from the learning phase (14 expected but disconfirmed endings, and 14 unexpected target endings), and 28 could be completed with control words that were not presented in the learning phase. The critical and control items were counterbalanced across participants such that the items that served as critical items for half of the participants served as control items for the remaining participants, and vice versa.

Design . This experiment used a 2 × 3 (Selection Condition: Case vs. Font × Item Type: Target, Distractor, Control) mixed analysis of variance (ANOVA), with selection condition varied between participants and item type varied within participants.

Procedure . Participants first read 28 normatively high-cloze sentences presented individually on a computer screen. Sentences were presented individually in white font on a black background in the center of a VGA monitor. Before each sentence was presented, a fixation cross was presented for 750 ms to help participants focus attention. After the fixation, the entire sentence frame appeared at once, minus the final word. The participant generated an ending into a microphone, and with the sentence frame still in view, two words appeared onscreen 400 ms after the ending was generated. These words were the participant-generated (expected) ending and a new (unexpected but plausible) ending. For each sentence, participants had to select the appropriate target item, which was the expected ending for filler sentences, and the unexpected ending for critical sentences. Half of the participants selected targets on the basis of letter case (easy selection condition), and the other half selected on the basis of letter font (difficult selection condition).

For filler sentences, the participant-generated ending was the to-be-remembered target item, and the less probable ending was the disconfirmed item. For critical sentences, the participant-generated ending was the to-be-forgotten disconfirmed item, and the less probable ending was the to-be-remembered target item. In the easy selection condition, the target appeared in capital letters and the distractor in lowercase letters; in the difficult selection condition, the target appeared in italic font and the distractor in normal font. For both conditions, participants were to name the target item as quickly and accurately as possible. A Lafayette voice key connected to the computer was used to log response times, and the experimenter recorded the participants' responses.

After participants selected the target item, the distractor ending immediately disappeared, and the sentence frame and target ending then remained onscreen for an additional 2,000 ms. For each sentence, participants were instructed to remember only the final target word for a later memory test. Participants received two practice sentences before beginning the learning phase.

When participants completed the learning phase, they were informed that several unrelated tasks would be given before the memory tests. Following a 5-min filled interval with an unrelated distractor task, participants completed the indirect memory test for the critical words from the learning phase. Participants were told that the purpose of the indirect test was to create stimuli for a future experiment.

For the indirect test, participants generated aloud the endings for 56 moderate-cloze sentences frames. They read each sentence aloud and generated the first word that came to mind as an ending for the sentence frame. Each sentence frame remained on screen until the participant responded, and the experimenter recorded the response. Participants advanced to the next sentence frame by pressing the space bar.

The question of central interest was whether participants showed priming for the target endings (e.g., lights ) and disconfirmed endings (e.g., stove ) from the learning phase. In other words, did participants produce the critical target and disconfirmed items more or less often than they produced the new, never-presented control items as endings for the moderate-cloze sentence frames, and did this priming vary across selection conditions? Also of importance was the time taken to select the target, with the anticipation that selection times would be longer for participants in the difficult (font) condition relative to the easy (case) condition. After completing the indirect memory test, all participants completed a health questionnaire and a questionnaire that assessed awareness of the relation between the sentence completion task and the indirect memory test.

Consistent with the findings that evening tends to be the optimal time of day for young college students, (e.g., May, Hasher, & Stoltzfus, 1993 ; May & Hasher, 1998 ) and that performance on the sentence completion task can be affected by time of testing ( May & Hasher, 1998 ), all participants in this study were tested in the afternoon and evening hours.

Sixteen participants (6 in the easy selection condition, and 10 in the difficult selection condition) reported some awareness of the relation between the learning phase and the test phase. Their data were omitted from further analyses, leaving 38 participants in the easy selection condition and 34 in the difficult selection condition. These 72 participants had a mean age of 19.6 years (range = 18—22), and an average of 12.6 years of education. There were no differences in either of these measures across selection difficulty conditions (F s < 1).

The alpha level was set at .05 for all analyses. To determine whether selection of the target item was indeed more demanding in the presumed difficult selection condition than in the presumed easy selection condition, we compared response time and error rates for target naming across the two conditions. As can be seen in Table 1 , participants were significantly faster to name the target words in the easy, case-based selection condition than in the difficult, font-based selection condition, t (70) = 3.83, p < .001. There was a tendency for the error rates to be higher in the difficult condition than in the easy condition, but the difference across conditions was not significant (p > .20). It is clear that the difference in response times across conditions was not the result of a speed—accuracy trade-off. Thus, we were successful in our manipulation of selection difficulty, with a more demanding selection task in the difficult relative to the easy selection condition, as assessed by response time for target naming.

Note that all trials in which participants made a selection error, as well as those trials in which participants failed to produce the expected ending, were omitted from further analyses. As both selection errors and generation errors were low (generation error rates were 10% and 8% in the easy and difficult conditions, respectively), few trials were omitted, and numbers were similar across conditions.

Completion rates for each of the three types of moderate-cloze sentence frames–control, target, and disconfirmed–were calculated for each participant. On the basis of these completion rates, we calculated target priming scores and disconfirmed priming scores by subtracting for each participant the control completion rate from the target and disconfirmed completion rates, respectively. Means for the completion rates and priming scores are displayed in Table 2 .

A 2 × 3 (Selection Difficulty × Item Type) mixed ANOVA was conducted on completion rates, with selection difficulty as a between-subjects variable and item type as a within-subjects variable. There was neither a main effect of selection difficulty nor a Selection Difficulty × Item Type interaction (F s < 1). There was, however, a main effect of item type, F (2, 140) = 9.6, MSE = 181.5, p < .05. Further analyses indicated a significant difference between target and control completion rates, F (1, 70) = 21.0, MSE = 316.9, p < .01, but no difference between disconfirmed and control completion rates (p > .12). Finally, the target priming effect was reliably greater than the disconfirmed priming effect, F (1, 70) = 6.4, MSE = 247.2, p < .05. Thus, in both the easy and the difficult selection conditions, participants showed priming for target but not for disconfirmed items. These data indicate that the pattern of priming for the indirect sentence completion task is determined by item relevance (as a target or disconfirmed item) rather than by the difficulty of target selection.

This study investigated the effect of selection difficulty on priming for unexpected target items and expected distractor items in a garden-path sentence completion task ( Hartman & Hasher, 1991 ). According to the integration—selection hypothesis, selection difficulty should interact with item expectancy to affect the pattern of priming observed, with priming for unexpected (target) items only in the easy selection condition, and priming for both unexpected (target) and expected (disconfirmed) items in the difficult selection condition. An inhibitory framework, by contrast, predicts that selection difficulty should not affect the pattern of priming; instead, priming should be determined by item relevance, with significant priming for target but not for disconfirmed items. The data are clearly in line with an inhibitory view: In both the easy and difficult selection conditions, younger adults showed priming for target items but not for disconfirmed items, despite significant differences in selection difficulty across conditions. In other words, when information is no longer relevant (as when a generated ending has been disconfirmed), it is efficiently suppressed and thus not easily accessible, even though it should have had increased activation from having been generated.

These findings were obtained with a design that, in contrast to Hartman's (1995) research, varied selection difficulty (and only selection difficulty), across conditions of a single experiment, and that included support for the presumed difficulty difference from an independent, external measure. In our experiment, the critical determinant of performance in the garden-path sentence completion task appears not to have been selection difficulty, but rather item relevance, that is, whether an item was a target or a disconfirmed word.

Note that in this study, item relevance was fully confounded with item expectancy. Critical target items were always unexpected words, and disconfirmed items were always expected words. One might argue, therefore, that the pattern of priming was determined by item expectancy, with priming for unexpected but not expected items. However, recall that on the basis of Masson and MacLeod's (1992) findings, one would anticipate priming for expected items in this task, as participants generated those items. The lack of priming for expected (disconfirmed) items therefore suggests that younger adults actively inhibited those items.

One other potential limitation of this study is that we used only one manipulation of selection difficulty, and that manipulation was perceptual in nature. It is possible that although selection was significantly more difficult in the font condition than in the case condition, the size of manipulation was not robust enough to affect priming, and that a comparison across types of selection, such as across perceptual and semantic selection, may be needed. Indeed, in the Hartman (1995) studies, perceptual selection and semantic selection conditions produced dramatically different patterns of priming. In Experiment 2, therefore, we explored the possibility that changes in priming are dependent on a semantic selection task. To this end, we compared performance on a difficult perceptual task (similar to the font condition in Experiment 1) with that on a semantic selection task modeled after that used by Hartman (1995 ; Experiments 3 and 4). In addition, we tested both younger and older adults to examine priming for people who are presumed to be efficient and inefficient selectors, respectively, in these challenging contexts.

With respect to younger adults, we expected performance in the perceptual selection condition to closely resemble that in the analogous font condition in Experiment 1, with reliable priming for targets only. In line with an inhibitory framework, we also expected a similar pattern of target priming but not disconfirmed priming for the semantic selection condition, as younger adults should actively inhibit the disconfirmed items regardless of selection difficulty. Note that the integration—selection hypothesis generates very different predictions for the semantic condition. As our semantic selection condition was similar to that of Hartman (1995 ; Experiment 3), the integration—selection hypothesis predicts that younger adults will show priming for both targets and disconfirmed items in this condition.

With respect to older adults, the inhibitory view predicts reliable priming for both target and disconfirmed items across the two selection conditions. The integration—selection view, by contrast, predicts a different pattern of priming for the perceptual and semantic conditions. By Hartman's (1995) view, perceptual selection is relatively easy, and so older adults in the perceptual selection condition should show priming for the unexpected, target items but not for the expected, disconfirmed items (as was found in her Experiments 1 and 2). For the semantic selection condition, which by Hartman's view is extremely difficult, the pattern of priming should be reversed for older adults: Reliable priming should be evident for the expected, disconfirmed items but not for the unexpected, target items (as was found in Hartman, 1995 ; Experiment 4).

Participants . Fifty-two college students (age range = 18—22 years) from the University of Arizona and 44 healthy, well-educated older adults (age range = 60—75 years) from the communities of Tucson, Arizona and East Lansing, Michigan participated in this study. Younger adults received course credit for their participation and older adults received monetary compensation plus parking fees. Half of the participants in each age group were randomly assigned to the perceptual selection condition, and half were randomly assigned to the semantic selection condition.

Materials and procedure . The materials and procedure in Experiment 2 were identical to those in Experiment 1, with the following exceptions: In the learning phase of this experiment (as in Experiment 1), participants generated endings to high-cloze sentence frames. However, after each expected ending was generated, a cue appeared on screen to indicate which of the upcoming items was to be selected as the target. For the perceptual selection condition, the cue was either the word italics or upright ; in the semantic selection condition, the cue was either the word expected or unexpected. The cue appeared for 500 ms and was followed by the presentation of the target and distractor items. These items remained on screen until the participant named the target into a microphone, and then the sentence frame and target ending appeared for an additional 2,500 ms.

As in Experiment 1, the participant-generated ending was the target for half of the trials (filler items), and the unexpected ending was the target for the remaining trials (critical items) across both perceptual selection and semantic selection conditions. For participants in the perceptual selection condition, half of the critical (unexpected) targets and half of the filler (expected) targets appeared in italics, and the remaining half of all targets appeared in upright font. Because item font and item expectancy were counterbalanced, participants were forced to use the cues on each trial to select the target rather than use guessing strategies (as might occur if the italicized items were always the expected endings). In a similar vein, participants in the semantic selection condition were cued to select the expected item for half of the trials (filler items) and to select the unexpected items for the other half of the trials (critical items). All target and distractor items were presented in upright font in the semantic condition.

As normative data suggest that younger adults often show eveningness tendencies whereas older adults often show morningness tendencies (e.g., Adan & Almirall, 1990 ; May & Hasher, 1998 ; Mecacci & Zani, 1983 ; Yoon & Lee, 1997 ), all younger adults in this study were tested in the afternoon and evening, and all older adults were tested in the morning.

Design . The design was a 2 × 2 × 3 (Age: Young vs. Old × Condition: Perceptual vs. Semantic × Item Type: Target, Distractor, Control) mixed ANOVA, with age and condition varied between participants, and item type manipulated within participants.

Eight young participants (5 in the perceptual selection condition, and 3 in the semantic selection condition) reported some awareness of the relation between the learning phase and the test phase. Their data were omitted from further analyses. The remaining 44 young adults had a mean age of 20.1 years (range = 18—21), an average score of 19.9 on the Extended Range Vocabulary Test (ERVT), and an average of 14.1 years of education. The 44 older adults had a mean age of 68.2 years (range = 62—73), a significantly higher score on the ERVT (M = 30.1), F (1, 84) = 76.4, MSE = 2651.8, and an average of 14.6 years of education, which did not differ from that of younger adults (F < 1.5). There were no effects of selection condition and no interactions for these variables (all F s < 1.5).

One initial comparison of interest was the selection difficulty in the perceptual condition versus the semantic condition. To assess selection difficulty, we compared mean response times and error rates for target naming across conditions (see Table 3 ). A 2 × 2 (Age × Selection Condition) ANOVA on response times indicated a main effect of age, F (1, 84) = 18.1, MSE = 29,352,534.1, p < .05, with younger adults responding significantly faster than older adults. There was no effect of selection condition (F < 1.5) and no Age × Selection Condition interaction (F < 1).

The error rate data show an identical pattern of age and condition effects, with a main effect of age, F (1, 84) = 5.6, MSE =320.3, p < .05, but no effect of selection condition (F < 1), and no reliable Age × Selection Condition interaction, F (1, 84) = 1.7, p > .19. Thus, the response time data and the error rate data join together to suggest that our font selection condition may be as difficult as the semantic selection condition.

As we had accepted Hartman's (1995) arguments about the relative difficulty of perceptual versus semantic selection tasks, we had expected slower response times or higher error rates, or both, in the semantic as compared with the font task. Although the response time means are in the expected direction, the differences are far from significant. There are a number of possible explanations for this outcome, including the high between-subjects variability in the response times, especially for older adults, and the possibility that our measure was relatively insensitive. It could also be that our intuitions (and Hartman's) about what makes a selection task especially difficult are in error. Regardless of which, if any, of these points is valid, it is still the case that by comparing the font and semantic selection tasks in Experiment 2, we are comparing qualitatively different selection tasks, in which the difference is similar to that between Hartman's (1995) Experiments 1 and 2 versus her Experiments 3 and 4. ⁶ As the degree of selection difficulty did not differ across our font and semantic conditions, the next question we addressed was whether the nature of the selection task, that is, perceptual versus semantic, had any impact on the pattern of priming observed.

To assess priming on the perceptual and semantic tasks, we first calculated completion rates and priming scores for each participant in the same manner as in Experiment 1. As in Experiment 1, all items for which participants made a selection error in the learning phase, as well as those trials in which participants failed to produce the expected ending in the learning phase, were excluded from these analyses. Generation error rates for young adults were 13% and 14% in the perceptual and semantic conditions, respectively; for older adults they were 11% in both conditions. Means for the completion rates and priming scores are displayed in Table 4 .

A 2 × 2 × 3 (Age × Selection Condition × Item Type) mixed ANOVA was conducted on completion rates. Results indicated a main effect of age, F (1, 84) = 6.8, MSE = 17,467.3, p < .05, and of item type, F (2, 168) = 11.2, MSE = 17,467.3, p < .05, and a reliable Age × Item Type interaction, F (2, 168) = 4.4, MSE = 17,467.3, p < .05. There were no effects of selection condition and no higher order interactions (F s < 1). To further assess the Age × Item Type interaction, we first examined performance on control items. There was no effect of age on control items (F < 1.5), with both age groups showing equivalent rates for completing sentence frames with the never-seen control items. The remaining comparisons were thus conducted on priming scores for the target and disconfirmed items.

As can be seen in Table 4 , the source of the Age × Item Type interaction is the priming observed for disconfirmed items. Target priming was robust both for younger adults, F (1, 43) = 22.2, MSE = 6,757.6, p < .05, and for older adults, F (1, 43) = 9.7, MSE = 10,941.0, p < .05, with no significant difference across groups (F < 1). However, older adults showed reliably greater priming for disconfirmed items than did younger adults, F (1, 84) = 6.4, MSE = 19,325.5, p < .05. Indeed, younger adults failed to show significant priming for disconfirmed items in either selection condition (F s < 1), whereas older adults showed reliable disconfirmed priming in both the perceptual, F (1, 21) = 10.3, MSE = 2,089.3, p < .05, and semantic conditions, F (1, 21) = 8.9, MSE = 6,171.8, p < .05. Finally, younger adults showed reliably greater target priming effects than disconfirmed priming effects, F (1, 43) = 9.01, MSE = 164.3, p < .05, but older adults did not (F < 1). Thus, what emerges is a pattern in which younger adults show priming for targets but not disconfirmed items (regardless of whether selection is based on perceptual or semantic stimulus attributes) and in which older adults show priming for targets and for disconfirmed items (again, regardless of selection condition).

These findings are fully in line with predictions from an inhibitory view. In both difficult perceptual and semantic selection tasks, younger adults are efficient at suppressing no-longer-relevant information and thus show priming for target items only. However, older adults, who suffer age-related deficits in inhibitory processing (e.g., Hasher et al., 1999 ; McDowd et al., 1995 ), show priming for both target items and disconfirmed items.

By contrast, these findings are problematic for an integration—selection view, which holds that (at least for the semantic selection condition) younger adults should show priming for both the unexpected (target) items and the expected (disconfirmed) items and that older adults should show priming for the expected (disconfirmed) items only. These data thus join with those from Experiment 1 to suggest that when the garden-path sentence completion task involves the generation and subsequent rejection of high-cloze endings, the primary determinants of the pattern of priming are not selection difficulty and item expectancy, but rather item relevance and inhibitory efficiency.

This research used both perceptual and semantic selection tasks to investigate the impact of selection difficulty on priming for target and distractor information in the Hartman and Hasher (1991) garden-path sentence completion task. Across two studies, the evidence strongly suggests that the pattern of priming demonstrated by younger and older adults is relatively unaffected by selection difficulty, at least when the task is implemented with learning procedures that require the generation of the to-be-disconfirmed endings. Instead, priming in this paradigm is determined by item relevance and by the ability to inhibit no-longer-relevant information. In Experiment 1, younger adults showed priming only for targets both in an easy perceptual selection condition and in a difficult perceptual selection condition. In Experiment 2, younger adults again showed priming only for targets across perceptual and semantic selection conditions, whereas older adults showed priming for both targets and distractors across selection conditions. These findings are consistent with an inhibitory framework that suggests that younger adults are efficient at deleting no-longer-relevant information from working memory, whereas older adults are less efficient at doing so (e.g., Hasher & Zacks, 1988 ; Hasher et al., 1999 ). Thus, we believe that the sentence completion task developed by Hartman and Hasher (1991) provides a measure of individuals' ability to suppress no-longer-relevant information from working memory.

Further evidence that this task does indeed gauge inhibitory functioning comes from a study by May and Hasher (1998) in which younger and older participants were tested at peak and off-peak times of day in the original garden-path sentence completion task. One key finding was that younger adults tested at their peak time showed reliable below-baseline priming for the disconfirmed items; that is, they were less likely to produce the disconfirmed items than the new, never-seen control items. It is difficult to see how a noninhibitory view could account for such a finding.

The data from these experiments and from related research therefore suggest that performance in the original garden-path sentence completion task ( Hartman & Hasher, 1991 ) reflects individuals' ability to inhibit no-longer-relevant material. One remaining puzzle with respect to the sentence completion task, however, concerns the difference in the pattern of priming found in our research versus that reported by Hartman (1995) , particularly for the semantic selection task. In Hartman's semantic selection condition, younger adults showed priming for expected and unexpected items, regardless of item relevance, and older adults showed priming only for expected items. Our younger adults, however, showed priming for target (unexpected) items only, and older adults showed priming for both item types. As our semantic selection task was modeled after that used by Hartman (1995) , it is puzzling that the data are so different across studies. However, we suggest that an important methodological difference between the Hartman (1995) study and our research may account for the differences in priming across the studies, at least for younger adults.

Consider Hartman's (1995) finding that younger adults show reliable priming for distractor items, a finding that appears to conflict with an inhibitory view. It is possible that a relative lack of competition from the distractors during the learning phase contributed to Hartman's results. Note that in our studies (and in the original Hartman & Hasher [1991] study), participants first generated the endings to sentences, and half of those generated endings were confirmed, whereas the other half were disconfirmed. In this way, participants could not discriminate between targets and disconfirmed items on the basis of expectancy, nor could they instantly reject the generated items. Each item that participants generated was thus a "potential" target, and participants may have temporarily committed themselves to the generated items until they were subsequently disconfirmed (or confirmed). By contrast, in the Hartman (1995) study, participants did not generate endings to the sentence frames; the target and distractor items were presented simultaneously below the sentence frame, and the distractor was always either the expected item or the unexpected item for a given participant. Thus, the distractor item was never a potential target because (a) the item was never generated and (b) the mapping of item relevance and item expectancy was consistent for a given participant. Relative to the disconfirmed items in our study, then, the distractor items in the Hartman (1995) study may have been weak competitors that were never held as a potential target candidate. Data elsewhere suggest that strong, interfering competitors are actively inhibited, whereas weak competitors are not ( Anderson, Björk, & Björk, 1994 ; Dagenbach & Carr, 1994 ), and that inhibition may be restricted to situations in which one candidate for response directly competes with another ( Simpson & Kang, 1994 ). Thus, because the distractor items in the Hartman (1995) study were relatively weak competitors, they may not have elicited inhibition and, as a result, showed significant priming effects.

The discrepancy in the pattern of priming for older adults across studies is a bit more puzzling. Although older adults in our study showed priming for targets (unexpected) and for disconfirmed (expected) items, those in the Hartman (1995 ; Experiment 4) study showed priming for disconfirmed (expected) items only. The lack of priming for target (unexpected) items in Hartman's study is not easily accounted for by inhibitory framework, but it is also difficult to accommodate within the integration—selection view. ⁷ The explanation posed by Hartman is that when selection difficulty is extremely high, older adults have difficulty integrating even expected items and are likely to suffer even greater deficits in encoding unexpected information. The lack of priming for unexpected items thus reflects an encoding deficit for those items under difficult selection conditions. We find this explanation to be somewhat ad hoc and unsatisfactory for two reasons. First, by an integration—selection view, it seems reasonable to anticipate that priming for unexpected items should increase rather than decrease in a difficult selection situation. If heightened selection difficulty impairs integration of items within a context, then unexpected items should be even more difficult to integrate, and thus, according to the integration hypothesis, one would expect hyperpriming, rather than reduced priming, for those items. Second, Hartman (1995) suggested that the enhanced selection difficulty in her semantic condition disrupted participants' ability to encode the unexpected items and hence resulted in reduced priming for those items relative to easy selection conditions. However, it is unlikely that either older or younger adults in Hartman's study failed to encode the unexpected items, as both age groups were quite successful in identifying those items as targets in a post-experiment target-identification test. Given that individuals were instructed to select the unexpected items as the targets and that they were successful in this task, a failure to encode those items does not seem like a reasonable explanation for the diminished priming observed. Thus, the selection difficulty hypothesis does not provide a satisfactory account of the reduced priming for unexpected items evidenced by older adults in Hartman's (1995) semantic selection task.

A final point to consider in reconciling the discrepancies in priming patterns across studies is that it is difficult to know whether the changes in priming observed internally in Hartman's (1995) experiments were in fact due to changes in selection difficulty alone, because several aspects of the task changed across the experiments (e.g., absence vs. presence of distractor items, manipulation of target expectancy within-participants vs. between-participants, and exposure duration of the sentence frames). What is especially troubling about the selection argument, however, is that there was no independent measure of selection difficulty in any of the experiments. Thus, the evidence for an increase in selection difficulty was indirect at best and circular at worst: It was the change in the pattern of priming, a change interpreted as due to an increase in selection difficulty. Our work attempted to address this limitation by examining priming in the garden-path sentence completion task when all parameters except selection difficulty were held constant and when selection difficulty was assessed with a measure independent of priming. In Experiment 1 we found reliable differences in selection difficulty with no effect on the pattern of priming shown by younger adults. In Experiment 2, the nature of the selection task (perceptual or semantic) did not affect selection difficulty measures as would have been expected on the basis of arguments raised in Hartman (1995) . However, task demands were very different across the perceptual and semantic conditions, yet there was no effect of the nature of the task on priming for either younger or older adults.

The data in our studies converge to indicate that the Hartman and Hasher (1991) garden-path sentence completion task, when implemented with learning procedures that require the generation of to-be-disconfirmed information, assesses individuals' ability to inhibit information that was once appropriate but becomes irrelevant. Although selection difficulty (or at least the nature of the selection task) may influence priming for variations of the garden-path paradigm (e.g., those that do not involve generation of disconfirmed endings, or those in which participants consistently select either the expected or the unexpected item), it does not appear to affect priming for either younger or older adults in the original garden-path task.

The usefulness of the garden-path task as an inhibitory index may hold particular significance to our understanding of cognitive aging, as inhibitory deficits have been proposed to account for age-related deficits in a number of cognitive domains, including selective attention, language comprehension and production, and memory retrieval (e.g., Arbuckle & Gold, 1993 ; Hasher & Zacks, 1988 ; Hasher et al., 1999 ; McDowd et al., 1995 ). Thus, the development of valid measures of inhibitory processing is critical to successful investigations of the inhibitory-deficit account of age-related cognitive declines. Indeed, this issue may become increasingly important both for cognitive aging and for general investigations of inhibitory processing because the validity of other measures (such as negative priming) as indexes of inhibitory efficiency has recently been called into question (e.g., May, Kane, & Hasher, 1995 ; Neill & Valdes, 1992 ).

1

Items are counterbalanced across participants so that each item serves as either a critical item (target or distractor) or a control item an equal number of times in the experiment.

At optimal times of day, younger adults actually show reliable below-baseline priming for disconfirmed items ( May & Hasher, 1998 ).

At nonoptimal times of day, older adults show increased priming for disconfirmed items and fail to show priming for the target items ( May & Hasher, 1998 ).

We note that this is an ad hoc assumption generated by Hartman (1995) as an explanation for her findings of reduced priming for unexpected items in her difficult selection conditions. Difficulties with this assumption are addressed in the General Discussion section of the article.

In the easy selection conditions of Hartman's (1995) experiments, targets either appeared alone at the end of each sentence (Experiment 1) or appeared in capital letters at the end of each sentence but were preceded by the disconfirmed items, which appeared in lowercase (Experiment 2). In the difficult selection conditions (Experiments 3 and 4), target and disconfirmed items were presented simultaneously below each sentence, and participants made selection decisions on the basis of item expectancy.

There is one notable difference between the semantic task used by Hartman (1995 ; Experiments 3 and 4) and the one in our study: In Hartman's study, participants were instructed to select either the expected or the unexpected item for all trials, whereas in our study, participants did not know which item would be designated as the target until the cue appeared on each trial. Note that the effect of this variation on selection difficulty, if any, should be an increase in difficulty level for the condition in our study relative to Hartman's. Thus we believe that the selection difficulty in our semantic condition was at least equivalent to, if not greater than, that in Hartman's (1995) semantic conditions.

Although unknown in 1995, recent evidence indicates that circadian arousal has a large impact on tasks involving inhibition and can have profound effects on the pattern of priming observed (e.g., May & Hasher, 1998 ). Indeed, one study with the original Hartman and Hasher (1991) task demonstrates that older adults tested at off-peak times show robust priming for expected distractors but not for unexpected targets ( May & Hasher, 1998 ). Of course, with respect to the Hartman (1995) studies, as with many other investigations in cognitive gerontology, it is difficult to estimate what contribution, if any, circadian arousal had to performance, as little is known about when during the day participants were tested. We emphasize, however, that given that approximately 75% of older adults are morning types, it is important to consider time of day in any investigation of inhibitory functioning in older adults.