Eating Up Cognitive Resources: Does Attentional Consumption Lead to Food Consumption?

Abstract

What is the role played by attentional load in eating? Does attending to an unrelated task generally lead to overeating, perhaps by preventing individuals from focusing on a goal to limit consumption? Or does such attentional diversion typically lead to reductions in eating, perhaps by preventing people from noticing tempting features of relevant food cues? Past research has supported each of these two propositions, but comparisons between existing studies have been hampered to the extent that various experimental manipulations differ in the degree to which they occupy attention, as well as differing in the particular type of attentional resources they exploit. To resolve existing discrepancies in the literature, in a series of studies, we made use of a working memory manipulation, the n-back task (Kirchner, 1958), that can be systematically modified to induce varying levels of cognitive load, allowing for rigorous comparisons of the effects of different levels of attentional load on eating. These studies revealed a complex pattern of results. Analysis of findings from three studies employing within-subjects designs documented a linear relationship, in that participants consumed less food when completing a higher cognitive-load task than when completing a lower cognitive-load task. Three studies employing between-subjects designs highlighted a less consistent pattern of results, but when combined in a mini-meta-analysis, suggested the opposite linear relationship, with participants assigned to higher cognitive-load conditions generally consuming more food than participants assigned to lower cognitive-load conditions. We conducted two additional studies to reconcile these conflicting patterns of data. Neither finding received unequivocal support, although both studies found that participants ate less when engaged in higher cognitive-load tasks than lower cognitive-load tasks. The precise nature of the relationship between attentional load and eating remains elusive.

1. Introduction

Consider two common experiences. In the first, you are working hard at your computer and suddenly realize that you have eaten an entire bowl of potato chips without even noticing. In the second, you stop after a period of hard work and notice that you have not even touched the snack you set out for yourself. The first example suggests that concentrating on a task may lead to overeating, perhaps because it prevents individuals from focusing on their goals or desires to limit how much they intend to eat. The second example suggests that such concentration may reduce eating, perhaps by preventing individuals from focusing on the tempting features of the food or their own hunger sensations. What then is the relationship between attention and eating? Is it necessary to focus on food to resist it, or does focusing on food make it all too easy to eat it?

To study the role played by attention in eating, researchers generally make use of manipulations of cognitive load, such as requiring individuals to hold a series of numbers in memory (Liguori et al., 2020), make arithmetic calculations or otherwise process stimuli (e.g., Ward & Mann, 2000), or, in an effort to more closely approximate typical daily experiences, watch an absorbing TV show or engage in a conversation (e.g., Ogden et al., 2013). If individuals’ working memories are occupied with any of these forms of cognitive load, they will likely have fewer mental resources available to focus on other external or internal stimuli (Oberauer, 2019), such as the features of the food (e.g., Liang et al., 2018), eating goals (e.g., Ward & Mann, 2000), feelings of hunger or fullness (Ogden et al., 2013), or the amount consumed (e.g., Braude & Stevenson, 2014).

The aforementioned cognitive load manipulations differ from each other in many ways, rendering comparisons between them difficult. For example, they almost certainly differ in the extent to which they occupy attention, as well as in the particular type of attention they occupy. However, it is not typically possible to quantify the degree of attentional resources consumed by such tasks or even to identify what particular aspect of attention is being expended. The tasks also differ in ways that do not pertain directly to attention but that still may influence eating, such as in the case of a manipulation that involves the presence of other people (e.g., a “confederate” participant who also engages in food consumption; Herman et al., 2003). Finally, the manner in which the cognitive load tasks are presented in the studies differs based on whether they are presented as the primary task for participants to focus on, or if instead eating is presented as the primary task and the cognitive task is merely secondary. Accordingly, in the present research, across eight studies, we made use of an attention task, the n-back task (Kirchner, 1958), that can be systematically manipulated to induce varying levels of cognitive load, thereby allowing for more precise comparisons of their effects on eating. In all cases, it was presented as the primary task for participants to engage in, with food offered incidentally.

1.1. Does Cognitive Load Affect Eating?

Given the differing attention tasks that have been employed in past studies, it is perhaps not surprising that no single clear pattern of consumption has emerged under conditions of cognitive load. In fact, the pattern is sometimes unclear even when comparing the results of studies that use the same types of tasks. For example, in two studies researchers manipulated cognitive load through the use of a dual-task procedure that required participants to listen to a recorded story, with the expectation of being tested for their memory of it later, while they were also asked to keep track of the number of times animals were mentioned in the story. One of these studies found that the task resulted in increased eating, as compared to a control condition (Boon et al., 2002), whereas the other failed to find such differences in eating (Boon et al., 1997).

In another set of studies that also made use of a dual-task cognitive-load procedure, participants viewed a series of art slides with the expectation that they would be tested on their memory for them, while at the same time performing a reaction time task. The results revealed increased food intake under these circumstances (although only among restrained eaters; Ward & Mann, 2000). However, a further set of studies conducted by those same researchers showed that it is possible to alter the relevant situation in such a way that participants will consume less under cognitive load (i.e., by placing salient reminders of dieting in the environment; Mann & Ward, 2004). In related research in which participants performed a memory task (listening to a stream of numbers and indicating whenever they heard three odd or three even numbers in a row), participants again ate less when in this cognitively demanding condition than when in a control condition (Liguori et al., 2020).

Additional studies have involved such cognitive-load manipulations as having participants watch television (e.g., Bellisle et al., 2004; Braude & Stevenson, 2014; Ogden et al., 2013) or listen to stories presented via the radio (Bellisle et al., 2004; Long et al., 2011), again resulting in increased eating, as compared to control conditions involving minimal cognitive load (but see Bellisle et al., 2009, in which neither watching TV nor listening to the radio led to increased eating). One study compared a television viewing task to one involving listening to a symphony and found that participants ate more while watching TV than listening to music (Blass et al., 2006). That study did not include a control condition, but a study in which participants reported their eating multiple times per day and answered questions about the circumstances surrounding their consumption found that the presence of music was associated with higher food intake than when music was not present (Stroebele & de Castro, 2006).

Overall, these studies suggest that engaging in tasks that presumably expend attentional resources, such as watching TV, often leads to increased food intake. However, a subset of the literature suggests that food intake may not change (e.g., Boon et al., 1997) or may even decrease (e.g., Liguori et al., 2020) when intake is accompanied by simultaneous engagement in a task that demands attention. To integrate these varied findings, it is helpful to consider how cognitive load specifically affects food consumption.

1.2. How Does Cognitive Load Affect Eating?

If, as many of the prior studies suggest, cognitive load leads to increased eating, it may do so by limiting attention that would otherwise be allocated to the self-regulatory tasks of monitoring and modifying food intake to align with one’s consumption goals. Without the ability to attend to these processes, people may end up overeating (Carver & Scheier, 1981). In partial support of this hypothesis, one study found that when people ate while watching TV, they later provided less accurate estimates of how much they had eaten (estimating they had eaten less than they actually had), as compared to people in a control group, but, interestingly, the two groups’ consumption levels did not actually differ (Moray et al., 2007). Moreover, in other research, when participants under conditions of high cognitive load ate more than did those under low cognitive load, they still provided accurate estimates of their consumption (Mann & Ward, 2004; Ward & Mann, 2000). It may well be the case, however, that food consumption levels depend not so much on a particular estimate of food intake but rather on the accuracy of the estimate in relation to one’s goals (see Carver & Scheier, 2001).

Another explanation for increased consumption under cognitive load comes from research on the sensory aspects of eating. Studies suggest that perceptions of the taste and odor of foods may be less intense under higher cognitive load (Hoffmann-Hensel et al., 2017; Liang et al., 2018; Van der Wal & van Dillen, 2013), and this reduced intensity could lead to increased consumption in order to feel satiated (van der Wal & van Dillen, 2013). Related findings come from studies in which participants are told to focus specifically on the sensory aspects of provided foods, which appears to reduce overeating (Long et al., 2011; but see Bellisle & Dallix, 2001). In addition, training people to eat mindfully, such that they devote close attention to multiple aspects of food, tends to reduce consumption (for a review, see O’Reilly et al., 2014). In sum, there appear to be several processes by which food intake may increase under conditions of higher cognitive load (and may similarly decrease in the presence of food-focused attention).

In contrast to the aforementioned findings, however, plausible explanations have also been offered for why cognitive load might result in decreased food intake. Researchers have argued that high levels of cognitive load might reduce food consumption by either distracting one away from either the act of eating or from feelings of hunger (Ogden et al., 2013); or it may prevent individuals from recognizing the food as tempting (Van Dillen, Papies, & Hofmann, 2013). It is not difficult to imagine that the reduction in intensity of flavors and smells, brought about by cognitive load (as in Hoffmann-Hensel et al., 2017 and Liang et al., 2018; above), could, under the right circumstances, lead to less consumption by reducing the pleasure or temptation of the relevant food items (such as when people who lose their sense of smell also lose their appetite and thereafter lose weight, Davidson et al., 1987). In fact, a series of studies showed that while individuals pay more attention to images of tempting stimuli than non-tempting stimuli in the absence of cognitive load, this difference does not occur when individuals are under cognitive load (Van Dillen, Papies, & Hofmann, 2013). The studies’ authors argued that a certain level of cognitive resources is necessary for people to recognize the temptation or “hedonic value” of an item; under sufficiently high cognitive load (as opposed to no load or moderate load), individuals may still notice the stimulus, but they won’t be tempted by it.

Finally, social factors may explain why people consume less under cognitive load. Research on eating that occurs during social interactions (which is sometimes considered a form of cognitive load) has produced mixed findings but frequently documents reduced consumption by study participants who are eating in the presence of others (Mori et al.,1987; Pliner & Chaiken, 1990). This result, however, may reflect social desirability concerns or the influence of social norms, rather than the effects of cognitive load per se.

In sum, based on the existing literature, it appears that a case could be made for either increased or decreased food intake under conditions of cognitive load, particularly because it is not always clear what might underlie these differing consumption patterns. And because studies have made use of a variety of cognitive load manipulations, it is difficult to account for the inconsistency in the results of those studies. Even when individual studies allow for comparisons between multiple cognitive-load conditions (e.g., Ogden et al., 2013, which compared watching TV to driving in a simulator to interacting with other people), it is not always possible to assess what aspect of the differences between those conditions may have led to the differing patterns of consumption. Indeed, to date, even manipulations considered to be highly ecologically valid provide little basis to argue that differing results reflect differences in the extent to which they tax participants’ cognitive capacity.

1.3. The Present Studies

All of the concerns discussed above point to the necessity of testing the role of attention in food intake by using a well-defined task that allows for experimental manipulations involving clear, systematic gradations of cognitive load. Moreover, if the effect of limited attention on consumption is to be properly assessed, it is essential that research participants treat the relevant cognitive task as their primary task, with eating as a secondary task. The task we selected, the n-back task (Kirchner, 1958), appears to be well-suited for these purposes. It is technically a test of working memory, which constitutes the aspect of memory that “hold[s] the mental representations currently most needed for an ongoing cognitive task available for processing” (Oberauer, 2019, p. 1). In addition, it has been used repeatedly in past research to systematically vary working memory load (Cohen et al., 1997; Evans et al., 2011; Nystrom et al., 2000).

In the field of cognitive psychology, there is consensus that working memory and attention are closely linked processes, such that tasks that occupy working memory reduce one’s ability to attend to other stimuli. However, the precise nature of that relationship remains unclear (see Oberauer, 2019, for a thorough discussion of this issue). Part of the confusion undoubtedly stems from multiple conceptualizations of attention, which is sometimes defined as a process for selecting information to be given priority in processing and sometimes defined as a limited resource for performing that processing (among several other conceptualizations). Either way, as classic work distinguishing between controlled and automatic processing (Shiffrin & Schneider, 1977) has made clear, if working memory is occupied, attention is less available for other tasks. In the studies we report here, technically, we manipulated working memory, but we have retained the term attention because we believe it best comports with how a typical reader conceives of these processes. In short, in the current studies, to limit participants’ attentional capacity, we made use of a task that occupies increasing amounts of working memory.

The task we chose, the n-back task, requires people to maintain and update information in working memory, with each successive level of the task further taxing those abilities. It has been used in functional imaging studies as a prototypical measure of working memory (Jacola et al., 2014), with such studies finding activation in brain areas that are thought to be the primary regions underlying working memory processing. Studies have also found that relevant patterns of activation vary as the task becomes more difficult (Smith & Jonides, 1998).

In the n-back task, participants are presented with a series of stimuli and asked to respond when the current stimulus matches the one presented “n” trials ago (e.g., for a 2-back version of the task, one would respond to the second A in the series ABA). This task can be used to systematically alter cognitive load simply by modifying “n,” with higher values resulting in greater cognitive demand, as they require maintaining and updating a longer series of stimuli in working memory. The n-back task also enjoys the advantage that the other aspects of the task, including the particular memory stimuli presented (e.g., letters vs. numbers) and the particular response required (e.g., hitting a button vs. providing answers orally), can be held constant across trials, serving to isolate the differences between conditions solely to differences in cognitive load. At the same time, the memory stimuli and type of response may also be manipulated between different studies, allowing for more generalizable conclusions that are not dependent on the particular type of stimuli or response in a given study.

Accordingly, across a series of eight studies, we examined how varying cognitive load, via an n-back task, affects food intake. Based on the general trend observed in much of the prior research (e.g., Ward & Mann, 2000), we hypothesized that greater cognitive load would disinhibit eating and thus result in increased food intake. However, as will become apparent, the first six studies we conducted in this investigation, which relied on two different experimental designs, revealed divergent patterns of results regarding food intake under cognitive load. We therefore conducted two additional studies to test hypotheses intended to resolve these conflicting accounts.

2. Studies 1, 2, and 3: Between-Subjects Design

The first three studies aimed to assess eating as a function of the level of working memory required by a task that participants completed while also engaging in food consumption. All three studies made use of between-subjects designs, such that each participant was randomly assigned to only one study condition and its accompanying level of cognitive load.

2.1. Method

2.1.1. Participants

A total of 187 participants took part in Study 1, 84 in Study 2, and 114 in Study 3. Studies 1 and 2 included only female participants. The age, gender, and ethnic breakdown of participants appears in Table 1. Participants received either credit toward required research participation or received extra credit associated with their introductory psychology course or other psychology courses. As past differences in eating studies involving cognitive load could have potentially stemmed from differences in the dietary status of participants, the three studies varied in the types of eaters they included, based on scores on two subscales of a shortened version of the Three Factor Eating Questionnaire (TFEQ; Stunkard & Messick, 1985). The questionnaire was included in a pre-screening survey presented several weeks before the study. In Study 1, only participants whose scores were above the sample median on both the restraint and disinhibition subscales (n = 135), or below the median on both subscales (n = 52), were recruited.¹ In Study 2, only individuals whose scores were above the medians on both subscales were recruited (n = 84). Study 3 did not make use of participants preselected based on these (or any other) measures (nor did Studies 4 through 8).

Table 1.

Demographics for Studies 1-8

	Study 1 n=187	Study 2 n=84	Study 3 n=114	Study 4 n=57	Study 5 n=115	Study 6 n=74	Study 7 n=64	Study 8 n=144
Age* M (SD)			22.29 (7.20)	22.84 (7.14)	19.79 (2.56)	23.24 (8.00)	19.58 (1.98)	20.03 (2.74)
Range			18-61	18-56	17-36	18-73	18-26	18-34
Gender n (%)
Female	187 (100%)	84 (100%)	76 (66.7%)	29 (50.9%)	74 (64.3%)	52 (70.3%)	41 (64%)	97 (67%)
Male	0 (0%)	0 (0%)	37 (32.5%)	28 (49.1%)	40 (34.8%)	22 (29.7%)	23 (36%)	46 (32%)
Other Gender Identity	0 (0%)	0 (0%)	1 (.9%)	0 (0%)	1 (.9%)	0 (0%)	0 (0%)	1 (.7%)
Ethnicity* n (%)
Black			6 (5.3%)	2 (3.5%)	1 (.9%)	3 (4.1%)	3 (4.7%)	9 (6%)
White			83 (72.8%)	38 (66.7%)	74 (64.3%)	50 (67.6%)	46 (71.9%)	90 (63%)
Asian			14 (12.3%)	13 (22.8%)	36 (31.3%)	18 (24.3%)	10 (15.6%)	43 (30%)
Hispanic			2 (1.8%)	2 (3.5%)	1 (.9%)	1 (1.4%)	1 (1.6%)	7 (45
Other			8 (7.0%)	2 (3.5%)	3 (2.6%)	1 (1.4%)	4 (6.3%)	4 (3%)
Missing			1 (.9%)	0 (0%)	0 (0%)	1 (1.4%)	0 (0%)	0 (0%)

^*These items were not assessed in Studies 1 and 2.

2.1.2. Procedures

Participants took part in individual sessions. Upon entering the laboratory, they were told by an experimenter that they were to take part in a study of the “effects of fats vs. sugars on memory” (Study 1); “effects of different types of food on memory” (Study 2); or “effects of different components of food on memory” (Study 3). They were told that they would therefore be asked to eat some food while performing a memory task at the same time. After being given the cover story, participants were provided with a consent form, permitted time to read it and ask questions, and once the experimenter was satisfied that they were willing to participate, they were asked to sign the form. Participants then completed pre-task measures of hunger and liking for the type of food they were about to be given. They were then randomly assigned to perform a particular version of the n-back task (as described below), trained in that task, and given ten practice trials to ensure they understood it. Before leaving participants alone in the lab room to complete the task, the experimenter placed the (pre-weighed) food directly in front of participants and instructed them that they were free to eat as much or as little of the food as they wanted, “but for the sake of the memory task, be sure to eat at least a little.” Participants then began their assigned task, and upon its conclusion, the experimenter returned to the room, casually moved the food away from the participants (so that they would not be able to continue eating after the task was over), administered a post-task questionnaire, and debriefed them. After participants left, the experimenter re-weighed the food in order to calculate the amount consumed during the task.

2.1.3. N-back Task

The n-back task (Kirchner, 1958) was used to systematically manipulate demands on the cognitive capacity of participants. In each version of the task, run using Medialab data collection software, participants heard a computerized audio recording that presented a series of 240 letters, with a new letter presented every 3 seconds. Participants were instructed to respond “yes” or “no” to each presented letter, via a foot-button on the floor, based on a decision rule specific to that version of the task. The foot-button served to keep participants’ hands available at all times for food consumption.

In the 0-back task condition, participants were instructed to respond “yes” whenever a particular letter, indicated to them ahead of time, was mentioned (e.g., A). In the 1-back task condition, participants were to respond “yes” whenever the current letter was the same as the letter mentioned just before it (the letter “one back”; e.g., the participant would say “yes” to the second A in B A A). In the 2-back task condition, participants were to respond “yes” only when the letter was the same as the letter two before it (e.g., the second A in A B A). Finally, in the 3-back task condition, participants were instructed to respond “yes” only when the letter was the same as the letter three before it (e.g., the second A in A B C A). In addition to the four levels of the n-back task, a control (no cognitive load) condition was run in which participants were simply given a one-digit number and asked to hold it in memory. They were not exposed to any recorded stimuli, and they were not asked to make any responses during the task.

2.1.4. Measures

2.1.4.1. Eating Measure

Before participants arrived, the experimenter portioned out onto a serving dish the food that would be given to participants to consume during the study task. The experimenter then weighed the food plus serving dish. To prevent participants from potentially feeling awkward or uncomfortable while eating, the food was given in sufficient quantity that an individual could eat substantial amounts without producing any obvious visual indication of consumption. The food provided in Study 1 consisted of potato chips (about 620 grams) and French onion dip (about 140 grams).² After Study 1, the food was altered to M&Ms because they were easier to serve, weigh, and store. Participants in Studies 2 and 3 were therefore provided M&Ms (about 320 grams and 240 grams, respectively). After the study was completed and participants had left the laboratory, the remaining food was re-weighed along with its serving dish. By subtracting the final weight from the original weight, we could calculate the amount of food consumed. Food was weighed by means of an EatSmart Precision Pro Digital Kitchen Scale, which, according to its user manual, is accurate to within 1 gram.

2.1.4.2. Validity Measures.

To assess the validity of the n-back tasks, we measured the accuracy of participants’ responses to each item (in Studies 1-3), along with their mean reaction times in responding to each item (in Studies 2 and 3 only). For accuracy, we calculated the number of items participants responded to incorrectly during each task. Reaction times in response to each item were recorded by the Medialab program that was used to run the study.

2.1.4.3. Self-Report Measures.

Self-report measures were administered before and after the eating task. Prior to the task, levels of hunger and liking for the provided type of food were assessed. After the task, participants completed items that assessed their subjective impressions of the study task in terms of how stressful, frustrating, engrossing, effortful, enjoyable, and difficult they found it. All self-report items were rated on 7-point Likert-type scales ranging from 1 (not at all) to 7 (extremely). In Study 1 only, participants were asked to self-report their height and weight.

2.1.5. Differences Between the Studies

As reported earlier, different foods were provided in Study 1 than in Studies 2 and 3. In addition, in Studies 1 and 2, the eating session lasted 9 minutes, whereas in Study 3 it lasted 12 minutes. Study 1 included all five conditions (control, 0-back, 1-back, 2-back, and 3-back), but to allow for more efficient use of our resources, no further studies included the 0-back condition, a condition that is often considered by researchers to be superfluous and whose omission is paralleled by other n-back studies (see, for example, Evan et al., 2011; Jaeggi et al., 2010). Indeed, in the studies reported here, participants performed similarly on the 0-back task (i.e., attained the same error rate) as on the 1-back task and rated it similarly to the 1-back condition on most of the subjective task ratings². Descriptive features of the studies appear in Table 2.

Table 2.

Features of Studies 1-8

Study	Manipulation of n-back Conditions	Stimuli	How Ps Respond	How Often Ps Respond	P is Alone or with RA	Time/Condition
1	Between	Audio	Foot	Every time	Alone	12 min
2	Between	Audio	Foot	Every time	Alone	12 min
3	Between	Audio	Foot	Every time	Alone	9 min
4	Within	Visual	Verbal	Hits only	RA	2 min
5	Within	Visual	Verbal	Hits only	RA	3 min
6	Within	Visual	Verbal	Hits only	RA	3 min
7	Within	Audio	Foot	Every time	Alone	3 min
8	Within*	Audio	Foot	Hits Only	Alone	6 min

P = participant

^*This study included a 2x2 design, and the additional independent variable, involving whether participants were informed of the study conditions or not, was manipulated in a between-subjects fashion.

2.2. Results

2.2.1. Validity of N-Back Tasks.

To assess whether the tasks became increasingly more difficult as they advanced from 0-back to 3-back, we calculated linear contrasts comparing accuracy as a function of n-back condition. In each study, the linear contrasts provided statistically significant support for this pattern, with participants responding incorrectly more often as the tasks progressed from 0-back to 3-back.³ The reaction time patterns were slightly different, revealing in Studies 2 and 3 a quadratic pattern rather than a linear one.³ However, reaction time measures are somewhat difficult to interpret, as participants sometimes give up or guess as the tasks become more difficult, which can lead to faster reaction times at the highest difficulty levels (Rose et al., 2006).

2.2.2. Subjective Experience of the Tasks

Based on self-reports from the post-task questionnaire, we assessed whether participants found that, as the task level increased from control and 0-back up to 3-back, the tasks became increasingly difficult, effortful, engrossing, stressful, and frustrating. To do so, we calculated linear contrasts for each item based on n-back condition. In each study, linear contrasts provided statistically significantly support for this trend.³

2.2.3. Hunger and Liking for the Foods.

Within each study, analysis of variance (ANOVA) revealed no significant differences between conditions with regard to participants’ level of hunger (all ps > .37) or liking for the presented foods (all ps > .50).³ Not surprisingly, including them as covariates in subsequent analyses did not alter any findings.⁴

2.2.4. Consumption

Means, standard deviations, and sample sizes for the amount consumed in each condition of Studies 1-3 appear in Table 3. These means are graphed in Figure 1.

Table 3.

Means (SD) of grams consumed by n-back condition for Studies 1-3, as well as linear contrast estimates and tests

	Study 1	Study 2	Study 3
Control	25.28 (17.00) n = 38	36.96 (18.31) n = 23	22.67 (15.17) n = 24
0-Back	34.69 (17.90) n = 32	N/A	N/A
1-Back	28.15 (16.74) n = 39	60.00 (29.08) n = 21	36.41 (15.90) n = 32
2-Back	30.46 (24.75) n = 39	37.05 (23.65) n = 20	40.38 (30.03) n = 29
3-back	28.67 (17.28) n=39	50.90 (31.83) n = 20	46.14 (45.49) n = 29
Linear Contrast Estimate (SE)	2.79 (3.10)	4.22 (5.64)	16.63 (5.74)
Linear Contrast t-test	t(151) = .90, p = .37	t(80) = .75, p = .46	t(110) = 2.90 p = .004
Linear Contrast r	.07	.08	.27

Figure 1.

Mean (+/−SE) grams consumed in between-Ss studies. (Note: Study 1 included a 0-back condition that the other studies did not include. The red diamond located between the Control and 1-back results indicates that value.)

Our primary aim was to explore the overall pattern of consumption in response to increasing levels of cognitive load (rather than, for example, comparing each particular condition to every other). To assess whether the relevant relationship formed a linear trend, we calculated a linear contrast for each study, using the option for polynomial contrasts accompanying the General Linear Model command in SPSS, with grams consumed as the dependent variable and n-back condition as the independent variable. The linear contrast estimate and standard error for each study appear in Table 3, as do t-tests for the contrasts. According to those t-tests, the linear contrast was not statistically significant for Study 1 or Study 2,⁴ but was significant for Study 3.

2.2.5. Meta-Analysis.

To assess the overall pattern and effect size, we then conducted a mini-meta-analysis of the three studies, as recommended by Goh, Hall, and Rosenthal (2016). To calculate effect sizes for each study, the t value from the linear contrast was transformed into r values by using the standard equation $r=\sqrt{\left[(t^2/\left(t^2+df\right)\right]}$. These r values (found in the bottom row of Table 3) indicate the extent to which each relationship represents a linear one, as well as the direction of the relevant effect (because they are calculated from the linear contrast t value, which includes direction, and not from an omnibus F statistic, which has no directionality). We then used published mini-meta-analysis procedures (Goh et al., 2016) to z-transform the r values (into r_zs), and then used those values to calculate the weighted mean correlation,⁵ i.e., weighted $\Sigma (\left[N-3\right]r_z/\Sigma \left(N-3\right)$. The value of this weighted mean correlation was .14 (and was also .14 when converted back to a regular, non-z-transformed r) and constitutes a small effect size for the linear relationship. To test for statistical significance of the linear relationship, we calculated Stouffer’s combined Z value (Mosteller & Bush, 1954) by dividing the weighted mean correlation by its standard error. The value of this combined Z value was 2.59, which is significant at p < .01, suggesting a statistically significant positive linear function.

2.3. Studies 1-3 Discussion

The findings from the mini-meta analysis of Studies 1-3 suggest a small effect of cognitive load on food intake, such that higher cognitive load leads to increased consumption. These findings are consistent with previous research indicating that participants tend to eat more during activities that require greater attentional resources than during activities that require fewer attentional resources (e.g., Bellisle & Dalix, 2004; Braude & Stevenson, 2014; Long et al., 2011). However, in the individual studies, only Study 3 demonstrated the increasing linear effect, and visual inspection of the results suggests noticeable differences across the three studies, with Study 2, in particular, revealing a rather perplexing pattern of results. This variability in effects and the overall small meta-analytic effect therefore provided only limited evidence in support of increasing food intake under increasing cognitive load. The next set of studies each utilized a within-subjects design, which was intended, in part, to provide greater power for detecting changes in food intake under different levels of cognitive load.

3. Studies 4, 5, and 6: Within-Subjects Design

The aim of these studies was again to assess the amount participants would eat as a function of the level of attentional processing required by a simultaneously occurring memory task. Because these latter studies all employed within-subjects designs, all participants were exposed to every study condition, although they were randomly assigned to the specific order in which they completed the conditions.

3.1. Method

3.1.1. Participants

A total of 57 participants took part in Study 4, 115 in Study 5, and 74 in Study 6. Participants received credit toward required or extra credit research participation for their introductory psychology course or another psychology course. Age, gender, and ethnic breakdown of the participants in each study appear in Table 1.

3.1.2. Procedures

In all three studies, participants arrived at the lab and were told they were taking part in a study of the effects of sugar on attention and that they would be eating some sugary food while playing four attention games using a deck of cards. After a consent process, participants completed the same pre-task measures as in Studies 1-3. The experimenter then explained all of the attention tasks to participants, as described below, and then participants completed all conditions (control, 1-back, 2-back, and 3-back) in a randomly assigned order,⁶ each for two minutes. Before initiating each condition, the experimenter provided the participant with an example of how to complete the task associated with that condition and then instructed participants that “for the sake of studying the instantaneous effect of sugar on attention, please eat some M&Ms during this task. Feel free to eat as many as you want, but at least have some during this task.” After each condition, the experimenter gave the participant a post-task self-report questionnaire probing their experience while completing the task. After all four conditions were complete, the participant filled out the same subscales of the Three Factor Eating Questionnaire as in Studies 1-3 and then was debriefed.¹

3.1.3. N-Back Task

The n-back task used in these studies was conceptually the same as the one used in Studies 1-3 but was administered differently to work in a within-subjects setting. Instead of listening to a stream of letters through earphones and responding to each letter with a foot pedal, participants viewed a stream of numbers on unusually large (about 3.5” x 5.75”) novelty playing cards and responded orally to the experimenter. Experimenters picked up cards and revealed them to participants one by one from a deck. Participants only had to respond to “hits” (i.e., they said “yes” when the number satisfied the n-back requirement for that condition), which occurred in about 10-15% of the trials, rather than responding “yes” or “no” to each item. As in Studies 1-3, participants’ hands again were free so as to facilitate food consumption, and since they did not have to offer many verbal responses, they were essentially unencumbered with regard to their ability to eat.

Because a typical deck of playing cards includes only four cards representing each number between 2 and 10, we constructed a special 56-card deck that consisted solely of eight instances of each number between two and eight. The experimenter showed cards to the participant according to the beat of a metronome set to 30 beats per minute, thus revealing each card for 2 seconds before moving to the next one. In the two-minute task, participants were shown 56 cards. While that particular sequence technically takes 8 seconds less than the full two minutes to complete, a small cushion of time (i.e., eight seconds) was included in case the experimenter fumbled with a card or encountered problems in properly flipping one over to show to the participant. For the control condition, participants were told that for that particular task, they needed to just remember the number “4” for the duration of the task. To ensure that the control task lasted the same length of time as the other conditions, the experimenter still showed each card in the deck at the same metronome pace but kept them face down, so participants could not see the numbers on the cards but rather just saw the “bicycle” design on the back.

3.1.4. Measures

3.1.4.1. Eating Measure

The amount of food consumed was assessed as in Studies 1-3, namely, by pre-weighing the food for each condition in its serving container and then re-weighing each container after the session was over. In Studies 4-6, the food consisted of M&M candies, and it was presented in separate small disposable plastic cups for each study condition. The pre-weight of each filled cup averaged 35 grams. The experimenter placed a new cup in front of each participant for each condition and then removed it when that condition ended and replaced it with another cup. All cups were discreetly numbered on the bottom so that experimenters could record which cup was used for each condition. Prior to running each participant, experimenters consulted a randomized list that informed them of the order in which to conduct the various conditions, at which point they labeled, prepared, and weighed all four cups of food.

3.1.4.2. Self-Report Measures

Participants responded to the same self-report measures as in Studies 1-3, with hunger and liking assessed before the commencement of any of the various conditions, along with five post-task items probing their views of each task (i.e., how stressful, frustrating, engrossing, effortful, and difficult they found it). All self-report items were rated on 7-point Likert-type scales ranging from 1 (not at all) to 7 (extremely).

3.1.5. Differences Between the Studies

For purposes of generalizability, slight variations were introduced into each of the three studies. In Study 4, each eating condition lasted two minutes, as described, but in Studies 5 and 6, each eating condition lasted three minutes. In Studies 5 and 6, each card was shown for three seconds rather than two, and the metronome was set to sound 20 times per minute. In Study 6, a 4-back condition was run in addition to the other conditions. Descriptive features of the studies appear in Table 2.

3.2. Results

3.2.1. Subjective Experience of the Tasks

We assessed whether participants reported that as the n-back level increased, the tasks became increasingly difficult, effortful, engrossing, stressful, and frustrating. To do so, we calculated linear contrasts for each item based on n-back condition. In each study, linear contrasts significantly supported this linear relationship.³

3.2.2. Consumption

Means, standard deviations, and sample sizes for each condition of Studies 4-6 appear in Table 4 (along with results for Study 7, which is introduced in the next section). The means and standard errors are also graphed in Figure 2. As with Studies 1-3, our interest was in the overall pattern of responses, rather than comparisons between any particular conditions. Accordingly, we calculated a linear contrast for each study, using the option for polynomial contrasts found in the repeated measures option of the General Linear Model command in SPSS, with grams consumed in each n-back condition as the repeated (within-subjects) measure. We then combined the studies in a mini-meta-analysis to examine the overall pattern and effect size.

Table 4.

Means (SD) of grams consumed by n-back condition for Studies 4-7, as well as linear contrast estimates and tests

	Study 4 n = 57	Study 5 n = 115	Study 6 n = 74	Study 7 n = 64
Control	7.89 (6.59)	7.33 (6.42)	10.35 (7.16)	17.19 (9.37)
1-Back	7.75 (7.16)	7.05 (6.21)	7.73 (6.79)	14.08 (9.66)
2-Back	6.54 (5.04)	7.17 (7.05)	6.70 (5.35)	12.89 (10.03)
3-back	5.32 (4.50)	5.45 (5.97)	6.22 (5.55)	10.58 (9.14)
4-back	N/A	N/A	5.90 (4.61)	N/A
Linear Contrast t-test	t(56) = 3.60, p <= .001	t(114) = 3.84, p <= .001	t(73) = 4.73, p <= .001	t(63) = 4.40, p <= .001
Linear Contrast r	−.43	−.34	−.48	−.48

Figure 2.

Mean (+/−SE) grams consumed in within-Ss studies.

For each of the three studies, the linear contrast was statistically significant, providing evidence for the linear trend visible in the graph. See Table 4 for the contrast t-tests. The direction of the linear trend was of decreasing magnitude, indicating that, in contrast to the overall findings of the between-subjects studies (i.e., Studies 1-3), the more difficult the n-back task, the less food the individual consumed.

3.2.3. Meta-Analysis

To assess the overall pattern and effect size, we then conducted a mini-meta-analysis of the three studies. We performed this analysis separately from the analysis of the first three studies because combining between and within-subjects studies is generally avoided in meta-analyses, as within-subjects studies tend to have smaller variances and stronger effect sizes than between subjects studies and are therefore generally viewed as improperly influential in integrated meta-analyses. To calculate effect sizes for each study, the linear contrast t value was transformed into an r value, as in Studies 1-3. These r values (found in the bottom row of Table 4) indicate the extent to which each relationship represents a linear one, as well as the direction of the relevant effect. We then used published mini-meta-analysis procedures (Goh et al., 2016), as in Studies 1-3.⁷ We first z-transformed the r values (into r_zs) and then calculated the overall weighted $r_z=\Sigma (\left[N-3\right]r_z/\Sigma \left(N-3\right)$. The value of this weighted mean correlation was −.43 (and was −.40 when converted back to a regular, non-z-transformed r) and constitutes a medium effect size for the linear relationship. To test that result for statistical significance, we calculated Stouffer’s combined Z value (Mosteller & Bush, 1954) by dividing the weighted mean correlation by its standard error. The combined Z value = −6.60, which is significant at p < .0001, suggesting a statistically significant negative linear function.

3.2.4. Role of Restrained Eating

As is evident from the above analyses, the within-subjects studies yielded findings that were discrepant from those of the earlier between-subjects studies. We therefore conducted exploratory analyses to assess whether the different patterns of results from Studies 1-3, as compared to Studies 4-6, could be accounted for by the different types of eaters in the respective studies. Participants in Studies 1 and 2 had been pre-selected based on scoring above or below the medians on the restraint and disinhibition subscales of the TFEQ. While we assessed participants in Studies 3-6 on those same subscales, we did not pre-select them based on those scores. Participants in all six studies were coded as one of three eater types: Not restrained or disinhibited (if they were below the median on each subscale), both restrained and disinhibited (if they were above the median on each subscale), and mixed (if they were above the median on one and below it on the other). We then re-ran the same analyses on consumption as before, but included eater type as an additional between-subjects factor, and to preserve sufficient power, combined the between-subjects participants into one sample and the within-subjects participants into another. This analysis revealed that the linear contrasts based on n-back condition were still statistically significant, and in the same direction, as reported in the previous sections: For the between-subjects studies, t(351) = 2.47, p = .014; and for the within-subjects studies, F(1, 243) = 46.71, p < .001. No other main effects or interactions were statistically significant, although two were marginally significant (at p = .09 or p = .10). In sum, a participant’s status as a restrained, disinhibited, or neither type of eater did not appear to explain their consumption patterns or the differences we found in the between-subjects studies as compared to the within-subjects studies.

3.3. Studies 4-6 Discussion

In these studies, each participant engaged in every version of the n-back task, resulting in a clear pattern of eating: the more challenging the task, the less participants ate. This finding represents a different pattern of eating than was seen in the studies in which participants engaged in only one task (i.e., the between-subjects studies). Before hypothesizing as to why the two study designs yielded different patterns of consumption, it is important to note that four other features of the studies were confounded with the differing study designs, and any of these features may have led to the different patterns of results.

First, there was an experimenter present in the room in the within-subjects studies, but in the between-subjects versions of the studies, participants were alone in the room. Having another person present may lead to less eating (Roth et al., 2001), perhaps by making people feel self-conscious, or by leading them to eat in ways they find self-enhancing (Salvy et al., 2007), which generally involves inhibiting consumption. In addition, the presence of another person may have served as a further source of distracted attention. Such a presence also may have led participants to behave according to whatever they considered appropriate norms for the situation (Herman et al., 2003), which in these studies, may have meant prioritizing success on the tasks, perhaps leading to increased attentional focus on the more difficult tasks at the expense of eating.

A second potentially confounding factor is that the within-subjects studies involved viewing the stimuli, which may have taken participants’ attention away from the food, whereas in the between-subjects studies, participants heard the stimuli through headphones, perhaps allowing their attention to remain focused on the food even during the more difficult n-back tasks. Third, participants in the within-subjects studies had to respond to the stimuli out loud, which, despite our precautions, may have inhibited their eating, particularly if they were concerned about possibly being caught with food in their mouth at the moment they intended to make an oral response. Recall that participants in the between-subjects study responded using a foot button, which left their oral consumption completely unencumbered. Finally, in the within-subjects studies, participants had to respond only to “hits,” which were infrequent, whereas in the between-subjects studies, participants had to respond to every stimulus using the foot button, indicating whether a particular trial constituted a hit or a miss. This frequent responding, and the need to remember which foot button represented a “yes” vs. “no” response, may have occupied further attention beyond just completing the n-back task. For all of these reasons, therefore, Study 7 was designed to unconfound these features from the study design.

4. Study 7: Unconfounding Study Design from Other Study Features

In this study, participants completed a within-subjects study that included all four of the key features of the between-subjects studies (see Table 2 for features of the study). That is, participants engaged in each n-back condition, but they did so alone in a room, while listening to the stimuli, which they responded to using a foot button, providing a response for every stimulus, not just hits. If those features were critical to the aforementioned findings, participants were expected to show the overall pattern revealed by the between-subjects studies, whereas if the primary study design (between vs. within) remained the critical distinction, participants should presumably show the within-subjects pattern of results.

4.1. Method

4.1.1. Participants

Study 7 included 64 participants. The age, gender, and ethnic breakdown of participants appear in Table 1. Participants received credit toward required or extra credit research participation for their introductory psychology course or another psychology course.

4.1.2. Procedures

Participants arrived at the lab and were told by the experimenter that the study was exploring the effects of sugar and/or caffeine on attention and that they would be asked to consume a food while performing an attention task. After a consent process, participants were seated at a computer and completed demographic and pre-test measures, as in the other studies. Then they were told that they were in the “sugar plus caffeine” condition, and therefore they would be asked to eat a food that contained both sugar and caffeine, namely, M&Ms, while doing the task. The experimenter then explained all four study conditions to participants, showed them the foot button they would use to respond to stimuli, and explained that they would eat from a different cup of food for each condition, rotating the cups when the computer instructions indicated that they should do so. After urging participants to read all instructions on the computer very carefully, the experimenter left the room for the duration of the study.

The instructions on the screen before each condition walked participants through an example of how to perform that particular n-back task (or the control task, in which participants were simply told to keep the number “4” in their memory for the duration of the task), including how to respond with the foot button. Then, on a new screen, the instructions stated, “For the sake of studying the instantaneous effect of sugar/caffeine on attention, please eat some M&Ms during the task. Feel free to eat as many as you want, but at least have some during the task.” On the next screen, another set of instructions told participants to take the first cup of M&Ms (labeled “1”) from the tray marked “new,” to place it by their dominant hand, to remove the lid, but not eat any until they started the task. Once ready, participants were instructed to advance to the next screen to begin the task. After completing the task, on-screen instructions told participants to move the cup aside onto the tray marked “old.” This entire procedure was followed for each of the four conditions.

4.1.3. N-Back Task

The letters for the n-back tasks were presented in an audio-only format using the Medialab stimulus-presentation software. The tasks were programmed so that each condition would present 60 letters, including 10 “hits” (except for the control condition, in which participants also heard a stream of letters but did not have to respond to any of them). Each condition lasted three minutes, at the pace of one letter presented every 3 seconds.

4.1.4. Self-Report Measures

Participants responded to the same self-report measures as in Studies 1-6, with hunger and liking assessed before the start of any of the various conditions, and five items about their subjective experience of the task assessed after each task (how stressful, frustrating, engrossing, effortful, and difficult they found it). All self-report items were rated on 7-point Likert-type scales ranging from 1 (not at all) to 7 (extremely).³

4.2. Results

4.2.1. Post-task Items

We assessed whether participants reported that as the task level increased, the task became increasingly difficult, effortful, engrossing, stressful, and frustrating. To do so, we calculated linear contrasts for each item based on n-back condition. Linear contrasts supported this trend for each of the items.³

4.2.2. Consumption

Means and standard deviations for each n-back condition are reported in Table 4, and the means for each condition are graphed in Figure 2. The same linear contrast was calculated for Study 7 as was calculated for Studies 4-6 (i.e., the prior studies that also employed within-subjects designs). As in those studies, the contrast revealed a significant linear pattern to the eating, and in the same direction as those earlier findings, as indicated in Table 4. It is evident from the table and figure that participants in this study ate more food overall than did participants in Studies 4-6. This pattern likely reflects the fact that participants were alone in the room during this study, whereas the experimenter was present during the earlier studies.

4.2.3. Meta-Analysis

Including Study 7 in the mini-meta-analysis with Studies 4-6 resulted in an overall mean correlation of −.45 (or a non-normalized r of −.48), and a Stouffer’s Z of −7.43, which is statistically significant at p <.0001.⁷ In sum, Study 7 revealed a significant linear relationship in the negative (decreasing) direction, such that as the tasks became more challenging, participants ate less.

4.3. Study 7 Discussion

In this study, a within-subjects design was used as in Studies 4-6, but it included all of the key features of the between-subjects studies conducted earlier in Studies 1-3. Participants were run through the procedure alone; listened to stimuli, rather than viewing them, allowing them to devote their full visual attention to the food; and responded to every stimulus with their feet, which left their hands free to facilitate eating. If those features accounted for the between-subjects pattern of results reported above, in which participants generally ate more under increasing cognitive load, that pattern should have been evident in this study. But it was the aforementioned within-subjects pattern of results that was obtained—participants ate less under increasing levels of cognitive load. Thus, rather than focus on potential confounds, we must look elsewhere to explain the differing patterns of results gathered from our between-subjects studies, and comparable studies reported in the literature, as compared to our within-subjects studies. Study 8 was designed to test a seemingly straightforward hypothesis that could potentially explain these differences.

5. Study 8: Testing an Alternative Explanation for Differing Pattern of Results

Study 7 failed to resolve the discrepancies found between Studies 1-3 and Studies 4-6. Altering particular attributes of the study design to match that of Studies 1-3, but with the within-subjects factor used in Studies 4-6, again revealed the effect from Studies 4-6 of decreasing intake with increasing cognitive load. With no clear explanation for the discrepancy, these studies languished in our file drawer. As the open science movement heated up and the issue of suppressing “failed” studies gained attention, we opened our file drawer and attempted to learn something from our prior research.

The possibility existed that the generally contradictory patterns of results obtained from the between-versus within-subjects designs reported here may have been traceable to differing participant expectations associated with the two designs, leading participants to engage in different strategies for approaching the study tasks. In particular, when encountering the within-subjects designs, individuals were aware that they would be performing several trials of the same task under varying conditions (e.g., 1-back, 2-back, and 3-back), and they may therefore have implicitly compared these different conditions. As a result, they may have attempted to manage their responses as efficiently as possible, reserving the bulk of their attention for the most cognitively demanding conditions, perhaps in an effort to conform to social norms and/or based on what they believed was expected of them in the study (see discussions of such demand characteristics, e.g., Nichols & Maner, 2008; Orne, 1962). In other words, in the context of these studies, they may have believed that it would be difficult or unusual to be eating during a task that requires extensive concentration (i.e., the 3-back task), and therefore they chose to eat very little during that task, knowing they would presumably be much freer to eat while performing a simpler task. Participants completing the relevant task presented in a between-subjects design, however, were, of course, unaware of the other conditions, and it was therefore presumably more difficult for them to plan their behavior or construct hypotheses regarding what response the experimenter might expect.

To pursue this hypothesized distinction, Study 8 aimed to contrast such putative within-versus between-subjects design effects by manipulating participant knowledge of the different study conditions (see Table 2 for study features). We did so by either informing or not informing participants beforehand of the tasks they would complete during the study. Participants in the informed condition were told before the tasks commenced that they would encounter both an “easy task” (which mirrored the control task used in Studies 1-7) and a “hard task” (which entailed completing a 3-back task), and they were further informed which task they would perform first. This condition was intended to mimic the knowledge level of participants in the prior within-subjects studies; we therefore hypothesized that these participants would demonstrate the documented within-subjects effect of decreasing food intake under higher cognitive load. The participants in the naïve condition were not told how many tasks they would perform, their relative ease, or the order in which they would perform them. This condition was designed to mimic the knowledge level of participants in a between-subjects version of the study; we therefore hypothesized that these participants would demonstrate the prior “between-subjects” effect of increased food intake under increased cognitive load. These hypotheses were pre-registered (https://osf.io/vyufg/), along with study procedures.

5.1. Methods

5.1.1. Participants

Participants (n = 144) were undergraduate students completing the study in exchange for extra credit research participation in psychology courses. See Table 1 for age, gender, and ethnic breakdown of participants.

5.1.2. Materials

5.1.2.1. Food

To support our cover story (see below), participants were presented with two bowls of candy. One bowl contained approximately 45 grams of pretzel M&Ms, and the other bowl contained approximately 55 grams of caramel M&Ms.

5.1.2.2. Study Tasks

As part of the cover story for the study, participants first completed a bogus taste test, supposedly comparing two new kinds of M&Ms. They then completed two computer tasks, a 3-back task and a control task, presented in a randomized order. These two tasks were chosen for this study because their results most reliably differed from each other in Studies 1-7, and using just two tasks was sufficient to examine whether more or less food was consumed under higher vs. low cognitive load. In the 3-back task, a stream of letters was presented in an auditory fashion using the program OpenSesame, with participants responding to “hits” using a foot pedal. In the control task, participants were presented with three letters on the screen and asked to remember those letters until the task ended. To keep the conditions as similar as possible, a stream of letters was presented auditorily, and at the same pace and duration, as in the 3-back condition. Each task lasted 6 minutes.

5.1.3. Procedures

The study employed a 2x2 mixed design, with knowledge of the tasks (naïve (n = 81) vs. informed (n = 63)) varying between participants and task difficulty (control task vs. 3-back) repeated within participants. Participants were randomly assigned to one of these four conditions (n = 46 naive, hard task first; n = 35 naive, easy task first; n = 30 informed, hard task first; n = 30 informed, easy task first). Participants were run in individual sessions, and the study made use of two separate lab rooms to allow for surreptitious food weighing between tasks. In the first room they entered, participants began the study by completing a consent form. They were then taken into a second room furnished with a computer for the n-back task, as well as two small bowls, each containing a different kind of candy. They then completed the bogus “taste test” of the candies, which consisted of tasting the foods and answering questions about various features of the two items. After completing the task, participants returned to the first room, where they responded to a survey seeking their feedback about the taste test they had just completed. The sole purpose of the survey was to occupy the participant so that the experimenter could return to the second room and surreptitiously weigh the bowls of candy under the guise of setting up the next part of the study. This weight provided a baseline measurement of the amount of food that would be present for the first legitimate study task.

The experimenter then brought the participant back to the second room. All participants were told that they could feel free to eat the remainder of the food from their prior taste test if they so desired, and that the next part of the study was going to involve responding using a computer. In an effort to conceal our interest in their food intake, participants were not explicitly instructed to eat but were instead simply given the option to do so. Participants in the informed condition were then instructed that they were going to complete two computer-based tasks, with one of the tasks being relatively easy and the other, relatively hard. They were also told which task they would perform first. Participants in the naïve condition were merely told that they were going to complete a computer-based task. Importantly, these instructions were all conveyed within the computer program; the experimenter remained blind to condition. The experimenter left the room while the participant completed the first task. When participants had completed that task, the program instructed them to return to the first room to fill out another survey about their experiences, thus allowing the experimenter to return to the second room to re-weigh the bowls of candy. The process then repeated for the second task.

After completing the second task, participants completed another survey about their experiences, along with demographic questions and the same restrained eating questionnaire as in Studies 1-7.¹ Finally, participants were probed for suspicion and expectations during the study before they were fully debriefed. After the participant left the lab, the experimenter weighed the food one final time.

5.2. Results

5.2.1. Power Considerations

Although our initial analysis anticipated collecting data for 266 participants for 90% power to detect a small effect of size f = .10 with alpha of .05 in a 2x2 mixed ANOVA design (assuming a correlation of .5 between repeated measurements), data collection was halted early because of the COVID-19 pandemic, and given the expectation that further data collection would not be possible for the foreseeable future, we decided to terminate the study. The achieved sample size of 144 provides 90% power to detect a relatively small effect size of f = .13 for that same test.

5.2.2. Food Intake

Examination of the amount of food consumed for each task revealed that many participants had not eaten any food during these tasks (n = 39 (27%) did not eat during either task, n = 8 (6%) ate during the hard task but not the easy task, n = 50 (35%) ate during the easy task but not the hard task, and n = 47 (33%) ate during both tasks)), as in this study, as opposed to the previous ones, participants were not told that they were required to eat some of the food during the tasks. The pre-registered analyses, which required normally distributed variables, were no longer suitable for this data, and we instead used analytic techniques deemed appropriate for zero-inflated data.

5.2.3. Effect of Knowledge Condition

We hypothesized that individuals in the naïve condition would engage in greater food consumption with increasing cognitive load, whereas individuals in the informed condition would show reduced consumption with increasing cognitive load. To handle the zero-inflated data, for each individual, we calculated a difference score in consumption between their two cognitive load conditions (with positive values reflecting greater intake under higher cognitive load) and then computed an independent samples t-test to determine if this score differed by knowledge condition. Because of this zero-inflation, initial distributions for intake during the hard task in particular were non-normal (skew = 2.61, kurtosis = 7.74); computing difference scores captures the hypothesized effect of interest, which is a crossover interaction, and improved normality of food intake distributions (skew = −.93, kurtosis = 3.4).

Participants in the naïve condition (M = −5.60, SD = 12.37) and the informed condition (M = −5.70, SD = 11.33) did not significantly differ from each other in their food consumption difference scores, t(142) = −.05, p = .957. Contrary to our hypothesis, prior knowledge of the study tasks did not affect whether individuals ate more during the control task than during the 3-back task, and in fact, in both conditions, the negative difference scores reveal that, on average, participants ate more during the easier task. The near-zero effect size between the two knowledge status conditions, d = .01 (or f = .005), suggests that even substantial additional statistical power would not have revealed a difference between the naïve and informed participants.

5.2.4. Effect of Task Difficulty Condition

Because there were a large number of instances in which participants ate nothing during the task (only 47 participants (33%) ate during both tasks), we also conducted a logistic regression to model whether knowledge condition, task difficulty condition, or their interaction affected whether or not participants consumed any food at all. This analysis revealed only a significant effect of task difficulty, OR = 5.68, p < .001. The effect of knowledge status (OR = 1.19, p = .705) and the interaction between task difficulty and knowledge status (OR = .90, p = .859) were not significant predictors of consumption. Consistent with the previous analyses, these findings indicate that individuals were more likely to eat during the easier control task than during the more difficult 3-back task, regardless of whether they were naïve or informed about the tasks they would perform. This pattern of effects was unchanged with the addition to this model of task order or self-rated hunger.³

5.2.5. Effect of Task Difficulty on Initial Consumption

To test for the hypothesized increase in food consumption under higher cognitive load among naïve individuals in a manner most closely paralleling the between-subjects studies previously reported, we also examined intake that occurred solely during the first task engaged in by naïve condition participants (i.e., before they were aware of any other task condition; n = 81). We again used logistic regression to predict whether or not participants ate during the task. Task difficulty was a significant predictor, but contrary to our hypothesis (and similar to the findings above that include both tasks completed by naïve participants), naïve participants were more likely to eat during the easy control task (74%) than during the more difficult 3-back task (48%), OR = 3.15, p = .018.

Across all of our analyses, the estimates of effects suggest we would have been unlikely to detect effects of knowledge condition even if the study had run to completion and we had greater statistical power. Indeed, to obtain effects consistent with our initial hypothesis, the results among naive participants would have had to reverse in direction.

5.3. Study 8 Discussion

We had hypothesized that participants who were informed of both of the tasks they were about to perform, along with the associated difficulty of those tasks, would eat more during the easier control task than during the more difficult 3-back task, paralleling what we observed in the within-subjects designs of the studies reported above. This hypothesis was supported. We also hypothesized that participants with no prior knowledge of the tasks would eat more during the more difficult 3-back task than the easier control task, mimicking the overall between-subjects effect observed earlier. This hypothesis was not supported. Results indicated that regardless of prior information about the tasks, participants were more likely to eat during the easier control task than the harder 3-back task. The effect of increasing food intake under increasing cognitive load, demonstrated in some of our earlier between-subjects designs, and in similar studies in the literature (e.g., Boon et al., 2002), did not replicate in this study, even when only considering the first task completed by participants in the naïve condition, the most direct parallel to those studies.

6. General Discussion

Across eight studies, we sought to answer a seemingly simple question: Do people eat more or do they eat less when their attention is primarily focused elsewhere? In Studies 1-3, using between-subjects designs, we showed that people generally (but not always) eat more, whereas in Studies 4-6, using within-subjects designs, we found the opposite result, namely, that people eat less when simultaneously performing an attention consuming task. We conducted two additional studies in an effort to reconcile our findings, but the results of those studies ultimately failed to support our hypothesized explanations for the earlier conflicting pattern of results. One of those studies (Study 7) was designed to test whether the discrepant results obtained in the between- vs. within-subjects studies were produced by other aspects of the studies that happened to be confounded with those differing study designs. Evidently, they were not. Using a study that included the design elements that had been confounded with the between-subjects format (e.g., being alone, hearing the stimuli, etc.) but that made use of a within-subjects design, we still obtained the pattern of findings shown in our original within-subjects studies. Accordingly, the procedural details that distinguished our earlier between- vs. within-subjects efforts do not appear to have played a role in producing the relevant differences in results.

In our second effort to reconcile our findings, we tested one more feature of our within-subjects designs that may have accounted for our results, namely, the fact that when taking part in the within-subjects studies, participants had knowledge of all the tasks they were to engage in before starting any of them. They may then have used that knowledge to make strategic choices (based on demand characteristics, social norms, or some other reason) regarding how to behave during the various tasks, including if and when to eat. These strategic choices would, of course, not have been available to a participant in a between-subjects study. To account for this potential confound, in our final study we manipulated whether or not participants were informed beforehand about the tasks that they would engage in, predicting that informed participants would show a pattern of eating consistent with our prior within-subjects studies, whereas naive participants would show a pattern echoing that of results garnered from our between-subjects studies. In fact, participants in both conditions showed the pattern observed in our within-subjects studies, eating less food under increasing levels of cognitive load. These results thus failed to confirm our hypothesis that knowledge of the upcoming tasks had led participants to strategically alter their behavior.

Difficulty determining consistent effects of cognitive load on food intake is not unique to this set of studies, and indeed, as outlined in the introduction, each set of findings has enjoyed at least some documented support in the literature. If we are to learn about the mechanisms involved in regulating food intake, as this line of research ultimately aims to do, it is necessary to further clarify what is actually manipulated in studies of attention and food consumption. With regard to cognitive load manipulations, it is still unclear whether inducing individuals to focus on the relevant cognitive load task prevents them from focusing on the food itself, the actions necessary to eat, goals and motivations regarding eating, sensations of hunger or fullness, or something else entirely. Without knowing more about which aspects of the process of eating are disrupted or disabled by these tasks, it is difficult to move this area of research forward. An important limitation of our studies, then, is that we simply do not know everything that participants may have been attending to while completing the cognitive load tasks. Future research may be able to rectify this limitation through studies involving eye tracking or perhaps asking participants to engage in a “think out loud” procedure, verbalizing their thoughts as they complete the pertinent tasks (but see Jansen et al., 1988, for difficulty validating such techniques).

Additional consideration of the impact of various features of a particular study design on the results of the study may eventually help disentangle seemingly conflicting findings. For example, Study 8 suggested that the key difference in findings was not accounted for by manipulating knowledge of the tasks to be completed. However, even this conclusion must be treated as tentative; in many studies, research participants are generally aware of how long they will participate in a study (e.g., 30 minutes v. 1 hour), and this awareness may provide another source of knowledge, such that participants may anticipate further tasks even if they are not explicitly told about them in advance.

Another limitation of the studies reported here is that we tested only for linear effects; it is possible that the relationship between attention and eating may not be linear. It has been argued, for example, that in order for eating to increase under conditions of high cognitive load, there must be sufficient cognitive resources remaining to engage in the process of food consumption (Ogden et al., 2013). Accordingly, a manipulation that requires too much cognitive load may restrict one from eating during the relevant task, whereas a more moderate amount of cognitive load could plausibly lead to increased eating. While this sort of pattern sounds theoretically plausible, eyeballing the graphs of our data does not reveal such a relationship, and only Study 2 seemingly shows a higher order pattern (and, incidentally, not a pattern that is easily explained). Our mini-meta-analyses also supported linear patterns.

This research is also limited in that BMI was measured only in one study, and two of the studies pre-selected participants for particular high or low levels of dietary restraint and disinhibition. However, ancillary analyses showed that these variables did not influence any of the primary outcomes. It must be acknowledged as well that we did not ask participants in the within-subjects studies to report their hunger and liking after each condition; nor did we ask participants to fast before the study. Such measures were implemented in order to keep participants as naive as possible with regard to our study hypotheses. Finally, we collected validity data for the n-back task only in Studies 1-3.

In contrast to the aforementioned limitations, a unique strength of our studies is that the task we used allowed us essentially to control the amount of working memory that was occupied, through exposure to a series of systematically more absorbing levels of the task. This approach represents an important advance over, and complement to, research that makes use of less well-defined cognitive load manipulations such as asking participants to watch TV, listen to news broadcast over the radio, or socialize. It is unclear precisely what sorts of processing are required by these other tasks, how much cognitive load they induce, or how one might systematically alter that amount (though such alteration may be possible, for example, by varying qualities such as the familiarity of a particular television program; Braude & Stevenson, 2014; Mathur & Stevenson, 2015). These latter tasks also introduce a variety of additional confounds to the situation. For example, watching TV might lead to enjoyment, stress reduction, feelings of nostalgia, or a variety of other phenomena that might affect food intake (Chapman et al., 2014; Mathur & Stevenson, 2015). Socializing may introduce normative pressures, self-presentational issues, or self-conscious emotions, among other confounds. In short, the n-back tasks used in our studies allowed for more precise isolation and manipulation of cognitive load.

6.1. Conclusion

Future research should aim to explain the divergent patterns of results we have obtained here, as we have surely overlooked some hypothesis, design feature, or moderator that may have played a critical role. Studies involving different types of systematic manipulations of cognitive load may be informative, as may studies that aim to assess the specific nature of participants’ thoughts, or the focus of their attention. Contemplating these different influences on behavior is particularly important given the disparate findings reported in this paper—findings that do not align, despite the use of essentially the same attentional manipulation in each study.

The factors that influence eating behavior are of great interest to scientists, clinicians, and individuals who seek to control their eating. Over eight studies, we consistently manipulated one factor, cognitive load, and found different patterns of food intake. It is not clear to what these varying patterns may be attributed, and our hypothesized explanations did not pan out. Despite the somewhat inconclusive nature of these findings, in five of the eight studies conducted here, increasing cognitive load led to decreased food consumption, a pattern in conflict with existing research using manipulations such as TV viewing, listening to a story, or socializing (e.g., Braude & Stevenson, 2014; Long et al., 2011; Ogden et al., 2013). That studies purportedly manipulating attention sometimes produce increased food consumption (e.g., Study 3 here; Braude & Stevenson, 2014) and sometimes produce decreased food consumption (e.g., Studies 4-8 here; Liguori et al., 2020) suggests that cognitive load does play some role in altering food consumption. We just do not know yet exactly what that role is.