2.1 Participants

54 participants (15 female) were recruited on Amazon Mechanical Turk, and were paid $1.80 for their participation. One participant was excluded from analysis for not being a native speaker of English.

2.2 Procedure and materials

Participants were directed to a web-based acceptability judgment task, hosted on Alex Drummond’s Ibex platform for psycholinguistic experiments.³ Participants were told that they would read sentences (about animal characters) produced by Zap, an alien learning English, and that they were to judge how these sentences sounded on a scale from 1 to 5, with 1 being completely odd and 5 being completely okay. Participants registered their ratings by clicking on the appropriately numbered box (see Figure 1 for an example of the response buttons).

Figure 1

Experiment 1: An example of a test item with the NPI any in an intervention configuration.

Participants first saw two practice trials, one involving a clearly well-formed sentence and one involving a clearly ill-formed sentence, accompanied by suggested ratings of 5 and 1, respectively. The purpose of these examples was to demonstrate to the participants that Zap was indeed capable of producing both acceptable and odd sentences. Participants then began the test phase of the experiment, the first two items of which were identical to the two practice trials. These were then followed by the 48 test trials schematically described in Table 1. We manipulated two factors, crossing NPI (present vs. absent) with the six item types listed in Table 1. The 48 trials were presented in randomized order.

As seen in Table 1, in addition to the target sentences, participants saw sentences involving DE and non-monotonic environments, as well as sentences involving different kinds of upward-entailing (UE) environments, which license inferences from sets to supersets. The UE sentences were not expected to license the NPI, and included simple UE sentences (UE-Simple), more complex UE sentences (UE-Complex), and UE sentences with a DE operator but in an irrelevant position (UE-Illusory). The particular number of repetitions of each of the non-target trial types was chosen in such a way as to balance them in terms of the overall number of degraded sentences (i.e. UE-simple, UE-complex, UE-illusory, NM[+NPI]) and non-degraded sentences (ie. DE[+NPI], all [–NPI] items).

The target sentences involved negated ditransitive constructions built on the following structural template: <Subject NP> didn’t give every <Indirect Object NP> <Direct Object NP>. For the subject and indirect object NPs, we randomly chose animals from the following list: lion, rabbit, cat, dog, giraffe, bear, monkey, elephant. The direct object NPs were randomly chosen from the following list: juice, tea, milk, ice cream, pie, cheese, cake, honey. We opted to use mass nouns, in order to avoid any confound related to the plurality of objects.

2.3 Results

2.3.2 Targets

Figure 2 presents the mean judgments across different environments from the remaining 52 participants (collapsing the three UE conditions); Figure 3 presents the same data in more detail for the critical sentences with an NPI. We are mostly interested in comparing the acceptability of the NPIs, in different environments. To do so, we computed a measure of this acceptability that factors out possible effects coming from the environment itself, independently of the NPIs. Specifically, for each participant and each environment, we calculated the log-ratio of the mean acceptability ratings for the [+NPI] targets over their [–NPI] counterparts, schematically: $\text{log}\frac{\text{mean}([+\text{NPI}])}{\text{mean}([-\text{NPI}])}$ . These log-ratios measure the degradation due to the NPI, for each participant in each environment; a comparison of log-ratios for two environments (below, through paired two-tailed t-tests) thus determines in which environment the introduction of an NPI yielded a higher degradation, i.e. which condition was judged worse.

Figure 2

Experiment 1: Average response for each sentence type: target, downward-entailing, combined upward-entailing, and non-monotonic sentence types.

Figure 3

Experiment 1: Distribution of responses (1–5) for each sentence type in the [+NPI] condition.

First, we observe, as we expect, that NPIs are rated better in DE environments than in NM environments (t(51) = 9.7, p < 0.001), and that they are more acceptable in NM environments than in each of the UE environments (UE-simple: (t(51) = 4.41, p < 0.001, UE-complex: t(51) = 3.61, p < 0.001, UE-illusory: t(51) = 2.45, p = 0.02). This is as it should be, given the known intermediate status of NPIs in NM environments (Rothschild 2006; Chemla et al. 2011; Crnič 2014).

Second, intervention effects were detected, in the sense that the critical target sentences were judged worse than the DE sentences (t(51) = 6.8, p < 0.001). Crucially, the degradation was not as strong as for plain violations in UE environments, with sentences in the target environments judged better than sentences in each of the UE [+NPI] conditions (UE-Simple (t(51) = –7.9, p < 0.001); UE-Complex (t(51) = –6.5, p < 0.001); UE-Illusory (t(51) = –5.45, p < 0.001)). Finally, the critical target sentences were also judged better than the NM sentences (t(51) = –3.39, p = 0.001).⁴⁵

2.4 Discussion

Intervention effects were found to be NPI licensing violations, as the NPIs in the critical target sentences were judged as less acceptable than those in the DE sentences. But the violations that they yielded were weaker than other licensing violations, as they were still judged better than NPIs in the plain UE or NM sentences. This previously unnoticed intermediate status of intervention sentences in fact eliminates the initial objection to the implicature theory: that implicatures are volatile while intervention effects are not. This judgment pattern of intervention effects is thus interestingly compatible with the SI theory; intervention configurations may create NPI violations (say, because they create NM environments, see footnote 1), but this violation should only occur when the implicature is derived. Testing the dependence of the violation on the presence of the implicature is the goal of the following experiments.

3.1 Participants

54 participants (24 female) were recruited on Amazon Mechanical Turk and were paid $1.80 for their participation. Two participants were excluded from analysis for not being native speakers of English.

3.2 Procedure and materials

Experiment 2 involved two tasks: a Picture Selection Task (Roeper 2007) was used to estimate the participants’ rates of implicature derivation, and an Acceptability Judgment Task was used to estimate the strength of the intervention effects. The order of the two tasks was counterbalanced across participants.

3.2.2 Picture Selection Task

In the Picture Selection Task of Experiment 2, participants saw two pictures on each trial with a very short introduction sentence at the top of the screen that provided a setting for the story, and read a sentence that they were told had been produced by a puppet named Raffie.⁷ Participants were instructed to evaluate Raffie’s sentence with respect to the pictures on the screen. They were told that Raffie’s sentence would sometimes be applicable to only one of the two pictures (Picture 1 on the left or Picture 2 on the right), and sometimes to both of the pictures. For each test sentence, participants were asked to decide from three response options: Picture 1, Picture 2, Both Pictures. An example item can be seen in Figure 4.

Figure 4

Experiment 2: An example of a Picture Selection Task trial.

The Picture Selection Task consisted of 31 items. Participants were first presented with three examples. Raffie’s sentence in the first example clearly matched Picture 1. Raffie’s sentence in the second example clearly matched Picture 2. Raffie’s sentence in the third example was harder to judge, and participants were told that some people would say it matched only one of the pictures, while others would say it matched both pictures, and they should simply follow their intuitions while doing the task. The three examples were then presented again as the first three items of the experiment. The remaining 28 items consisted of eight target items and 20 controls, presented in randomized order.

The target items in the Picture Selection Task were exactly parallel to those in the Acceptability Judgment Task, the only difference being that there was no NPI in the target sentences in the Picture Selection Task. An example target item from the Picture Selection Task is provided in Figure 4. In this example, the target sentence corresponds to (11a). Its logically stronger scalar alternative is in (11b). The scalar implicature of (11a) is thus the negation of (11b), i.e. the resulting interpretation of (11a) enriched with its scalar implicature is (11c).

(11)	a.	Sentence: Monkey didn’t give every rabbit juice.
	b.	Alternative: Monkey didn’t give any rabbits juice.
	c.	Overall meaning with SI: Monkey gave some but not every rabbit juice.

In Picture 1, Monkey gave two of the four rabbits juice, and in Picture 2 Monkey dropped all of the juiceboxes and none of the rabbits got juice. Now, if the participant derived the implicature, they would understand the sentence as (11c) (by the negation of the logically stronger alternative Monkey didn’t give any rabbits juice). They would thus opt for the response “Picture 1”. We will refer to this type of response as a some-response (as some of the animals in the picture received juice). On the other hand, if the participant didn’t derive the implicature, they would interpret the sentence literally, true both in the situation in which Monkey gave only some of the rabbits juice (Picture 1) and in the situation in which he gave none of them juice (Picture 2); in this case, the participant would opt for the response “Both pictures”.⁸ Based on their response on target items, we can thus evaluate whether the participant interpreted the sentence with or without the implicature.

In addition to the targets, participants also received 20 control sentences. Four had the same construction as the target items, but their meaning was compatible with only one of the pictures. An example of this type of control is provided in (12a), which was paired with a picture in which two out of four elephants got cheese, and a picture in which four out of four elephants got cheese. Participants also saw four positive ditransitive control sentences, as in (12b), four negative ditransitive sentences with the universal quantifier in direct object position, as in (12c), and four negative ditransitive sentences with a definite noun phrase in indirect object position and a mass noun in direct object position, as in (12d). For half of the controls, the correct response was “Picture 1”, and for the other half the target response was “Picture 2”. Finally, there were four sentences like (12e), which were compatible with both of the pictures they were presented with.

(12)	a.	Dog didn’t give every elephant cheese.
	b.	Elephant gave every dog cake.
	c.	Giraffe didn’t give Lion every strawberry.
	d.	Rabbit didn’t give Cat tea.
	e.	Giraffe got every flower.

3.3 Results

3.3.4 Correlation

The main goal of Experiment 2 was to assess whether there was a correlation between implicature derivation and intervention effects.¹⁰ First, we normalized each participant’s responses on target items in the Acceptability Judgment Task by scaling each target response within the particular participant’s extreme judgments of the good and bad controls, as in: $\frac{\text{target}-{\text{mean}}_p(\text{bad})}{{\text{mean}}_p(\text{good})-{\text{mean}}_p(\text{bad})}$ , where mean_p represents the mean of responses in a particular condition for a given participant p. This normalization corrects for different uses of the response scale by matching extreme values across participants. In Figure 5 below, each participant is represented at the height corresponding to the mean of these normalized responses for that given participant $\frac{{\text{mean}}_p(\text{target})-{\text{mean}}_p(\text{bad})}{{\text{mean}}_p(\text{good})-{\text{mean}}_p(\text{bad})}$ . This is a value between 0 (intervention effects are as bad as bad controls) and 1 (intervention effects are as good as good controls), since we excluded participants who did not satisfy mean_p (good) ≥ mean_p (target) ≥ mean_p (bad). To assess the correlation between the strength of intervention effects so represented and implicature derivation, we computed an implicature index for each participant: the proportion of SI+ responses to targets in the Picture Selection Task by that given participant. This implicature index is represented on the x-axis of Figure 5.

Figure 5

Results from Experiment 2: Individuals as a function of responses to implicature and intervention targets.

The first thing to note about this graph is that there is a wide range of perceived strength of intervention effects (participants are distributed all along the range of the y-axis), which replicates the finding from Experiment 1 that there is a lot of variation in the perception of intervention effects. The implicature index also shows variability, but in the form of bimodality (participants fall either to the far left or to the far right of the graph).

Turning to the correlation, mixed effects linear regression models were fitted to the normalized responses to the Acceptability Judgment Task targets, with mean response to the implicature targets as a fixed effect, and random by-participant intercepts. Comparisons with the models containing only the random by-participant intercepts revealed no significant effect of implicature response on the intervention judgments (χ² (1) = 0.35, p = .55). Experiment 2 thus reveals no correlation between an individual tendency to derive implicatures and individual perception of intervention effects.¹¹

3.4 Discussion

In this experiment, we measured participants’ propensity to derive SIs and their sensitivity to intervention effects. While we observed variability in both, there was no observed correlation between the two indices.

4.1 Participants

55 participants (27 female) were recruited through Amazon Mechanical Turk and were paid $1.80 for their participation. All participants reported English as their native language.

4.2 Procedure and materials

The procedure was the same as in Experiment 2, comprising two tasks. In addition to the Acceptability Judgment Task, participants completed a Covered Picture Task. In this task, they saw two pictures on the screen, one visible and the other covered, and a short introduction sentence at the top of the screen to introduce the story, followed by a target sentence produced by a puppet called Raffie. The participants were instructed to evaluate the visible picture with respect to Raffie’s sentence. They were told that Raffie’s sentence could apply to only one of the pictures — either the visible or the covered one. The participants had to click on the visible picture if they thought that Raffie’s sentence was describing this picture, and on the covered picture if they thought that Raffie’s sentence was describing a different picture from the visible one. An example of a target item is provided in Figure 6.

Figure 6

Experiment 3: An example of a Covered Picture Task trial.

The visible picture that appeared with the target items was one that was incompatible with the derivation of the implicature. In the visible picture in Figure 6, we can see that none of the rabbits got juice. The participant would thus opt for the visible picture if they had access to the implicature-less reading, and for the covered picture if they expected, based on the target sentence, that at least some of the rabbits got juice, which is the reading of the target sentence enriched with an implicature, as in (11c). As in the Picture Selection Task of Experiment 2, we can evaluate based on responses to the target items whether the participant has interpreted the sentence with or without the implicature.

The Covered Picture Task contained 29 items. Participants were first presented with three examples. The first one involved a correct description of the visible picture; the immediate feedback to participants stated that in that case most participants would select the visible picture. The second example involved an incorrect description of the visible picture; here the feedback to participants was that most participants would select the covered picture. The third example was harder to judge; for this item, the participants were told that some people would select the visible picture and others the covered picture, and that they should follow their intuitions while doing the task. These three examples were then presented again as the first three items of the experiment. The remaining 26 items consisted of eight target items and 18 controls, presented in randomized order.

Among the 18 control items, 10 were compatible with the visible picture, and eight were not. They were very similar to the controls in the Picture Selection Task of Experiment 2 (but adapted for the Covered Picture Task).

4.3 Results

4.3.4 Correlation

As in Experiment 2, we calculated each participant’s intervention index, represented on the y-axis of Figure 7, and the proportion of implicature derivation, represented on the x-axis.¹³ The closer the value on the x-axis to 1, the more implicatures the participant derived in the course of the Covered Picture Task; the closer the value on the y-axis to 1, the greater the judgments provided by the participant in response to the intervention targets in the Acceptability Judgment Task.

Figure 7

Results from Experiment 3: Individuals as a function of responses to implicature and intervention targets.

A mixed effects linear regression model was fitted to the normalized¹⁴ responses to the Acceptability Judgment Task targets, with mean response to the implicature targets as a fixed effect and random by-participant intercepts. A comparison with the model containing only the random by-participant intercepts revealed no significant effect of covered picture implicature responses on the intervention judgments (χ² (1) = 1.18, p = .28).

4.4 Discussion

The results of Experiment 3 replicate those of Experiment 2: whether one uses a Picture Selection Task or a Covered Picture Task to estimate participants’ rates of implicature derivation, we find that this estimate does not predict participants’ acceptability judgments of intervention sentences.

5.1 Participants

54 participants (25 female) were recruited through Amazon Mechanical Turk and were paid $1.80 for their participation. All participants reported English as their native language.

5.2 Procedure and materials

The experiment was divided into two parts. The first task was designed to train half of the participants to derive SIs and the other half to not derive SIs. In the second part, participants received an Acceptability Judgment Task (designed to measure intervention effects), as in Experiments 2 and 3.

In the implicature training task, participants completed a Truth Value Judgment Task (TVJT) (Crain & Thornton 1998) (see Figure 8) in which they received immediate feedback after each trial. Participants saw a series of pictures with a short sentence at the top of the screen that introduced the story, followed by a target sentence produced by Raffie. Participants were instructed to evaluate the picture with respect to Raffie’s sentence: they had to decide whether the descriptions were right or wrong. The response options were binary: the participants could either choose a “yes” or a “no” response.

Figure 8

Experiment 4: An example of an implicature training TVJT trial.

If the participant derived the implicature, they would consider Raffie’s sentence false in a situation in which none of the lions got ice cream (Figure 8), and would therefore select the “no” response. On the other hand, if the participant didn’t derive the implicature, they would consider Raffie’s sentence to be an acceptable description of the picture, and would select the “yes” response. What distinguished the group that was trained to derive implicatures (SI+ training group) from the group that was trained not to derive implicatures (SI- training group) was the feedback after the participant had provided a response on each of the target items. For example, for the experimental item in Figure 8, participants in the SI- training group received the feedback in (13), and participants in the SI+ training group received the feedback in (14).

The task consisted of 28 items. Participants first saw two example items, the first a clearly true target and the second a clearly false target. For both example items feedback was provided about which response was correct. The example items were then presented again as the first two items of the implicature training TVJT. The remaining 26 items consisted of eight target items and 18 control items, presented in randomized order.

The target sentences on the implicature training TVJT were the same as those from Experiments 2 and 3. The pictures that accompanied these sentences were incompatible with the derivation of implicatures. The control items in this experiment were similar to those in Experiments 2 and 3. Of the 18 control items, 10 items had a clear “yes” target, and eight items had a clear “no” target. The main difference between these controls and those of the previous experiments was that the control trials in this experiment contained feedback, so as to be parallel with the target trials.

5.3 Results

5.3.2 Training results

In the implicature training TVJT, the proportion of SI+ responses from the group of participants that received the SI+ training was 0.58 (SD = 0.25), while the group that received the SI- training gave no SI+ responses. This suggests that (i) the implicature-less response is salient from the start but (ii) the training is effective. To see this in more detail, Figure 9 displays the proportion of SI+ responses to the critical cases, arranged by the order in which they were presented. This graph allows us to visualize the effect of the SI+ training: all of the participants started off without implicatures, but they became more and more likely to derive the implicature with each subsequent SI+ training target.¹⁵

Figure 9

Experiment 4: Proportion of SI+ responses to implicature targets for each training group, by order of target appearance.

5.4 Discussion

The training administered in Experiment 4 succeeded in leading participants to either derive implicatures or to not derive implicatures. However, this training effect did not subsequently affect their acceptability judgments of the intervention targets. While the SI theory might have led us to expect that the SI+ group would judge the intervention targets worse than the SI- group, this prediction was not borne out.

6.1 Participants

52 participants (24 female) were recruited through Amazon Mechanical Turk and were paid $1.80 for their participation. All participants reported English as their native language.

6.2 Procedure and materials

The experiment consisted of a single task, which was equivalent to the Picture Selection Task from Experiment 2. Participants were randomly assigned to one of two groups, which differed only in the target sentences they saw. The “regular” group saw the standard 8 [–NPI] target sentences from Experiment 2 (Appendix B.2, (6)). These sentences did not contain an intervention configuration, and therefore there was no need for any repair strategies. The “intervention” group saw almost the same 8 target sentences, except that these contained the NPI any in an intervention configuration. If participants in the intervention group detected an intervention effect, they were expected to repair the sentence in order to carry out the task. For example, the intervention group might see a target like that in Figure 10, while the regular group would see a target like Figure 4 from Experiment 2 (which is identical except that it does not include the NPI any).

Figure 10

Experiment 5: An example of a target item administered to the intervention group.

In addition to the critical targets, both groups received the same 20 control items from the Picture Selection Task of Experiment 2 (Appendix B.2, (7)).

The intervention group (but not the regular group) was also asked to fill out a small questionnaire at the end of the experiment. In this questionnaire, the participants were asked to motivate their response choice on the target items: this questionnaire was there to make sure that participants had actually assigned a meaning to the target sentences, and had not simply responded at random.

6.3 Repair strategies

Before discussing the results, let us consider some possible repair strategies for a sentence like (15).

One possibility would be to simply drop the NPI, and interpret the sentence as in (16a). Another possibility would be to reinterpret the universal quantifier as a definite description, as in (16b). Yet a third possibility would be to assign wide scope to the universally quantified noun phrase, above negation, as in (16c). The question of interest is whether there is a fourth possibility: to interpret the sentence as it is, without the implicature, as in (16d).

What would be the expected responses on the Picture Selection Task under each of the above repair strategies? Table 2 summarizes possible repair strategies for (15) and their predictions, which we will now spell out. We will continue to refer to selections of pictures in which no animals got the object in question (e.g., a Picture 1 response in Figure 10) as none-responses,¹⁷ and to selections of pictures in which some animals got the object in question as some-responses (e.g., a Picture 2 response in Figure 10).

Participants who adopt the strategy that leads to the interpretation in (16a), i.e. dropping the NPI, might respond in one of two ways, depending on whether they compute an indirect scalar implicature from (16a). If they compute the implicature Bear gave some of the lions ice cream, they are expected to give a some-response. If they do not compute an implicature from (16a), they are expected to select “Both pictures”, since the implicature-less reading of (16a) is compatible with both pictures. Notice that if dropping the NPI is the strategy adopted by the intervention group, we may expect to see parallel performance in the two groups, since the reanalyzed (16a) is identical to the target sentences that the regular group is presented with.

Participants who employ the strategy of reinterpreting the universal quantifier as a definite description, as in (16b), or of giving the universal wide scope, as in (16c), would be expected to give a none-response. If the intervention group makes use of either of these two strategies, we expect to see a greater proportion of none-responses in the intervention group than in the regular group.

Finally, participants who employ the strategy of implicature suppression as in (16d) are expected to select “Both pictures”. Since none of the above repair strategies lead to an increase in “Both pictures” responses, an increase of such responses in the intervention group could be interpreted as the propensity to employ this strategy (to be discussed further below).

6.4 Results

6.4.2 Targets

Figure 11 displays the counts of the different response types for the implicature targets from the intervention group and the regular group. Recall that a “Both pictures” response corresponds to a non-implicature interpretation of the target sentence. To determine whether group (intervention vs. regular) had an effect on the type of response to the implicature targets, we recoded the responses in binary terms (as “Both pictures” and “Other”). We then fitted a mixed effects logistic regression model on these responses, with group as a fixed effect and random by-participant intercepts. A comparison of this model with a reduced model without the group fixed effect revealed a significant effect of group on response type (χ² (1) = 6.05, p = .01), with more “Both pictures” (non-implicature) responses in the intervention group than in the regular group.

Figure 11

Experiment 5: Distribution of responses to implicature targets from the intervention and the regular group.

6.5 Discussion

One interpretation of these results is that they are consistent with the SI theory: if scalar implicatures cause intervention effects, it is natural to expect that when faced with an NPI in an intervention configuration, participants may choose to repair the sentence by cancelling the derivation of the scalar implicature.

However, there are two alternative explanations for this effect. First, note that there is evidence that at least some of our participants repaired the target sentences by replacing the universal quantifier with a definite description or by means of wide-scoping the universal quantifier: this is suggested by the increase in the amount of none-responses in the intervention group compared to the regular group. As these strategies seem to be available to some of our participants, the participants who selected “Both pictures” could have entertained not one but two possible repair strategies simultaneously, namely, one of the strategies leading to a none-response, in combination with the strategy of dropping the NPI. Such participants might simply not have known which of the two interpretations should be chosen. For example, such a participant might think that it is as probable that Raffie wanted to use an indefinite instead of the NPI, and derive the implicature (which would make the sentence compatible with Picture 2 of Figure 10), as it is that the universal quantifier should be assigned wide scope with respect to negation (which would make Raffie’s sentence compatible with Picture 1 of Figure 10). The participants could therefore be unsure which of the two meanings Raffie intended to convey, which is why they might opt for “Both pictures” more often in the intervention group than in the regular group.

An argument against this explanation comes from the questionnaire that the intervention group was asked to complete at the end of the experiment. The explanations that participants gave for their picture selections can be found in Appendix B.5.2. These explanations confirm that for the most part, participants were not merely responding at random to the target items, but rather repaired the sentences to assign some meaning to them. The crucial question for us was whether participants who selected “Both pictures” might report confusion or uncertainty about the intended meaning of the sentence, and whether this uncertainty could have driven their picture selection. Upon close inspection, however, few responses were indicative of such confusion, with only one of the 10 “Both pictures” justifications referring to an ambiguity (i.e. The sentence is a bit ambiguous and could either mean that the elephant didn’t give any tea to the giraffes, or that he didn’t give tea to all the giraffes). The questionnaire responses therefore speak against an alternative explanation of the data according to which people were unsure about the intended meaning of the sentence due to the existence of more than one repair strategy.

While we do not find the multiple repair strategy explanation very strong in light of the questionnaire data, it is worth noting that we cannot at present completely rule it out either. Even when participants’ explanations for “Both pictures” selections are compatible with the non-implicature reading (e.g., In both pictures not every giraffe has tea, so both are correct), we cannot definitively know whether they chose “Both pictures” because they did not derive the implicature, or because there were multiple repair strategies that would render the sentence true as soon as not all of the giraffes got tea.

A second alternative explanation is that participants in the repair group were under more cognitive load, which is why they derived fewer implicatures. Previous experimental work has shown that in a dual task setting people tend to prefer readings without scalar implicatures (De Neys & Schaeken 2007; Marty & Chemla 2013). Facing a broken sentence (with an intervention effect) and perhaps trying to interpret it could require additional cognitive load leading to fewer implicatures, not because implicature suppression is a possible repair strategy, but because participants are expending effort to repair the sentence.

Overall, one might have considered this experiment to hold the greatest chance of yielding an observable correlation between intervention effects and SIs, because it allowed both to be measured at a single point in time. Compared to the other experiments, this one yields the results most compatible with a positive conclusion in favor of a correlation, although there remain alternative explanations that require further investigation.