5.1 Participants

We recruited 1344 participants in total through Prolific: 112 for each of the twelve experiments. They were paid $2.75 USD for their participation. To identify native speakers of US English or European Spanish, we used three tools. First, we used Prolific’s prescreening tools to identify individuals who considered the target language their first language and had mostly lived in either the US or Spain before turning 18. Second, we asked all participants questions about where they lived from birth until age 13, and about the languages that were spoken in their home as children. Finally, we included two trials within the experiment where we asked participants to read a short description of an ethically challenging situation and write at least one complete sentence in the target language about how they would respond. These open-ended questions were included to identify both bots and potentially uncooperative or non-native participants (Chmielewski & Kucker 2020; Dennis et al. 2020). We excluded participants from analysis if they reported not living in the target country, not speaking the target language (or speaking it as a non-dominant language) in their home, or if their responses to the morality trials appeared uncooperative or non-native to us, e.g., “mostly likely let my friend known about see he/she” (only six participants across the twelve experiments were excluded based on their responses to the morality trials). We also removed participants from analysis if they responded to four or more fillers with a rating that was more than 2 standard deviations away from the mean rating for that filler. The final sample sizes of each experiment are shown in Table 3.

5.2 Materials

The materials for both languages followed the 2 × 2 factorial design for island effects described in Section 2, where island effects are identified statistically as a ‘superadditive’ Position × Structure interaction and measured by means of DD scores as in (5). Example (7) illustrates the 2 × 2 design for Spanish, with the alternative verbs and wh-phrases separated by slashes. The English design is illustrated in the translations.

For each language, we created sixteen item sets based on the four conditions. These sixteen sets were used in all experiments (with the island type and verb modifications). The English and Spanish sentences were translation-equivalents and as lexically-matched as possible. We did have to make several choices to ensure a clear test of island effects. First, the embedded subject in English was always an overt pronoun (either they or you), but a null pronoun in Spanish. We chose null pronouns in Spanish because we perceived them as more natural than overt pronouns in our sentences, and because their position relative to the verb is not observable, meaning that participants can posit it as preverbal or postverbal, as they prefer. We could have alternatively used an overt determiner phrase subject in a fixed position, but we decided not to do so because Torrego (1984) observed that there may be independent factors that influence the acceptability of the subject-verb vs verb-subject orderings, which in turn may be confounded with, or even interact with, island effects. While this is an interesting question in its own right, we abstract away from it here, allowing participants to posit the most acceptable subject position in their grammar, giving us a test of island effects alone. Second, in the know experiments, we used want to know/querer saber (rather than simply know) because want to know seems to highlight the interrogative nature of the complement wh-clause, similar to ask. Third, in the embedded clause, we used verbs that we perceived to be transitively-biased to facilitate the interpretation of the wh-word as an embedded object. Finally, we presented embedded verbs in the conditional tense (e.g., would reject) to make a modifier interpretation of the wh-questions (particularly cuándo/when-islands) less likely (such an interpretation would make them an adjunct island rather than a wh-island). The full set of materials is available at the Open Science Framework Repository.⁴

In addition to the 16 experimental items, we selected 32 pre-tested fillers and seven pre-tested burn-in items per language that were distributed equally across the full range of acceptability from previous studies. The fillers and burn-in items were sentences from the theoretical syntactic literature – or modelled after its examples – that instanced different degrees of acceptability according to formal studies. For English, we took the items from Sprouse et al. (2013), who tested sentences on a 7-point scale and computed the most frequent score for each item. Based on that measure, we selected a set of sentences that evenly represented all seven ratings: there was one burn-in item and four or five fillers by rating. For Spanish, we took the sentences from Ortega-Santos (2020), who tested the acceptability of a number of sentences in Chilean, Venezuelan and Puerto Rican Spanish and provides the mean z-scores for each item. Given that our participants were speakers of yet another variety, we arbitrarily selected the Chilean variety as a reference and picked a set of sentences with a widespread distribution of z-scores: for the fillers, the range was –1.449 to 1.097 and the average was 0.108; for the burn-in items, the range was –1.545 to 1.108 and the average was 0.056.

5.3 Presentation

Items were distributed across four lists using a Latin Square procedure, such that participants rated each of the four conditions four times, and each time the item was from a unique item set (no repetitions of lexical items). Participants first saw instructions with three example items explicitly given ratings of 1, 4, and 7 to demonstrate the task. They then rated items themselves. The first seven items spanned the full range of acceptability based on our expectations, and they were presented in the same (pseudorandom) order for each participant. These items were not analyzed, as they were burn-in items presented to make sure that the participants saw the full range of possible acceptability ratings prior to judging any trials that we would analyze. The next two items were two of the filler items (a very low and very high rating). The rest of the experiment contained the 16 experimental items for that particular list and 30 remaining fillers in a pseudorandom order such that there was at least one filler between two experimental items. The experiments were run using Qualtrics (Provo, Utah). It took participants about 10 minutes to complete an experiment.

6.1 The presence of island effects

Our first question is whether each of the six island effects is present in the two languages. Figure 3 shows the interaction plots for each wh-island under each verb in the two languages, along with the differences-in-differences or DD scores, an estimate of the interaction term that we calculated as in (5). The mean z-scores by condition are shown in Table 4. To assess the presence of island effects statistically, we constructed linear mixed effects models crossing Structure × Position for each of the twelve island effects, with the maximal random effect structure that did not result in convergence failure. The presence of an island effect would show up as a significant Structure × Position interaction. Figure 3 also shows the p-values and Bayes factors for the interaction term in these models. The full results are shown in Table 5 for English and Table 6 for Spanish.

In all cases, the island/embedded condition was much less acceptable than the other three conditions, resulting in the visual patterns indicative of superadditive Structure × Position interactions, which were confirmed statistically (all p < .001 and all BFs > 100). This suggests that all six island effects are present in both languages.

6.2 The size of island effects

Though there does not appear to be any cross-linguistic variation in the presence of island effects (a binary question), there could still be variation in the size of the island effects (a gradient question). For example, wh-island effects could be present but smaller in Spanish. To test this possibility, we constructed linear mixed effects models for each of the six island types, but crucially combined the results of the two languages. These models crossed Structure × Position × Language and included the maximal random effect structure that converged. A difference in island effect size between the two languages would show up as a significant three-way interaction. A summary of the three-way interaction effects is shown in Table 7. The full results of the models are available as “Additional Results” at the Open Science Framework Repository. The p-values and Bayes factors of the critical three-way interaction term are also shown in Figure 3 above the arrows.

Figure 3: Interaction plots for whether, why and when islands under both know and ask in English and Spanish. For each island in each language, we show the differences-in-differences (DD) score, the p-value of the superadditive Structure × Position interaction indicating an island effect and the Bayes Factor (BF). Above the arrows, we also show the p-value and the BF for the three-way Structure × Position × Language interactions, which assess cross-linguistic differences in size (see Section 6.2).

We find a statistically significant size difference for five out of the six island types by both null hypothesis testing (p < .05) and Bayes factor (BF > 3): know whether, know why, ask why, know when, and ask when. From these, all BFs are robust to prior widths except for know why, where the BF is inconclusive with a wide (BF = 1.65) and an ultrawide prior (BF = 1.56). As for ask whether, it is not significant by p-value (p = .308) and shows evidence against a size difference by BF (BF = .12). It is important to note that the direction of the size differences we observed is opposite to the one that might be expected under a weaker version of the original cross-linguistic claim – Spanish island effects are in fact larger than English island effects by about .3 z-units on average. This suggests that while there is evidence of a gradient form of cross-linguistic variation for four or five out of six island types, it is not in line with the original observation.

6.3 Individual variation

Our third question is whether there is more individual variation in Spanish than English. We ask this because there can be individual variation, representing different idiolects, that is obscured by focusing on sample means. Idiolectal variation could explain the apparent discrepancy between our results and Torrego’s (1984) observations. In other words, even if, overall, we find island effects, some speakers of Spanish may manifest no or smaller island effects while others show large island effects (i.e., between-participant variation). It is also possible that some speakers of Spanish show more variability within their own judgments than English speakers (i.e., within-participant variation), suggesting that they may have two grammars at their disposal. To be clear, we expect some amount of variation both between- and within-participants. That is the nature of behavioral studies (they are inherently noisy). The critical question is whether Spanish shows more variation than English in one or both of these ways.

To look for between-participant variation, we plot two sets of distributions: the island effect sizes calculated as by-participant DD scores (Figure 4), and the by-participant z-score means of the island/embedded conditions (Figure 5; the non-island/embedded conditions are also shown as a control; see Kush et al. 2018; 2019; Bondevik et al. 2021 for similar analyses). We can look for a visual pattern suggesting two or more populations of speakers, which would appear as a bimodal (or multimodal) distribution. We see no obvious sign of bimodality in the Spanish experiments, either in the DD scores or in the island/embedded conditions. Some signs of bimodality are instead observable in the English experiments (e.g., the ask whether DD scores and the ask when DD scores and island/embedded conditions).

Figure 4: Distribution of by-participant DD scores in the English and Spanish experiments. The dashed vertical line marks the limit between DD scores higher than 0 (indicative of an island effect) and DD scores lower than 0 (indicative of no island effect).

Figure 5: Distribution of by-participant z-score means in the island/embedded and non-island/embedded conditions of the English and Spanish experiments.

We tested for bimodality statistically using the multimode package (Ameijeiras-Alonso et al. 2021) in R (R Core Team 2023). We chose three tests that instantiate the most common approaches to identifying multimodality: the dip test (Hartigan & Hartigan 1985), the excess mass test (Müller & Sawitzki 1991; Cheng & Hall 1998; Ameijeiras-Alonso et al. 2019), and the bandwidth test (Silverman 1981; Hall & York 2001). What we looked for is a statistically significant effect of multimodality in Spanish but not in English in either the DD scores or the individual condition z-scores. The full list of p-values for each of these tests is shown in Table 8. Crucially, there are no statistically significant effects for any of the Spanish islands under any of the tests, corroborating the visual inspection of Figure 4 and Figure 5, and suggesting that there is no evidence of two or more populations of speakers of Spanish in our studies. In English, there are significant effects in the ask whether DD scores and in the know whether and ask when island/embedded condition, both under the excess mass test (a significant effect is also observed in the English ask when non-island/embedded condition). If anything, these results suggest that English is more variable than Spanish, contrary to the hypothesis we set out to explore.

Within-participant variation is displayed in Figures 6 and 7. Figure 6 plots each of the four ratings of the island/embedded condition for each participant: each horizontal line represents a participant; their individual ratings are represented by the colored dots and their mean rating is represented by the black circles. We categorized the four ratings of the island/embedded condition given by each participant as either below or above their mid-point of the scale (0). Then, we counted how many participants rated all four tokens below 0, how many rated three out of four below 0, etc. The end result is counts for five categories of speakers for each of the two languages, which we then converted to proportions of the total sample for each experiment. These proportions are shown as a stacked bar plot in Figure 7. Visual inspection of the proportions in Figure 7 suggests that the two languages show similar amounts of within-participant variability.

Figure 6: Distribution of by-participant z-scores in the island/embedded conditions of all the experiments. Each horizontal line represents a participant; the dots on each line indicate the location of each of the participant’s observations on the z-score scale, and the black circles show the participants’ mean (for each experiment, participants are ordered by their mean z-score).

Figure 7: Proportion of participants that rated zero, one, two, three or four out of four tokens below their midpoint of the z-score scale (0) in each of the experiments.

We tested whether the counts of the five categories differed statistically between the two languages by using Fisher’s exact test (on a 2 × 5 contingency table crossing language and the five count categories). We found that only ask when yielded a significant effect (p = .028), while the other five island types were non-significant (know when: p = .415; ask whether: p = .429; know whether: p = .922; ask why: p = .477; know why: p = .282). As we can see in Figure 7, the significant difference for ask when appears to be driven by English showing more within-participant variability (fewer participants with 4/4 ratings below 0) than Spanish. This runs contrary to the hypothesis that perhaps Spanish shows more within-participant variability.

Taken together, the results of these two analyses suggest that there is no more between- or within-participant variation in Spanish than in English, at least for the participants in our experiments.

6.4 Post-hoc analyses

A reviewer helpfully asked whether our data set could be explored for additional facts about island effects, in particular (i) whether there are differences in effect size between the island types within a language, and (ii) whether there are differences in individual variation between island types within a language. Though these questions are not part of our theoretical exploration (we are focused on cross-language, not within-language, differences), we are happy to provide these analyses for researchers who may be working on these questions.

To assess differences in effect size between island types, we constructed 2 × 2 × 2 linear mixed effects models for each possible pair of island types (within each verb), and we calculated the p-value of the three-way interaction term as an indicator of a significant difference in the size of the island effects. Table 9 reports these results for the twelve possible comparisons. Because this is a post-hoc exploration, we present the uncorrected p-values and three possible corrections: Dunn correction of every term in the models (81 comparisons; the most conservative correction); Dunn correction of the interaction terms alone (twelve comparisons; a less conservative correction); and false discovery rate (Benjamini and Hochberg 2000) of the interaction terms (twelve comparisons; the least conservative correction). We use the Neyman-Pearson asterisk to indicate when the correction would yield significance at an alpha criterion of .05.

We know of no theory that predicts differences in effect sizes between island types (or, differences in effect sizes more generally), so we leave this as information for other researchers to use if they are pursuing such a theory.

To assess differences in individual variation between island types within a language, we calculated Fisher’s exact tests on 2 × 5 contingency tables comparing each pair of island types (within language and verb) and the number of tokens that were rated below the midpoint of the scale (zero through four). Table 10 reports these results for the twelve possible comparisons. Because this is a post-hoc exploration, we present the uncorrected p-values and two possible corrections: Dunn correction of the twelve comparisons (the more conservative correction), and false discovery rate (Benjamini and Hochberg 2000) of the twelve comparisons (the less conservative correction).

Again, we know of no theory that predicts differences in the amount of variation between island types, so these results are presented for informational purposes only. We also note that the categorization that we use (above/below the midpoint) means that island types with higher ratings for the island/embedded condition are more likely to show variability. This likely explains why whether islands tend to show more variation than the other island types. So these results should be interpreted with caution. (This is less of a concern for the cross-linguistic comparison because we were comparing the same island type, which under default assumptions, should show similar absolute ratings.)

7.1 The empirical facts of wh-island effects in Spanish and English

In a series of twelve acceptability judgment experiments, we examined three types of wh-island effects (whether, why and when island effects) under two embedding verbs (know and ask) in English and Spanish translation-matched sentences with wh-dependencies. The first empirical question we sought to address was whether wh-extraction showed island effects in these contexts not only in English but also in Spanish. The answer to this question appears to be yes – Spanish shows clear wh-island effects in all three wh-clauses under the conditions in which these effects were also observed in English.

Our findings replicate previous work showing that there are island effects in both English (Sprouse 2007; Sprouse et al. 2011; 2012; 2016; Almeida 2014; Michel 2014; Aldosari 2015; Ortega-Santos et al. 2018; Pham et al. 2020) and Spanish (López-Sancio 2015; Ortega-Santos et al. 2018; Pañeda et al. 2020; Rodríguez & Goodall 2020; Stigliano & Xiang 2021; Pañeda & Kush 2022), but they also extend it: previous studies usually tested fewer types of wh-islands and fewer verbs, they had a smaller sample size and, crucially, they generally examined the two languages separately, with different materials and, sometimes, under different conditions that hamper any cross-linguistic comparisons. In this context, our study significantly broadens our knowledge about how wh-islands compare in English and Spanish.

The second empirical question we sought to address was whether there is a gradient difference in effect size between wh-islands in English and Spanish. One possibility we considered was that the locus of cross-linguistic variation lies in effect size rather than the categorical presence/absence of island effects, perhaps with Spanish showing smaller island effects than English. What we found is that there is a difference in effect size for four or five out of the six island types that we tested, but it goes in an unexpected direction – Spanish island effects tend to be slightly larger than English island effects (with translation-matched materials). Thus, not only are there wh-island effects in Spanish, but they also do not appear to be mild, in contrast to what previous work has suggested (Pañeda et al. 2020; Pañeda & Kush 2022). We note that the experimental findings supporting this possibility were obtained under different conditions that may impact effect sizes. For instance, Pañeda & Kush (2022) obtained very small whether island effects (DD = 0.22, 0.38), but their sentences did not contain bare extractees, like ours, but rather complex or “d-linked” extractees, which are known to reduce wh-island effects (Pesetsky 1987; Goodall 2015; Atkinson et al. 2016; Villata et al. 2016).⁵ Similarly, Pañeda et al.’s (2020) island effect sizes were based on binary judgments obtained in a speeded task, in contrast to our effect sizes, which were based on 7-point scale judgments obtained in an untimed task. Because of these differences, our study is not directly comparable to those previous ones.

An anonymous reviewer asks if we can propose a theory for the surprising observation that island effects are larger in Spanish than English. While the experimental literature has cataloged a number of results showing minor variation in the island effect sizes across languages (Sprouse & Villata 2021 for a review), we know of no theories of differences in effect sizes for island effects (or any other acceptability judgment effect) that we can use to interpret the contrast. While there are frameworks that can capture differences in effect sizes (e.g., Keller’s 2000 Linear Optimality Theory or Featherston’s 2005 Decathlon model), our understanding is that these frameworks do not currently predict those differences, but are rather at an earlier stage: namely, they are setting the constraint values (based on judgment studies) necessary to make such predictions in the future. We are therefore reluctant to propose an entirely new theory based on this surprising result.

That said, we can rule out some possible explanations. For example, one possibility would be that island effects are equal in size in the two languages, but the English island/embedded condition hit the floor of the scale, yielding an underestimation of the effect size. However, the English island/embedded conditions are all at or less than –1 z-scores on average, while some of the English fillers are rated on average as low as –1.5. This suggest that the island/embedded conditions did not hit the floor, and therefore that the English effect is not underestimated.

Another possibility is that the Spanish violation conditions were perceived as less acceptable than their English counterparts due to differences in the acceptability of the fillers that we used for each language, which were not translation-matched. Filler sentences can act as a benchmark against which the experimental conditions are evaluated (Cowart 1997). Thus, if the fillers were less acceptable in English than in Spanish, this might have caused the island/embedded conditions to be perceived as more acceptable in English than Spanish. We do not think this was the case, though, because, impressionistically, the fillers seem to span a similar range of acceptability in both languages (Figure 2).

Finally, a reviewer wonders whether the larger island effects in Spanish could be related to cross-linguistic differences regarding the interpretation of the embedded third person plural subject that appears in half of the items (overt they in English and its null counterpart in Spanish). The difference is that, in the matrix conditions, this subject might be interpreted as co-referential with the extractee who in English but not in Spanish. The reviewer notes that co-referentiality could help process the long-distance dependency, making the sentences easier to process and more acceptable in English than in Spanish. If English they facilitated processing and increased the acceptability of the sentences due to co-referentiality, the subset of the English items that contained they should be more acceptable than the other half of the items, where the embedded subject was you and co-referentiality with who was not possible. This was not the case: overall, the English they and you items obtained very similar ratings in the two relevant matrix conditions (non-island/matrix they: mean = 0.749, SD = 0.402; non-island/matrix you: mean = 0.770, SD = 0.368; island/matrix they: mean = 0.567, SD = 0.494; island/matrix you: mean = 0.597, SD = 0.479). Thus, the presence of overt they in English did not affect ratings, and therefore it is unlikely that the larger island effects in Spanish are attributable to the contrast between English they and its Spanish null counterpart.

The final empirical question that we sought to address was whether there is increased individual variation in Spanish as compared to English (both between- and within-participants), as this could be a potential explanation for some of the variability in the observations reported in the literature on Spanish. We find no evidence of greater between- or within-participant variability in Spanish (and, in fact, the only island that showed increased variability was in English). This provides additional support for the similarity between the two languages. The question whether there could be more variability in Spanish than English was reasonable given the contrast between Torrego’s (1984) observations and recent experimental findings, and given that experimental studies have found increased between and within-participant variability, both in other languages (e.g., Norwegian; see Kush et al. 2018; 2019; Kobzeva et al. 2022; Kush & Dahl 2022) and in Spanish when wh-islands with “d-linked” fillers (Pañeda & Kush 2022). However, the Spanish speakers in our study do not show more variation than the set of English speakers. This finding helps to underscore that the island effects that we observed are likely a robust part of Spanish for these speakers. On that note, it seems valuable for future work to continue to probe individual variation because it can potentially reveal more information on factors that govern acceptability (cognitive, dialectal, etc.) or help to establish that the effect is indeed robust.

7.2 Theoretical implications

We have shown that wh-extraction gives rise to island effects in both English and Spanish and that these effects are not smaller in Spanish. The interpretation of our results and their implications for different theories depend on two different assumptions: (i) whether one assumes a single source for wh-island effects, and (ii) whether one assumes a common source for the effects in English and Spanish.

If one assumes a single, shared source for wh-island effects, then our findings simplify the theories that we considered above by removing the need to account for variation in Spanish. For the Subjacency approach, this would entail that the critical bounding node for Spanish is IP, similar to English. For the Phase-Impenetrability approach, this would entail that the C head in Spanish embedded questions licenses a single specifier position that hosts the wh-items that introduce embedded questions, just as in English (and that the position is, crucially, a phase edge). For Relativized Minimality, our results would entail that the wh-interveners that we tested in Spanish all share a critical movement feature with the extracted wh-items, and that the structural position of the intervener is the same type as the landing position of the extracted wh-items. For the Information-Structure-based approach, our results suggest that the embedded wh-questions that we tested are all likely to be considered backgrounded by speakers of Spanish (and therefore would show as such under the backgroundedness/focus diagnostics that Erteschik-Shir 1973 identifies). For working memory-related processing-difficulty approaches, our results suggest that the encoding and/or retrieval cues in Spanish are similar to those in English, such that the processing of island structures while processing a long-distance dependency creates an overload in the working memory system. All of these accounts predict that wh-island effects should be present across comparable configurations in Spanish and English beyond the sentences we tested. Whether there is such cross-linguistic uniformity is an empirical question for future research.

Relaxing the single source assumption while maintaining the common source assumption would admit more flexibility regarding the possibility of cross-linguistic variation. Under a multiple constraint approach, the fact that Spanish and English exhibit island effects entails that the test sentences violate at least one shared constraint, but still allows variation among any remaining constraints included in the model. The abstract logic holds regardless of the set of constraints one adopts, but the implications vary depending on the constraints chosen by the analyst. We offer a hypothetical example where a multiple constraint approach would lead to different conclusions than a single constraint one. Suppose that Subjacency exists alongside a universal semantic/Information-Structural constraint that independently blocks wh-extraction from embedded questions. In this scenario, our test sentences would be ruled out in English and Spanish even if the two languages differed in their bounding nodes. Thus, the presence of island effects in both languages would not necessarily correspond to alignment in underlying grammar – a conclusion that might be welcomed by proponents of parameterized Bounding or Phase Theory. Importantly, the multiple constraint model in the above scenario predicts that if the contribution of the shared constraint(s) was neutralized or factored out, cross-linguistic differences would emerge: wh-island effects would persist in English, but not be found in Spanish. Once again, we leave testing these predictions to future experiments. Before moving on, it is important to note that a multiple constraint approach in which both/all constraints are acquired from evidence during language acquisition would likely create a difficult learning problem for the child, as it would not be clear how to apportion responsibility across the constraint set. As such, the multiple constraint approach might only be tractable if one or more of the constraints is universal (as in the hypothetical example above).

Finally, we could relax the assumption that the source of the wh-island effects is the same in both languages. It is possible that the English and Spanish wh-island effects arise for different reasons. This style of reasoning has been called the eclectic approach by Chaves & Putnam (2020). An example of such an approach can be found in Christensen et al. (2013), who observed low acceptability for extraction out of a wh-structure in Danish (about 1.2 z-units). They argue based on evidence from an fMRI experiment that the low acceptability is driven by sentence processing difficulty rather than a grammatical constraint, in contrast to English (e.g., Sprouse et al. 2012; Yoshida et al. 2014). Eclectic approaches allow for more degrees of freedom than common source approaches. Theories of the acquisition of eclectic approaches to islands would likely entail the child tracking multiple pieces of evidence (perhaps across dependency types) to determine the sources of the constraints.

Broadly, our results suggest new research directions for two of the major overarching theoretical questions surrounding island effects: the extent of the cross-linguistic variation of island effects, and the source of island effects. For cross-linguistic variation, our results join a growing number of experimental studies that have shown that wh-island effects do in fact exist in languages that were thought to not have them, such as Brazilian Portuguese (Almeida 2014), Italian (Sprouse et al. 2016), Norwegian (Kush et al. 2018; 2019; Kobzeva et al. 2022), and Mandarin (Chen 2024). Italian and Spanish together formed the initial empirical argument for a highly constrained theory of cross-linguistic variation, instantiated at the time through parameters (Rizzi 1982; Torrego 1984), whereas Scandinavian languages raised the possibility that languages with wh-movement could exhibit even fewer island effects (Engdahl 1982), and Mandarin raised the possibility that the syntactic operations underlying wh-in-situ, such as covert movement (Huang 1982) and/or unselective binding (Nishigauchi 1986; Pesetsky 1987; Cheng 1991; Tsai 1994), could also be immune to island effects. The results that have been building within the experimental syntax literature challenge all of these conclusions. This raises the question of what exactly the empirical contours of cross-linguistic variation are in (wh- and other) island effects. Though there has been progress on this front in recent years (see Sprouse & Villata 2021 for a review up to that date), the field has still only tested a small number of languages compared to the number that have contributed to the development of previous theories of cross-linguistic variation of island effects. We see our study as a small contribution to this effort; and we see our results as demonstrating the value in systematically re-testing the languages that have contributed to these theories. While our results and the results of the empirical studies listed above have uncovered island effects where there were presumed to be none, they do not conclusively determine what set of factors bear causal responsibility for the effects themselves. Future work should systematically explore competing predictions of proposals already on the market and explore new proposals to account for effects that traditional accounts have difficulty explaining.