Using multivariate statistical methods, the authors deal with the interplay of an impressive twenty-two presumably universal and chiefly semantic and syntactic constraints influencing the variation of subject-controlled (marked) to-infinitives and prepositional gerunds complementing mainly three selected adjectives (primarily in present-day British and North American English). In the field of variable non-finite complementation, one of the two authors, Juhani Rudanko, is a highly productive and long-time expert who has been concentrating for years on a special set of matrix adjectives, including afraid and accustomed.

Large-scale electronic corpus analyses now revolve around afraid, accustomed and prone with the investigation of afraid (partly alongside other fear-adjectives) making up nearly two-thirds of the book. The text collections include (i) The Strathy Corpus of Canadian English, (ii) a sample of the British National Corpus (BNC) to match the Strathy corpus in size and genre composition, (iii) the News On the Web (NOW) corpus from twenty English-speaking countries and (iv) the British Hansard corpus. The authors engage in a great deal of statistical work: in contrast to traditional univariate studies, multivariate regression analyses, holding all other explanatory variables constant, are able to show whether a certain factor has an independent influence on the choice of the variants. This is particularly helpful with a high number of potential determinants.

The seven chapters of the book are organised as follows: all of them are preceded by an abstract, and chapters 2–6 also contain introductory and concluding sections as well as sections with information on the corpus, data or method and, apart from chapter 3, on model selection and model interpretation. Chapter 1 gives a general introduction presenting the notion of a head-based point of departure as the major type of approach (to the study of complementation) taken by all but one of the thematic chapters. It is only chapter 3 that examines the heads selected by a particular pattern or complement type (pattern-based approach). Chapters 2–4 are concerned with the adjective afraid as in ‘I was afraid to hang up / of hanging up’. In particular, chapter 2 focuses on afraid in the Strathy corpus and the BNC sample. Contrasting the spoken registers with other categories shows only negligible effects. Cases of afraid embedded within a degree complement construction (DCC) are excluded from the analyses and saved for chapter 3, which, in turn, is a spin-off from chapter 4 in that it looks into the syntactic and semantic peculiarities of afraid and other fear-adjectives (scared, terrified, fearful and frightened) nested within DCCs in the NOW corpus. Here, the adjectives are preceded by a degree adverb like too (see example (1)):

The figures concerning only the adjective afraid clearly differ from those obtained from the corpora used in chapter 2: within a DCC, afraid takes an infinitival complement in only 32 per cent of all cases as compared to roughly three-quarters without DCCs. The question whether this result is due to an (anticipatory) horror aequi effect is answered in chapter 4. Moreover, the authors raise the question as to whether the distance between the two complements plays a role in such an environment. In fact, this is suggested in the previous literature (cf. Vosberg Reference Vosberg2006: 207, 229, 275). Chapter 4 now examines afraid in the very sizable NOW corpus. Using a purposeful sampling method, the authors collect a randomised sample of 250 cases involving the adjective complemented by a marked infinitive and another 250 occurrences of prepositional gerunds, which, of course, leads to an oversampling of gerundial complements (see chapter 2). The aim of using this corpus and this method is to retrieve a statistically sufficient number of hits especially for low-frequency factors such as insertions or horror aequi (see below). Chapter 5 studies the adjective accustomed (as in ‘People are accustomed to disregard / to disregarding political prejudices’) in a subsection of the Hansard corpus covering the period 1945–64, a span which has been identified as showing the greatest variation between infinitives and prepositional gerunds. Chapter 6 reports on non-finite complementation of the adjective prone (as in ‘He is prone to stress / to stressing the agony of composition’) in the NOW corpus. Prone is semantically highly complex in that it is increasingly used as a characterising, non-propositional and dynamic modal with a concrete subject involving temporally stable inherent properties and a repeatable non-stative complement situation in the past. The authors conclude that ‘the more the context … differs from these prototypical features, the likelier the selection of the infinitival variant’ (p. 144). The prepositional gerund may have outstripped the to-infinitive complement only in recent years: therefore, 750 occurrences for the years 2014–17 in British English (BrE) and 750 occurrences for 2015–17 in American English (AmE) were extracted from the corpus, and the analysis yields 924 to -ing patterns. Chapter 7 provides a general conclusion summarising the main findings chapter by chapter and pointing out that the methodological significance of the contribution to the research field was to show that univariate analyses can be misleading in that the observed effect of a particular constraint is overruled by another.

Let us now come to the results of the statistical analyses with regard to the individual factors and their bearing on the distribution of infinitives and prepositional gerunds complementing the three adjectives. The determinants are here categorised according to the following concepts: (I) Cognitive complexity, (II) Horror aequi, (III) Semantics, (IV) Text register, (V) Diachrony, (VI) National variety, (VII) Interactions.

(I) As stated by Rohdenburg's Complexity Principle, the (structurally) more explicit grammatical option will tend to be favoured in cognitively more complex environments. In the case of non-finite complement variation between to-infinitives and prepositional gerunds, the question arises as to which of the two types is more explicit (pp. 84, 142), and, in line with Rohdenburg (Reference Rohdenburg2016: 472) and Vosberg (Reference Vosberg2006: 277), the authors adopt the view that ‘the to infinitive has a higher cue validity as a sentential structure than gerunds’ (p. 103). As a special manifestation of the Complexity Principle, the Extraction Principle describes the tendency to prefer the infinitive to the gerund if a post-verbal element (e.g. an object or an adjunct) is fronted by extraction out of the complement clause. The Hansard corpus is large enough to demonstrate that both extractions out of complements (mostly objects) as well as extractions out of adjuncts tend to trigger an infinitival complement (of accustomed). A further manifestation of cognitive complexity is represented by structural discontinuities: insertions are characterised by (normally adverbial) material intervening either (a) between the matrix expression and its complement or (b) between the element licensing the complement (here: the infinitive marker or the preposition) and the verb within the complement. As to (a) pre-complement insertions, we know that the higher the number of intervening words, the likelier an infinitival complement. The distinction between clausal and non-clausal types of insertions is dropped from the analyses (of accustomed) because it proves to be immaterial. With afraid (NOW), the infinitive-favouring effect of (a) pre-complement insertions is found to be stronger than that of (b) insertions within the complement. Another complexity factor is usually represented by the passive voice. However, contrary to the Complexity Principle, all three adjectives here show a slight preference for the gerundial variant, and the effect turns out to be either statistically non-significant (prone) or otherwise inconsequential (afraid: Strathy, BNC sample, NOW; accustomed). Plural and non-count (mass) matrix subjects can also be classified as a complexity factor (cf. Vosberg Reference Vosberg2006: 54, 153, 156; Rohdenburg Reference Rohdenburg, Rohdenburg and Mondorf2003: 221–2), which increases the chances of an infinitival complement after prone. A negative polarity (including both not- or no-negation) of either the superordinate or the subordinate clause is another factor featuring an increased cognitive complexity (cf. Vosberg Reference Vosberg2006: 53). On a statistically non-significant basis, the effect of matrix no-negation is slightly stronger than that of not-negation (for afraid: NOW). Surprisingly, not-negation of the complement shows a non-significant tendency to take the prepositional gerund complementing the adjective afraid (NOW). The Complexity Principle might be overruled by Hawkins’ Domain Minimisation Principle here, which implies that the human processor prefers to minimise the domain between connected sequences. Therefore, the infinitival complement is dispreferred because the negative particle not as in ‘He's afraid not to do it’ maximises the domain between afraid and the licensing element (the infinitive marker to) of the dependent clause. The two complexity factors of matrix modalisation and matrix past tense (or perfective aspect) both show a statistically non-significant predilection towards the infinitive complement of prone; with accustomed the effects are immaterial. The same holds true for matrix subject length and type (noun vs pronoun) because these two features also have no bearing on the complement choice following accustomed. Interestingly, post-nominal (‘zero-linked’) uses of accustomed exhibit a statistically significant tendency to take a gerundial complement, which is difficult to explain. The authors wonder whether ‘more nominalized clauses are associated with backgrounding’ (p. 217), and they are in doubt as to whether post-nominal uses are ‘universally inclined toward less explicit clausal structures’ (p. 115). In fact, however, Rohdenburg (Reference Rohdenburg and Kühlwein1998: 74–5) has found that post-nominal uses are associated with complex environments, which, in turn, tend to prefer the more explicit structure according to the Complexity Principle (also cf. Vosberg & Rohdenburg Reference Vosberg, Rohdenburg, Claridge and Bös2019: 293, 302–3). Furthermore, there is a statistically significant predilection of polysyllabic complement verbs for the infinitive variant following accustomed and prone, which is a clear corroboration of the Complexity Principle. In the case of accustomed (see p. 140), this has already been pointed out in Vosberg (Reference Vosberg2006: 240–1, table 94). Finally, various complement verb-specific idiosyncrasies are detected to the effect that the subordinate verb itself has an influence on the choice of the complement options. Remarkably, discuss (realised as a prepositional gerund) and debate (in the form of an infinitive) as complements of accustomed show opposite biases although they are both disyllabic. The authors unfortunately failed to analyse the overall frequency of these verbs. We know that, according to the Complexity Principle, less frequent items tend to favour the structurally more explicit variant (cf. Vosberg Reference Vosberg2006: 241–4 for accustomed) because they are more marked and have a higher load of cognitive processing complexity.

(II) According to Rohdenburg's horror aequi Principle, the use of formally (near-)identical and (near-)adjacent (non-coordinate) grammatical elements or structures is likely to be avoided. The effect can either be (a) anaphoric or (b) cataphoric. (a) If the superordinate adjective occurs in a to-infinitival construction itself, there is a statistically non-significant tendency to avoid another marked infinitive in the complement of afraid (NOW) and accustomed. By contrast, the avoidance strategy is indeed statistically significant if afraid (NOW) itself occurs in a gerundial construction introduced by the preposition of. (b) Predictably at first sight, a univariate analysis shows that complements of afraid (NOW) embedded within a DCC, as in example (1) above, tend to favour the prepositional gerund (35/52 = 67%). However, nearly all of the gerundial instances (32 out of 35) appear in a [–Choice] context (see below) so that the distribution is due to the Choice Principle rather than horror aequi. Unexpectedly, but on a statistically non-significant level, the multivariate analysis then reveals that DCC horror aequi contexts tend to choose infinitival complements.

(III) Rudanko's Choice Principle predicts that agentive ‘[+Choice]’ subjects (involving volition, control and responsibility) tend to favour the infinitival complement type whereas subjects showing the semantic feature [non-agentive] or [–Choice] are likely to prefer the gerund. On a statistically significant level, this principle is corroborated across the board with all adjectives and all datasets. Other semantic properties, such as the following four, have been studied, all of which tend to trigger the infinitive after prone: (i) a stative meaning of the complement denoting a static quality (as opposed to a dynamic meaning signifying a potential action); (ii) a propositional/wide scope meaning of the adjective (involving a prediction or what is likely as a consequence rather than a physical, inherent property or a characterising behaviour of the subject); (iii) a future reference of the adjective (also including non-propositional/narrow scope meanings); (iv) abstract inanimate subjects (e.g. history). Unexpectedly, concrete inanimate subjects (like tree) exhibit a tendency to take the prepositional gerund after prone probably because they are paradigmatic of the patient role and therefore subject to the Choice Principle. Likewise, easy repeatability of the complement situation involves iterative, recurring or habitual meanings and also has a proclivity for the to -ing pattern following prone.

(IV) Concerning text registers, the authors ‘expect fictional narratives to use complement clauses to refer to specific actions, which have been associated with the infinitive’ (p. 39). This assumption is actually substantiated with afraid (Strathy). Furthermore, the following two conclusions can only be drawn on a statistically non-significant level: while academic prose is apt to take the infinitive (after afraid: Strathy), learned registers are likely to prefer the prepositional gerund (afraid: BNC sample) because the latter are ‘expected to focus on generalities’, which are often associated with the semantic features of the gerund, the authors claim (p. 39).

(V) In line with the Great Complement Shift, a diachronic evolution on the way to the prepositional gerund in complement selection is only found for accustomed.

(VI) Interestingly, AmE is lagging behind BrE in the trend towards the prepositional gerund complementing the adjective prone. By and large, however, it is AmE that has been observed to be more progressive and more inclined to use the gerundial complement (cf. Vosberg Reference Vosberg, Rohdenburg and Schlüter2009: 226–7 and Rohdenburg & Schlüter Reference Rohdenburg, Schlüter, Rohdenburg and Schlüter2009: 421). Moreover, a univariate analysis of afraid (Strathy and BNC sample) reveals that the overall share of the infinitival complement (78%) is somewhat higher in Canadian English (CanE) than in BrE (72%) with the difference being statistically significant. The authors believe that the manifestation of other explanatory determinants might be responsible in a stronger way than the regional factor (p. 35). Finally, outer-circle varieties of English are significantly more likely to use the of -ing pattern in non-finite complements of afraid (NOW) than inner-circle varieties.

(VII) A special interaction is reported to exist between the semantic Choice Principle and the syntactic Complexity Principle: The observation is that there is no additional infinitive-favouring influence of matrix not-negation in [+Choice] contexts so that the complexity effect disappears if the semantic feature of [+agentive] comes in (although both factors each show a preference for the infinitive when regarded in the absence of the other). Even more surprisingly (but on a statistically non-significant level), with accustomed, matrix not-negations in [+Choice] contexts should theoretically tend to trigger the gerundial complement.

Both head-based and pattern-based summaries of the statistically significant findings are now provided in the following paragraphs.

For afraid, the identified constraints include the following. (A) Infinitive-favouring: (I) extractions (in NOW), pre-complement insertions (NOW), within-complement insertions (NOW), matrix not-negation (BNC sample, NOW), matrix no-negation (NOW), (II) gerundial aequi (NOW), (III) [+Choice] contexts (Strathy, BNC sample, NOW), (IV) fictional register (Strathy), CanE (Strathy, BNC sample), (VII) interaction [+Choice] and matrix not-negation (BNC sample, NOW); (B) Gerund-favouring: (I) complement verb-specific idiosyncrasy (with get in BNC sample), (VI) outer-circle variety (NOW).

For accustomed, the (A) Infinitive-favouring constraints are: (I) complement/object extractions, adjunct extractions, pre-complement insertions, multiple word insertions, polysyllabic complement verb, complement verb-specific idiosyncrasy with expect, (III) [+Choice] contexts; (B) Gerund-favouring: (I) post-nominal adjective, complement verb-specific idiosyncrasy with discuss, (V) diachrony.

For prone, the (A) Infinitive-favouring constraints are: (I) extractions, plural and non-count matrix subjects, polysyllabic complement verb, verb-specific idiosyncrasy, (III) [+Choice] contexts, stativity, propositional scope, future orientation, abstract inanimate subjects, (VI) US-AmE; (B) Gerund-favouring: (III) concrete inanimate subjects, easy repeatability.

Generally, the infinitive-favouring factors include: (I) extractions (afraid: NOW, accustomed, prone), complement/object extractions (accustomed), adjunct extractions (accustomed); pre-complement insertions (afraid: NOW, accustomed), multiple word insertions (accustomed), insertions within complement (afraid: NOW); plural or non-count subject (prone); matrix not- and no-negation (afraid: NOW), matrix not-negation (afraid: BNC sample); polysyllabic complement verb (accustomed, prone); (II) gerundial aequi in of being afraid (NOW); (III) [+Choice] contexts (afraid: Strathy, BNC sample and NOW, accustomed, prone); stativity (prone); propositional scope (prone); abstract inanimate subject (prone); future orientation (prone); (IV) fictional register (afraid: Strathy); (VI) CanE (afraid: Strathy, BNC sample), US-AmE (prone); (VII) interaction [+Choice] and matrix not-negation (afraid: BNC sample and NOW).

Gerund-favouring factors are: (I) post-nominal adjective (accustomed); (III) concrete inanimate subject (prone); easy repeatability (prone); (V) diachrony (accustomed); (VI) outer-circle variety (afraid: NOW).

In conclusion, what makes this book so valuable is, firstly, its examination of an enormously large number of interacting and mostly universal factors, secondly, its statistical justification of conjectures, some of which are already well-known, and thirdly, its attempt to clarify the question of whether syntactic or semantic influences on the choice of non-finite complement variants prevail. Both the Choice Principle and the Extraction Principle are confirmed as the most consistently significant predictors across the board. In particular, the Choice Principle was found to be more pronounced with afraid than with accustomed or prone while, vice versa, the Extraction Principle was stronger with accustomed and prone rather than with afraid.

Results of univariate analyses are only presented in the cases of accustomed and prone, but most of them are highly suggestive in that they foreshadow the results obtained from the multivariate analyses. This is the case for the following infinitive-favouring determinants: extractions, [+Choice] contexts, polysyllabic complement verb, pre-complement insertions (only accustomed) as well as AmE, stative complements, propositional/wide scope of the adjective, abstract inanimate subjects and future orientation (only prone). However, the multivariate regression analyses have been able to demonstrate that univariate analyses are indeed sometimes deceptive. For example, although the overall proportion of infinitives complementing afraid (in the Strathy corpus and the BNC sample) is approximately three-quarters at first sight, it is actually the of -ing structure that is the predominant option if all influencing factors under investigation are eliminated. Additionally, the results have shown that the semantic Choice Principle, if in conflict, often outranks the syntactic Complexity Principle or the horror aequi Principle. This is the case with passivisation and DCC-nesting, respectively. Closer inspection of the passive complement clauses reveals that most of them are associated with [–Choice] contexts because the main ‘function of the passive voice is to defocus the agent of a transitive clause’ (p. 19), and subjects of passive constructions are prototypically assigned the semantic role of a non-agentive patient, which, according to the Choice Principle, tends to trigger gerundial complements rather than infinitives as is expected by the Complexity Principle. Therefore, multivariate statistical methods are necessary to estimate the added value of each of the relevant determinants.

The vast majority of the constraints known so far (including the relevant authors) are mentioned and discussed in the book, which, in many chapters, also suggests opportunities and avenues for further research, e.g. the influence of transitivity (in the case of afraid, p. 42); see Rohdenburg (Reference Rohdenburg, Kaunisto, Höglund and Rickman2018: 143). Frequency of use (of the dependent verb, see above) could also be added to the list of determinants. The same might possibly apply to the Principle of Rhythmic Alternation (cf. Schlüter Reference Schlüter, Vogel and de Vijver2015): if, for instance, the word following a finally stressed complement verb is initially stressed (as in ‘afraíd to gét / of gétting wét’), we would expect the gerundial complement to be preferred in order to avoid a stress clash with the -ing suffix serving as an unstressed buffer. Lastly, with chapters 2 and 4 making up almost half of the book, the main results are presented only as coefficients and intervals. In order to assess the benefit of the multivariate studies, however, it would be helpful for readers in general and especially for those who are not familiar with estimated logit-scale effects and log odds to provide absolute figures, which, unfortunately, are either missing completely or merely hidden in the running text and not presented in a reader-friendly table or chart form.