Suspended Affixation as Morpheme Ellipsis: Evidence from Ossetic Alternative Questions

This paper provides novel evidence that ellipsis can target bound morphemes. The evidence comes from suspended affixation of case markers in alternative questions in Digor and Iron Ossetic. The current literature on alternative questions (e.g. Does Mary like coffee or tea? ) proposes that in many languages they are derived by disjunction of and ellipsis in constituents as large as a vP or even as a CP. Language-specific evidence in favor of such structure of alternative questions is available for Ossetic as well. Accordingly, the ostensible disjuncts coffee or tea do not actually form a constituent and case must be separately assigned to each of the DPs. Therefore, a case suffix shared under suspended affixation cannot attach to the orP as a whole. A deletion-based analysis can successfully derive the properties of suspended affixation in Ossetic alternative questions. I advance a specific proposal that incorporates ellipsis into the Distributed Morphology derivation.


Introduction
How similar is morphology to syntax? Can processes usually thought of as exclusively syntactic occur in morphology as well? The expected answer to this question depends on one's theoretical stance: if one pursues a strictly lexicalist theory, e.g. such as developed in Di Sciullo & Williams (1987) or Bresnan & Mchombo (1995), or a realizational theory of morphology in the spirit of Stump (2001), the expected answer is negative. On the other hand, if one assumes that morphology should be essentially similar, or even reduced, to syntax, as the current work in Distributed Morphology does, see a.o. Halle & Marantz (1993;1994); Marantz (1997); Noyer (1997); Embick & Noyer (2001); Siddiqi (2010); and Arregi & Nevins (2012), the expected answer is affirmative. In this paper, I will present novel evidence that in Ossetic one such process, namely ellipsis, may target bound morphemes (specifically, case markers) to produce suspended affixation. Finally, if one chooses to pursue a prosodic deletion analysis of the facts, one would be forced to explain why it is only morphemes, and, moreover, only some morphemes that can be targeted by this deletion process.
Suspended Affixation (which will be abbreviated here as SA) is a phenomenon, or a class of superficially similar phenomena, when an affix only appears on the edgemost coordinand (or disjunct), but takes scope over all the coordinands (or disjuncts) (1a). SA is possible both with suffixes and prefixes, but, in this paper, I will focus on suffixes. I will consider only examples with two coordinands in this paper (the properties of coordinations with more coordinands are completely identical in the relevant respects). In pretheoretical terms, I will say that the "suspended" affix is shared between the conjuncts (or disjuncts). In the examples in (1b-c), the shared affixes are marked in boldface.
The latter work provides a number of examples of this phenomenon from the world's languages.
As has been observed in Kornfilt (2012), suspended affixation is similar to the right node raising (2): in the case of the RNR, instead of a shared suffix, two conjuncts share an entire DP or more syntactic material, Hartmann (2000); Chaves (2014); and Sabbagh (2014). For instance, in (2), the shared material is the DP the beans.
(2) Mary cooked, and John ate the beans.
A number of scenarios are imaginable by which suspended affixation comes about. Some of them are common with analyses of the right node raising, which has been argued to not be a uniform phenomenon, Barros & Vicente (2011); Chaves (2014). One logically possible analysis proposes that the "suspended" suffix attaches to the entire coordinate phrase (3). For expository purposes, I use a symmetric, non-binary branching structure for coordination and disjunction here and elsewhere in the paper.
Turkish yılan(-dan) ve köpek-ten korkuyorum. snake(-ABL) and dog-ABL I.fear 'I fear snakes and dogs.' The term "suspended affixation" was apparently coined by Lewis (1967). Other terms used in the literature are brachylogy, morphological ellipsis, both used by Pounder (2006), coordination reduction, Kenesei (2007), and unbalanced coordination, Johannessen (1998). The latter work provides a number of examples of this phenomenon from the world's languages.
As has been observed in Kornfilt (2012), suspended affixation is similar to the right node raising (2): in the case of the RNR, instead of a shared suffix, two conjuncts share an entire DP or more syntactic material, Hartmann (2000); Chaves (2014); and Sabbagh (2014). For instance, in (2), the shared material is the DP the beans.
(2) Mary cooked, and John ate the beans.
A number of scenarios are imaginable by which suspended affixation comes about. Some of them are common with analyses of the right node raising, which has been argued to not be a uniform phenomenon, Barros & Vicente (2011); Chaves (2014). One logically possible analysis proposes that the "suspended" suffix attaches to the entire coordinate phrase (3). For expository purposes, I use a symmetric, non-binary branching structure for coordination ; 2012) and Broadwell (2008) although Broadwell's and Belayev's formulations were couched in the LFG terms. A reformulation of Broadwell's (2008) analysis in terms of multidominance is given in Weisser & Guseva (2016). The analysis of Korean in Yoon & Lee (2005) is also similar in spirit, modulo that in their system case heads take DPs as complements.
A different approach, advanced in Erschler (2012a) for suspended affixation in Ossetic and Eastern Armenian, and in Weisser & Guseva (2016) for Mari, assumes that the absent affixes on the non-final conjuncts get deleted by means of a process similar or identical to the syntactic ellipsis (4). Trommer (2008) proposed that, in Hungarian, ellipsis is not restricted to free morphemes either. Kornfilt (2012: 190) mentions a possibility of "backward gapping applying in syntax" as the source of the SA, although this is not the analysis she eventually opts for.
This is essentially what has been proposed by Kornfilt (2000; and Broadwell (2008) for Turkish, and, with certain morphological complications, by Belyaev (2014) for Ossetic, although Broadwell's and Belayev's formulations were couched in the LFG terms. A reformulation of Broadwell's (2008) analysis in terms of multidominance is given in Weisser & Guseva (2016). The analysis of Korean in Yoon & Lee (2005) is also similar in spirit, modulo that in their system case heads take DPs as complements.
A different approach, advanced in Erschler (2012a) for suspended affixation in Ossetic and Eastern Armenian, and in Weisser & Guseva (2016) for Mari, assumes that the absent affixes on the non-final conjuncts get deleted by means of a process similar or identical to the syntactic ellipsis (4). Trommer (2008) proposed that, in Hungarian, ellipsis is not restricted to free morphemes either. Kornfilt (2012: 190) mentions a possibility of "backward gapping applying in syntax" as the source of the SA, although this is not the analysis she eventually opts for. (4) The analyses schematized in (3) and (4), that is, affixation to &P and affixation to each coordinand followed by deletion, are not easy to tell apart empirically. In this paper, I observe that they make sharply different predictions if the shared affix is a case marker, and XP1 and XP2 do not belong to the same domain of case assignment. One possible environment when this situation can obtain are alternative questions. A large body of the literature, starting from Han & Romero (2004), argues for derivation of alternative questions as a combination of large chunk disjunction and deletion 1 . How large the disjoined constituents are still remains a debated question, see, e.g. the discussion in Uegaki (2014a;b). The answer observe that they make sharply different predictions if the shared affix is a case marker, and XP 1 and XP 2 do not belong to the same domain of case assignment. One possible environment when this situation can obtain are alternative questions. A large body of the literature, starting from Han & Romero (2004), argues for derivation of alternative questions as a combination of large chunk disjunction and deletion. 1 How large the disjoined constituents are still remains a debated question, see, e.g. the discussion in Uegaki (2014a;b). The answer probably varies depending on the language and the specific construction, but they are at least the size of a VP. What is crucial for analyses of suspended affixation, is that, according to such analyses, the ostensibly disjunct DPs in an alternative question do not actually form a constituent: they belong to separate VPs, or, perhaps, even larger constituents as schematically shown in (5) Now, if suspended affixation of case markers is observed in a given language in alternative questions and it is possible to show that alternative questions are obtained by ellipsis, we obtain a strong argument in favor of a deletion analysis. Namely, in such an architecture and under the standard assumptions about case assignment, the case must be assigned to each of the DPs separately within the respective disjunct. In (6), I schematically indicate case assignment: for the purposes of my argument, it does not matter whether the accusative is actually assigned by agreement with some functional projection, Chomsky (1981;, or by a case assigning algorithm ("dependent case theory"), Yip et al. (1987); Marantz (1991); Levin & Preminger (2015); Levin (2017), or perhaps by both, as was argued in Baker & Vinokurova (2010 If SA arises through case assignment to the entire disjunction phrase, this analysis predicts that in alternative questions it must be ungrammatical: two DPs that are ostensibly disjoint in an alternative question, say 'cats' and 'dogs' in the sentence above, do not actually form a constituent to which the shared affix would attach.
On the other hand, if SA arises through post-syntactic deletion of exponents, it could be grammatical even in alternative questions. This of course does not rule out the possibility that even in such a language SA in alternative questions will still be ungrammatical or at least somewhat degraded for some independent reasons. However, if SA is allowed in alternative questions, it provides a strong argument in favor of the deletion theory.
In this paper, I show that exactly such a situation obtains in Digor and Iron Ossetic. I show that suspended affixation occurs in Ossetic alternative questions and provide evidence that these questions are indeed obtained by disjunction of large constituents and ellipsis in them. The sentences in (7) are alternative questions and they do exhibit SA: in (7a), the allative marker is optional on the first conjunct, 'Sarmat', whereas in the attested sentence in (7b), the ablative marker is absent from the first conjunct arv-ə c'ɐχ 'sky blue'.
Sarmat (-all) or.q Uruzmag-all you.called 'Did you call Sarmat or Uruzmag?' b. Iron Ossetic (Galuanty 2008) adɐjmag kʷəd fɐžənd? arv-ə c'ɐχ(-ɐj) ɐvi šəʤət-ɐj rajgʷərd? human how appeared sky-obl blue-abl or.q clay-abl was.born 'How did the humans appear? Were they born from the sky blue or from clay?' Accordingly, suspended affixation of case markers in Ossetic alternative questions has to be analyzed as morpheme ellipsis. The rest of the paper develops this argument in more detail. A word about the theoretical assumptions of this study is in place here. As I have mentioned already, I use the ternary branching structure for coordination (8).
Iron Ossetic (Galuanty 2008) adɐjmag kʷəd fɐžənd? human how appeared arv-ə c'ɐχ(-ɐj) ɐvi šəʤət-ɐj rajgʷərd? sky-OBL blue-ABL or.Q clay-ABL was.born 'How did the humans appear? Were they born from the sky blue or from clay?' Accordingly, suspended affixation of case markers in Ossetic alternative questions has to be analyzed as morpheme ellipsis. The rest of the paper develops this argument in more detail. A word about the theoretical assumptions of this study is in place here. As I have mentioned already, I use the ternary branching structure for coordination (8).
Nothing in the argument below would change substantially if one uses an asymmetric structure (9) in the spirit of Munn (1993); Zoerner (1995); Johannessen (1998);and Hartmann (2000), and adopted by much of the literature since then 2 . In the case of disjunction, this structure was adopted by den Dikken (2006).
2 If, on the other hand, one adopts the proposal of Camacho (2003) that only sentence-sized constituents can be coordinated, the argument presented here will only be strengthened: ostensibly coordinated DPs will never belong to a same case assignment domain.
Nothing in the argument below would change substantially if one uses an asymmetric structure (9) in the spirit of Munn (1993); Zoerner (1995); Johannessen (1998);and Hartmann (2000), and adopted by much of the literature since then. 2 In the case of disjunction, this structure was adopted by den Dikken (2006).
Iron Ossetic (Galuanty 2008) adɐjmag kʷəd fɐžənd? human how appeared arv-ə c'ɐχ(-ɐj) ɐvi šəʤət-ɐj rajgʷərd? sky-OBL blue-ABL or.Q clay-ABL was.born 'How did the humans appear? Were they born from the sky blue or from clay?' Accordingly, suspended affixation of case markers in Ossetic alternative questions has to be analyzed as morpheme ellipsis. The rest of the paper develops this argument in more detail. A word about the theoretical assumptions of this study is in place here. As I have mentioned already, I use the ternary branching structure for coordination (8).
Nothing in the argument below would change substantially if one uses an asymmetric structure (9) in the spirit of Munn (1993); Zoerner (1995); Johannessen (1998);and Hartmann (2000), and adopted by much of the literature since then 2 . In the case of disjunction, this structure was adopted by den Dikken (2006).
2 If, on the other hand, one adopts the proposal of Camacho (2003) that only sentence-sized constituents can be coordinated, the argument presented here will only be strengthened: ostensibly coordinated DPs will never belong to a same case assignment domain.
I adopt a derivational approach to the post-spellout stages of computation, as represented, for instance, by Embick & Noyer (2001); Embick (2010) and Arregi & Nevins (2012). The paper is organized as follows: in Section 2, I overview the basic background facts about Ossetic, focusing on case marking, coordination, and disjunction in these languages. In Section 3, I focus on alternative questions: I provide evidence for a "big disjunct" analysis of alternative questions in Ossetic and overview some minor types of alternative questions. Section 4 provides a description of suspended affixation in Ossetic. In Section 5, I formulate the proposal about how to integrate morpheme deletion in the framework of Distributed Morphology. In Section 6, I show how the proposal derives the observed properties of SA in Ossetic. In Section 7, I address several ostensible challenges to the current proposal and show that it can actually handle the facts. In Section 8, I compare the current proposal to several alternative ones. In Section 9, I discuss the relation between suspended affixation and right node raising. Section 10 concludes.

Background on Ossetic
Here, I provide some background information on Ossetic and Ossetic grammar, specifically, I address case marking in Section 2.1; and the properties of coordination and disjunction in Section 2.2.
Digor and Iron Ossetic are closely related East Iranian languages 3 spoken in the Central Caucasus. These languages are loosely head-final: noun phrases and non-finite clauses are rigidly head-final, whereas for finite clauses, verb-finalness is merely a statistical tendency. The SOV order is the most frequent one, but scrambling is fairly free. The alignment is nominative-accusative. Both languages have a large system of Wackernagel or almost Wackernagel clitics, a circumstance that will turn out useful for us when arguing for a large disjunct analysis of alternative questions. The languages exhibit pro-drop of subjects. For non-nominative pronouns, clitic and non-clitic forms alternate. For more background on these languages and their grammar, see e.g. Erschler (2012b) and references there.

Case marking
In this subsection, I describe the morphophonology of the Ossetic markers in some detail. The upshot of the discussion is that the case marking is nearly agglutinative for nouns and is based on several stems for pronouns.
In a DP, the case and number are only marked on the head noun, no overt agreement of any kind exists (10). Case suffixes only undergo little morphonological variation and attach directly to the stem (10a), only preceded by the plural marker (10b). A peculiar behavior of DPs with numerals will be addressed in Section 7.1: for the time being, it is not important for our discussion. (10) Digor Ossetic a. ači ustur bel this big spade 'this big spade' b. ači ustur bel-tɐ-bɐl this big spade-pl-sup 'on these big spades' The case inventory of Ossetic comprises the nominative, accusative, genitive, 4 dative, ablative/instrumental, allative, superessive, inessive, and equative. Traditional descriptions of Iron Ossetic also list the comitative case, which probably should be analyzed as a postposition. In this subsection, I provide some typical paradigms; see Axvlediani (1963); Abaev (1964); Isaev (1966);and Takazov (2009) for more data. The behavior of stems is different in nouns and in pronouns. For nouns, changes in the stem that can be triggered by a case suffix are minimal, whereas for pronouns, two or three different stems occur in case forms.
For nouns, the shape of case suffixes depends on whether the stem ends in a consonant, as illustrated for Iron and Digor bɐχ 'horse' in Tables 1 and 2, in a vowel other than ɐ, as illustrated for Iron gɐdǝ 'cat' in Table 1 and Digor k'ibila 'bucket' in Table 2. In Iron, singular nouns in ɐ do not differ from other vowel-final stems, as the paradigm of žɐrdɐ 'heart' in Table 1 shows. I postpone a systematic discussion of the properties of ɐ-final stems, and, in particular of plural forms, until Section 7.2. 3 As an anonymous reviewer correctly remarks, they are called dialects of a single language in much of the literature. This is not, however, how they are perceived by the speakers of Digor Ossetic, nor are they mutually intelligible. 4 The accusative and the genitive differ for very few lexical items, these include the pronominal clitics and the wh-word 'what': the genitive clitics procliticize to possessed noun phrases, while the accusative ones occupy the Wackernagel position in the clause. For the wh-word 'what' the accusative form coincides with the nominative one. The inessive differs from the oblique for enclitic pronouns and, in Digor, for numerals and numeral phrases. I use the label 'oblique' whenever these morphological forms coincide.
A remark about syncretism patterns is in place here. The suffix -i in Digor and its cognate -ǝ in Iron serve as the marker of the accusative, genitive, and inessive for lexical nouns. Moreover, it appears as the numeral suffix with nouns in the nominative. While in Digor its phonological properties are unremarkable, in Iron it causes affricativization in velar stops: it turns k into ʧ and g into ʤ. This occurs in all the four functions of the suffix, as is illustrated for k-final stems 5 in (11).
For pronouns, the situation is more complex: each pronoun can exhibit one, two, or three inflectional stems, as illustrated in Tables 3-4 for personal pronouns and in Table 5 for wh-pronouns. The inessive form is not used with pronouns and, accordingly, the respective row is missing from the tables, see the discussion in Footnote 14 below.
If the case markers in Ossetic were clitics, and not affixes, the fact that they may undergo ellipsis might have been less surprising. 6 However, Ossetic case markers are definitely affixes: this claim is substantiated, for instance, by the fact that case suffixes can attach to the stems that are not independent words, contrary to what would be expected for clitics. 7 6 A connection between the ability to undergo SA and clitic, rather than affix status, has been proposed by Good & Yu (2005: 321) for certain verbal endings in Turkish. Similarly, in Hungarian, the ability or inability of certain morphemes to undergo SA has been connected to their being case markers (affixes) or postpositions (clitics), see Kiss (2002: 184). 7 That is to say, case markers show a high degree of selection, as affixes should according to diagnostic (A) in Zwicky & Pullum's (1983) list of properties that distinguish clitics and affixes. I am not aware of arbitrary gaps in the set of combinations, diagnostics (B). For examples of morphophonological idiosyncrasies, diagnostic (C), see the discussion of epenthesis in Section 7.2. As an example of a semantic idiosyncrasy, one can name the ban for animate nouns to stand in the inessive, diagnostic (D).

(12)
Iron Ossetic rɐšuʁd čǝžʤǝ-tɐ ɐmɐ tǝχʤǝn lɐppu-tɐ beautiful girl-pl and strong boy-pl 'beautiful girls and strong boys' Besides that, in Iron, DPs can be coordinated by means of the negative marker nɐ, which is placed in front of each coordinand to express the meaning 'neither … nor' (13).

Derivation of alternative questions
Starting with Han & Romero (2004), evidence has accumulated that in many languages alternative questions are always derived by conjoining much larger constituents and then performing ellipsis in one of them, even when ostensibly only DPs are disjoined as in (17a). The actual parse of (17a) is assumed to be (17b) or (17c). Ellipsis, however, creates an illusion of direct DP disjunction in such questions.  The languages where this has been argued for so far are English, Hindi, and Korean in Han & Romero (2004); Japanese in Uegaki (2014a; b); Serbo-Croatian in Gračanin-Yüksek (2016a); and Turkish in Gračanin-Yüksek (2016b). Arguments for this type of analysis are, however, language specific. In this subsection, I provide evidence that this analysis is correct for Ossetic as well.

Big disjunct analysis in Ossetic
Ossetic-specific evidence for an ellipsis analysis of alternative questions comes from the behavior of Wackernagel clitics in alternative questions. Ossetic has a large number of Wackernagel and almost Wackernagel clitics. In Iron, they obligatorily occupy the appropriately defined second position of the clause, Lowe & Belyaev (2015), while in Digor placement of the cluster is somewhat freer, Erschler (2010). In Iron, the clitic cluster is placed after the first word of the clause, (18a-b). If the first constituent is an NP or a DP, however, enclitics attach to the right edge of the noun phrase, (18c-d). In Digor, (extended) noun phrases are impenetrable for clitics as well, (19a) but the clitic cluster may be placed further from the sentence left edge than in Iron, although acceptability of such sentences decreases the further the clitics are from the second position, (19b-c).
apple=dat.3sg they.gave 'They gave him a beautiful large tasty red apple.' b. [mink'ij k'ɐbis]=dan =in aboni ravardtoncɐ. little puppy=quot =dat.3sg today they.gave '(They say), they gave him/her a little puppy today.' c. ?[mink'ij k'ɐbis] aboni=dan =in ravardtoncɐ. little puppy today=quot =dat.3sg they.gave Idem Now, in Iron, ta 'again', 9 which we have already seen in (18c), is a Wackernagel clitic. In (20), it encliticizes to ɐvi 'or', which shows that the constituent 'or again a potato pie' forms a separate clause: otherwise we would expect ta to cliticize in the second position of the whole sentence, after the constituent Alan-dat. Furthermore, semantically the clitic ta 'again' modifies the event of making a potato pie in (20a) and of biting Khetag in (20b). It is hard to see how to derive these meaning unless the conjunction 'or' introduces an entire clause rather than a DP. 10 (20) Iron Ossetic a. alan-ɐn jɐ=mad fəččən škodta ɐvi=ta kartofʤən? Alan-dat his=mother meat.pie made or.q=again potato.pie 'Did Alan's mother cook him a meat pie or [did she again cook him a potato pie]?' If, on the other hand, =ta attaches to the first constituents in (20a-b), we obtain two different readings: Iron Ossetic a. alan-ɐn=ta jɐ=mad fəččən škodta ɐvi kartofʤən? Alan-dat=again his=mother meat.pie made or.q potato.pie Reading 1: 'What happened again: did Alan's mother cook him a meat pie or a potato pie?' Reading 2: '[Did Alan's mother cook him a meat pie again] or did she cook him a potato pie (this time The scope evidence can be replicated for Digor as well. In the Digor sentence in (22a), babɐj 'again', an almost Wackernagel clitic, only takes scope 11 over the clause 'made a potato pie', which shows that 'for Alan, his mother a meat pie' is a separate clause, an outcome of backward gapping, (22b).

(22)
Digor Ossetic a. alan-ɐn ɐ=madɐ fidgun ɐvi=babɐj kartofgun iskodta? Alan-dat his=mother meat.pie or.q=again potato.pie made 'Did Alan's mother cook him a meat pie or again a potato pie?' b. alan-ɐn ɐ=madɐ fidgun iskodta ɐvi=babɐj kartofgun iskodta? Alan-dat his=mother meat.pie made or.q=again potato.pie made Finally, under the word order illustrated in (16a), with the second alternative following the entire question, the ellipsis is rather obvious given the standard assumptions about coordination (23): otherwise, a sentence would be coordinated with a DP.

(23)
Iron Ossetic kɐlmǝ-t-ɐj tɐršǝš ɐvi kʷǝj-t-ɐj tɐršǝš? snake-pl-abl you.fear or.q dog-pl-abl you.fear 'Do you fear snakes or dogs?' As an additional piece of evidence, 12 one can add that non-constituents may be ostensibly disjoined in Ossetic alternative questions. In (24), soslan-i ʁɐwungɐ-bɐl Soslan-acc street-sup 'Soslan in the street' is clearly not a constituent: 'in the street' is supposed to modify the event of seeing rather than the DP Soslan.

Minor varieties of alternative questions
Even speakers who prefer alternative questions without ostensible disjunction of DPs as in (23), allow ostensible disjunction in fragment questions and split questions, as is schematically shown in (25a) and (25b), respectively. The term "split questions" was introduced by Arregi (2010). (25) a. Rice or beans? b. What did Mary cook, rice or beans?
As argued in Arregi (2010), derivation of split questions involves deletion of the rest of the material in a full-fledged alternative question. This analysis (and the arguments in its favor) can be applied verbatim to isolated fragment alternative questions, which, to the best of my knowledge have not been addressed in the literature: the bulk of the research on fragments focuses on fragment answers, see Merchant (2005); Weir (2014), and references there. An additional variety of alternative questions are "or-sluices": embedded alternative questions where everything but the ostensible disjuncts is deleted. In other words, these are embedded fragment alternative questions. Such constructions are relatively common crosslinguistically. In the German sentence in (26a), the or-sluice is (dem) Uwe oder (dem) Jan def.dat Uwe or def.dat Jan 'Uwe or Jan'. In the Polish sentence in (26b), the or-sluice is (czy) ryż czy kasz-ę gryczan-ą q rice.acc q buckwheat-acc porridge-acc 'rice or kasha'.

b. Polish
Zosia coś ugotowała, ale nie wiem, (czy) ryż czy Zosia something she.cooked but neg I.know whether rice whether kasz-ę gryczan-ą. porridge-acc of.buckwheat-acc 'Zosia cooked something, but I don't know whether (she cooked) rice or buckwheat.' (Stanisław Dunin-Horkawicz, p.c.) Note that in both sentences in (26), the DPs in the sluices stand in the case assigned by the verb in the antecedent, the dative by schmeicheln 'to flatter' in (26a), and the accusative by ugotować 'to cook' in (26b). In the same way as for regular sluicing, Ross (1969Ross ( [2012); Merchant (2001), this fact serves as evidence for ellipsis derivation of or-sluices.
Ossetic languages have all these varieties of reduced alternative questions: fragment questions (27a), split questions (27b), and or-sluices (27c). As we will see in the next section, SA of case markers is possible in all of them. For the sake of clarity, now I only give the variants without SA.

Suspended affixation in Ossetic: Descriptive generalizations
With the background facts about case marking, coordination, disjunction, and alternative questions at hand, we can proceed to discussion of suspended affixation in Ossetic. In this section, I first formulate the overall descriptive generalizations about SA in Ossetic (Section 4.1) and illustrate them for regular coordination, and then focus on SA in alternative questions (Section 4.2). In Section 4.3, I argue that availability of SA in alternative questions together with the fact that alternative questions are derived by ellipsis compel us to a deletion analysis of SA, no matter what the precise technical implementation of this idea is.

Descriptive generalizations
In the descriptive literature, suspended affixation in Ossetic was addressed in Kulaev (1981); two different theoretical analyses (with additional descriptive details) were proposed in Erschler (2012a) and Belyaev (2014). Descriptively, suspended affixation in Ossetic has the following properties.
(A) Suspended affixation occurs with any type of coordinator. (B) SA only proceeds backwards. (C) Only case markers, and any case markers, may be suspended. (D) Remnants under suspended affixation construction must be substrings of the respective full forms (modulo phonological readjustments, and one very interesting exception that concerns numeral phrases, which will be discussed below). (E) A remnant under SA substring must be an actual independent word. (F) This word should not have an accidental idiosyncratic lexical meaning. (G) When both conjuncts are pronouns or are marked with the enclitic =dɐr, suspended affixation is ungrammatical.
To comment upon property (A), suspended affixation freely occurs under coordination with 'and', (28a), negative coordination (28b), and under disjunction, (28c). The DPs that participate in SA are bolded. I postpone the discussion of SA in alternative questions until the next subsection.

(29)
Digor Ossetic a. alan(-ɐn) ɐma soslan-ɐn sɐ=madɐ Alan-dat and Soslan-dat 3pl=mother 'Alan and Soslan's mother' b. *alan-ɐn ɐma soslan-ɐn sɐ=madɐ Alan-dat and Soslan-dat 3pl=mother Idem (intended) Property (C) states that only case suffixes (and any case markers) can be suspended in Ossetic. Unlike in Turkish (see Lewis 1967;Kornfilt 1996;and Kabak 2007) and in some other Turkic languages, the nominal plural marker cannot be suspended, nor is SA available for any verbal suffixes. SA is possible for any morphological case in Ossetic (30), no matter as structural as the case of the possessors in (30a), lexically assigned by the verb as the allative in (30b) or the ablative in (28b), or semantic, as the inessive (which is expressed the oblique suffix in this instance) in (30c), the allative in (28a), or the superessive in (28c).
Soslan (-all) and Alan-all waiting we.wait 'We are waiting for Soslan and Alan.' c. Digor Ossetic ači kiwunugutɐ balχɐdton mɐsku(-j) ɐma boston-i. this book.pl I.bought Moscow(-loc) and Boston-loc 'I bought these books in Moscow and Boston.' Property (D), that remnants under SA must be substrings of the respective full forms, describes the contrast of the type illustrated in (31). While dɐw, the stem of dɐw-bɐl and an independent word, the oblique (i.e. the accusative and genitive) form of du 'you.sg' is possible in SA (31a), the nominative form du, which is not a substring of dɐw-bɐl, is judged ungrammatical 13 (31b). Readjustments will be discussed in sections 7.2 and 7.3 below.
Property (E) is that a stem that cannot function as an independent word may not appear as a remnant under SA. For instance, in Iron Ossetic, the reciprocal stem kɐrɐzi is a word, whereas its Digor cognate, kɐrɐʤe, cannot be used without a case suffix. Accordingly, in Iron but not in Digor the reciprocal stem may serve as a non-last conjunct in suspended affixation constructions. The two sentences in (32a-b), the grammatical Iron one and the ungrammatical Digor one (32b), are fully parallel. The grammatical Digor sentence in (32c) shows that (32b) becomes grammatical once the case marker is restored on kɐrɐʤe.

(32)
Iron Ossetic a. ?nɐ=dəwɐ gɐdə-jə kɐrɐzi ɐmɐ nɐ=kwəz-ɐj tɐršync. our=two cat-obl each.other and our=dog-abl fear.prs.3pl 'Two our cats are afraid of each other and of our dog.' Digor Ossetic b. *nɐ=duwɐ tikiš-i kɐrɐʤe ɐma nɐ=kuj-ɐj tɐrsuncɐ. our=two cat-obl each.other and our=dog-abl fear.prs.3pl Idem (intended) c. nɐ=duwɐ tikiš-i kɐrɐʤe-jɐ ɐma nɐ=kuj-ɐj poss.1pl=two cat-obl each.other-abl and poss.1pl=dog-abl tɐrsuncɐ. fear.prs.3pl Idem Property (F) requires that the remainder under SA may not be a word with an idiosyncratic meaning, which accidentally or for diachronic reasons coincides with a conjunct minus the suspended affix. This property can be illustrated by the behavior of the Iron wəm 13 An anonymous reviewer observes that examples such as (i) are encountered in written texts, where both conjuncts are pronouns and the first pronoun in the nominative. Such examples are problematic for any account that proposes that case is assigned to all the conjuncts. My consultants, however, find such examples ungrammatical.
(i) Iron Ossetic ɐž ɐmɐ dɐw-ɐn I.nom and you-dat 'for me and you' 'there', historically the inessive of the third person singular pronoun. As it has acquired an idiosyncratic lexical meaning, 14 it cannot serve as the remainder in (33a). 15
there and M-dat flowers gave 'S/he gave flowers to her and Madina.' b. wəm-ɐn ɐmɐ mɐdinɐ-jɐn didinʤətɐ ratta. s/he-dat and M-dat flowers gave 'S/he gave flowers to her and Madina.
To illustrate the working of (G), the ban on emphasis on both disjuncts, consider the pairs of sentences in (34). In (34a), SA with two coordinated pronouns is attempted. The sentence is well-formed with respect to constraints (A-F). It is nevertheless ungrammatical, while its counterpart without SA (34b) is fine. Likewise, (34c) is an attempt to do SA with two DPs marked with the enclitic =dɐr and coordinated by the conjunction 'and'. The control (34d) is the same sentence without =dɐr, where SA is fully grammatical. Finally (34e-f) illustrate the same effect for 'neither … nor' coordination.
The reasons for this phenomenon will be addressed in Section 5.2.

Suspended affixation in alternative questions
Crucially for the argument of this paper, SA in Ossetic is also possible in alternative questions, which, as we have seen, are derived by ellipsis in these languages. In written texts, suspended affixation is admittedly rare in such contexts, however, it is judged possible by native speakers both in unreduced alternative questions, 16 (35), and different varieties of reduced alternative questions, (36-38).
In (35a), the ablative marker can be omitted on the DP 'dog'. The ablative is lexically assigned by the verb 'to fear' to the cause of fear. In (35b), the oblique, which in this case functions as the inessive, can be omitted on the DP 'Vladikavkaz'.
Soslan-abl or.q Khetag-abl 'Of Soslan or of Khetag?' To illustrate SA in split questions, in the question in (37a), the oblique marker, which can be omitted on the first disjunct ' b. kɐmɐj tɐršəš alan(-ɐj) ɐvi šošlan-ɐj? who.abl you.fear Alan-abl or.q Soslan-abl 'Who are you afraid of, Alan or Soslan?' To illustrate SA in or-sluices, in the sluice 'in Matsuta or in Dzinagha' in (38a), the oblique serves as the inessive. It can be dropped on 'Matsuta'. In (38b), the sluice is 'Madina or Fatima' and the suspended case is the allative, which is lexically assigned by the verb 'to wait'. In (38c), the sluice is 'Madina or Fatima', and the oblique, which can be suspended, functions as the accusative. Finally, in (38d), the sluice is 'to Sarmat or to Uruzmag' and the suspended case, the allative, is assigned to the addressee of the verb 'to talk'.

Suspended affixation in alternative questions: Implications for analyses
Let us now see why existence of SA in alternative questions that are derived by ellipsis compels us to a deletion analysis of the question. Given that the DPs ostensibly disjoint in an alternative question in actuality do not belong to the same VP, any approach that analyzes SA as assignment of case to the entire &P will wrongly predict SA to be impossible in alternative questions. Now let us see how the exponent deletion analysis derives suspended affixation in Ossetic alternative questions. For instance, consider the derivation of the sentence in (35a). At the first stage, an alternative question will be derived with case affixes on both DPs, (39). To repeat, the precise mechanism of case assignment is immaterial for our current purposes. To be specific, I represent the structure as a disjunction of two VPs, but nothing will substantially change in the argument if larger constituents are to be disjoined. As I have argued in section 3.1, the verb deletes in the first of the VPs in (39).
(39) 21 At the first stage, an alternative question will be derived with case affixes on both DPs, (39). To repeat, the precise mechanism of case assignment is immaterial for our current purposes. To be specific, I represent the structure as a disjunction of two VPs, but nothing will substantially change in the argument if larger constituents are to be disjoined. As I have argued in section 3.1, the verb deletes in the first of the VPs in (39 , modulo an additional step, ellipsis of everything except the DP in the other disjunct.
Accordingly, no matter what a concrete technical implementation, suspended affixation in alternative questions can only be achieved by morpheme deletion 17 . In the next section, I lay out a specific proposal to this effect.

Proposal
In this section, I formulate the proposal regarding the place of ellipsis in the syntaxmorphology interface and address the licensing conditions that need to hold for this variety of ellipsis to occur. In the next section, I will show how this proposal deduces the properties of SA listed in Section 4.

Place of deletion in derivation
I adopt the standard assumptions of distributed morphology, Halle & Marantz (1993;1994); Halle (1997);Noyer (1997); Embick & Noyer (2001); Siddiqi (2010); and Arregi & Nevins (2012). I assume thus that syntactic computations operate with feature bundles, while vocabulary insertion of the exponents occurs late in the derivation. Vocabulary insertion is followed by phonological readjustments (application of morpheme-specific rules in terms of Arregi & Nevins 2012). The existence of this stage is motivated by intuitively clear difference between the choice of phonologically unmotivated allomorphs (e.g. I vs. me) and phonologically motivated changes (e.g. the English indefinite article taking the form a before consonant-initial words and an before vowel-initial ones).
Under these assumptions, it makes sense to ask when deletion occurs. The logically possible options are deletion of feature bundles prior to vocabulary insertion 18 , deletion of inserted exponents before phonological readjustments, and deletion after the readjustments have taken place.
I propose that morpheme deletion proceeds after vocabulary insertion but prior to application of morpheme specific rules. Other possibilities necessarily lead to incorrect empirical predictions. (35b) and (36-38) is analogous to the one shown above, modulo an additional step, ellipsis of everything except the DP in the other disjunct.

Derivation of sentences in
Accordingly, no matter what a concrete technical implementation, suspended affixation in alternative questions can only be achieved by morpheme deletion. 17 In the next section, I lay out a specific proposal to this effect.

Proposal
In this section, I formulate the proposal regarding the place of ellipsis in the syntax-morphology interface and address the licensing conditions that need to hold for this variety of ellipsis to occur. In the next section, I will show how this proposal deduces the properties of SA listed in Section 4.

Place of deletion in derivation
I adopt the standard assumptions of distributed morphology, Halle & Marantz (1993;1994); Halle (1997);Noyer (1997); Embick & Noyer (2001); Siddiqi (2010); and Arregi & Nevins (2012). I assume thus that syntactic computations operate with feature bundles, while vocabulary insertion of the exponents occurs late in the derivation. Vocabulary insertion is followed by phonological readjustments (application of morpheme-specific rules in terms of Arregi & Nevins 2012). The existence of this stage is motivated by intuitively clear difference between the choice of phonologically unmotivated allomorphs (e.g. I vs. me) and phonologically motivated changes (e.g. the English indefinite article taking the form a before consonant-initial words and an before vowel-initial ones).
Under these assumptions, it makes sense to ask when deletion occurs. The logically possible options are deletion of feature bundles prior to vocabulary insertion, 18 deletion of inserted exponents before phonological readjustments, and deletion after the readjustments have taken place.
I propose that morpheme deletion proceeds after vocabulary insertion but prior to application of morpheme specific rules. Other possibilities necessarily lead to incorrect empirical predictions.
If we assume that deletion precedes VI, we would expect allomorphy contexts to be destroyed, and, in particular, the default allomorphs of pronouns, rather than their oblique forms, to surface under SA, as shown in (31), repeated here as (41), for the 2sg pronoun.

(41)
Digor Ossetic dɐw(-bɐl)/*du ɐma mɐdinɐ-bɐl isɐmbalttɐn. you.obl-sup/you.nom and M-nom-sup I.met 'I met you and Madina.' If deletion had occurred after the application of morpheme specific rules, we would expect results of readjustments to remain visible after SA. For instance, as we have seen in Section 2.1 the oblique suffix in Iron causes affricativization of velar stops (42a). However, under SA, the respective stems surface in their underlying form: in (42b) it is park rather than *parʧ in the first conjunct. The derivation is thus proposed to proceed in the following stages.
Step I. Ellipsis that forms alternative questions occurs.
Step II. Vocabulary insertion.
Step III. Checking the licensing condition and morpheme deletion Step IV. Phonological readjustment Given that gapping, including backwards gapping, is insensitive to the PF identity, it is reasonable to assume that it occurs prior to VI. If one adopts the proposal of Murphy (2016) who treats gapping as VI of null exponents, one can unify Stage I and Stage II of this derivation.

Licensing conditions
I propose that two conditions need to hold in order for morpheme deletion to proceed in Ossetic: the underlying form of the deleted affix must be identical to its overt correlate, and additionally, the coordinands or disjuncts should not all bear [+EMP] feature. Full pronouns carry this feature as part of their lexical specification; while on lexical DPs, it is marked with the enclitic =dɐr. I stay agnostic as to the precise semantic interpretation of [+EMP] feature. The reason to postulate it is that both =dɐr marking and use of full pronouns is associated with some kind of prominence that is hard to precisely capture at the present stage of research.
As we have seen in Section 2, Ossetic languages use several series of personal pronouns. For subjects, Romance-type pro-drop occurs. For non-subject pronouns, the clitic forms are used by default. A large amount of generative literature has strived to pin down the difference between clitic and non-clitic pronouns in such languages, starting at least with Cardinaletti & Starke (1999), see also Neeleman & Szendrői (2007) and Sigurðsson (2011). Although the difference is often identified with something like contrast ("strong" pronouns tend to appear when some contrast is present), see, a.o., Öztürk (2001);Frascarelli (2007); Frascarelli & Hinterhölzl (2007); and Sigurðsson (2011), it is not clear whether they only appear in the presence of contrast, see examples in Cardinaletti & Starke (1995: 163) for Romance languages and in Bohnacker (2013) for Swabian, a Southern German variety, which, as Bohnacker argues, exhibits a Romance-type pro-drop.
The idea to include the respective prima facie information-structural feature in the lexical makeup of a pronoun was advanced by Herbeck (2016) who proposes that, in Spanish, overt subject pronouns have feature [+Foc] in their feature bundle. I propose that a purely formal [+EMP] feature underlies the occurrences of strong pronouns in Ossetic.
As for the enclitic =dɐr, besides its occurrences in coordinations, it also may be used in the capacity of scalar particles 'too' and 'even'. Additionally, it marks universal quantifiers in the absence of the clausal negation. Provisionally, I propose that all these uses involve the same feature [+EMP] as overt pronouns.
If this proposal is on the right track, a plausible reason why a considerable number of speakers disallow SA in alternative question can be that, for such speakers, the disjuncts of an alternative question both carry [+EMP] feature.

Directionality of morpheme deletion
In Ossetic, suffix deletion only proceeds backwards, nor am I aware of any cross-linguistic evidence for forward suffix deletion, as schematized in (43).

(43) *X-aff & Y-aff
At present, SA is the only known ellipsis variety in Ossetic that may only proceed backwards. Gapping and sluicing that can proceed in both directions in Ossetic. On the other hand, stripping can only proceed forward, as the contrast illustrates between the grammatical sentence in (44a) and the ungrammatical one in (44b). This shows that Ossetic in principle allows unidirectional ellipsis types, and it is not particularly surprising that there exists one that only can proceed backwards. Correctly predicting the cross-linguistic variation of ellipsis directionality is a problem for all existing accounts of ellipsis, and I leave the matter for the further research.

Deriving the generalizations
With a proposal at hand about the process in the narrow grammar that is responsible for SA, I now proceed to showing that the properties of SA in Ossetic (described in Section 4) are predicted by my proposal coupled with some standard assumptions about the syntaxphonology interface and processing requirements. Property (B) was discussed in Section 5.3 above.
Property (C) that only case markers can undergo SA, needs a longer discussion and is addressed in Section 6.2 below.
Property (D) that remnants under SA are substrings of the respective full forms, follows immediately from the assumption that deletion occurs after VI.
Property (E) that requires a remnant under SA to be a free-standing word is discussed in Section 6.1.
Property (F), the ban on words with an idiosyncratic meaning as SA remnants, has to do with exigencies of processing rather than with grammar in the proper sense, cf a similar proposal of Kabak (2007) regarding the ban on bare verb stems to occur as remnants under SA in Turkish.
Property (G), the ban for all coordinands to be pronouns or be marked with the clitic =dɐr, was addressed in Section 5.2 and was argued to follow from the licensing requirements on ellipsis.

Ban on non-words as remnants under suspended affixation
I propose that the ban for non-words to serve as SA remnants ensues from a general property of the syntax-phonology interface: I take being able to occur as a free-standing word a lexical diacritic, and only free-standing words can be coordinated in Ossetic.
Languages vary on whether this constraint is absolute: Polish, for instance is reported to have constructions where a verbal prefix, for instance prze-in (45), is ostensibly coordinated with a phrase.

(45)
Polish (Citko 2017) Jan prze-a Piotr pod-pisał list do prezydenta. Jan prv-ctr Peter prv-wrote letter to president 'Jan copied and Peter signed the letter to the president.' Ossetic, although possesses a system of verbal prefixes similar to a Slavic one, disallows such constructions. The ungrammatical (46) is a direct counterpart of (45). In (46), a stranded preverb nǝ-stands for the phrase nǝ-ffǝšta pišmo prezident-mɐ 'wrote a letter to the president', and is coordinate with the phrase ra-fǝšta pišmo prezident-mɐ 'copied a letter to the president'.

Restriction of suspended affixation to case markers
As we have seen, suspended affixation in Ossetic can only target case markers: unlike in Turkish, the nominal plural marker or the verb agreement suffixes cannot undergo SA.
The ban on suspended affixation of plural markers is motivated by the requirement for an SA remnant to be a word: on their own, plural stems are not independent words.
The plural is formed from the plural stem and the plural suffix -tɐ. Plural stems often undergo small, but unpredictable changes as compared to singular stems. For instance, in Iron, qug 'cow' yields in the plural quʦ:itɐ 'cows' with a change g > ʦ:i; but dug 'epoch' becomes dugtɐ 'epochs', without any change in the stem. Likewise, in the plural stem of gal 'bull' no changes occur: the plural is galtɐ 'bulls', but in a word with a similar final syllable a change does occur: ɐmbal 'friend', ɐmbɐlttɐ 'friends: the a of the final syllable becomes ɐ and the -t-of the plural marker reduplicates. Accordingly, the plural stems must be inserted as allomorphs at the stage of vocabulary insertion, and phonological readjustment cannot undo these changes.
Suspended affixation is impossible for any verb forms either (47). I argue that this follows from the requirements for the remnant to be an extant word (already discussed in Section 6.1) and to not have an idiosyncratic meaning of its own.
(47) a. Iron Ossetic sǝd-*(tɐn) ɐmɐ kʷǝd-tɐn go.pst-pst.1sg and cry.pst-pst.1sg 'I was walking and crying' b. Digor Ossetic cɐw-*(gɐ) =ma kɐw-gɐ go.prs-cvb and cry-prs.cvb 'walking and crying' All the verb forms are based on either the present stem or the past stem. For a given verb, its past stem is not synchronically predictable as illustrated by the three Iron forms in Table 6: the present stems only differ in the initial consonant, but the past stems are formed differently for all the three. The natural conclusion is the relationship between a present stem and the past stem is listed lexically. The status of present stems is somewhat different in Digor and Iron: in Digor, they cannot function as independent words, while in Iron they coincide with the second person imperative form. 19 Accordingly, what prevents SA with present stem verbs in Digor is that the would-be remnant would not be a word, while in Iron, the remnant would be interpreted as the imperative.
The past stem, for the vast majority of the verbs, coincides with the past participle, which obviously is a word. However, the past participle on its own has the distribution of a noun. I propose that this is what prevents SA with verb forms created from the past stem.

Challenges
In this section, I discuss several finer points of Ossetic morphology that ostensibly are problematic for the analysis advocated for in this paper and show that they can be accounted for by the current proposal. 19 In Digor, the imperative carries a dedicated suffix -ɐ: kɐn-ɐ do-imp.2sg.

Suspended affixation and numeral phrases
Belyaev 20 (2014) noticed an interesting pattern in SA with numeral phrases, which, as an anonymous reviewer suggests, can be a problem for the analysis presented in this paper. Before presenting the relevant data, and arguing why they do not actually pose a problem for my analysis, I need to recall some facts about numerals and numeral phrases in Ossetic already touched upon in Section 4.1.
The numeral precedes the noun phrase. When the entire phrase stands in the nominative, the (phrase-final) noun gets marked with a suffix that is homophonous to the oblique case marker, -ə in Iron and -i in Digor (48a, c). The usual plural marking of the noun is impossible in the presence of a numeral (48 b, d).
(49) Iron Ossetic a. dəwɐ bɐχ-ɐj two horse-abl 'from two horses' b. dəwɐ bɐχ-ɐn two horse-dat 'for two horses' c. dəwɐ bɐχ-aw two horse-equ 'as two horses' In Digor, on the other hand, numerals have a separate inflectional paradigm, and it is an exponent from this paradigm that appears on the noun in a numeral phrase, as illustrated in Table 7. Specifically, a suffix -e-or -em-, depending on the case, is inserted between the stem and the case marker. Now, as was observed in Belyaev (2014), when two numeral phrases are coordinated, for many speakers it is the numeral suffix (i.e. the one homophonous to the oblique case) rather than the nominative (or null) marking that emerges under SA. In (50a) and (50c), the superessive gets assigned by the verbs 'to meet ' and 'to bite', respectively; and in (50b) the ablative is assigned by the preposition 'without'.

(50)
Digor Ossetic a. ɐrtɐ kižg-i/?kizgɐ ɐma duwɐ biččew-e-bɐl isɐmbaldtɐn. three girl-num/girl.nom and two boy-num-sup I.met 'I met three girls and two boys.' b. ɐnɐ cuppar bɐχ-i/ ?bɐχ/ bɐχ-e-mɐj ɐma gal-ɐj. without four horse-num horse/ horse-num-abl and bull-abl 'without four horses and a bull.' c. Iron Ossetic asə kʷəz faron ɐrtɐ lɐppu/-jə/-jəl ɐmɐ čəžg-əl fɐχasədi. this dog last.year three boy/-num/-sup and girl-sup bit 'This dog last year bit three boys and a girl.' Given that the numeral suffix is homophonous with the oblique (i.e. genitive/accusative) case marker (-i in Digor and -ǝ in Iron), it seems prima facie natural to identify it with this case marker. Under this assumption, its behavior under SA is not what the system I have laid out here predicts. For instance, in (50a), when the superessive case is assigned to both conjuncts, we obtain (without suspended affixation): Digor Ossetic ɐrtɐ kižg-e-bɐl ɐma duwɐ biččew-e-bɐl three girl-num-sup and two boy-num-sup 'on three girls and two boys' Under the assumption that the marker -ǝ/-i assigned by the numeral is case, we must conclude that, as a result of deletion, the first conjunct would have to lose overt case marking (52a), which is not what happens for many speakers (52b): Digor Ossetic a. ?ɐrtɐ kizgɐ ɐma duwɐ biččew-e-bɐl three girl and two boy-num-sup b. ɐrtɐ kižg-i ɐma duwɐ biččew-e-bɐl three girl-num and two boy-num-sup However, it is plausible that the numeral suffix is not a form of case marking, but a different entity, 22 specifically, a form of plural marking that is only used in the presence of 21 As was mentioned above, the numeral paradigm in Digor is one of the very few instances where the inessive form differs from the oblique one. 22 Assuming which I concur with Belyaev (2014). numerals. Accordingly, SA is not expected to apply to it: (52a) will be ill formed because of the absence of the plural marking. 23 Another way to explain the grammaticality of (50) could be to assume that the external case is only assigned to the second conjunct in (50a-b). However, this is not what happens in Ossetic. As we have seen in Section 4.1, if the case could be assigned only to the second conjunct, we would expect to obtain, alongside with (53a), also (53b), with the pronoun in the nominative. However, such sentences are ungrammatical. To conclude, the idiosyncratic behavior of numeral morphology does not present a problem to deletion-based analyses of SA in Ossetic.

Behavior of ɐ-final stems
When a noun with an ɐ-final stem participates in SA, the remnant can ostensibly violate generalization (D) that the remnant under SA must be a substring of the corresponding full form (54). Such nouns include plural stems in both languages and ɐ-final singular stems in Digor. 24 This is one important situation where the relative timing of ellipsis and phonological readjustments plays a role is SA with ɐ-final stems. Under SA, when a noun of this type is (the head of) the first conjunct, the final -ɐ obligatorily resurfaces (54).
(54) a. Digor singular zɐrdɐ (ɐ)ma wod-i heart and soul-obl 'in the soul and the heart' b. Digor plural kižgi-tɐ (ɐ)ma biččew-t-ɐn girl-pl and boy-pl-dat 'for girls and boys' c. Iron plural k'oʁo-tɐ (ɐ)mɐ žok'o-t-ǝl mosquito-pl and mushroom-pl-sup 'about mosquitoes and mushrooms' On my proposal, this is accounted for by the assumption that deletion of a case morpheme occurs before readjustments take place, and that deletion of the stem-final -ɐ in front of a vowel-initial suffix is a readjustment that occurs to resolve hiatus. If the case suffix is deleted, no need for a readjustment arises, and the final -ɐ is retained.
When discussing morphophonology of case marking in Section 4.1, we have postponed examination of ɐ-final stems. Now is the time to resume this discussion. When such a noun takes a vowel-initial case suffix, the final -ɐ of the stem disappears, Table 8.
With consonant-initial case stems, that is, the allative -mɐ in Digor and -m in Iron, and the superessive -bɐl in Digor, the final -ɐ is retained, Table 9.
Under the current proposal, the derivation of the normal Digor oblique zɐrdi heart.obl and of its occurrence in the suspended affixation in (54a) proceeds in the following steps. In the absence of suspended affixation, vocabulary insertion produces zɐrdɐ-i, which undergoes readjustment to result in zɐrdɐ-i. In a suspended affixation configuration, the derivation proceeds as shown in (55): vocabulary insertion again produces zɐrdɐ-i, after which morpheme deletion applies.
(55) a. Vocabulary insertion zɐrdɐ-i (ɐ)ma wod-i heart-obl and soul-obl b. Deletion zɐrdɐ-i (ɐ)ma wod-i heart-obl and soul-obl This is not the only situation when a morpheme-final -ɐ gets deleted to resolve hiatus: the same process occurs, for instance, with the prefix ɐnɐ-'without': it surfaces as ɐnɐ-before consonants and as ɐn-before vowels, Table 10.   Accordingly, interpreting the deletion of -ɐ before vowel-initial suffixes as a phonological readjustment is independently motivated.

Case marking of reflexives
Finally, Belyaev (2014) reports that SA is possible for some speakers of Iron Ossetic 25 when the first conjunct is a reflexive pronoun. An anonymous reviewer suggests that these data are problematic for the current proposal given the peculiar morphology of the case marking in reflexives, see Table 3 below. In this section, I advance a proposal about the representation of the reflexive morphology in the grammars of such speakers.
For speakers who allow SA with reflexives the bare stem χiemerges in the first conjunct. For instance, in (56a), the first conjunct is mɐ=χi 1sg=refl instead of the expected ablative-marked form mɐ=χisaj 1sg=refl.abl. Likewise, in (56b), the first conjunct is the same bare stem, instead of the superessive-marked mɐ=χiwǝl 1sg=refl.sup. These facts are surprising because case paradigms of reflexives in Ossetic are somewhat peculiar. Reflexives are formed by the reflexive stem χe-(Digor) / χi-(Iron) and a possessive prefix that expresses the phi-features of the binder: dɐ=χi 2.sg=refl.obl 'yourself' (Iron). The inflection paradigms of the reflexive in Iron and Digor are given in Table 11.
These paradigms show two remarkable features: first, in the dative and the ablative, -c-/-s-appears between the stem and the case marker; second, in the Iron superessive, -wemerges in the same position.
Prima facie, there are three possible ways to account for this phenomenon. First, one can posit stem allomorphs χis-/χecand χiwfor the respective cases. Second, one can posit the existence of allomorphs -sɐj/-cɐj of the ablative suffix; -sɐn/-cɐn of the dative suffix and -wǝl of the superessive suffix. Third, insertion of -c-/-s-and -w-can be treated as epenthesis 26 that resolves hiatus that is created between the reflexive stem χiand the 25 No one among my Digor consultants allows SA in this case. 26 Epenthesis of a non-homorganic segment, in particular of a coronal such as -s-is not uncommon crosslinguistically, see a discussion and examples in Lombardi (2002). vowel-initial case suffixes -ɐn, -ɐj, and -ǝl. It is possible, moreover, that different speakers can have different representations of these consonants in their grammars.
If the consonants -c-/-s-and -w-are parts of the stem, the current account predicts SA with reflexives to be impossible for these cases. We would have expected to obtain χec-/χisand χiwas the remnants under SA, but they are not independent words and the derivation would have crashed.
Therefore, it occurs at the stage of phonological readjustment, and, if the case suffix is deleted at the preceding stage, epenthesis is no longer motivated and does not occur at all.
The interpretation of -c-/-s-as an epenthetic consonant is independently motivated by the fact that, in Iron, an epenthetic -s-also appears between certain vowel-final preverbs 27 and the stem in some a-initial verbs in Iron, Table 12.
A reviewer objects that this leads to a single root χi using two different epenthetic consonants, -w-and -s-. Diachronic reasons for this are clear: historically, as the reviewer correctly remarks, -w-is a reflex of the initial -b of the superessive suffix. As long as the case suffix was stop-initial, no epenthetic consonant was needed. Synchronically, it is reasonable to assume that -w-has gotten reanalyzed as an epenthetic glide: for the reciprocal kɐrɐzi it is already in free variation with the default glide -j-. Additionally, w is in variation with j or null in widon; (j)idon 'bridle' widaz/(j)idaz 'reins ' Abaev (1989: 105-106). It also appears in the onset in wasin 'ladder' in the Kudar dialect of Iron spoken in South Ossetia, as opposed to ašin in North Ossetian Iron, (Cheung 2008: 215, the editor's footnote).
To recapitulate, I propose that speakers of Iron who allow SA with reflexives treat -wand -s-that emerge in different paradigm cells as epenthetic consonants and insert them at the stage of phonological readjustments.

Alternative analyses
In Section 4.3, I have already argued against one type of possible analyses of SA in Ossetic, namely the one that is based on case assignment to the entire &P or orP. I have shown that this is incompatible with the data from alternative questions brought up in this paper. I overview here, in Section 8.1, a specific implementation of this idea by Belyaev (2014). In this section, I address two more possible alternative analyses of SA in Ossetic. In Section 8.2, I explore the option that single features rather than morphemes can be deleted under 27 An anonymous reviewer objects to this interpretation on the grounds that -s-is inserted after the consonantfinal preverb ɐr-as well. However, historically this preverb apparently was r̩ -, i.e. it consisted of a syllabic /r̩ /, which was phonemic at that stage, Cheung (2002). At that stage, insertion of an epenthetic after /r̩ / was phonologically motivated. Digor completely lost -c-epenthesis after preverbs, and -j-or -ʔ-are inserted there instead.  Mukai (2003) and by  for certain ellipsis-like constructions in Japanese and Korean.

Two-layer case assignment to &P or orP
I overview here the analysis of Belyaev (2014)  Ungrammatical coordinations of a nominative pronoun with a non-nominative DP, such as illustrated in (53b) above, will be ruled out because the coordinands will not match in the feature [Obl]/ [Dir].
Under the assumption that case in SA construction is only exponed on the rightmost edge of &P, the retention of the numeral morphology under SA discussed in section 7.1 above is a priori not a problem.
Finally, to the best of my understanding, the approach of Belyaev (2014) does not predict the ungrammaticality of SA for two coordinated pronouns, or two lexical DPs marked with =dɐr. Probably, the proposal could be modified to account for this fact, but the fact that the outer case attaches to &P as a whole appears to be crucial for Belyaev (2014).

Deletion of features
On the current proposal, deletion targets morphemes. One can envisage an alternative system where features are allowed to delete one by one, irrespective of whether the deleted feature corresponds to a vocabulary item. Such a system, however, will vastly Erschler: Suspended Affixation as Morpheme Ellipsis Art. 12, page 31 of 41 overgenerate. For Ossetic, one immediate prediction of a deletion by feature analysis concerns the nominal plural. If the deletion had targeted single features and thus were unrelated to vocabulary insertion, we would expect that the nominal plural be deletable. However, as we have seen in Section 6.2, this is not the case. Further examples of overgeneration require a more careful examination of the relevant feature geometry. I will address here two such instances: of possessive marking in Turkish compounds, where Kharytonava (2012) proposed a feature deletion analysis, and, closer home, of SA in Ossetic.
The relevant Turkish facts are the following. In a non-coordinated compound without a possessor, the 3rd person possessive suffix appears on the head of the compound, -i in (58a). In possessed compounds, it gets replaced by the possessive suffix whose phifeatures match those of the possessor, -iniz in (58b).
(60) Turkish a. Kharytonava (2012: 175) (siz-in) doǧum yer-i ve tarih-iniz you.pl-gen birth place-poss.3sg and date-poss.2pl 'your birth place and date' b. Kharytonava (2012: 175) (siz-in) doǧum yer ve tarih-iniz you.pl-gen birth place and date-poss.2pl 'your birth place and date' Kharytonava (2012) proposes a natural feature representation of the Turkish possessive suffixes. Underlyingly, the head yer 'place' of the first coordinand in (59) are assumed to carry the features corresponding to the 3rd person singular and the 2nd person plural. The 2nd person plural suffix is assumed to be inserted as more fully specified, according to the standard Panini principle. (60a) is derived if the features corresponding to the 2nd person plural suffix are all deleted simultaneously, while in (60b) all the possessive phi-features have deleted.
However, under such an analysis, the fact is accidental that both the features corresponding to the 2nd person and the plural are deleted simultaneously. Should deletion indeed proceed feature by feature, we would expect that the possessor's feature [+pl] could be deleted on its own. In the case of the sentence in (59), this would produce the sentence in (61) with the 2nd person singular suffix on the first conjunct. However, this sentence is ungrammatical.

(61)
Turkish *(siz-in) doǧum-in yer ve tarih-iniz you.pl-gen birth-poss.2sg place and date-poss.2pl 'your birth place and date' (intended) To ensure that only grammatical sentences in (60) be derived, we must require that ellipsis may only target combinations of features that correspond to actual morphemes. This phenomenon provides evidence for the reality of morphemes, contrary to what is proposed in purely realizational, in the sense of Stump (2001), theories of morphology.
Turning to the case of the SA in Ossetic, let us assume, for the sake of the argument, that feature deletion may follow gapping that creates alternative questions. To compare morpheme-based and feature-based deletion approaches, we need to specify a feature decomposition of cases in Ossetic. Recall that the case inventory of Ossetic is comprised of the nominative; accusative and genitive, which coincide for most lexical items; inessive; dative; ablative; superessive; and equative.
Given that it is the nominative forms that are used as the default forms (in the sense of Schütze 2001), it is natural to consider them unmarked for morphological case, 28 that is to assume that they do not bear any morphological case feature.
Then, if we assume that each non-nominative case corresponds to an atomic feature of its own, say [+Dat] for the dative, [+Sup] for the superessive, etc, we predict that, under SA, the nominative forms of pronouns will surface. As we have seen, this prediction is not borne out.
Therefore, it is necessary to posit a richer structure of case features as shown in Table 13. One must stipulate then that only the external case feature may be deleted (to avoid surfacing of nominative pronominal forms as first conjuncts), and, moreover, that this must be the only type of feature in Ossetic that can undergo deletion, to avoid predicting SA of anything other than case suffixes.
The first conjunct under SA will then only carry the case feature [+NNom], and this will be the only situation when nominals with such a feature specification surface. 29 With this stipulation, this system would be essentially equivalent to the one developed in this paper.

One fell swoop deletion of a verb and a case marker
Crucially, the current proposal assumes that SA in alternative questions proceeds in two stages: verb deletion (which occurs relatively early in derivation) and affix deletion (which occurs after VI). This is what distinguishes the current proposal from some analyses of somewhat similar phenomena in Japanese and Korean. 28 I adopt the approach that distinguishes the abstract case and the morphological case that surfaces on the nominals, McFadden (2004); Legate (2008), and the ensuing literature. To repeat, I stay agnostic as to the way case is assigned in syntax. 29 Belyaev (2014) argues that this form occurs as the complement of the preposition ɐnɐ 'without'. For lexical nouns, the complement of this preposition either lacks overt case marking ɐnɐ bɐχ without horse 'without a horse' or is marked with the ablative ɐnɐ bɐχ-ɐj without horse-abl. The ablative marking is optional. For personal pronouns, the choice is between the oblique form ɐnɐ mɐn without I.nnom 'without me' and the ablative form ɐnɐ mɐn-ɐj without I.nnom-abl. However, for all wh-based items the case assigned by ɐnɐ is obligatorily ablative: ɐnɐ kɐmɐj without who.abl 'without whom'; ɐnɐ iš-kɐmɐj without idf-who.abl, etc. However, if ɐnɐ were able to only assign the feature [+NNom], we would have wrongly expected the nonnominative stems of wh-words to be able to combine with it: *ɐnɐ kɐ(m).