2.1 Network dynamics and negotiation timing

Civil wars are increasingly multi-actor conflict systems rather than simple state–rebel dyads. Armed actors enter, split, align, and redirect violence, and these shifts can reshape bargaining conditions (Cunningham et al., Reference Cunningham, Bakke, Seymour, Pearlman and Cunningham2012; Dorff et al., Reference Dorff, Gallop and Minhas2020; Metternich et al., Reference Metternich, Dorff, Gallop, Weschle and Ward2013). Network approaches capture these interdependencies: when one dyad changes, other actors update strategies in response to shifting coalitional balance and local rivalries (Bakke et al., Reference Bakke, Cunningham and Seymour2012; Rudloff and Findley, Reference Rudloff and Findley2016).

A growing literature examines the conditions under which negotiations begin, linking onset to military stalemates, third-party intervention, rebel organizational characteristics, disasters, and the strategic calculations of internally divided movements (Cunningham, Reference Cunningham2011; Kaplow, Reference Kaplow2016; Rosvold and Rustad, Reference Rosvold and Rustad2026). Peace-process scholarship also shows that conflict resolution unfolds across three distinct stages: entering talks, reaching agreement, and implementing terms. The correlates of each can differ. Walter (Reference Walter2002); Findley (Reference Findley2013). Research on fragmentation shows that internal divisions can intensify commitment and information problems, but fragmentation does not uniformly prolong war or preclude negotiated outcomes (Bakke et al., Reference Bakke, Cunningham and Seymour2012; Cunningham, Reference Cunningham2013). Yet, this work has largely focused on dyadic or actor-level predictors rather than system-wide network reorganization. Our contribution is to identify a network-level micro-sequence within the transition to talks: not whether fragmentation helps or hurts, but when and how the timing and sequence of network reorganization create a window for onset.

Despite advances in mapping conflict networks, we lack systematic evidence on how changes in network structure shape the timing of negotiation onset. Bargaining and ripeness traditions offer apparently opposing predictions. Bargaining theory starts from the puzzle that mutually preferable bargains should often exist, yet actors still fail to reach them; in civil war, the core barriers are private information and credible-commitment problems (Fearon, Reference Fearon1995; Walter, Reference Walter1997). In a multi-actor extension of that logic, more legible and organizationally coherent conflict environments should reduce barriers to negotiation by clarifying who can commit and who can implement, though bargaining logic alone does not specify when actors will choose to incur the costs of opening talks (Kaplow, Reference Kaplow2016). Ripeness theory addresses timing directly: serious negotiations become likely when parties perceive both a mutually hurting stalemate and a “way out” (Zartman, Reference Zartman1985, Reference Zartman, Stern and Druckman2000). Our network measures are not direct observations of those perceptions but observable proxies for conditions likely to shift actors’ assessments of stalemate and feasibility in multi-actor settings.

These approaches ask fundamentally different questions. Bargaining asks what prevents settlement given willingness; ripeness asks what prevents willingness given the ability to settle. We argue below that sequencing the two logics resolves the apparent contradiction. Disruption generates the pressure side of ripeness (creating motivation), while recovery provides the network form of a way out by restoring the legibility that bargaining logic identifies as necessary.

We conceptualize durability as the extent to which dyadic conflict structure persists over time. High durability means that the set of fighting dyads is similar across periods and that actors’ relative conflict involvement is broadly consistent. Low durability indicates turnover in who fights whom and changes in relative engagement. Importantly, instability can reflect two distinct features: a low level of durability (a system that is persistently fluid) and a trajectory of change (a system that is deteriorating or recovering). A conflict can be chronically low in durability yet improving or relatively durable yet deteriorating. We therefore distinguish durability level from durability trajectory and test whether negotiation onset is more closely associated with the state of the conflict system or its direction of movement.

2.2 Pressure: How deterioration shocks create incentives

We outline two complementary forces that shape negotiation timing in multi-actor conflicts: pressure from deterioration shocks and feasibility from subsequent recovery. The framework is probabilistic since it identifies conditions under which onset is more likely, not conditions that make onset inevitable.

Durable conflict networks support entrenched patterns of violence. When dyadic structure is predictable, actors can adapt militarily and organizationally, develop routines, and pursue strategies calibrated to a known environment (Kalyvas, Reference Kalyvas2006; Wood, Reference Wood2003). Over time, these routines can sustain continued fighting even when peace would be preferable in the abstract. Entrenchment is not an information problem (actors in stable conflicts may have clear information about capabilities and partners) but a reservation-value problem. In other words, continuation payoffs in the known environment exceed expected payoffs from uncertain negotiations, so nobody comes to the table.

A key implication is that incentives to reconsider strategies may arise less from chronic disorder than from deterioration shocks that invalidate recent expectations. This is the logic that ripeness accounts identify: disruption breaks the sustainability of the status quo, lowers continuation payoffs, and makes actors more willing to consider alternatives (Zartman, Reference Zartman1985, Reference Zartman, Stern and Druckman2000). When durability declines relative to the recent past, actors face sharper reassessment problems. Coalition partners are less reliable, recent bargaining expectations become harder to extrapolate, and the set of active opponents may shift quickly. These shifts increase uncertainty about future outcomes and weaken confidence in battlefield victory, widening the perceived space for bargaining (Fearon, Reference Fearon1995; Powell, Reference Powell2004). They can also impose distributed costs across multiple actors when realignments and rivalries intensify, creating conditions that resemble the mutually hurting stalemates emphasized in the ripeness literature. The mechanism is not chronic disorder but a shock relative to the recent past. More precisely, deterioration shocks capture the pressure component of ripeness rather than ripeness in full; they can make continuation less attractive without yet establishing that actors perceive a feasible exit. In our empirical tests, we operationalize pressure as a decline in durability into $t{-}2$ relative to $t{-}3$ ( $D_{t-2} \lt D_{t-3}$ ).

2.3 Feasibility: How recovery restores legibility

Pressure alone is not sufficient. Ripeness theory itself recognizes this. Zartman (Reference Zartman1985) requires not only a mutually hurting stalemate but also a perceived “way out,” a sense that a negotiated settlement is feasible. Bargaining accounts formalize that feasibility requires legibility and organizational capacity (Fearon, Reference Fearon1995; Walter, Reference Walter1997). When dyadic structure is more stable, parties can form clearer assessments of who matters, who can implement, and what kinds of guarantees might be necessary (Cunningham, Reference Cunningham2013; Staniland, Reference Staniland2014). In multi-actor settings, feasibility is not only about relative power; it is also about whether bargaining partners can be identified and whether commitments can plausibly be delivered by the relevant organizations and coalitions. The environment also shapes incentives for brief cessation of fighting. Disasters can create opportunities for ceasefires where rebels want to regroup or gain public support (Rosvold and Rustad, Reference Rosvold and Rustad2026). If disruption alone were sufficient, we would expect chaotic conflicts to produce negotiations, but they often do not, precisely because the uncertainty that motivates talks also undermines the conditions for initiating them.

We propose that a network’s trajectory informs feasibility. A recovering network (improving durability) suggests that a period of restructuring is ending and that dyadic conflict structure is becoming more legible. Even if the system remains far from fully stable, improvement can signal consolidation and provide a clear enough set of salient dyads and coalitional alignments for negotiations to begin.

The crucial distinction is between post-disruption recovery and undisrupted stability. An undisrupted stable conflict has legibility but also entrenchment. Actors can read the landscape but have optimized for it, and their continuation payoffs exceed their negotiation payoffs. A post-disruption recovery has legibility without entrenchment. Actors can read the new landscape (salient dyads and coalitional alignments have clarified) but have not yet had time to dig in and optimize for continued fighting. This provides the clarity that bargaining theory identifies as necessary for talks to work, alongside the motivation that ripeness theory identifies as necessary for talks to start. The window exists because legibility and motivation rarely co-occur. Stable systems have legibility but not motivation, chaotic systems have motivation but not legibility, and post-disruption recovery is the configuration where both are present. It is also temporary, because actors re-entrench as the new configuration stabilizes. We operationalize feasibility as a rebound in durability into $t{-}1$ relative to $t{-}2$ ( $D_{t-1} \gt D_{t-2}$ ).

2.4 The window of opportunity

The window of opportunity hypothesis combines these stages. Pressure without feasibility describes periods in which durability has declined but continues to deteriorate; costs may be rising, but partner identification and credible commitment are less clear. Feasibility without pressure describes periods in which durability is improving without a preceding deterioration shock; bargaining conditions may be clearer, but incentives to initiate talks may be weaker. This is the entrenchment problem in a different guise, as recovery without prior disruption resembles ordinary stability.

The window is the sequence in which prior disruption is followed by recovery. In this configuration, a deterioration shock creates the motivation that ripeness accounts identify as necessary (the pressure stage), and subsequent recovery creates the legibility that bargaining accounts identify as necessary (the feasibility stage). A pure legibility account would predict that recovery should help regardless of prior disruption; a pure pressure account would predict that disruption should matter even absent recovery. Classic ripeness is closer to our full argument because it already requires a perceived way out, but it does not specify what that way out looks like in network terms. Our framework specifies the network-level content of that way out: recovery is the moment when salient dyads, coalition structure, and likely implementers become legible enough for actors to treat negotiation as a viable exit rather than a leap into organizational uncertainty. The sequencing framework thus predicts that recovery matters only following disruption and that disruption is most consequential when followed by recovery. Negotiations should be concentrated in this configuration, though onset remains possible in other configurations. Figure 1 displays the window of opportunity framework.

Figure 1.

Window-of-opportunity framework. The framework distinguishes prior disruption ( $D_{t-2} \lt D_{t-3}$ ) and subsequent recovery ( $D_{t-1} \gt D_{t-2}$ ). The window is the configuration in which disruption is followed by recovery. The framework is probabilistic since onset is expected to be more likely in the window configuration than in the other quadrants.

2.5 Illustrative cases

We illustrate these patterns with three conflicts in Africa. Two cases fall into the window of opportunity framework (Mali 2012 and CAR 2011), and one shows an exception falling outside the window quadrant (Sudan 2012). Of the 9 negotiation onsets in the estimation sample, 5 (55.6%) occur in the window configuration (prior disruption followed by recovery) while 4 occur in other quadrants. Figure 2 displays two window cases and one exception.

Figure 2.

Durability trajectories preceding negotiation onset. Panel A: CAR (2011) illustrates the window pattern with within-year disruption in 2009 followed by cross-year recovery into 2010. Panel B: Mali (2012) illustrates another window pathway with within-year disruption in 2010 followed by cross-year recovery into 2011. Panel C: Sudan (2012) illustrates a case falling in the “neither” quadrant under the window framework. Dashed vertical line indicates negotiation onset.

Mali 2012. Mali’s conflict landscape underwent dramatic transformation in the years preceding negotiation onset (Lecocq et al., Reference Lecocq, Mann, Whitehouse, Badi, Pelckmans, Belalimat, Hall and Lacher2013; Wing, Reference Wing2013). Northern Mali had experienced recurring Tuareg insurgencies since independence (Lecocq and Klute, Reference Lecocq and Klute2013), but the conflict network restructured significantly within 2010 as AQIM (Al-Qaeda in the Islamic Maghreb) consolidated its presence in the north – conducting kidnappings and provoking government and international military responses that altered existing fighting configurations. Within-year network stability collapsed ( $D$ fell from 0.509 to 0.180), reflecting this restructuring of conflict relationships across the northern theater.

By the cross-year transition into 2011, the network had reconsolidated ( $D$ recovered to 0.564) as conflict relationships clarified into more stable configurations. The window pattern is visible: disruption within 2010 created pressure as the old equilibrium broke, and recovery crossing into 2011 provided feasibility as the conflict landscape became newly legible. The situation then escalated sharply. As Libya collapsed into civil war in 2011, heavily armed Tuareg fighters returned to Mali, a process that accelerated after Gaddafi’s fall in October (Lecocq and Klute, Reference Lecocq and Klute2013). The MNLA (National Movement for the Liberation of Azawad) launched a coordinated offensive in January 2012, quickly overrunning government positions. Islamist groups – Ansar Dine, AQIM, and MUJAO (Movement for Oneness and Jihad in West Africa) – cooperated with the MNLA in seizing northern cities but pursued distinct objectives, and by mid-2012 Ansar Dine controlled Kidal and Timbuktu while MUJAO held Gao (Lecocq et al., Reference Lecocq, Mann, Whitehouse, Badi, Pelckmans, Belalimat, Hall and Lacher2013; Wing, Reference Wing2013). Following the March 2012 military coup in Bamako and the MNLA’s declaration of Azawad independence in April, ECOWAS appointed a mediator, and negotiations began in 2012.

Central African Republic 2011. The CAR conflict had reached a violent equilibrium following the 2007 Birao peace agreement. The government of François Bozizé controlled Bangui and major towns while three main rebel coalitions held distinct territories in the north. The Union of Democratic Forces for Unity (UFDR) maintained strongholds in the Vakaga prefecture in the far northeast near the Sudanese border. The Convention of Patriots for Justice and Peace (CPJP), based around Ndele, contested diamond mining areas in the north-central and eastern prefectures. The People’s Army for the Restoration of Democracy (APRD) operated in the northwest (Lombard, Reference Lombard2016; Mehler, Reference Mehler2011). This configuration produced persistent violence that all parties had adapted to survive.

The network disrupted dramatically within 2009. The CPJP, formed in late 2008 by ethnic Runga fighters in opposition to the Goula-led UFDR, escalated its challenge to both the government and rival rebel groups over control of diamond mining areas around Sam Ouandja and Ndele (Spittaels and Hilgert, Reference Spittaels and Hilgert2009). New self-defense militias emerged in response to increased banditry as the security situation deteriorated. Armed groups clashed in new configurations as they competed for control of mining sites and trade routes ( $D$ fell from 0.466 to 0.131). The government, facing proliferating threats on multiple fronts, could no longer maintain its strategy of containing rebels to peripheral zones (Agger, Reference Agger2015). The window pattern is visible in the durability trajectory: within-year stability collapsed as the network restructured, creating a deterioration shock that increased pressure.

By the cross-year transition into 2010, durability rebounded ( $D$ recovered to 0.576) as the network began to consolidate. The within-year disruption operated through the ripeness logic: fragmentation made existing strategies untenable for both the government and rebel leaders, who recognized that continued splintering was eroding their political leverage and military (International Criminal Court, Office of the Prosecutor, 2014). The cross-year recovery operated through the bargaining logic: as the network began to consolidate, the military situation clarified enough to identify who could negotiate and what commitments were feasible, but the new configuration was too recent for actors to have re-entrenched into it. The window opened, and the CPJP entered negotiations and eventually signed a ceasefire in June 2011 (Reuters, 2011; United Nations Secretary-General, 2009).

Sudan 2012. Following South Sudan’s independence in July 2011, the Sudanese government faced insurgencies across multiple fronts: escalating fighting in Darfur and new conflicts in South Kordofan (from June 2011) and Blue Nile (from September 2011) (International Crisis Group, 2013). In November 2011, four rebel groups – the Justice and Equality Movement (JEM), two Sudan Liberation Movement/Army factions (SLM-Abdel Wahid and SLM-Minnawi), and SPLM-North – formed the Sudan Revolutionary Front (SRF), a coalition that sought to coordinate opposition across these theaters (Gramizzi and Tubiana, Reference Gramizzi and Tubiana2013). Despite the SRF’s aspiration to negotiate as a unified bloc, the government pursued bilateral channels. Talks with a JEM splinter faction began in Doha in late 2012, and the AU High-Level Implementation Panel brokered direct negotiations between the government and SPLM-North beginning in 2013 (Knopf, Reference Knopf2011; Permanent Mission of South Sudan to the United Nations, 2012). Under the window framework, Sudan falls in the “neither” quadrant because the network did not exhibit the within-year disruption followed by cross-year recovery that characterizes the window. This exception suggests that onset can occur outside the window configuration. Here, the political dynamics surrounding the SRF’s formation and the post-independence crises created pressure for talks through channels that our network recovery measure does not capture (Zartman, Reference Zartman1985). Sudan thus clarifies a scope condition: the window framework describes the typical path, but alternative sources of pressure and feasibility can operate through political rather than network-structural mechanisms.

Across the window cases, the common thread is not disruption or recovery alone but the sequence: pressure first, then legibility, then talks. The Sudan exception shows that alternative pathways exist, but the window cases illustrate the typical pattern. Semiannual measurement of network dynamics is important for capturing these sequences, since the within-year disruption-recovery patterns unfolded over months, not years, and annual network measures would have smoothed over the window dynamics that preceded negotiations (Tamm, Reference Tamm2016).

2.6 Hypotheses

The sequencing framework generates specific predictions about each stage and their combination. If disruption creates motivation through the ripeness mechanism and recovery creates legibility through the bargaining mechanism, then the strongest signal should come from the sequence in which both operate. This implies a two-stage pattern in durability changes prior to onset: (i) a decline in durability into $t{-}2$ (pressure), followed by (ii) an improvement into $t{-}1$ (feasibility), both observed before onset at $t$ .

H1 (Window sequence): Negotiation onset is more likely when a durability decline into $t{-}2$ ( $D_{t-2}\lt D_{t-3}$ ) is followed by a durability rebound into $t{-}1$ ( $D_{t-1}\gt D_{t-2}$ ).

The constituent stages yield further predictions that distinguish the sequencing framework from either tradition alone:

H1a (Pressure stage): A durability decline into $t{-}2$ is associated with a higher hazard of onset. This captures the pressure component of the ripeness logic. Disruption creates motivation by breaking entrenched equilibria and lowering continuation payoffs.

H1b (Feasibility stage): Conditional on prior decline, a durability rebound into $t{-}1$ is associated with a higher hazard of onset. Importantly, recovery without prior disruption should not predict onset on its own, because recovery in the absence of a preceding shock resembles ordinary stability, that is, legibility without motivation. This distinguishes the sequencing framework from a pure bargaining account, which would predict that recovery helps regardless of prior disruption.

Our second hypothesis addresses whether the same network dynamics that facilitate negotiation onset also promote peace agreement conclusion. The sequencing framework predicts an asymmetry; the factors that create the window (pressure from disruption, feasibility from recovery) are about getting parties to the table, not about what happens once they are there. Agreement conclusion depends on a different set of factors: bargaining dynamics during negotiations, third-party guarantees, and agreement design and implementation credibility that network-level measures do not capture (Mattes and Savun, Reference Mattes and Savun2009; Walter, Reference Walter2002). Fragmented networks can create commitment problems as parties cannot credibly promise to control all armed actors within their nominal coalitions (Cunningham, Reference Cunningham2006). Thus, we anticipate a null relationship between network dynamics and agreement conclusion despite effects on negotiation onset. The onset–agreement asymmetry is itself a prediction of the sequencing theory, not merely an empirical surprise.

H2 (Agreements): Network dynamics do not show a detectable association with peace agreement conclusion.

Our third hypothesis examines how networks evolve after negotiations begin. We expect that networks undergo reorganization following negotiation onset, though the direction is theoretically ambiguous. Several processes drive post-onset change. The announcement of negotiations often triggers defections as hardline elements reject talks and form splinter groups (Kydd and Walter, Reference Kydd and Walter2002). Excluded actors may escalate violence to gain seats at the negotiating table or spoil processes that threaten their interests (Nilsson, Reference Nilsson2008; Stedman, Reference Stedman1997). The reduction in violence between negotiating parties can create security vacuums that new armed groups exploit. Yet, consolidation is also possible as weaker actors align with stronger ones or exit the conflict entirely.

H3 (Reorganization): Negotiation onset is associated with network reorganization, with direction and magnitude varying across cases.

The asymmetry between these hypotheses reflects a tension in peace processes. The conditions that make conflicts ripe for negotiation may simultaneously make them difficult to resolve. Recovery creates feasibility for talks, but the underlying instability that created pressure may reassert itself through post-onset reorganization. This distinction suggests that models of negotiation onset and agreement conclusion may require different explanatory variables.

3.1 Data

We analyze civil conflicts from 1999 to 2013 using semiannual temporal resolution. Our primary data come from the Armed Conflict Location and Event Data (ACLED) project, which provides georeferenced information on political violence events (Raleigh et al., Reference Raleigh, Linke, Hegre and Karlsen2010). From ACLED, we extract battle events that form the basis of our network construction. These battles represent direct combat interactions between armed actors, excluding one-sided violence against civilians and remote violence. ACLED identifies actors by organizational name and codes battle events between specific pairs, enabling network construction. Our analysis draws on ACLED’s Africa coverage, which is the most comprehensive and temporally consistent during our study period; our analysis window begins in 1999 to allow lag construction.

Negotiation data come from the Peace Negotiations in Civil Conflicts (PNCC) dataset, which systematically codes formal bilateral talks between governments and rebel groups (Arı, Reference Arı2022). PNCC codes negotiations at the dyad-year level, recording whether a government–rebel dyad was in the formal negotiation phase in each year. We use this annual coding directly as our outcome measure. Negotiation onset is coded as the first year in which any government–rebel dyad enters negotiations after a period with no active negotiations in that country; continuing negotiations in subsequent years are excluded from the risk set to ensure proper hazard-model interpretation. Countries can experience multiple onsets if negotiations lapse and later resume.

To examine whether network dynamics affect not just negotiation onset but also successful resolution, we incorporate peace agreement data from the PA-X Peace Agreement Database (Bell and Badanjak, Reference Bell and Badanjak2019). PA-X defines agreements as “formal, publicly-available documents, produced after discussion with conflict protagonists and addressing conflict with a commitment to peace.” This includes ceasefires, framework agreements, comprehensive settlements, and implementation protocols, a heterogeneous set capturing the full peace process. This dataset identifies 30 agreement events within the analysis sample used for our window specification. The higher frequency of agreements compared to negotiation onsets reflects both the inclusion of agreements reached without formal negotiations and the reality that some negotiation processes yield multiple agreements over time.

The temporal structure of our data deserves emphasis. While ACLED coverage extends beyond our study period, we restrict analysis to 1999–2013 given constraints of the other datasets. We construct networks at semiannual frequency, dividing each year into two six-month periods: a first half (January–June) and a second half (July–December). This semiannual measurement provides critical advantages. It captures within-year dynamics that annual network measures would smooth over, including the alliance shifts and group fragmentation documented in qualitative accounts (Bakke et al., Reference Bakke, Cunningham and Seymour2012; Tamm, Reference Tamm2016). The outcome, however, is modeled at the annual level, using PNCC’s native annual coding directly and avoiding the need to assign onsets to sub-annual periods. This combination – semiannual network measurement with annual outcome modeling – captures within-year disruption-recovery sequences in the predictors. We model the outcome annually because PNCC records negotiation onset by calendar year rather than by month or half-year; the annual outcome reflects this feature of the available data rather than a claim that negotiations begin on an annual rhythm.

The unit of analysis is the country-year. After applying the network thresholds and lag requirements and excluding continuation spells (ongoing negotiations), the estimation sample comprises 97 country-years drawn from 21 African countries (one multi-actor conflict system per country). The entire onset analysis therefore rests on 9 negotiation onsets among these 97 country-years. Appendix Table A2 details sample construction.

3.2 Network construction

We construct conflict networks to capture multi-actor dynamics that characterize contemporary civil wars. Unlike dyadic approaches that examine government-rebel interactions in isolation, network analysis reveals how armed groups interact within broader conflict systems (Bakke et al., Reference Bakke, Cunningham and Seymour2012; Cunningham et al., Reference Cunningham, Bakke, Seymour, Pearlman and Cunningham2012). Nodes represent armed actors in conflicts such as national armies, paramilitary forces, ethnic militias, rebel organizations, pro-government militias, and foreign interveners. Edges capture direct combat relationships through which these actors contest territorial control, compete for resources, and pursue political objectives.

Each edge between actors $i$ and $j$ is weighted by battle frequency ( $w_{ij,t} = \sum _{e \in E_{ij,t}} 1$ ), capturing conflict intensity that varies across dyads and time. A government may simultaneously engage multiple rebel groups with different intensities while those rebels fight among themselves for dominance within the opposition. These patterns matter for understanding when negotiations become possible by helping us to understand the central government’s relationship with the main rebel group exists within a broader ecosystem of violence that shapes bargaining dynamics (Johnston, Reference Johnston2014; Prorok, Reference Prorok2016).

We construct networks at semiannual frequency.Footnote ² We focus exclusively on battle events because they capture the reciprocal violence central to bargaining theory. Battles represent contested interactions where both parties choose to engage, revealing information about capabilities and resolve that one-sided violence against civilians cannot provide. This distinction matters theoretically. While civilian victimization may signal desperation or strength, only direct combat tests relative military capacity, the core uncertainty driving bargaining failures in civil war (Fearon, Reference Fearon1995; Walter, Reference Walter2009).

The requirement of $|V| \geq 3$ actors and $|E| \geq 3$ edges ensures we analyze genuine multi-actor conflicts rather than simple bilateral confrontations. This threshold captures the complexity that distinguishes network wars from traditional insurgencies. Conflicts with government forces battling multiple rebel groups, rebels fighting each other for territorial control, or ethnically fragmented conflicts with shifting militia alliances. Country-semesters that fall below the threshold are excluded from durability calculations and thus from the estimation sample; our window specification requires valid semiannual durability values from the two years preceding each country-year observation, so a country-year enters the analysis only if the relevant preceding semesters also meet the network threshold. Countries may re-enter the sample if they cross back above the threshold after a gap. Networks meeting our threshold average 24 actors, enough to generate the fragmentation and coordination problems central to multi-party bargaining challenges (Cunningham, Reference Cunningham2006, Reference Cunningham2013).

3.3 Measuring network durability

Our durability measure operationalizes conflict system stability relevant to both bargaining and ripeness theories. High durability indicates an entrenched conflict equilibrium where actors have settled into predictable patterns of violence. Governments control urban centers while rebels hold rural territory, militias maintain ethnic enclaves, and all parties calibrate violence to sustainable levels (Kalyvas, Reference Kalyvas2006; Wood, Reference Wood2003). Low durability signals structural disruption, for example, when defections split rebel coalitions, new militias emerge to exploit power vacuums, or when external interventions alter the balance of forces.

We capture two dimensions of network change that theory identifies as consequential for negotiation dynamics. Each component reveals distinct mechanisms through which stability and change affect bargaining.

Edge persistence measures relationship turnover through the Jaccard coefficient:

(1) \begin{equation} J = \frac {|E_{t-2} \cap E_{t-1}|}{|E_{t-2} \cup E_{t-1}|} \end{equation}

where $E_t$ is the set of edges (fighting dyads) at time $t$ .Footnote ³ This captures whether the same pairs of actors continue fighting. In our sample, edge persistence averages 0.21; less than a quarter of fighting relationships persist across semiannual periods. This turnover reflects the realignments that reshape conflict dynamics. When former allies turn against each other or new actors redirect existing conflicts, these changes force all parties to reassess their strategic positions (Tamm, Reference Tamm2016). Such relationship changes matter because they affect who can negotiate with whom and whether coalitions can make credible commitments.

Intensity correlation tracks volatility in violence levels through the Pearson correlation of actor strengths:

(2) \begin{equation} \rho = \text{Cor}(s_{i,t-2}, s_{i,t-1}) \end{equation}

where $s_{i,t}$ represents actor $i$ ’s total battle involvement (weighted degree) at time $t$ , computed over all actors present in either period (with zeros assigned to actors absent from one period). The correlation averages 0.61 in our data, indicating moderate stability in relative intensity rankings. Governments may escalate operations against one rebel group while reducing pressure on another. Militias intensify violence to claim negotiation seats. Rebels reduce attacks when resources dwindle. Such intensity shifts may theoretically signal changes in capabilities and resolve (Akcinaroglu, Reference Akcinaroglu2012; Weinstein, Reference Weinstein2007), though as we show below, structural continuity proves more consequential for negotiation timing.

Composite index: $D = (J + |\rho |)/2$ , bounded in $[0,1]$ . We use the absolute value of $\rho$ so that both perfectly stable and perfectly inverted intensity profiles indicate structural persistence; robustness checks using signed $\rho$ yield similar results (Appendix Table A11). Component disaggregation shows that neither edge persistence nor intensity correlation is individually significant for the window specification; the composite index captures the disruption-recovery signal more effectively than either component alone (Appendix Table A14). This suggests that coordinated change across both dimensions – whether the same dyads continue fighting and how consistently actors fight – jointly defines the network dynamics associated with negotiation timing. Stable relationship structures signal who the relevant bargaining partners are, enabling credible negotiation.

Semiannual measurement of network dynamics is important for detecting these patterns in our data. Alliance realignments in conflicts like the Democratic Republic of Congo or Central African Republic often unfold over months rather than years (Tamm, Reference Tamm2016). By measuring at six-month intervals, we capture within-year disruption-recovery sequences that annual network measures would smooth over. The window specification relies on comparing semiannual durability values within and across years; annual network aggregation would collapse these distinct phases into a single value, obscuring the within-year dynamics that carry the signal. Our system-wide approach captures broad network restructuring, whether arising from endogenous cascades (defections triggering realignments) or exogenous shocks (external interventions affecting multiple actors simultaneously). The durability measure reflects the outcome, network change, without distinguishing between these sources, which we acknowledge as a limitation.

3.4 Model specification

We estimate an annual discrete-time hazard model where each country-year represents a “trial” in which negotiation either begins or does not (Beck et al., Reference Beck, Katz and Tucker1998; Carter and Signorino, Reference Carter and Signorino2010). The binary outcome approach matches our theoretical focus on negotiation onset (the decision to begin talks) rather than the continuous process of negotiation itself. Countries enter the risk set in the first year they meet the network thresholds and contribute one country-year observation for each subsequent year they remain at risk of onset. A country leaves the risk set in the year an onset occurs and re-enters if negotiations later lapse, so a single country can contribute multiple spells. A country-year is dropped from the sample in any year the network falls below the multi-actor threshold or lacks the lagged semiannual values the window requires; the country re-enters in later years once these conditions are again met.

The core specification operationalizes the two-stage mechanism through the window indicator:

(3) \begin{equation} \Pr (\text{Onset}_{it}) = \text{logit}^{-1}\left [\beta _0 + \beta _1 \text{Window}_{it} + f(\text{Duration}_{it})\right ] \end{equation}

where $\text{Window}$ indicates that the network exhibited both within-year disruption and subsequent cross-year recovery in the two years preceding onset, and $f(\!\cdot \!)$ is a cubic spline for duration dependence.

We measure network dynamics at semiannual frequency but model onset at the annual level. For each country-year, we measure how much the network changed across three consecutive half-year transitions in the two years before onset: from the second half of $t{-}3$ to the first half of $t{-}2$ , from the first half of $t{-}2$ to the second half of $t{-}2$ , and from the second half of $t{-}2$ to the first half of $t{-}1$ . Each transition yields a durability value, with higher values indicating that the network changed less. Prior disruption means the network became less durable across the second transition than across the first; recovery means it became more durable across the third transition than across the second. The window is the configuration where both hold: the network disrupted during year $t{-}2$ and then recovered crossing into $t{-}1$ , before onset at $t$ .

Duration is measured in years since the country first met network thresholds. The spline specification follows recommendations for discrete-time hazard models (Beck et al., Reference Beck, Katz and Tucker1998), allowing flexible baseline hazards without imposing parametric assumptions. Critically, we exclude country-years with ongoing (continuation) negotiations from the risk set, since these periods are not “at risk” of onset; they already have active talks. This adjustment ensures proper hazard-model interpretation. The analysis sample contains 97 country-years with 9 negotiation onsets across 21 countries.

We also estimate specifications with separate prior disruption and recovery indicators to test the constituent stages. We report Firth penalized likelihood estimates as robustness checks for the rare-events setting (9 onsets in 97 observations), which adjusts for small-sample bias in maximum likelihood estimation (Firth, Reference Firth1993).