Human shields

Motivation

Fighters mingle with civilians:

• for protection from camouflage logic
• because they trust opponents are reluctant to kill civilians, and perhaps to
• to force revelation about how little the other side values noncombatant life (Uhler et al. 1994; Pilloud et al. 1987).

Key results:

• Violence requires uncertainty (given linear utility model)
• With uncertainty, actors trade off revelation and material gains: separation is possible with learning (especially about most abusive types)
• Reputational concerns cut both ways and can increase or reduce equilibrium violence
• When rebels are informed, they have incentives to deliberatly draw fire but only if (a) they value reputational concerns more than states (b) they are relative popular to begin with and (c) beliefs are not zero sum

Set up

A baseline game: rebels choose mingling intensity r; the state chooses violence s in populated areas; civilians update.

Strategies:

• Mingling r\in[0,1]: share of fighters among civilians
• Violence s\in[0,1]: force in populated areas

Parameters:

• identifiability k: mapping from civilian to fighter casualties conditional on attack.
• outside option q: mortality risk when not mingling.
• distributions on types Types \theta_S,\theta_R:
• \beta weights reputational concern through citizens’ beliefs \hat{\theta}_S (and, in Case 4, \hat{\theta}_R).

Utilities

\begin{align*} u_R = &+r(1-sk)+(1-r)(1-q) - \beta \hat{\theta}_S - \theta_Rs\\ u_S = &-r(1-sk)-(1-r)(1-q) + \beta \hat{\theta}_S - \theta_Ss \end{align*}

Note there is a strong zero sum component but if \theta_R, \theta_S both positive there is also a common interest component.

If there is no uncertainty then the \beta\hat\theta_S terms drop out.

Some models let \beta vary over actor and also include \beta\hat\theta_R

Case 1: full information

Proposition (benchmark). Unique equilibrium: the state attacks if and only if r > \theta_S/k; rebels choose r^{*} = \min\{\theta_S/k,\,1\}.

• Implication: if \theta_S > 0, rebels mingle but there are no civilian deaths: violence is off-path deterrence, not equilibrium punishment.

Note: not exactly knife edge. WIth convex costs there is equilibrium violence; but concave costs there is not. Which is more likely here?

Case 2: partial revelation by states

Citizens and rebels share a prior over \theta_S and observe (r,s).

Equilibrium typically exhibits partial separation: more abusive types are separated by violence; more moderate types pool on s = 0.

The cutoff type \theta_S^{c} is indifferent between the separating branch \hat{s}(\theta_S,r) and pooling at s = 0. Types on either side select into separation or pooling.

Revelation idea: more abusive states gain more from violence, are more tempted by violence, and more willing to separate. Creates a backwards mapping from violence to type.

Case 2: rebel incentives

Rebels essentially select the separating continuation game (by their choice of r).

• Higher r strengthens low-\theta_S types’ incentive to attack: violence is a costly signal.

• Rebels trade fighter safety (higher r) against the probability and level of violence and against reputational shifts in \hat{\theta}_S.

• The level of q (riskiness of outside options) determines whether rebels remain at the pooling boundary r = \underline{r} or move into semi-separation.

Case 3: informed rebels

When only civilians are uninformed, separation can sometimes be implemented by mingling levels that map \theta_S into distinct r, with incentives disciplined at the attack threshold.

Case 3 — work in progress. The informed-rebels extension is under revision in the draft; this talk emphasizes Case 2 comparative statics and Case 4 two-sided uncertainty.

Case 4: Rebels are informed, citizens are unsure

Binary types \theta \in \{0,1\}; prior cells p_{jk}.

\Pr(R = H | S = L)=\frac{p_{HL}}{p_{LL}+p_{HL}}, \qquad \Pr(R = L | S = H)=\frac{p_{LH}}{p_{LH}+p_{HH}}.

These objects enter reputational payoffs when citizens are uncertain about both sides but learn something directly about one side in equilibrium.

Case 4: main results

Especially interested in whether rebels mingle even though—or indeed because—they expect a violent response.

• Symmetric reputational weights: no always-mingle exposure equilibrium: the forces that reward mingling also support restraint in a way that empties the feasible k-interval.

• Asymmetric weights (\beta_R sufficiently large relative to \beta_S): exposure can return, but prior association matters: citizens must not infer from a humane state that rebels must therefore be inhumane.

Conjecture: Appears to also require high prior on rebels type than state’s type.

Case 4c: prior slice

Fixing \mathrm{Pr}(S=H)=\frac{1}{2}, equilibria as a function of (p_{LH},p_{HL})..

Figure 3: Regions where subsets of the three conditions hold; star where all three hold.

Case 4c: boundary shifts

Figure 4: Small shifts in k, q, beta_S, and beta_R: each condition’s initial (solid) and shifted (dashed) boundary share one colour; labels and arrows match that colour.

Takeaways

1. Uncertainty links mingling to on-path civilian risk; full information keeps violence off path as a threat only.

2. Case 2: mingling generates partial revelation (tail separated, body pooled); k shifts both targeting conditional on attack and equilibrium mingling.

3. Case 4: reputation is relative in posterior space; asymmetric \beta and (non zero sum priors) can yield equilibria that symmetric \beta rules out.

Closing

• Human shields bundle distinct strategic mechanisms (protection, restraint of the opponent, and forced revelation). The same observed mingling can mean different things depending on information and beliefs.

• Uncertainty is what generates civilian harm on path in the linear benchmark: full information keeps violence purely off-path as a threat that pins mingling.

• Partial-revelation regimes separate abusive governments through violence while moderate types pool; comparative statics in targeting technology and outside options move both the attack probability and the severity conditional on attack.

• Two-sided citizen uncertainty makes reputation a relative posterior object; deliberate (costly) exposure equilibria possible once concern is sufficiently asymmetric and the prior does not load zero-sum correlation across sides.

• Policy and legal inference need the mapping from information structure to equilibrium, not a single reduced-form sign on “shields.”

Next steps: Connect to our changing times

Seek now to connect this to seeming reduced sanctioning of civilian deaths:

Say now state’s costs are: (\theta_s + \pi)s where \pi is a marginal policy cost. But reputational concerns remain the same.

Expectations (conjectures) from reduced \pi:

• reduced mingling, no voilence in perfect information case case 1
• more violence for each revelation type in case 2
• (possibly) increased incentives to reveal \theta_S, which is obscured by \pi

So likely bad news but depends on information environment again. To be seen.

Help

Empirical component?

1. Needed?

2. Correlational analysis perhaps possible. Mingling and violence join outcomes so treat as jointly distributed.

3. Or primarily qualitative. What scope.

Literature

Mea culpa. Not currently well connected to current work. All pointers welcome!