Appendix
We now set up and solve a decision-theoretical model of exit in continuous time. The model builds on Dixit (
1989) and Dixit and Pindyck (
1994), who study firms’ decisions to enter and exit markets with changing output prices. However, since output prices are normally positive, they focus on geometric Brownian motions. In contrast, the model presented here assumes a standard Brownian motion because the benefits from a federation can become negative.
In what follows, we assume for simplicity that exit is definitive and that there is no possibility of re-entering later. However, even if re-entry is free the qualitative conclusions from our model would still hold, but the solution process would be complicated as the optimal exit decisions and optimal entry decisions would be mutually interdependent.
First, assume that member
\(i\)’s net benefit flow from being in the federation at time
\(t\) is given by
$${\pi }_{i}\left(t\right)={x}_{i}\left(t\right)+{\theta }_{i}$$
(11)
where
\({x}_{i}(t)\) represents member
\(i\)’s state at time
\(t\), and
\({\theta }_{i}\) its type, as in the main text.
The state of the world captures the changes in benefits over time, which are uncertain. For simplicity, assume that member
\(i\)’s state
\({x}_{i}(t)\) evolves according to a Brownian motion without drift and with variance
\({\sigma }_{i}^{2}\). A Brownian motion is a mathematical representation of a random process over time. Over any given time period
\(dt\), it can go up or down. The increments
\(d{x}_{i}\) have a normal distribution characterized by variance
\({\sigma }_{i}^{2}\) (Mörters and Peres
2010). A higher variance
\({\sigma }_{i}^{2}\) means that member
\(i\)’s benefits are expected to fluctuate more. Formally, a Brownian motion is characterized by
$$d{x}_{i}={\sigma }_{i}{\epsilon }_{it}\sqrt{dt}$$
(12)
where
\({\epsilon }_{it}\) has a standard normal distribution and is serially uncorrelated so that for all
\(t>0\)$$E\left[{\epsilon }_{it}\right]=0,\quad Var\left({\epsilon }_{it}\right)=E\left[{\epsilon }_{it}^{2}\right]=1,\quad E\left[{\epsilon }_{i{t}_{1}}{\epsilon }_{i{t}_{2}}\right]=0,\quad {t}_{1}\ne {t}_{2}$$
(13)
This implies that
\(E[d{x}_{i}]=0\) and
\(E[{(d{x}_{i})}^{2}]={\sigma }_{i}^{2}dt\): benefits are expected to remain the same, but have a variance of
\({\sigma }_{i}^{2}\) per unit of time.
Without loss of generality, suppose that the union starts at
\(t=0\) and set
\({x}_{i}\left(0\right)=0\). Since we have assumed no drift in the states
\({x}_{i}\), their ex ante expected value is 0 for any future date
\(T\); mathematically,
\(E\left[{x}_{i}\left(T\right)\right]=0, \forall T>0\). Similarly, the expected value at time
\(t\) for a later time
\(T\) is simply the value at the time of the expectation:
\({E}_{t}\left[{x}_{i}\left(T\right)\right]={x}_{i}(t)\). At time
\(t\) the probability density for time
\(T\) at state
\({x}_{i}\) is
$${P}_{t}\left({x}_{i}\left(T\right)={x}_{i}\right)=\phi \left({x}_{i};{x}_{i}\left(t\right),{\sigma }_{i}^{2}(T-t)\right)=\phi \left(\frac{{x}_{i}-{x}_{i}\left(t\right)}{{\sigma }_{i}\sqrt{T-t}}\right)$$
(14)
where
\(\phi (x;\mu ,{\sigma }^{2})\) is the probability density function (pdf) of the normal distribution with mean
\(\mu \) and variance
\({\sigma }^{2}\), and
\(\phi (x)\) is the pdf of the standard normal distribution. Substantively, this means that the expected change in
\({x}_{i}\) over any time period
\(T-t\) is zero and that large changes are less likely than small changes. However, since the variance increases linearly with the time period
\(T-t\) the expected magnitude of changes increases with time.
Assume that the members have a common discount rate
\(r\), so that the expected discounted benefit from the union for member
\(i\) until time
\(T\) is
\(E\left[{\int }_{0}^{T}{\pi }_{i}\left(t\right){e}^{-rt}dt\right]\). The ex-ante expected value for member
\(i\) from a perpetual union is
$$E\left[{\int }_{0}^{+\infty }{\pi }_{i}\left(t\right){e}^{-rt}dt\right]= \underset{T\to +\infty }{\mathrm{l}\mathrm{i}\mathrm{m}}\left(\frac{{\theta }_{i}}{r}-\frac{{\theta }_{i}}{r}{e}^{-rT}\right)=\frac{{\theta }_{i}}{r}$$
(15)
In the decision-theoretical version of the model, we assume that only one member can exit from the union, and only consider that member. Hence we drop the index \(i\) to lighten notation.
When thinking about exit in a continuous time set-up, a member compares the expected value of maintaining the union to the value of exiting right now. Assuming rationality, the expected value of maintaining the union right now should take as a given optimal exit behavior in the future. This is captured by the notion of continuation value: the continuation value
\(V(x)\) is the expected discounted benefit from maintaining the union when the current state is
\(x\), assuming optimal exit behavior in the future (Dixit and Pindyck
1994).
At any point in time, in order for a member to be willing to maintain the union for
\(dt\) longer, the expected change in the continuation value combined with the benefit flow (
11) over
\(dt\) should add up to the normal return
\(rV\left(x\right)dt\):
$$E\left[dV\right]+\left(x+\theta \right)dt=rV\left(x\right)dt$$
(16)
Determining the continuation value is just like pricing a stock with expected appreciation
\(E[dV]\) and dividend flow
\(x+\theta \). By Ito’s lemma,
\(dV=\frac{\partial V}{\partial t}dt+\frac{\partial V}{\partial x}dx+\frac{1}{2}\frac{{\partial }^{2}V}{\partial {x}^{2}}{\left(dx\right)}^{2}\); see Øksendal (
1991) for the theory of stochastic differential equations and a discussion of Ito’s lemma. The benefit flow does not depend on calendar time directly:
\(\frac{\partial V}{\partial t}=0\). Together with (
12),
\(E\left[dx\right]=0\) and
\(E[{(d{x}_{i})}^{2}]={\sigma }_{i}^{2}dt\), this implies
$$dV={V}^{\prime}\left(x\right)dx+\frac{1}{2}{V}^{\prime \prime}\left(x\right){\sigma }^{2}{\epsilon }_{t}^{2}dt\Rightarrow E[dV]=\frac{1}{2}{V}^{\prime \prime}\left(x\right){\sigma }^{2}dt$$
(17)
By substituting (
17) in (
16), we find
$$\frac{1}{2}{\sigma }^{2}{V}^{\prime \prime}\left(x\right)-rV\left(x\right)=-x-\theta $$
(18)
This is a second order differential equation in
\(x\): we solve it by first identifying the solution
\({V}_{h}(x)\) of the homogeneous equation
\(\frac{1}{2}{\sigma }^{2}{V}^{\prime \prime}\left(x\right)-rV\left(x\right)=0\). Next we identify a particular solution
\({V}_{p}\left(x\right)\) which satisfies the equation
\(\frac{1}{2}{\sigma }^{2}{V}^{\prime \prime}\left(x\right)-rV\left(x\right)=-x-\theta \). The general solution is given by the particular solution plus any linear combination of solutions to the homogeneous equation. For the homogeneous part, we try a solution of the form
\(V\left(x\right)={e}^{\lambda x}\). This yields
$${V}_{h}\left(x\right)=A{e}^{\alpha x}+B{e}^{\beta x}, \quad \alpha =\sqrt{\frac{2r}{{\sigma }^{2}}}, \beta =-\sqrt{\frac{2r}{{\sigma }^{2}}}$$
(19)
with constants
\(A\) and
\(B\) to be identified. Note that
\(\alpha >0\) and
\(\beta =-\alpha <0\). For a particular solution, we try
\(V\left(x\right)=ax+b\). This yields the solution
\({V}_{p}\left(x\right)=\frac{x+\theta }{r}\). This is the expected perpetuity value from the union starting from state
\(x\). Combining the particular solution with the homogeneous part, the general solution is
$$V\left(x\right)={V}_{p}\left(x\right)+{V}_{h}\left(x\right)=\frac{x+\theta }{r}+ A{e}^{\alpha x}+B{e}^{\beta x}$$
(20)
Since
\({V}_{p}(x)\) represents the value from maintaining the union perpetually,
\({V}_{h}\left(x\right)\) represents the option value of exit, which should be positive. As the state improves, the value of the exit option should converge to 0: the better the state, the higher the benefit flow and the less valuable the exit option. This implies that
\(A=0\) since
\(\alpha >0\). We now have an expression for the continuation value
\(V(x)\) up to the constant
\(B\).
$$V\left(x\right)=\frac{x+\theta }{r}+B{e}^{\beta x} , \quad \beta =-\sqrt{\frac{2r}{{\sigma }^{2}}}$$
(21)
The continuation value is the value of maintaining the union for now, assuming optimal behavior for the future. Hence the constant
\(B\) depends on the optimal exit state
\({x}^{e}\). Two conditions are needed for optimal exit in a continuous time stochastic model (Dixit and Pindyck
1994). The first is
Value Matching (VM): exit should occur when
\(V(x)\) drops to the value of the outside option. In our model, the outside option consists of paying the one-off costs and the penalty
\(c\), which corresponds to a value of
\(-k-c\). The second condition is
Smooth Pasting (SP): optimal stopping requires that
\(V^{\prime}(x)\) be equal to the derivative of the value of being outside of the union—which is 0 in our case, since the exit value does not depend on the state.
Grouping the two conditions we obtain a system with two equations and two unknowns (
\({x}^{e},B){:}\)$$V\left({x}^{e}\right)=-k-c\iff \frac{{x}^{e}+\theta }{r}+B{e}^{\beta {x}^{e}}=-k-c \quad\quad{\varvec{V}}{\varvec{M}}$$
(22)
$${V}^{\prime}\left({x}^{e}\right)=0\iff \frac{1}{r}+\beta B{e}^{\beta {x}^{e}}=0 \quad\quad{\varvec{S}}{\varvec{P}}$$
(23)
The solutions for (
\({x}^{e},B)\) are.
$${x}^{e}=-\theta -r\left(k+c\right)+\frac{1}{\beta },\quad B=-\frac{{e}^{-\beta {x}^{e}}}{r\beta }$$
(24)
The optimal exit state consists of three components. The first one, \(-\theta \), is the most intuitive. When \(x\) reaches \(-\theta \), the benefit flow drops to zero: \(\pi \left(t|x\left(t\right)=-\theta \right)=-\theta +\theta =0\). The higher a member’s type \(\theta \), the longer it is optimal to stay in the union, i.e. the more negative the optimal exit state: \(\frac{\partial {x}^{e}}{\partial \theta }<0\). The second term, \(-r(k+c)\), reflects the deterring effect of exit costs since \(\frac{\partial {x}^{e}}{\partial k}=\frac{\partial {x}^{e}}{\partial c}<0\). The benefit flow needs to drop to \(-r\left(k+c\right)\) to make exit worth considering, since the corresponding expected perpetuity value at that state would be equal to minus the exit costs \(-(k+c)\).
The third term, \(1/\beta \), is the least intuitive but can be interpreted as the optimal forbearance level. The equivalent of this term does not feature in the discrete time model presented in the main text. It reflects sophisticated rational behavior under the option logic: one should be willing to sustain some losses in the hope that the state improves again. Intuitively, the higher the variance of the benefits, the higher the possibility that a bad state turns around, and the more reluctant one should be to exit.