# On-the-Job Search¶

## Overview¶

In this section we solve a simple on-the-job search model

### Model features¶

- job-specific human capital accumulation combined with on-the-job search
- infinite horizon dynamic programming with one state variable and two controls

## Model¶

Let

- \(x_t\) denote the time-\(t\) job-specific human capital of a worker employed at a given firm
- \(w_t\) denote current wages

Let \(w_t = x_t(1 - s_t - \phi_t)\), where

- \(\phi_t\) is investment in job-specific human capital for the current role
- \(s_t\) is search effort, devoted to obtaining new offers from other firms.

For as long as the worker remains in the current job, evolution of \(\{x_t\}\) is given by \(x_{t+1} = G(x_t, \phi_t)\)

When search effort at \(t\) is \(s_t\), the worker receives a new job offer with probability \(\pi(s_t) \in [0, 1]\)

Value of offer is \(U_{t+1}\), where \(\{U_t\}\) is iid with common distribution \(F\)

Worker has the right to reject the current offer and continue with existing job.

In particular, \(x_{t+1} = U_{t+1}\) if accepts and \(x_{t+1} = G(x_t, \phi_t)\) if rejects

Letting \(b_{t+1} \in \{0,1\}\) be binary with \(b_{t+1} = 1\) indicating an offer, we can write

Agent’s objective: maximize expected discounted sum of wages via controls \(\{s_t\}\) and \(\{\phi_t\}\)

Taking the expectation of \(V(x_{t+1})\) and using (1), the Bellman equation for this problem can be written as

Here nonnegativity of \(s\) and \(\phi\) is understood, while \(a \vee b := \max\{a, b\}\)

### Parameterization¶

In the implementation below, we will focus on the parameterization

with default parameter values

- \(A = 1.4\)
- \(\alpha = 0.6\)
- \(\beta = 0.96\)

The Beta(2,2) distribution is supported on \((0,1)\). It has a unimodal, symmetric density peaked at 0.5.

### Back-of-the-Envelope Calculations¶

Before we solve the model, let’s make some quick calculations that provide intuition on what the solution should look like.

To begin, observe that the worker has two instruments to build capital and hence wages:

- invest in capital specific to the current job via \(\phi\)
- search for a new job with better job-specific capital match via \(s\)

Since wages are \(x (1 - s - \phi)\), marginal cost of investment via either \(\phi\) or \(s\) is identical

Our risk neutral worker should focus on whatever instrument has the highest expected return

The relative expected return will depend on \(x\)

For example, suppose first that \(x = 0.05\)

- If \(s=1\) and \(\phi = 0\), then since \(G(x,\phi) = 0\), taking expectations of (1) gives expected next period capital equal to \(\pi(s) \mathbb{E} U = \mathbb{E} U = 0.5\)
- If \(s=0\) and \(\phi=1\), then next period capital is \(G(x, \phi) = G(0.05, 1) \approx 0.23\)

Both rates of return are good, but the return from search is better

Next suppose that \(x = 0.4\)

- If \(s=1\) and \(\phi = 0\), then expected next period capital is again \(0.5\)
- If \(s=0\) and \(\phi = 1\), then \(G(x, \phi) = G(0.4, 1) \approx 0.8\)

Return from investment via \(\phi\) dominates expected return from search

Combining these observations gives us two informal predictions:

- At any given state \(x\), the two controls \(\phi\) and \(s\) will function primarily as substitutes — worker will focus on whichever instrument has the higher expected return
- For sufficiently small \(x\), search will be preferable to investment in job-specific human capital. For larger \(x\), the reverse will be true

Now let’s turn to implementation, and see if we can match our predictions.

## Implementation¶

The QuantEcon package provides some code for solving the DP problem described above

See in particular `jv.jl`

, which is repeated here for convenience

```
#=
@author : Spencer Lyon <spencer.lyon@nyu.edu>
@date: 2014-06-27
References
----------
Simple port of the file quantecon.models.jv
http://quant-econ.net/jl/jv.html
=#
using Distributions
using QuantEcon
# NOTE: only brute-force approach is available in bellman operator. Waiting on a simple constrained optimizer to be written in pure Julia
"""
A Jovanovic-type model of employment with on-the-job search.
The value function is given by
\\[V(x) = \\max_{\\phi, s} w(x, \\phi, s)\\]
for
w(x, phi, s) := x(1 - phi - s) + beta (1 - pi(s)) V(G(x, phi)) +
beta pi(s) E V[ max(G(x, phi), U)
where
* `x`: : human capital
* `s` : search effort
* `phi` : investment in human capital
* `pi(s)` : probability of new offer given search level s
* `x(1 - \phi - s)` : wage
* `G(x, \phi)` : new human capital when current job retained
* `U` : Random variable with distribution F -- new draw of human capital
##### Fields
- `A::Real` : Parameter in human capital transition function
- `alpha::Real` : Parameter in human capital transition function
- `bet::Real` : Discount factor in (0, 1)
- `x_grid::Vector` : Grid for potential levels of x
- `G::Function` : Transition `function` for human captial
- `pi_func::Function` : `function` mapping search effort to the probability of getting a new job offer
- `F::UnivariateDistribution` : A univariate distribution from which the value of new job offers is drawn
- `quad_nodes::Vector` : Quadrature nodes for integrating over phi
- `quad_weights::Vector` : Quadrature weights for integrating over phi
- `epsilon::Float64` : A small number, used in optimization routine
"""
type JvWorker
A::Real
alpha::Real
bet::Real
x_grid::Vector{Float64}
G::Function
pi_func::Function
F::UnivariateDistribution
quad_nodes::Vector
quad_weights::Vector
epsilon::Float64
end
"""
Constructor with default values for `JvWorker`
##### Arguments
- `A::Real(1.4)` : Parameter in human capital transition function
- `alpha::Real(0.6)` : Parameter in human capital transition function
- `bet::Real(0.96)` : Discount factor in (0, 1)
- `grid_size::Int(50)` : Number of points in discrete grid for `x`
- `epsilon::Float(1e-4)` : A small number, used in optimization routine
##### Notes
There is also a version of this function that accepts keyword arguments for
each parameter
"""
function JvWorker(A=1.4, alpha=0.6, bet=0.96, grid_size=50, epsilon=1e-4)
G(x, phi) = A .* (x .* phi).^alpha
pi_func = sqrt
F = Beta(2, 2)
# integration bounds
a, b, = quantile(F, 0.005), quantile(F, 0.995)
# quadrature nodes/weights
nodes, weights = qnwlege(21, a, b)
# Set up grid over the state space for DP
# Max of grid is the max of a large quantile value for F and the
# fixed point y = G(y, 1).
grid_max = max(A^(1.0 / (1.0 - alpha)), quantile(F, 1 - epsilon))
# range for linspace(epsilon, grid_max, grid_size). Needed for
# CoordInterpGrid below
x_grid = collect(linspace(epsilon, grid_max, grid_size))
JvWorker(A, alpha, bet, x_grid, G, pi_func, F, nodes, weights, epsilon)
end
# make kwarg version
JvWorker(;A=1.4, alpha=0.6, bet=0.96, grid_size=50, epsilon=1e-4) = JvWorker(A, alpha, bet, grid_size, epsilon)
"""
Apply the Bellman operator for a given model and initial value, returning only the value function
##### Arguments
- `jv::JvWorker` : Instance of `JvWorker`
- `V::Vector`: Current guess for the value function
- `new_V::Vector` : Storage for updated value function
##### Returns
None, `new_V` is updated in place with the value function.
##### Notes
Currently, only the brute-force approach is available. We are waiting on a simple constrained optimizer to be written in pure Julia
"""
function bellman_operator!(jv::JvWorker, V::Vector, new_V::Vector)
# simplify notation
G, pi_func, F, bet, epsilon = jv.G, jv.pi_func, jv.F, jv.bet, jv.epsilon
nodes, weights = jv.quad_nodes, jv.quad_weights
# prepare interpoland of value function
Vf = LinInterp(jv.x_grid, V)
# instantiate the linesearch variables
max_val = -1.0
cur_val = 0.0
max_s = 1.0
max_phi = 1.0
search_grid = linspace(epsilon, 1.0, 15)
for (i, x) in enumerate(jv.x_grid)
function w(z)
s, phi = z
function h(u)
out = similar(u)
for j in 1:length(u)
out[j] = Vf(max(G(x, phi), u[j])) * pdf(F, u[j])
end
out
end
integral = do_quad(h, nodes, weights)
q = pi_func(s) * integral + (1.0 - pi_func(s)) * Vf(G(x, phi))
return - x * (1.0 - phi - s) - bet * q
end
for s in search_grid
for phi in search_grid
if s + phi <= 1.0
cur_val = -w((s, phi))
else
cur_val = -1.0
end
if cur_val > max_val
max_val, max_s, max_phi = cur_val, s, phi
end
end
end
new_V[i] = max_val
end
end
"""
Apply the Bellman operator for a given model and initial value, returning policies
##### Arguments
- `jv::JvWorker` : Instance of `JvWorker`
- `V::Vector`: Current guess for the value function
- `out::Tuple{Vector, Vector}` : Storage for the two policy rules
##### Returns
None, `out` is updated in place with the two policy functions.
##### Notes
Currently, only the brute-force approach is available. We are waiting on a simple constrained optimizer to be written in pure Julia
"""
function bellman_operator!(jv::JvWorker, V::Vector, out::Tuple{Vector, Vector})
# simplify notation
G, pi_func, F, bet, epsilon = jv.G, jv.pi_func, jv.F, jv.bet, jv.epsilon
nodes, weights = jv.quad_nodes, jv.quad_weights
# prepare interpoland of value function
Vf = LinInterp(jv.x_grid, V)
# instantiate variables
s_policy, phi_policy = out[1], out[2]
# instantiate the linesearch variables
max_val = -1.0
cur_val = 0.0
max_s = 1.0
max_phi = 1.0
search_grid = linspace(epsilon, 1.0, 15)
for (i, x) in enumerate(jv.x_grid)
function w(z)
s, phi = z
function h(u)
out = similar(u)
for j in 1:length(u)
out[j] = Vf(max(G(x, phi), u[j])) * pdf(F, u[j])
end
out
end
integral = do_quad(h, nodes, weights)
q = pi_func(s) * integral + (1.0 - pi_func(s)) * Vf(G(x, phi))
return - x * (1.0 - phi - s) - bet * q
end
for s in search_grid
for phi in search_grid
if s + phi <= 1.0
cur_val = -w((s, phi))
else
cur_val = -1.0
end
if cur_val > max_val
max_val, max_s, max_phi = cur_val, s, phi
end
end
end
s_policy[i], phi_policy[i] = max_s, max_phi
end
end
function bellman_operator(jv::JvWorker, V::Vector; ret_policies=false)
if ret_policies
out = (similar(V), similar(V))
else
out = similar(V)
end
bellman_operator!(jv, V, out)
return out
end
```

The code is written to be relatively generic—and hence reusable

- For example, we use generic \(G(x,\phi)\) instead of specific \(A (x \phi)^{\alpha}\)

Regarding the imports

`fixed_quad`

is a simple non-adaptive integration routine`fmin_slsqp`

is a minimization routine that permits inequality constraints

Next we build a type called `JvWorker`

that

- packages all the parameters and other basic attributes of a given model
- Implements the method
`bellman_operator`

for value function iteration

The `bellman_operator`

method
takes a candidate value function \(V\) and updates it to \(TV\) via

where

Here we are minimizing instead of maximizing to fit with SciPy’s optimization routines

When we represent \(V\), it will be with a Julia array `V`

giving values on grid `x_grid`

But to evaluate the right-hand side of (3), we need a function, so
we replace the arrays `V`

and `x_grid`

with a function `Vf`

that gives linear
interpolation of `V`

on `x_grid`

Hence in the preliminaries of `bellman_operator`

from the array

`V`

we define a linear interpolation`Vf`

of its values`c1`

is used to implement the constraint \(s + \phi \leq 1\)`c2`

is used to implement \(s \geq \epsilon\), a numerically stablealternative to the true constraint \(s \geq 0\)

`c3`

does the same for \(\phi\)

Inside the `for`

loop, for each `x`

in the grid over the state space, we
set up the function \(w(z) = w(s, \phi)\) defined in (3).

The function is minimized over all feasible \((s, \phi)\) pairs, either by

- a relatively sophisticated solver from SciPy called
`fmin_slsqp`

, or - brute force search over a grid

The former is much faster, but convergence to the global optimum is not guaranteed. Grid search is a simple way to check results

## Solving for Policies¶

Let’s plot the optimal policies and see what they look like

The code is in `this file`

and looks as follows

```
V = compute_fixed_point(f, v_init, max_iter=200)
s_policy, phi_policy = bellman_operator(wp, V, ret_policies=true)
# === plot solution === #
plot(wp.x_grid, [phi_policy s_policy V],
title = ["phi policy" "s policy" "value function"],
color = [:orange :blue :green], width = 3,
xaxis = ("x", (0.0, maximum(wp.x_grid))),
yaxis = ((-0.1, 1.1)),
layout = (3,1), bottom_margin = 20mm, show = true)
```

It produces the following figure

The horizontal axis is the state \(x\), while the vertical axis gives \(s(x)\) and \(\phi(x)\)

Overall, the policies match well with our predictions from section.

- Worker switches from one investment strategy to the other depending on relative return
- For low values of \(x\), the best option is to search for a new job
- Once \(x\) is larger, worker does better by investing in human capital specific to the current position

## Exercises¶

### Exercise 1¶

Let’s look at the dynamics for the state process \(\{x_t\}\) associated with these policies.

The dynamics are given by (1) when \(\phi_t\) and \(s_t\) are chosen according to the optimal policies, and \(\mathbb{P}\{b_{t+1} = 1\} = \pi(s_t)\).

Since the dynamics are random, analysis is a bit subtle

One way to do it is to plot, for each \(x\) in a relatively fine grid
called `plot_grid`

, a
large number \(K\) of realizations of \(x_{t+1}\) given \(x_t =
x\). Plot this with one dot for each realization, in the form of a 45 degree
diagram. Set:

```
K = 50
plot_grid_max, plot_grid_size = 1.2, 100
plot_grid = linspace(0, plot_grid_max, plot_grid_size)
fig, ax = subplots()
ax[:set_xlim](0, plot_grid_max)
ax[:set_ylim](0, plot_grid_max)
```

By examining the plot, argue that under the optimal policies, the state \(x_t\) will converge to a constant value \(\bar x\) close to unity

Argue that at the steady state, \(s_t \approx 0\) and \(\phi_t \approx 0.6\).

### Exercise 2¶

In the preceding exercise we found that \(s_t\) converges to zero and \(\phi_t\) converges to about 0.6

Since these results were calculated at a value of \(\beta\) close to
one, let’s compare them to the best choice for an *infinitely* patient worker.

Intuitively, an infinitely patient worker would like to maximize steady state wages, which are a function of steady state capital.

You can take it as given—it’s certainly true—that the infinitely patient worker does not search in the long run (i.e., \(s_t = 0\) for large \(t\))

Thus, given \(\phi\), steady state capital is the positive fixed point \(x^*(\phi)\) of the map \(x \mapsto G(x, \phi)\).

Steady state wages can be written as \(w^*(\phi) = x^*(\phi) (1 - \phi)\)

Graph \(w^*(\phi)\) with respect to \(\phi\), and examine the best choice of \(\phi\)

Can you give a rough interpretation for the value that you see?