Single Agent Dynamics

Introduction

Motivation

IO: role of market structure on equilibrium outcomes.

Dynamics: study the endogenous evolution of market structure.

• Supply side dynamics
  • Irreversible investment
  • Entry sunk costs
  • Product repositioning costs
  • Price adjustment costs
  • Learning by doing
• Demand side dynamics
  • Switching costs
  • Durable or storable products

Bonus motivation: AI literature studies essentially the same set of problems with similar tools (Igami 2020)

• Irony: niche topic in IO (super niche in econ), but at the core of the frontier in computer science
  • Why? Computation is hard, estimation harder, but extremely powerful prediction tool
  • The world is intrinsecally dynamic

Examples (1)

Some examples in empirical IO

• Investment
  • Rust (1987): bus engine replacement decision
• Durable goods
  • Gowrisankaran and Rysman (2012): consumer demand in the digital camcorder industry
• Stockpiling
  • Erdem, Imai, and Keane (2003): promotions and stockpiling of ketchup
  • Hendel and Nevo (2006): stockpiling of laundry detergents
• Learning
  • Erdem and Keane (1996): brand learning in the laundry detergent industry
  • Crawford and Shum (2005): demand learning of anti‐ulcer drug prescriptions
• Switching costs
  • Handel (2013): inertia in demand for health insurance

Examples (2)

But also in other applied micro fields:

• Labor economics
  • Should you go to college? (Keane and Wolpin 1997)
• Health economics

  • Which health insurance to pick given there are switching costs? (Handel 2013)

  • Addiction (Becker and Murphy 1988)

• Public finance
  • How should you set optimal taxes in a dynamic environment? (Golosov et al. 2006)

Do we really need dynamics?

In some cases, we can reduce a dynamic problem to a:

1. Static problem
2. Reduced-form problem

E.g., Investment decision

• Dynamic problem, as gains are realized after costs

• “Static” solution: invest if $\mathbb{E}(NPV)>TC$

• Action today ($a_t=0$ or $1$) does not affect the amount of future payoffs (NPV)

But many cases where it’s hard to evaluate dynamic questions in a static/reduced-form setting.

• Typically, cases where decision today would affect payoffs tomorrow
• And you care about those payoffs ($\ne$ myopia)

“A dynamic model can do anything a static model can.”

New Empirical IO

So-called New Empirical IO (summary in Bresnahan (1989))

• Some decisions today might affect payoffs tomorrow
• But the decision today depends on the state today
• And the state today might have been the result of a decision yesterday
• Etc…
• Need dynamics to study these questions
• Where does it all start?
  • Pakes (1986)
  • Berry (1992)

Pros and Cons

Advantages

• We can adress intertemporal trade-offs

  • Flow vs stock stocks and benefits

• We can examine transitions and not only steady states

• We are able to address policy questions that cannot be addressed with reduced-form methods

  • Standard advantage of structural estimation
  • But in a context with relevant intertemporal trade-offs / decisions

Disadvantages

• We typically need more assumptions

  • Robustness testing will therefore be important

• Identification in dynamic models is less transparent

  • Thus time should be spent articulating what variation in the data identifies our parameters of interest)

• It is often computationally intensive (i.e., slow / unfeasible)

From Statics to Dynamics

Typical steps

1. Specify the primitives of the model
   • Static: single period agents’ payoff functions (utility or profit)
   • Dynamic: static payoffs + evolution of state variables
     • Can be exogenous
     • … or endogenous: decision today has an effect on the state tomorrow
2. Solve for optimal behavior
   • Static: tipically agents maximize current utility or profit
   • Dynamic: agents maximize present discounted value of future utilities or profits
3. Search for parameter values that result in the “best match” between our model predictions and observed behavior

1st year Macro Recap

Markov Decision Processes

Formally, a discrete-time MDP consists of the following objects

• A discrete time index $t\in \{0,1,2,\dots ,T\}$, for $T\le \mathrm{\infty }$

• A state space $\mathcal{S}$

• An action space $\mathcal{A}$

  • and a family of constraint sets $\{{\mathcal{a}}_t(s_t)\subseteq \mathcal{A}\}$

• A family of transition probabilities $\{\underset{t}{Pr}(s_{t+1}|s_t,a_t)\}$

• A discount factor, $\beta$

• A family of single-period reward functions $\{(u_t(s_t,a_t)\}$

  • so that the utility functional $U$ has an additively separable decomposition $$U(\mathit{s},\mathit{a})=\sum _{t=0}^T{\beta }^t{u}_t(s_t,a_t)$$

MDP (2)

In words

• The state space $\mathcal{S}$ contains all the information needed to

  • compute static utilities $u_t(s_t,a_t)$
  • compute transition probabilities $\{\underset{t}{Pr}(s_{t+1}|s_t,a_t)\}$

• The (conditional) action space $\mathcal{A}(s_t)$ contains all the actions available in state $s_t$

  • How can it be different by state? E.g. entry/exit decision if you’re in/out of the market

• The transition probabilities $\{\underset{t+1}{Pr}(s_{t+1}|s_t,a_t)\}$ define the probabilities of future states $s_{t+1}$ conditional on

  • Present state $s_t$
  • Present decision $a_t$

• The discount factor $\beta$ together with the static reward functions $\{(u_t(s_t,a_t)\}$ determines the objective function $${\mathbb{E}}_{{\mathit{s}}^{\prime }}[\sum _{t=0}^T{\beta }^t{u}_t(s_t,a_t)]$$

Notation

Brief parenthesis on notation

• I have seen states denoted as

  • $s$ (for state)
  • $x$
  • $\omega$
  • others, depending on the specific context, e.g. $e$ for experience

  I will try to stick to $s$ all the time

• I have seen decisions denoted as

  • $a$ (for action)
  • $d$ (for decision)
  • $x$
  • others, depending on the specific context, e.g. $i$ for investment

  I will try to stick to $a$ all the time

Maximization Problem

The objective is to pick the decision rule (or policy function) $P={\mathit{a}}^{\ast }=\{a_1^{\ast },\dots ,a_t^{\ast }\}$ that solves $$\underset{\mathit{a}}{max}{\mathbb{E}}_{{\mathit{s}}^{\prime }}[\sum _{t=0}^T{\beta }^t{u}_t(s_t,a_t)]$$ Where the expectation is taken over transition probabilities generated by the decision rule $\mathit{a}$.

Stationarity

In many applications, we assume stationarity

• The transition probabilities and utility functions do not directly depend on $t$

  • i.e., are the same for all $t$
    • $\underset{t}{Pr}(s_{t+1}|s_t,a_t)\to Pr(s_{t+1}|s_t,a_t)$
    • $u_t(s_t,a_t)\to u(s_t,a_t)$

• Uncomfortable assumption?

• You think there is some reason (variable) why today’s probabilities should be different from tomorrow’s?

  • If observable, include that variable in the state space
  • If unobservable, integrate it out

Stationarity (2)

• In the finite horizon case ($T\le \mathrm{\infty }$), stationarity does not help much

  • $\sum _{t=0}^T{\beta }^{t}u(s_t,a_t)$ still depends on $t$, conditional on $s_t$
  • Why? Difference between $t$ and $T$ matters in the sum

• In infinite-horizon problems, stationarity helps a lot

  • Now the difference between $t$ and $T$ is always the same, i.e. $\mathrm{\infty }$

  • $\sum _{t=0}^{\mathrm{\infty }}{\beta }^{t}u(s_t,a_t)$ does not depend on $t$, conditional on $s_t$

  • The future looks the same whether the agent is in state $s_t$ at time $t$ or in state $s_{t+\tau }=s_t$ at time $t+\tau$

Value Function

Consider a stationary infinite-horizon problem

• The only variable which affects the agent’s view about the future is the current value of the state, $s_t$

• We can rewrite the agent’s problem as $$V_0(s_0)=\underset{\mathit{a}}{max}{\mathbb{E}}_{{\mathit{s}}^{\prime }}[\sum _{t=0}^{\mathrm{\infty }}{\beta }^{t}u(s_t,a_t)]$$ where

  • $a_t\in \mathcal{A}(s_t)\mathrm{\forall }t$
  • The expectation is taken over future states ${\mathit{s}}^{\prime }$
    • that evolve according to $\{Pr(s_{t+1}|s_t,a_t)\}$
  • $V(\cdot )$ is called the value function

How to solve?

• One could try to solve it by brute force
  • i.e. try to solve for the structure of all of the optimal decisions, ${\mathit{a}}^{\ast }$
  • Indeed, for finite-horizon problems, that might be necessary
• For stationary infinite-horizon problems, the value and policy function should be time invariant
  • $V_t(s_t)=V(s_t)$
  • $P_t(s_t)=P(s_t)$
• What do we gain?

Bellman Equation

$$\begin{array}{rl}V(s_0)& =\underset{\mathit{a}}{max}{\mathbb{E}}_{{\mathit{s}}^{\prime }}[\sum _{t=0}^{\mathrm{\infty }}{\beta }^{t}u(s_t,a_t)]=\\ & =\underset{\mathit{a}}{max}{\mathbb{E}}_{{\mathit{s}}^{\prime }}[u(s_0,a_0)+\sum _{t=1}^{\mathrm{\infty }}{\beta }^{t}u(s_t,a_t)]=\\ & =\underset{\mathit{a}}{max}\{u(s_0,a_0)+{\mathbb{E}}_{{\mathit{s}}^{\prime }}[\sum _{t=1}^{\mathrm{\infty }}{\beta }^{t}u(s_t,a_t)]\}=\\ & =\underset{\mathit{a}}{max}\{u(s_0,a_0)+\beta {\mathbb{E}}_{{\mathit{s}}^{\prime }}[\sum _{t=1}^{\mathrm{\infty }}{\beta }^{t-1}u(s_t,a_t)]\}=\\ & =\underset{a_0}{max}\{u(s_0,a_0)+\beta \underset{\mathit{a}}{max}{\mathbb{E}}_{{\mathit{s}}^{\prime }}[\sum _{t=1}^{\mathrm{\infty }}{\beta }^{t-1}u(s_t,a_t)]\}=\\ & =\underset{a_0}{max}\{u(s_0,a_0)+\beta \int V(s_1)Pr(s_1|s_0,a_0)\}\end{array}$$

Bellman Equation (2)

We have now a recursive formulation of the value function: the Bellman Equation $$V(s_0)=\underset{a_0}{max}\{u(s_0,a_0)+\beta \int V(s_1)Pr(s_1|s_0,a_0)\}$$ Intuition

• The Bellman Equation is a functional equation
  • Has to be satisfied in every state
  • Can be written as $V=T(V)$
  • We are actually looking for a fixed point of $T$

The decision rule that satisfies the Bellman Equation is called the policy function $$a(s_0)=\arg \underset{a_0}{max}\{u(s_0,a_0)+\beta \int V(s_1)Pr(s_1|s_0,a_0)\}$$

Contractions

Under regularity conditions

• $u(s,a)$ is jointly continuous and bounded in $(s,a)$
• $\mathcal{A}(s)$ is a continuous correspondence

It is possible to show that $$T(W)(s)=\underset{a\in \mathcal{A}(s)}{max}\{u(s,a)+\beta \int W(s^{\prime })Pr(s^{\prime }|s,a)\}$$ is a contraction mapping of modulus $\beta$.

• Contraction Mapping Theorem: then $T$ has a unique fixed point!

Solving for the Value Function

How do we actually do it in practice?

• For finite horizon MDPs: backward induction
  • Start from the last period: static maximization problem
  • Move backwards taking the future value as given
• For infinite horizon MDPs: different options
  • value function iteration
    • most common
  • policy function iteration
  • successive approximations

Difference with 1st year Macro

So what’s going to be new here?

1. Estimation: retrieve model primitives from observed behavior
   • And related: uncertainty
2. Strategic interaction: multiple agents taking dynamic decisions
   • Next lecture

Rust (1987)

Setting

Rust (1987): An Empirical Model of Harold Zurcher

• Harold Zurcher (HZ) is the city bus superintendant in Madison, WI

• As bus engines get older, the probability of malfunctions increases

• HZ decides when to replace old bus engines with new ones

  • Optimal stopping / investment problem

• Tradeoff

  • Cost of a new engine (fixed, stock)
  • Repair costs, because of engine failures (continuous, flow)

• Do we care about Harold Zurcher?

  • Obviously not (and neither did Rust), it’s a method paper
  • But referee asked for an application

Data

Units of observation

• Rust observes 162 buses over time

Observables: for each bus, he sees

• monthly mileage (RHS, state variable)
• and whether the engine was replaced (LHS, choice variable),
• in a given month

Variation

• on average, bus engines were replaced every 5 years with over 200,000 elapsed miles
• considerable variation in the time and mileage at which replacement occurs

Idea

• Construct a (parametric) model which predicts the time and mileage at which engine replacement occurs
• Use the model predictions (conditional on parameter values) to estimate parameters that “fit” the data
  • predicted replacements, given mileage VS observed replacements, given mileage
• Ideally use the estimates to learn something new
  • e.g. the correct dynamic demand curve for bus engine replacement

Static Alternative

What would you do otherwise?

• You observe replacement decisions
• … and replacement costs
• $\to$ Regress replacement decision on replacement costs

Problem

• Replacement benefits are a flow (lower maintenance costs)
• … while the cost is a stock

Outcome

• We expect the overestimate demand elasticity. Why?
• Overpredict substitutions at low costs
• and underpredict substitution at high cost

Model

Assumptions of the structural model

• State: $s_t\in \{0,\dots ,s_{max}\}$
  • engine accumulated mileage at time $t$
  • Note: “continuous” in the data but has to be discretized into bins
• Action: $a_t\in \{0,1\}$
  • replace engine at time $t$
• State transitions: $Pr(s_{t+1}|s_0,\dots ,s_t;\theta )=Pr(s_{t+1}|s_t;\theta )$
  • mileage $s_t$ evolves exogenously according to a 1st-order Markov process
  • The transition function is the same for every bus.
  • If HZ replaces in period $t$ ($a_t=1$), then $s_t=0$

Model (2)

HZ static utility function (for a single bus) $$u(s_t,a_t;\theta )=\{\begin{array}{ll}-c(s_t;\theta )& \text{ if }{a}_t=0\text{ (not replace) }\\ -R-c(0;\theta )& \text{ if }{a}_t=1\text{ (replace) }\end{array}$$ where

• $c(s_t;\theta )$: expected costs of operating a bus with mileage $s_t$
  • ​ including maintenance costs & social costs of breakdown
  • We would expect $\frac{\partial c}{\partial s}>0$
• $R$ is the cost of replacement (i.e., a new engine)
  • Note that replacement occurs immediately
• $u(s_t,a_t;\theta )$: expected current utility from operating a bus with mileage $s_t$ and making replacement decision $a_t$

Model (3)

HZ objective function is to maximize the expected present discounted sum of future utilities $$V(s_t;\theta )=\underset{\mathit{a}}{max}{\mathbb{E}}_{s_{t+1}}[\sum _{\tau =t}^{\mathrm{\infty }}{\beta }^{\tau -t}u(s_{\tau },a_{\tau };\theta )|s_t,\mathit{a};\theta ]$$ where

• The expectation $\mathbb{E}$ is over future $x$, which evolve according to Markov process
• $max$ is over future choices $a_{t+1},\dots ,a_{\mathrm{\infty }}$,
  • because HZ will observe future states $s_{\tau }$ before choosing future actions $a_{\tau }$, this is a functional

Notes

• This is for one bus (but multiple engines).
• HZ has an infinite horizon for his decision making
• $s_t$ summarizes state at time $t$, i.e., the expected value of future utilities only depends on $s_t$

Bellman Equation

This (sequential) representation of HZ’s problem is very cumbersome to work with.

We can rewrite $V(s_t;\theta )$ with the following Bellman equation $$V(s_t;\theta )=\underset{a_t}{max}\{u(s_t,a_t;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[V(s_{t+1};\theta )|s_t,a_t;\theta ]\}$$ Basically we are dividing the infinite sum (in the sequential form) into a present component and a future component.

Notes:

• Same $V$ on both sides of equation because of infinite horizon - the future looks the same as the present for a given $s$ (i.e., it doesn’t matter where you are in time).
• The expectation $\mathbb{E}$ is over the state-transition probabilities, $Pr(s_{t+1}|s_t,a_t;\theta )$

Order of Markow Process

Suppose for a moment that $s_t$ follows a second-order markov process $$s_{t+1}=f(s_t,s_{t-1},\epsilon ;\theta )$$ Now $s_t$ is not sufficient to describe current $V$

• We need both $s_t$ and $s_{t-1}$ in the state space (i.e., $V(s_t,s_{t-1};\theta )$ contains $s_{t-1}$, too),
• and the expectation is over the transition probability $Pr(s_{t+1}|s_t,s_{t-1},a_t;\theta )$

Parenthesis: State Variables

Which variables should be state variables? I.e. should be included in the state space?

General rule for 1st order markow processes: variables need to

• define expected current payoff, and
• define expectations over next period state (i.e., distribution of $s_{t+1}$)

What do you do otherwise? Integrate them out! Examples

• Weather affects static utitilies but not transition probabilities
  • More annoying to replace the engine if it rains
  • Integration means: “compute expected utility of Harold Zurcher before he opens the window”
• Month of the year affects transition probabilities but not utilities
  • Buses are used more in the winter
  • Integration means: “compute average transition probabilities over months”

Note: you can always get the non-expected value function if you know the probability of raining or the transition probabilities by month

Policy Function

Along with this value function comes a corresponding policy (or choice) function mapping the state $s_t$ into HZ’s optimal replacement choice $a_t$ $$P(s_t;\theta )=\underset{a_t}{max}\{u(s_t,a_t;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[V(s_{t+1};\theta )|s_t,a_t;\theta ]\}$$ Given $\frac{\partial c}{\partial s}>0$, the policy function has the form $$P(s_t;\theta )=\{\begin{array}{ll}1& \text{ if }{s}_t\ge \gamma (\theta )\\ 0& \text{ if }{s}_t<\gamma (\theta )\end{array}$$ where $\gamma$ is the replacement mileage.

How would this compare with the optimal replacement mileage if HZ was myopic?

• Answer: HZ would wait until $R\le c(s)$ for the replacement action

Solving the Model

Why do we want to solve for the value and policy functions?

• We want to know the agent’s optimal behavior and the equilibrium outcomes
• and be able to conduct comparative statics/dynamics (a.k.a. counterfactual simulations)

We have the Bellman Equation $$V(s_t;\theta )=\underset{a_t}{max}\{u(s_t,a_t;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[V(s_{t+1};\theta )|s_t,a_t;\theta ]\}$$ Which we can compactly write as $$V(s_t;\theta )=T(V(s_{t+1};\theta ))$$ Blackwell’s Theorem: under regularity conditions, $T$ is a contraction mapping with modulus $\beta$.

Contraction Mapping Theorem: $T$ has a fixed point and we can find it by iterating $T$ from any starting value $V^{(0)}$.

Value Function Iteration

What does Blackwell’s Theorem allow us to do?

1. Start with any arbitrary function $\text{Math input error}$
2. Apply the mapping $T$ to get $\text{Math input error}$
3. Apply again $\text{Math input error}$
4. Continue applying $T$ , and $V^{(k)}$ will converge to the unique fixed point of $T$
   • i.e., the true value function $V(s_t;\theta )$
5. Once we have $V(s_t;\theta )$, it’s fairly trivial to compute the policy function $P(s_t;\theta )$
   • Static optimization problem (given $V$)

This process is called value function iteration

How to Reconcile Model and Data?

Ideal Estimation Routine

1. Pick a parameter value $\theta$
2. Solve value and policy function (inner loop)
3. Match predicted choices with observed choices
4. Find the parameter value $\hat{\theta }$ that best fits the data (outer loop)
   • Makes the observed choices “closest” to the predicted choices
   • (or maximizes the likelihood of the observed choices)

Issue: model easily rejected by the data

• The policy function takes the the form: replace iff $s_t\ge \gamma (\theta )$

• Can’t explain the coexistence of e.g. “a bus without replacement at 22K miles” and “another bus being replaced at 17K miles” in the data

• We need some unobservables in the model to explain why observed choices do not exactly match predicted choices

Rust (1987) - Estimation

Uncertainty

How can we explain different replacement actions at different mileages in the data?

1. Add other observables
2. Add some stochastic element

But where? Two options

1. Randomness in decisions
   • I.e. “Harold Zurcher sometimes would like to replace the bus engine but he forgets”
   • Probably still falsifiable
   • Also need “Harold Zurcher sometimes would like not to replace but replacement happens” 🤔🤔🤔
2. Randomness in the state
   • Harold Zurcher knows something that we don’t
   • He always makes the optimal decision but based on somethig we don’t observe

Unobservables

Rust uses the following utility specification: $$u(s_t,a_t,{ϵ}_t;\theta )=u(s_t,a_t;\theta )+{ϵ}_{a_{t}t}=\{\begin{array}{ll}-c(s_t;\theta )+{ϵ}_{0t}& \text{ if }{a}_t=0\\ \\ -R-c(0;\theta )+{ϵ}_{1t}& \text{ if }{a}_t=1\end{array}$$

• The ${ϵ}_{it}$ are components of utility of alternative $a$ that are observed by HZ but not by us, the econometrician.
  • E.g., the fact that an engine is running unusually smoothly given its mileage,
  • or the fact that HZ is sick and doesn’t feel like replacing the engine this month
• Note: we have assumed addictive separability of $ϵ$
• The ${ϵ}_a$s also affect HZ’s replacement decision
• ${ϵ}_{it}$ are both observed and relevant $\to$ part of the state space

Can we still solve the model? Can we estimate it?

Unobservables (2)

The Bellman Equation becomes $$V(\{s_{\tau }{\}}_{\tau =1}^t,\{{ϵ}_{\tau }{\}}_{\tau =1}^t;\theta )=\underset{a_t}{max}\{u(s_t,a_t;\theta )+{ϵ}_{it}+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{it+1};\theta )|\{s_{\tau }{\}}_{\tau =1}^t,\{{ϵ}_{\tau }{\}}_{\tau =1}^t,a_t;\theta ]\}$$ Issues

• The problem is not Markow anymore
  • Is ${ϵ}_t$ correlated with ${ϵ}_{t-\tau }$? How?
  • Is ${ϵ}_t$ correlated with $s_t$? And $s_{t-\tau }$? How?
• Dimension of the state space has increased
  • From $k=(k\text{ points}{)}^{1\text{ variable}\times 1\text{ period}}$ points, to $\mathrm{\infty }=(k\text{ points}{)}^{3\text{ variables}\times \mathrm{\infty }\text{ periods}}$ 🤯🤯
  • Assuming all variables assume $k$ values
• Number of variables to integrate over to compute expectation $\mathbb{E}$ has increased
  • From one variable, $s$, to three, $(s,{ϵ}_0,{ϵ}_1)$

Assumptions

Rust makes 4 assumptions to make the problem tractable:

1. First order Markow process of $ϵ$
2. Conditional independence of ${ϵ}_t|s_t$ from ${ϵ}_{t-1}$ and $s_{t-1}$
3. Independence of ${ϵ}_t$ from $s_t$
4. Logit distribution of $ϵ$

Assumption 1

A1: first-order markov process of $ϵ$ $$Pr(s_{t+1},{ϵ}_{t+1}|s_1,\dots ,s_t,{ϵ}_1,\dots ,{ϵ}_t,a_t;\theta )=Pr(s_{t+1},{ϵ}_{t+1}|s_t,{ϵ}_t,a_t;\theta )$$

• What it buys

  • $s$ and $ϵ$ prior to current period are irrelevant

• What it still allows:

  • allows $s_t$ to be correlated with ${ϵ}_t$

• What are we assuming away

  • Any sort of longer run dependence
  • Does it matter? If yes, just re-consider what is one time period
  • Or make the state space larger (as usual in Markow processes)

Assumption 1 - Implications

The Bellman Equation becomes $$V(s_t,{ϵ}_t;\theta )=\underset{a_t}{max}\{u(s_t,a_t;\theta )+{ϵ}_{a_{t}t}+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{t+1};\theta )|s_t,a_t,{ϵ}_t;\theta ]\}$$

• Now the state is $(s_t,{ϵ}_t)$
  • sufficient, because defines both current utility and (the expectation of) next-period state, under the first-order Markov assumption
  • ${ϵ}_t$ is now analogous to $s_t$
  • State space now is $k^3=(k\text{ points}{)}^{3\text{ variables}\times 1\text{ period}}$
    • From $\mathrm{\infty }=(k\text{ points}{)}^{3\text{ variables}\times \mathrm{\infty }\text{ periods}}$
• Now we could use value function iteration to solve the problem
  • If ${ϵ}_t$ is continuous, it has to be discretised

Assumption 1 - Issues

Open issues

1. Curse of dimensionality in the state space: ($s_t,{ϵ}_{0t},{ϵ}_{1t}$)

   • Before, there were $k$ points in state space (discrete values of $x$)
   • Now there are $k^3$ : $k$ each for $s$, ${ϵ}_0$, ${ϵ}_1$
     • (Assuming we discretize all state variables into $k$ values)
   • Generally, number of points in state space (and thus computational time) increases exponentially in the number of variables

2. Curse of dimensionality in the expected value: ${\mathbb{E}}_{s_{t+1},{ϵ}_{0,t+1},{ϵ}_{1,t+1}}$

   • For each point in state space (at each iteration of the contraction mapping), need to compute

   $${\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{t+1};\theta )|s_t,a_t,{ϵ}_t;\theta ]$$

   • Before, this was a 1-dimensional integral (or sum), now it’s 3-dimensional

3. Initial conditions

Assumption 2

A2: conditional independence of ${ϵ}_t|s_t$ from ${ϵ}_{t-1}$ and $s_{t-1}$ $$Pr(s_{t+1},{ϵ}_{t+1}|s_t,{ϵ}_t,a_t;\theta )=Pr({ϵ}_{t+1}|s_{t+1};\theta )Pr(s_{t+1}|s_t,a_t;\theta )$$

• What it buys

  • $s_{t+1}$ is independent of ${ϵ}_t$
  • ${ϵ}_{t+1}$ is independent of ${ϵ}_t$ and $s_t$, conditional on $s_{t+1}$

• What it still allows:

  • $ϵ$ can be correlated across time, but only through the $s$ process

• What are we assuming away

  • Any time of persistent heterogeneity

  • Does it matter? Easily yes

  • There are tons of applications where the unobservables are either fixed or correlated over time

    • If fixed, there are methods to handle unobserved heterogeneity (i.e. bus “types”)

Assumption 2 - Implications

The Bellman Equation is $$V(s_t,{ϵ}_t;\theta )=\underset{a_t}{max}\{u(s_t,a_t;\theta )+{ϵ}_{a_{t}t}+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{t+1};\theta )|s_t,a_t;\theta ]\}$$

• Now ${ϵ}_t$ is noise that doesn’t affect the future
  • That is, conditional on $s_{t+1}$, ${ϵ}_{t+1}$ is uncorrelated with ${ϵ}_t$

Remeber: if $ϵ$ does not affect the future, it should’t be in the state space!

How? Integrate it out.

Rust Shortcut: ASV

Rust: define the alternative-specific value function $$\begin{array}{rl}& {\overline{V}}_0(s_t;\theta )=u(s_t,0;\theta )+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{t+1};\theta )|s_t,a_t=0;\theta ]\\ & {\overline{V}}_1(s_t;\theta )=u(s_t,1;\theta )+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{t+1};\theta )|s_t,a_t=1;\theta ]\end{array}$$

• ${\overline{V}}_0(s_t)$ is the present discounted value of not replacing, net of ${ϵ}_{0t}$

• The state does not depend on ${ϵ}_t$!

• What is the relationship with the value function? $$V(s_t,{ϵ}_t;\theta )=\underset{a_t}{max}\{\begin{array}{l}{\overline{V}}_0(s_t;\theta )+{ϵ}_{0t};\\ {\overline{V}}_1(s_t;\theta )+{ϵ}_{1t}\end{array}\}$$

• We have a 1-to-1 mapping between $V(s_t,{ϵ}_t;\theta )$ and ${\overline{V}}_a(s_t;\theta )$ !

  • If we have one, we can get the other

Rust Shortcut

Can we solve for $\overline{V}$?

Yes! They have a recursive formulation $$\begin{array}{rl}& {\overline{V}}_0(s_t;\theta )=u(s_t,0;\theta )+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[\underset{a_{t+1}}{max}\{\begin{array}{c}{\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1};\\ {\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\end{array}\}|s_t,a_t=0;\theta ]\\ & {\overline{V}}_1(s_t;\theta )=u(s_t,1;\theta )+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[\underset{a_{t+1}}{max}\{\begin{array}{c}{\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1};\\ {\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\end{array}\}|s_t,a_t=1;\theta ]\end{array}$$

• Rust (1988) shows that it’s a joint contraction mapping
• Memo: the state space now is $2k=(2\text{ actions})\times (k\text{ points}{)}^{1\text{ variables}\times 1\text{ period}}$
  • instead of $3^k=(k\text{ points}{)}^{3\text{ variables}\times 1\text{ period}}$
  • Much smaller!
• Lesson: any state variable that does not affect continuation values (the future) does not have to be in the “actual” state space

Assumption 2 - Implications

We can also split the expectation in the alternative-specific value function $$\begin{array}{rl}& {\overline{V}}_0(s_t;\theta )=u(s_t,0;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[{\mathbb{E}}_{{ϵ}_{t+1}}[\underset{a_{t+1}}{max}\{\begin{array}{c}{\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1};\\ {\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\end{array}\}|s_t]|s_t,a_t=0;\theta ]\\ & {\overline{V}}_1(s_t;\theta )=u(s_t,1;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[{\mathbb{E}}_{{ϵ}_{t+1}}[\underset{a_{t+1}}{max}\{\begin{array}{c}{\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1};\\ {\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\end{array}\}|s_t]|s_t,a_t=1;\theta ]\end{array}$$ This allows us to concentrate on one single term $${\mathbb{E}}_{{ϵ}_{t+1}}[\underset{a_{t+1}}{max}\{\begin{array}{l}{\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1};\\ {\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\end{array}\}|s_t]$$ Open issues

• Distribution of ${ϵ}_{t+1}$ has to be simulated
• Distribution of ${ϵ}_{t+1}$ depends on $s_t$

Assumption 3

A3: independence of ${ϵ}_t$ from $s_t$ $$Pr({ϵ}_{t+1}|s_{t+1};\theta )Pr(s_{t+1}|s_t,a_t;\theta )=Pr({ϵ}_{t+1}|\theta )Pr(s_{t+1}|s_t,a_t;\theta )$$

• What it buys

  • $ϵ$ not correlated with anything $${\mathbb{E}}_{{ϵ}_{t+1}}[\underset{a_{t+1}\in \{0,1\}}{max}\{\begin{array}{l}{\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1};\\ {\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\end{array}\}]$$

• What are we assuming away

  • Some state-specific noise… probably irrelevant

• Open Issues

  • Distribution of ${ϵ}_{t+1}$ has to be simulated

Assumption 4

A4: $ϵ$ is type 1 extreme value distributed (logit)

• What it buys

  • Closed form solution for ${\mathbb{E}}_{{ϵ}_{t+1}}$

• What are we assuming away

  • Different substitution patterns

  • Relevant? Maybe, if there are at least three options (here binary choice)

    • As logit assumption in demand estimation

Logit magic 🧙🪄 $${\mathbb{E}}_{ϵ}[\underset{n}{max}(\{{\delta }_n+{ϵ}_n{\}}_{n=1}^N)]=0.5772+\ln (\sum _{n=1}^N{e}^{{\delta }_n})$$

where $0.5772$ is Euler’s constant

Assumption 4 - Implications

The Bellman equation becomes $$\begin{array}{rl}& {\overline{V}}_0(s_t;\theta )=u(s_t,0;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[0.5772+\ln (\sum _{a^{\prime }\in \{0,1\}}{e}^{{\overline{V}}_{a^{\prime }}(s_{t+1};\theta )})|s_t,a_t=0;\theta ]\\ & {\overline{V}}_1(s_t;\theta )=u(s_t,1;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[0.5772+\ln (\sum _{a^{\prime }\in \{0,1\}}{e}^{{\overline{V}}_{a^{\prime }}(s_{t+1};\theta )})|s_t,a_t=1;\theta ]\end{array}$$

• We got fully rid of $ϵ$!
  • How? With a lot of assumptions

Estimation

So far we have analysized how the 4 assumptions help solving the model.

• What about estimation?

Maximum Likelihood

• For a single bus, the likelihood function is

$$\mathcal{L}=Pr(s_1,\dots ,s_T,a_0,\dots ,a_T|s_0;\theta )$$

• i.e. probability of observed decisions $\{a_0,\dots ,a_T\}$
• and sequence of states $\{s_1,\dots ,s_T\}$
• conditional on the initial state $s_0$
• and the parameter values $\theta$

What is the impact of the 4 assumptions on the likelihood function?

Likelihood Function (A1)

A1: First order Markow process of $ϵ$

• We gain independence across time
• We can decompose the joint distribution in marginals across time

$$\begin{array}{rl}\mathcal{L}(\theta )& =Pr(s_1,\dots ,s_T,a_0,\dots ,a_T|s_0;\theta )\\ & =\prod _{t=1}^{T}Pr(a_{t+1},s_{t+1}|s_t,a_t;\theta )\end{array}$$

Likelihood Function (A2)

A2: independence of ${ϵ}_t$ from ${ϵ}_{t-1}$ and $s_{t-1}$ on $s_t$

• We can decompose the joint distribution of $a_t$ and $s_{t+1}$ into marginals $$\begin{array}{rl}\mathcal{L}(\theta )& =\prod _{t=1}^{T}Pr(a_{t+1},s_{t+1}|s_t,a_t;\theta )=\\ & =\prod _{t=1}^{T}Pr(a_t|s_t;\theta )Pr(s_{t+1}|s_t,a_t;\theta )\end{array}$$

• $Pr(s_{t+1}|s_t,a_t;\theta )$ can be estimated from the data

  • we’ll come back to it

• for $Pr(a_t|s_t;\theta )$ we need the two remaining assumptions

Likelihood Function (A3)

A3: Independence of ${ϵ}_t$ from $s_t$

• No need to condition on $s_t$
• E.g. probability of replacement

$$\begin{array}{rl}Pr(a_t=1|s_t;\theta )& =Pr({\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\ge {\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1}|s_t;\theta )=\\ & =Pr({\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\ge {\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1}|\theta )\end{array}$$

• In words: same distribution of shocks in every state

Likelihood Function (A4)

A4: Logit distribution of $ϵ$

• E.g. probability of replacement becomes

$$\begin{array}{rl}Pr(a_t=1|s_t;\theta )& =Pr({\overline{V}}_1(s_{t+1};\theta )+{ϵ}_{1t+1}\ge {\overline{V}}_0(s_{t+1};\theta )+{ϵ}_{0t+1}|\theta )=\\ & =\frac{e^{{\overline{V}}_1(s_{t+1};\theta )}}{e^{{\overline{V}}_0(s_{t+1};\theta )}+e^{{\overline{V}}_1(s_{t+1};\theta )}}\end{array}$$

• We have a closed form expression!

Likelihood Function

The final form of the likelihood function for one bus is $$\mathcal{L}(\theta )=\prod _{t=1}^{T}Pr(a_t|s_t;\theta )Pr(s_{t+1}|s_t,a_t;\theta )$$ where

• $Pr(s_{t+1}|s_t,a_t;\theta )$ can be estimated from the data
  • given mileage $x$ and investment decision $a$, what are the observed frequencies of future states $x^{\prime }$?
  • does not have to depend on $\theta$
• $Pr(a_t|s_t;\theta )$ depends on ${\overline{V}}_a(s;\theta )$
  • ${\overline{V}}_a(s;\theta )$ we know how to compute
  • given a value of $\theta$
  • solve by value function iteration

Likelihood Function (2)

Since we have may buses, $j$, the likelihood of the data is $$\mathcal{L}(\theta )=\prod _j{\mathcal{L}}_j(\theta )=\prod _j\prod _{t=1}^{T}Pr(a_{jt}|s_{jt};\theta )Pr(s_{j,t+1}|s_{jt},a_{jt};\theta )$$ And, as usual, we prefer to work with log-likelihoods $$\log \mathcal{L}(\theta )=\sum _j\sum _{t=1}^T(\log Pr(a_{jt}|s_{jt};\theta )+\log Pr(s_{j,t+1}|s_{jt},a_{jt};\theta ))$$

Estimation

Now we have all the pieces to estimate $\theta$!

Procedure

1. Estimate the state transition probabilities $Pr(s_{t+1}|s_t,a_t;\theta )$
2. Select a value of $\theta$
3. Init a choice-specific value function ${\overline{V}}_a^{(0)}(s_{t+1};\theta )$
   1. Apply the Bellman operator to compute ${\overline{V}}_a^{(1)}(s_{t+1};\theta )$
   2. Iterate until convergence to ${\overline{V}}_d^{(k\to \mathrm{\infty })}(s_{t+1};\theta )$ (inner loop)
4. Compute the choice probabilities $Pr(a_t|s_t;\theta )$
5. Compute the likelihood $\mathcal{L}=\prod _j\prod _{t=1}^{T}Pr(a_t|s_t;\theta )Pr(s_{t+1}|s_t,a_t;\theta )$
6. Iterate (2-5) until you are have found a (possibly global) minimum (outer loop)

Results

What do dynamics add?

• Static demand curve ($\beta =0$) is much more sensitive to the price of engine replacement. Why?
  • Compares present price with present savings
• If you compare present price with flow of future benefits, you are less price sensitive
  • More realistic

Extensions

Main limitation of Rust (1987): value function iteration

• Costly: has to be done for each parameter explored during optimization
• Particularly costly if the state space is large

Solutions

1. Solve the model without solving a fixed point problem
   • Hotz and Miller (1993)
2. Solve the model and estimate the parameters at the same time
   • Inner and outer loop in parallel
   • Imai, Jain, and Ching (2009)
3. Treat the estimation as a constrained optimization problem
   • MPEC, as for demand
   • Use off-the-shelf optimization algorithms
   • Su and Judd (2012)

We’ll cover Hotz and Miller (1993) since at the core of the estimation of dynamic games.

Hotz & Miller (1993)

Motivation

Setting: Harold Zurcher problem

• same model
• same assumptions
• same notation
• same objective

Problem: computationally intense to do value function iteration

Can we solve the model without solving a fixed point problem?

Estimation in Rust

How did we estimate the model in Rust? Two main equations

1. Solve the Bellman equation of the alternative-specific value function $$\overline{V}(s;\theta )=\tilde{f}(\overline{V}(s;\theta ))$$

2. Compute the expected policy function $$P(\cdot |s;\theta )=\tilde{g}(\overline{V}(s;\theta );\theta )$$

3. Maximize the likelihood function

$$\mathcal{L}(\theta )=\prod _j\prod _{t=1}^{T}Pr(a_{jt}|s_{jt};\theta )Pr(s_{j,t+1}|s_{jt},a_{jt};\theta )$$

Can we remove step 1?

Hotz & Miller Idea(s)

Idea 1: it would be great if we could start from something like $$P(\cdot |s;\theta )=T(P(\cdot |s;\theta );\theta )$$

• No need to solve for the value function
• But we would still need a to solve a fixed point problem
• Back from the start? No

Idea 2: could replace the RHS element with a consistent estimate $$P(\cdot |s;\theta )=T(\hat{P}(\cdot |s;\theta );\theta )$$ And this could give us an estimating equation!

Unclear? No problem, let’s go slowly step by step

Two Main Equations

1. Bellman equation $${\overline{V}}_a(s_t;\theta )=u(s_t,a;\theta )+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[\underset{a^{\prime }}{max}\{{\overline{V}}_{a^{\prime }}(s_{t+1};\theta )+{ϵ}_{a^{\prime },t+1}\}|s_t,a_t=a;\theta ]$$

2. Expected policy function

$$Pr(a_t=a|s_t;\theta )=Pr({\overline{V}}_a(s_{t+1};\theta )+{ϵ}_{a,t+1}\ge {\overline{V}}_{a^{\prime }}(s_{t+1};\theta )+{ϵ}_{a^{\prime },t+1},\mathrm{\forall }{a}^{\prime }|\theta )$$

Expected decision before the shocks ${ϵ}_t$ are realized

• Not the policy function
  • The policy function maps $s_t\times ϵ\to \{0,1\}$
  • The expected policy function maps $s_t\to [0,1]$
• Easier to work with: does not depend on the shocks
  • Not a deterministic policy, but a stochastic one

Hotz & Miller - Idea 1

How do we get from the two equations $$\begin{array}{rl}\overline{V}(s;\theta )& =\tilde{f}(\overline{V}(s;\theta ))\\ P(\cdot |s;\theta )& =\tilde{g}(\overline{V}(s;\theta );\theta )\end{array}$$ To one? $$P(\cdot |s;\theta )=T(P(\cdot |s;\theta );\theta )$$ If we could express $\overline{V}$ in terms of $P$, … $$\begin{array}{rl}\overline{V}(s;\theta )& =\tilde{h}(P(\cdot |s;\theta ))\\ P(\cdot |s;\theta )& =\tilde{g}(\overline{V}(s;\theta );\theta )\end{array}$$

…. we could then substitute the first equation into the second …

But, easier to work with a different representation of the value function.

Expected Value Function

Recall Rust value function (not the alternative-specific $\overline{V}$) $$V(s_t,{ϵ}_t;\theta )=\underset{a_t}{max}\{u(s_t,a_t;\theta )+{ϵ}_{a_{t}t}+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{t+1};\theta )|s_t,a_t;\theta ]\}$$ We can express it in terms of expected value function

$$V(s_t;\theta )={\mathbb{E}}_{{ϵ}_t}[\underset{a_t}{max}\{u(s_t,a_t;\theta )+{ϵ}_{a_{t}t}+\beta {\mathbb{E}}_{s_{t+1}}[V(s_{t+1};\theta )|s_t,a_t;\theta ]\}]$$

• Value of being in state $s_t$ without knowing the realization of the shock ${ϵ}_t$

  • “Value of Harold Zurcher before opening the window and seeing if it’s raining or not”

• Analogous to the relationship between policy funciton and expected policy function

• Note

  • expectation of future value now is only over $s_{t+1}$
  • $V(s_t;\theta )$ can be solved via value function iteration as the operator on the RHS is a contraction

Representation Equivalence

Recall the alternative-specific value function of Rust

$$\begin{array}{rl}{\overline{V}}_a(s_t;\theta )& =u(s_t,d;\theta )+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[\underset{a^{\prime }}{max}\{{\overline{V}}_{a^{\prime }}(s_{t+1};\theta )+{ϵ}_{a^{\prime },t+1}\}|s_t,a_t=a;\theta ]\\ & =u(s_t,a;\theta )+\beta {\mathbb{E}}_{s_{t+1},{ϵ}_{t+1}}[V(s_{t+1},{ϵ}_{t+1};\theta )|s_t,a_t=a;\theta ]\\ & =u(s_t,a;\theta )+\beta {\mathbb{E}}_{s_{t+1}}[V(s_{t+1};\theta )|s_t,a_t=a;\theta ]\end{array}$$

Relationship with the value function

$$V(s_t,{ϵ}_t;\theta )=\underset{a_t}{max}\{{\overline{V}}_0(s_t;\theta )+{ϵ}_{0t},{\overline{V}}_1(s_t;\theta )+{ϵ}_{1t}\}$$

Relationship with the expected value function $$V(s_t;\theta )={\mathbb{E}}_{{ϵ}_t}[V(s_t,{ϵ}_t;\theta )|s_t]$$

Goal

We switched from alternative-specific value function $\overline{V}(s_t;\theta )$ to expected value function $V(s_t;\theta )$

• But the goal is the same

Go from this representation $$\begin{array}{rl}V(s;\theta )& =f(V(s;\theta ))\\ P(\cdot |s;\theta )& =g(V(s;\theta );\theta )\end{array}$$ To this $$\begin{array}{rl}V(s;\theta )& =h(P(\cdot |s;\theta );\theta )\\ P(\cdot |s;\theta )& =g(V(s;\theta );\theta )\end{array}$$ I.e. we want to express the expected value function (EV) in terms of the expected policy function (EP).

**Note **: the $f$, $g$ and $h$ functions are different functions now.

Express EV in terms of EP (1)

First, let’s ged rid of one operator: the max operator $$V(s_t;\theta )=\sum _{a}Pr(a_t=a|s_t;\theta )\ast \left[\begin{array}{c}u(s_t,a;\theta )+{\mathbb{E}}_{{ϵ}_t}[{ϵ}_{at}|a_t=a,s_t]\\ +\beta {\mathbb{E}}_{s_{t+1}}[V(s_{t+1};\theta )|s_t,a_t=a;\theta ]\end{array}\right]$$

• We are just substituting the $max$ with the policy $Pr(a_t=a|s_t;\theta )$

• Important: we got rid of the $max$ operator

• But we are still taking the expectation over

  • Future states $s_{t+1}$
  • Shocks ${ϵ}_t$

Express EV in terms of EP (2)

Now we get rid of another operator: the expectation over $s_{t+1}$ $${\mathbb{E}}_{s_{t+1}}[V(s_{t+1};\theta )|s_t,a_t=a;\theta ]\to \sum _{s_{t+1}}V(s_{t+1};\theta )Pr(s_{t+1}|s_t,a_t=a;\theta )$$ where

• $\sum _{s_{t+1}}$ is the summation over the next states
• $Pr(s_{t+1}|s_t,a_t=a;\theta )$ is the transition probability (conditional on a particular choice)

so that the expected value function becomes $$V(s_t;\theta )=\sum _{a}Pr(a_t=a|s_t;\theta )\ast \left[\begin{array}{c}u(s_t,a;\theta )+{\mathbb{E}}_{{ϵ}_t}[{ϵ}_{at}|a_t=a,s_t]\\ +\beta \sum _{s_{t+1}}V(s_{t+1};\theta )Pr(s_{t+1}|s_t,a_t=a;\theta )\end{array}\right]$$

Express EV in terms of EP (3)

The previous equation, was defined at the state level $s_t$

• system of $k$ equations, 1 for each state (value of $x$)

If we stack them, we can write them as $$V(s;\theta )=\sum _{a}Pr(a|s;\theta ).\ast [u(s,a;\theta )+{\mathbb{E}}_{ϵ}[{ϵ}_a|a,s]+\beta T(a;\theta )V(s;\theta )]$$ where

• $T(a)$: $k\times k$ matrix of transition probabilities from state $s_t$ to $s_{t+1}$, given decision $a$
• $.\ast$ is the dot product operator (or element-wise matrix multiplication)

Express EV in terms of EP (4)

Now we have a system of $k$ equations in $k$ unknowns that we can solve.

Tearing down notation to the bare minimum, we have $$V=\sum _a{P}_a.\ast [u_a+\mathbb{E}[{ϵ}_a]+\beta T_{a}V]$$ which we can rewrite as $$V-\beta (\sum _a{P}_a.\ast T_a)V=\sum _a{P}_a.\ast [u_a+\mathbb{E}[{ϵ}_a]]$$

and finally we can solve for $V$ through the famous Hotz and Miller inversion $$V={[I-\beta \sum _a{P}_a.\ast T_a]}^{-1}\ast (\sum _a{P}_a.\ast [u_a+\mathbb{E}[{ϵ}_a]])$$ Solved? No. We still need to do something about $\mathbb{E}[{ϵ}_a]$.

Express EV in terms of EP (5)

What is $\mathbb{E}[{ϵ}_a]$?

Let’s consider for example the expected value of the shock, conditional on investment $$\begin{array}{rl}\mathbb{E}[{ϵ}_{1t}|a_t=1,\cdot ]& =\mathbb{E}[{ϵ}_t|{\overline{V}}_1(s_t;\theta )+{ϵ}_{1t}>{\overline{V}}_0(s_t;\theta )+{ϵ}_{0t}]\\ & =\mathbb{E}[{ϵ}_{1t}|{\overline{V}}_1(s_t;\theta )-{\overline{V}}_0(s_t;\theta )>{ϵ}_{0t}-{ϵ}_{1t}]\end{array}$$ with logit magic 🧙🪄 is $$\mathbb{E}[{ϵ}_{1t}|a_t=1,s_t]=0.5772-\ln \left(P(s_t;\theta )\right)$$

• where $0.5772$ is Euler’s constant.

We again got rid of another $max$ operator!

Express EV in terms of EP (6)

Now we can substitute it back and we have an equation which is just a function of primitives $$\begin{array}{rl}V(\cdot ;\theta )=& [I-(1-P(\cdot ;\theta ))\beta T(0;\theta )-P(\cdot ;\theta )\beta T(1;\theta ){]}^{-1}\\ \ast & \left[\begin{array}{c}(1-P(\cdot ;\theta ))[u(\cdot ,0;\theta )+0.5772-\ln (1-P(\cdot ;\theta ))]\\ +P(\cdot ;\theta )[u(\cdot ,1;\theta )+0.5772-\ln (P(\cdot ;\theta ))]\end{array}\right]\end{array}$$

Or more compactly $$V={[I-\beta \sum _a{P}_a.\ast T_a]}^{-1}\ast (\sum _a{P}_a.\ast [u_a+0.5772-\ln (P_d)])$$

First Equation

What is the first equation? $$V={[I-\beta \sum _a{P}_a.\ast T_a]}^{-1}\ast (\sum _a{P}_a.\ast [u_a+0.5772-\ln (P_a)])$$ Expected static payoff: $\sum _a{P}_a.\ast [u_a+0.5772+\ln (P_a)]$

• Is the expected static payoff of choice $a$ in each state, $u_a+0.5772+\ln (P_a)$
• … integrated over the choice probabilities, $P_a$
• It’s a $k\times 1$ vector

Unconditional transition probabilities: $\sum _a{P}_a.\ast T_a$

• Are the transition probabilities conditional on a choice $a$ for every present and future state, $T_a$
• … integrated over the choice probabilities, $P_a$
• It’s a $k\times k$ matrix

Recap

We got our first equation $$V={[I-\beta \sum _a{P}_a.\ast T_a]}^{-1}\ast (\sum _a{P}_a.\ast [u_a+0.5772-\ln (P_a)])$$

I.e. $$\begin{array}{rl}V(s;\theta )& =h(P(s;\theta );\theta )\end{array}$$

What about the second equation $P(\cdot |s;\theta )=g(V(s;\theta );\theta )$?

From V to P

In general, the expected probability of investment is $$P(a=1;\theta )=Pr\left[\begin{array}{c}u(\cdot ,1;\theta )+{ϵ}_{1t}+\beta \mathbb{E}[V(\cdot ;\theta )|\cdot ,a_t=1;\theta ]>\\ u(\cdot ,0;\theta )+{ϵ}_{0t}+\beta \mathbb{E}[V(\cdot ;\theta )|\cdot ,a_t=0;\theta ]\end{array}\right]$$

With the logit assumption, simplifies to $$P(a=1;\theta )=\frac{\exp (u(\cdot ,1;\theta )+\beta T(1;\theta )V(\cdot ;\theta ))}{\sum _{a^{\prime }}\exp (u(\cdot ,a^{\prime };\theta )+\beta T(a^{\prime };\theta )V(\cdot ;\theta ))}=\frac{\exp (u_1+\beta T_{1}V)}{\sum _{a^{\prime }}\exp (u_{a^{\prime }}+\beta T_{a^{\prime }}V)}$$

Now we have also the second equation! $$\begin{array}{rl}P(s;\theta )& =g(V(s;\theta );\theta )\end{array}$$

Hotz & Miller - Idea 2

Idea 2: Replace $P(\cdot )$ on the RHS with a consistent estimator $\hat{P}(\cdot )$ $$\overline{P}(\cdot ;\theta )=g(h(\hat{P}(\cdot );\theta );\theta )$$

• $\overline{P}(\cdot ;{\theta }_0)$ will converge to the true $P(\cdot ;{\theta }_0)$, because $\hat{P}(\cdot )$ is converging to $P(\cdot ;{\theta }_0)$ asymptotically.

  • Note: pay attention to ${\theta }_0$ vs $\theta$ here: $\overline{P}(\cdot ;\theta )$ does not generally converge to $P(\cdot ;\theta )$for arbitrary $\theta$, because $\hat{P}(\cdot )$ is converging to $P(\cdot ;{\theta }_0)$ but not $P(\cdot ;\theta )$ with any $\theta$.

How to compute $\hat{P}(\cdot )$?

• From the data, you observe states and decisions

• You can compute frequency of decisions given states

  • In Rust: frequency of engine replacement, given a mileage (discretized)

• Assumption: you have enough data

  • What if a state is not realised?
  • Use frequencies in observed states to extrapolate frequencies in unobserved states

Recap

Steps so far

1. Estimate the conditional choice probabilities $\hat{P}$ from the data

   • Nonparametrically: frequency of each decision in each state

2. Solve for the expected value function with the inverstion step $$\hat{V}={[I-\beta \sum _a{\hat{P}}_a.\ast T_a]}^{-1}\ast (\sum _a{\hat{P}}_a.\ast [u_a+0.5772-\ln ({\hat{P}}_a)])$$

3. Compute the predicted CCP, given $V$ $$\overline{P}(a=1;\theta )=\frac{\exp (u_1+\beta T_1\hat{V})}{\sum _{a^{\prime }}\exp (u_{a^{\prime }}+\beta T_{a^{\prime }}\hat{V})}$$

What now? Use the estimated CCP to build an objective function.

Objective Function

We have (at least) 2 options

1. Hotz and Miller (1993) use GMM

$$\mathbb{E}[a_t-\overline{P}(s_t,\theta )|s_t]=0\text{ at }\theta ={\theta }_0$$

1. Aguirregabiria and Mira (2002) use MLE
   • by putting $\overline{P}(s_t,\theta )$ in the likelihood function instead of $P(s_t,\theta )$

We will follow the second approach

Pseudo-Likelihood

The likelihood function for one bus is $$\mathcal{L}(\theta )=\prod _{t=1}^T(\widehat{\mathrm{Pr}}(a=1\mid s_t;\theta )\mathbb{1}(a_t=1)+(1-\widehat{\mathrm{Pr}}(a=0\mid s_t;\theta ))\mathbb{1}(a_t=0))$$ where $\widehat{Pr}(a_t|s_t;\theta )$ is a function of

• CCPs $\hat{P}$: estimated from data
• transition matrix $T$: estimated from the data, given $\theta$
• static payoffs $u$: known, given $\theta$
• discount factor $\beta$ : assumed

Why pseudo-likelihood? We have inputed something that is not a primitive but a consistent estimate of an equilibrium object, $\hat{P}$

Comments

Now a few comments on Hotz and Miller (1993)

• Computational bottleneck
• Aguirregabiria and Mira (2002)
• Importance of the T1EV assumption
• Data requirements
• Unobserved heterogeneity
• Identification

Computational Bottleneck

There is still 1 computational bottleneck in HM: the inversion step $$V={[I-\beta \sum _a{P}_a.\ast T_a]}^{-1}\ast (\sum _a{P}_a.\ast [u_a+0.5772-\ln (P_a)])$$ The $[I-\beta \sum _a{P}_a.\ast T_a]$ matrix has dimension $k\times k$

• With large state space, hard to invert
• Even with modern computational power
• Hotz et al. (1994): forward simulation of the value function
  • You have the policy, the transitions and the utilities
  • Just compute discounted flow of payoffs
  • Core idea behind the estimation of dynamic games

Aguirregabiria, Mira (2002)

Hotz and Miller (1993) inversion gets us a recursive equation in probability space

• instead of the Bellman Equation in the value space

$$\overline{P}(\cdot ;\theta )=g(h(\hat{P}(\cdot );\theta );\theta )$$

Idea

• Do you gain something by iterating $K$ imes?
  • $K=1$: Hotz and Miller (1993)
  • $K\to \mathrm{\infty }$: Rust (1987)
• Monte Carlo simulations: finite sample properties of K−stage estimators improve monotonically with K
  • But especially for $K=2$!
  • Really worth iterating once

Type 1 EV errors

Crucial assumption

• Without logit errors, we need to simulate their distribution
• True also for Rust
• But it’s generally accepted
  • doesn’t imply it’s innocuous

Data Requirements

For both Hotz et al. (1994) and Rust (1987), we need to discretize the state space

• Can be complicated with continuous variables
• Problem also in Rust
• But particularly problematic in Hotz et al. (1994)
  • Relies crucially on consistency of CCP estimates
  • Need sufficient variation in actions for each state

Unobserved Heterogeneity

Hotz et al. (1994) cannot handle unobserved heterogeneity or “unobserved state variables” that are persistent over time.

Example

• Suppose there are 2 bus types $\tau$: high and low quality

• We don’t know the share of types in the data

• With Rust

  • Parametrize the effect of the difference in qualities

    • E.g. high quality engines break less often

  • Parametrize the proportion of high quality buses

  • Solve the value function by type $V(s_t,\tau ;\theta )$

  • Integrate over types when computing choice probabilities $$P(a|s)=\int P(a|s,\tau )P(\tau )=Pr(a|s,\tau =0)\ast Pr(\tau =0)+Pr(a|s,\tau =1)\ast Pr(\tau =1)$$

Unobserved Heterogeneity (2)

What is the problem with Hotz et al. (1994)?

• The unobserved heterogeneity generates persistency in choices

  • I don’t replace today because it’s high quality, but I also probably don’t replace tomorrow either
  • Decisions independent across time only conditional on type

• Likelihood of decisions must be integrated over types $$\mathcal{L}(\theta )=\sum _{{\tau }_a}\prod _{t=1}^{T}Pr(a_{jt}|s_{jt},{\tau }_a)Pr({\tau }_a)$$

• Hotz & Miller needs consistent estimates of $P(a,s,\tau )$

• Difficult when $\tau$ is not observed!

Identification

Work on identification

• Rust (1994) and Magnac and Thesmar (2002)
  • Rust (1987) is non-paramentrically underidentified $\to$ parametric assumptions are essential
• Aguirregabiria and Suzuki (2014)
• Kalouptsidi, Scott, and Souza-Rodrigues (2017)
• Abbring and Daljord (2020)
  • Can identify discount factor with some “instrument” that shifts future utilities but not current payoff
• Kalouptsidi et al. (2020)

Appendix

References [references]