Identification in Models with Discrete Variables

Abstract

This paper provides a novel, simple, and computationally tractable method for determining an identified set that can account for a broad set of economic models when the economic variables are discrete. Using this method, we show using a simple example how imperfect instruments affect the size of the identified set when the assumption of strict exogeneity is relaxed. This knowledge can be of great value, as it is interesting to know the extent to which the exogeneity assumption drives results, given it is often a matter of some controversy. Moreover, the flexibility obtained from our newly proposed method suggests that the determination of the identified set need no longer be application specific, with the analysis presenting a unifying framework that algorithmically approaches the question of identification.

Appendices

A Proofs

1.1 A.1 Proof of Lemma 1

Proof

We need to show that there exists satisfying:

if and only if there exists satisfying:

$$\begin{aligned}&\sum _{i=1}^{n} \sum _{j=1}^{m} \pi _2(y_i,u^*_j)1(y_i \in G(u^*_j))du = 1, \end{aligned}$$

(D1)

$$\begin{aligned} \forall i = 1,...,n:&\sum _{j=1}^{m} \pi _2(y_i,u^*_j) = p_i, \end{aligned}$$

(D2)

$$\begin{aligned} \forall j = 1,...,m:&\sum _{i=1}^{n}\pi _2(y_i,u^*_j) = \nu ^*(u^*_j), \end{aligned}$$

(D3)

$$\begin{aligned} \forall I \in \mathbb {I};\forall j = 1,...,m:&\left| \sum _{i \in I}\pi _2(y_i,u^*_j) - \sum _{i \in I} p_i \nu ^*(u^*_j) \right| \le \alpha \sum _{i \in I} p_i \nu ^*(u^*_j). \end{aligned}$$

(D4)

$$\begin{aligned} \forall i = 1,...,n;\forall j = 1,...,m&: \pi _2(y_i,u^*_j) \ge 0 \end{aligned}$$

(D5)

“$(\Rightarrow )$” - Given $\pi _1$, we construct $\pi _2$ according to:

and this will ensure that {(C1),(C2),(C3M),(C4M),(C5)} implies {(D1),(D2),(D3M),(D4M),(D5)} as shown below:

“$(\Leftarrow )$” - If we know $\pi _2$, we obtain $\pi _1$ using:

(note that ($\Pi _1$) implies ($\Pi _2$)) and we now show that {(D1), (D2), (D3M), (D4M), (D5)} implies {(C1), (C2), (C3M), (C4M), (C5)}:

1.2 A.2 Proof of Lemma 2

Proof

Similarly to the proof of Lemma 1, we need to show that there exists satisfying (C1), (C2), (C5) and:

if and only if there exists satisfying (D1), (D2), (D5) and:

“$(\Rightarrow )$” - Given $\pi _1$, we construct $\pi _2$ according to:

and this will ensure that {(C1), (C2), (C3M), (C4M), (C5)} imply {(D1), (D2), (D3M), (D4M), (D5)}. Because the partitioning of the space using (PartU2) is finer than that using (PartU1), we find that {(C1), (C2), (C5)}, implying {(D1), (D2), (D5)} immediately using the proof of Lemma 1. It is therefore sufficient to show that {(C3M), (C4M)} imply {(D3M), (D4M)}:

“$(\Leftarrow )$” - Knowing $\pi _2$, we obtain $\pi _1$ using:

where $\gamma$ is an arbitrary strictly positive probability density function. It is now sufficient to show that {(D3M),(D4M) (D5)} imply {(C3M),(C4M),(C5)}, because the proof of Lemma 1 reveals that {(C1),(C2)} imply {(D1),(D2)} and (PartU2) provides a finer discretization of than does (PartU1):

B Technical Details on the Presented Examples

1.1 B.1 Example 1

1.1.1 B.1.1 Chesher’s Approach

In order to present the identification result from Chesher (2009), we first introduce the basic definitions. The notation used differs from that in GH that is employed in the present study.

• A model is defined as (10) with $U \sim Unif(0,1)$ and for all .

• A structure$S \equiv \{t, F_{UX|Z}\}$ is a pair of a threshold-crossing function t and a cumulative distribution function of the conditional distribution of U and X given Z.

• A structure S is said to be admitted by a model if $F_{UX|Z}$ respects the independence property, that is $F_U(u|z)\equiv F_{UX|Z}(u,\bar{x}|z)=u$ for all $u \in (0,1)$ and all , where $\bar{x}$ is the upper bound of X.

• A structure Sgenerates the joint distribution of Y and X given Z if $F_{YX|Z}(0,x|z)=F_{UX|Z}(t(x),x|z)$.

• Two structures $S^* \equiv \{t^*, F^*_{UX|Z}\}$ and $S^0 \equiv \{t^0, F^0_{UX|Z}\}$ are said to be observationally equivalent if they generate the same distribution of Y and X given Z for all , that is if $F^*_{YX|Z}(0,x|z) \equiv F^*_{UX|Z}(t^*(x),x|z) = F^0_{YX|Z}(0,x|z) \equiv F^0_{UX|Z}(t^0(x),x|z)$ for all and for all .

Theorem 1 from Chesher (2009) states that having a structure $S_0$ admitted by the model that generates the conditional distribution of Y and X given Z with cumulative distribution function $F^0_{YX|Z}$ and if this threshold-crossing function t is in structure S admitted by model that is observationally equivalent to $S^0$, then t satisfies:

(19)

(20)

where $Pr_0$ states that probabilities were calculated using the measure that was generated by $S^0$, that is using $F^0_{YX|Z}$ and l and u stand for the lower and upper bound, respectively.

Given the continuity of X, the converse is also true. This is equivalent to saying that the set of all functions p satisfying the above set of inequalities is a sharply defined identified set. In Chesher (2010), this theorem is proven, even for a more general setup. It is important to note that the proof is constructive, so that for a given threshold-crossing function t, a suitable distribution function $F_{UX|Z}$ is constructed such that $\{t,F_{UX|Z}\}$ is admitted by the model and generates the $F_{YX|Z}$ observed in the data. This highlights the link to the GH setup, as the aim there is to find the joint probability distribution that satisfies the independence restriction, has correct marginals, and places all the probability on those combinations of variables that are compatible with the data.

1.1.2 B.1.2 Illustration: Discrete Endogenous Variable

1.1.3 Construction of True Data-generating Process

The following example is taken from Chesher (2010). Suppose that both Y and X are binary; $Y \equiv 1(Y^* \ge 0)$ and $X \equiv 1(X^* \ge 0)$, where $Y^*$ and $X^*$ were generated in the following way:

(21)

with parameters:

$$\begin{aligned} (\theta _0,\theta _1,b_0,b_1,r) = (0,0.5,0,1,-0.25). \end{aligned}$$

(22)

and the instrument Z takes values in .

However, the econometrician does not know how the data were generated. She only assumes (10) and , $U \sim Unif(0,1)$, $t(X) = \Phi (-\theta _0 - \theta _1 X)$, and observes the distribution of the observable variables $p_{ijk}$.²² Even though it is impossible to recover the true value of $\theta = (0,0.5)$ exactly, it is possible to at least create informative bounds for it.

As the X threshold-crossing function t attains only two values, $t(0)=\Phi (-\theta _0) = 0.5$ and $t(1)=\Phi (-\theta _0-\theta _1)=0.308$.

1.1.4 B.1.3 Illustration: Continuous Endogenous Variable

1.1.5 Construction of the True Data-Generating Process

Suppose that the economic model is described by (10) and the data-generating process by (21) with the parameters:

$$\begin{aligned} (\theta _0,\theta _1,b_0,b_1,s_{wv},s_{vv}) = (0,-1,0,0.3,0.5,1). \end{aligned}$$

(23)

as before, the only difference being that X is no longer binary ($X=X^*)$.

The distribution of the observable variables $(Y^*,X|Z=z)$ ($Y^*$ and X given $Z=z$) is given by $N(\mu (z),\Sigma )$, where:

$$\begin{aligned} \mu (z) = \left[ \begin{array}{cc} \theta _0 + \theta _1 b_0 + \theta _1 b_1 z\\ b_0 + b_1 z \end{array}\right] \ \ \ \Sigma = \left[ \begin{array}{cc} 1+2\theta _1s_{wv}+\theta _1^2s_{vv} &{} s_{wv}+\theta _1 s_{vv}\\ s+{wv} + \theta _1 s_{vv} &{} s_{vv}. \end{array}\right] \end{aligned}$$

We provide details of the simulations here. Because of the continuity of X, the unobservable U was discretized as the equidistant point masses on [0, 1]. The distribution of observables is given by:

$$\begin{aligned} p_{ijk} = Pr(Y = y_i \cap X = x_j \cap Z=z_k) = Pr(Y = y_i \cap X =x_j | Z=z_k) Pr(Z=z_k). \end{aligned}$$

It is known that $(Y^*,X|Z) \sim N(\mu (z),\Sigma )$ and a suitable discretization of X is needed. It is easy to show that the density of $(Y^*|X=x,Z=z)$ is:

$$\begin{aligned} N\left( \mu (z)_{1}+ \frac{\Sigma _{21}}{\Sigma _{22}}(x-\mu (z)_{2}),\left( 1-\sqrt{\frac{\Sigma _{21}^2}{\Sigma _{11}\Sigma _{22}}}\right) \Sigma _{11}\right) . \end{aligned}$$

Integrating the corresponding probability density function at ($-\infty$,0) gives us $Pr(Y=0|X=x,Z=z)$. The distribution of X given $Z=z$ is $N(b_0+b_1 z,s_{vv})$, but now the question is how to discretize the support of X, which is $\mathbb {R}$. If the number of nodes is $n_x$, then one suggestion would be to set the z to its mean value, that is 0, and set the values of the discretized support of X to $n_x$ equidistant quantiles.²³ Even though this discretization appears natural, it brings some degree of arbitrariness to the problem.

Finally, taking all the pieces together yields:

$$\begin{aligned} p_{ijk} = Pr(Y = y_i| X = x_j,Z=z_k)Pr(X = x_j|Z=z_k)Pr(Z=z_k), \end{aligned}$$

where all quantities on the right-hand side are known.

1.2 B.2 Example 2

1.2.1 B.2.1 Illustration

1.2.2 True Data-Generating Process

For the illustration, $(\epsilon _1,\epsilon _2)$ are assumed to be $N(0,I_2)$. This assumption, together with (13) and (14), generates the distribution of Y and D given X and Z. The support of Z is assumed to be $\{-1,1\}$ and the support of X is either $\{0\}$ or $\{-2,-1,0,1,2\}$. (X, Z) are assumed to be uniformly distributed.²⁴

1.3 B.3 Example 3

1.3.1 B.3.1 Balke and Pearl’s Approach

Balke and Pearl (1997) made use of the fact that these restrictions impose the following decomposition on the joint distribution of (Y, D, Z, U):

$$\begin{aligned} Pr(Y,D,Z,U) = Pr(Y|D,U)Pr(D|Z,U)Pr(Z)Pr(U). \end{aligned}$$

(24)

There exist four different functions from Z to D and four different functions from D to Y, hence 16 different types of individuals that we can consider. Hence, one can think of U as having a discrete support with 16 points, each point representing a pair of functions, one from Z to D and the second from D to Y. For instance, one type u may be persons who always accept treatment and who do not display a positive outcome irrespective of treatment. The bounds on (15) are found using linear program searching through the space of distributions of the types (U) subject to the joint distribution to be compatible with observed data Pr(y, d|z). The full setup, together with discussion, is in Balke and Pearl (1997, 1994).

1.4 B.4 Example 4

1.4.1 B.4.1 Komarova’s Approach

Following Manski and Thompson (1986):

$$\begin{aligned} Pr(Y=1|X =x) = 1 - Pr(U <-x\beta |X=x), \end{aligned}$$

together with the zero-median restriction (17), implies:

$$\begin{aligned} Pr(Y=1|X=x) \ge 0.5 \ \Leftrightarrow \ x\beta \ge 0. \end{aligned}$$

(25)

Therefore, the bounds on the parameter vector $\beta$ are obtained as an intersection of linear half spaces. In Komarova (2013), a recursive procedure is proposed that translates this set of linear inequalities into bounds on the parameters.

C Implementation Issues

1.1 C.1 Extended GH Framework

Following routines were used and compared in order to solve linear program (12).

• linprog²⁵—Matlab built in function from Optimization Toolbox. Interior point method was superior to simplex method because of the computational time. Since the objective value is not minimized to exact zeros, certain threshold had to be employed. Natural choice is the tolerance level of the optimization routine ($10^{-8}$ for $n_x = n_u = 40$ was used). Results for the two approaches were identical.

• GNU Linear Programming Kit (GLPK)—Modified simplex method from Matlab MEX interface for the GLPK library.²⁶ Significantly faster than linprog with similar results.

Linear program is an old and well understood problem however if the discretization of X and U is large then the matrix that encodes the restrictions for the joint distribution can reach the limits of Matlab’s largest array that can be created.²⁷ For instance if the sizes of supports are $n_x = n_u = 50$ together with $n_y=2$ and $n_z=30$, then the joint probability $\pi _{ijkl}$ has 150, 000 elements. So that the matrix that carries the information about restrictions on $\pi _{ijkl}$ will have 150, 000 columns. This is not a large problem, but the complexity grows exponentially with additional covariates. The most time consuming part is creating the matrices of equalities and inequalities that define the linear program. The computational burden is further increased when resampling methods, such as subsampling, are used for statistical inference. All these reasons allow us only to consider small problems, those with one or two parameters of interest.

1.2 C.2 Original GH Approach

Since the formulation is also a linear program both linprog and GLPK can be used. In addition to these two functions optimal transportation structure can be exploited and more efficient algorithms can be applied. Galichon and Henry (2011) refer to the link between optimal transportation and maximum flow problem studied in Papadimitriou and Steiglitz (1998). If the capacities $\{\nu _j\},\{p_i\}$ of the arcs are the probabilities of the observables and unobservables respectively and the arcs with infinite capacity are between the values of the observables and unobservables that satisfy the economic model ($y_i \in G_\theta (u_j)$ or $c_{ij}=0$), then the question of feasibility of the economic model reduces to checking whether the maximum flow is equal to one. The maximum flow formulation is illustrated on Fig. 15.

Fig. 15

Reproduced from the Figures 7–8 from Papadimitriou and Steiglitz (1998) p. 145

These two algorithms are readily available

• maxflow of M. Rubinstein implements Boykov and Kolmogorov max-flow/min-cut algorithm (Boykov and Kolmogorov 2001).²⁸

• max_flow from Matlab Boost Graph Library written by D. Gleich implements push-relabel maximum flow algorithm of Goldberg and Tarjan (1986).²⁹

These combinatorial algorithms are computationaly very effective, indeed, the most challenging task is to efficiently implement the creation of the matrix that defines the graph. Since this matrix is smaller that the one from linear program, finer discretization can be employed if necessary. Even for very fine discretization maximum flow was calculated in a fraction of a second.