Supplementary Appendix to - Montana State University

3 downloads 8 Views 314KB Size Report
May 19, 2015 - pants at Illinois State University, Montana State University, the University of Colorado, the University of ..... are from the World Bank (2015).

Supplementary Appendix to “Actual and Potential Competition in International Telecommunications”∗ Jason Pearcy†

Scott J. Savage‡

May 19, 2015

Abstract By allowing carriers to route telephone calls over low-cost private lines, international simple resale (ISR) makes it possible for carriers to provide international telephone service without owning an international circuit. When approved, ISR reduces entry barriers and can increase competition. Using data from US markets from 1995 to 2004, we estimate the effects of ISR on entry and retail prices. Results show that ISR has a limited (and imprecisely estimated) impact on entry and actual competition. However, controlling for actual competition, ISR authorization causes an average reduction in prices of 32.7 percent. Markets with relatively high carrier surplus experience an additional reduction in the price by 0.4 percent, and prices are 3.4 percent lower in markets with relatively high private line capacity. Our findings suggest that ISR promotes potential competition and lower prices in markets where the threat of hit-and-run entry is more credible. JEL Classification: C21, D04, L1, L13, L96. Keywords: Contestable Markets, Barriers to Entry, Competition, Policy Evaluation, Treatment Effects.



We thank Mark Anderson, Matt Backus, Pierre Dubois (the editor), two reviewers, seminar participants at Illinois State University, Montana State University, the University of Colorado, the University of Redlands, Telcordia, and the 2014 International Industrial Organization Conference for comments. We acknowledge financial support from the NET Institute Summer Grants Program at http://www.NETinst.org. Andrew Kurtz and John Fuhrman were excellent research assistants. The usual disclaimer applies. † Montana State University, [email protected] ‡ University of Colorado at Boulder, [email protected]

1

A

The Econometric Model

The y2 δ2 and z2 β2 terms in the firm count equation and the price equation are interacted with R, the observed ISR authorization decision by the FCC. When y1 = y2 and z1 = z2 , a switching regression model is estimated (Maddala, 1983; Roy, 1951, Chap. 8).1 An alternative approach is a treatment effects model with y2 = 1 and z2 = 1, where one equation is estimated with pooled data and an ISR indicator (Greene, 1997; Heckman, 1978; Maddala, 1983). The model we use is a hybrid of the switching regression model and a treatment effects model. Our hybrid model has advantages over both a switching regression model and a treatment effects model. An advantage over a switching regression model is that we estimate a single pooled equation in the second stage rather than two separate equations: one for observations with Rct = 1 and one for observations with Rct = 0. The cost of this advantage is additional complexity when estimating the variance of parameters. An advantage over a treatment effects model is that our hybrid model allows for the treatment effects to depend on the model covariates. This allows us to estimate heterogeneous treatment effects we call the marginal average treatment effects (MATE) introduced in Section 3.2. 1

In the second-stage of a switching regression model, two different equations are estimated: one equation for bilateral markets where ISR has been authorized, and one equation for bilateral markets without ISR authorization.

2

B

The Covariance Matrices ΣN and Σp

The covariance matrices are  ΣN

2 σN





2  σp0

σN 01 θN 0      2 = σ σ θ N 1  N 01 N   θ N 0 θN 1 1

 Σp =  σp01  θp0

 σp01 θp0   2 σp1 θp1  .  θp1 1

Note that for equation (3), the variance of up1 may differ from the variance of up0 , but for equation (2) it is assumed that the variance of uN 1 is equal to the variance of uN 0 . Restricting the variance of the firm count equation simplifies the selection bias correction term, ψ(θN 1 , θN 0 , α).2 Non-zero θ terms in the covariance matrices indicate endogeneity between ISR authorization and the unobserved components of the firm count and price equations.

C

Expected Values of Equations (2) and (3)

Both the firm count and price equations are potentially complicated by selection issues mentioned above. In equation (2), uN i is unobservable and due to selection E[uN i |x, y, R] 6= 0. Likewise in equation (3), upi is unobservable and if selection is relevant E[upi |x, z, R] 6= 0. We determine the conditional means of the firm count and price equations, which are then used to motivate a consistent estimation of our empirical model. From equation (2), the expected firm count conditional on observables is

E[N |y, x, R] = exp (y1 δ1 + Ry2 δ2 ) E [exp (uN 0 + R(uN 1 − uN 0 )) |y, x, R] . 2

See the derivation of equation (1) in Section C for details.

3

Rewrite the expectation as

E [exp (uN 0 + R(uN 1 − uN 0 )) |y, x, R] = (1 − R)E [E [exp (uN 0 ) |uR , y, x] |uR < −xα, y, x] + RE [E [exp (uN 1 ) |uR , y, x] |uR ≥ −xα, y, x] .

Let f (uR1 , uR0 , uR ) be the joint pdf of the unobserved terms which is a trivariate normal with means of zero and a covariance matrix of ΣN . One of the properties of a multivariate normal distribution is that the marginal joint distribution of (uRi , uR ) is a bivariate normal with means of zero and a covariance matrix of 



2  σN

θN i  . 1

 θN i As shown by Terza (1998)

 1 2 1 2 E [exp (uN i ) |uR , y, x] = exp θN i uR + σN − θN i , 2 2 

which allows us to rewrite the expectation as 

1 2 σ 2 N



E [exp (uN 0 + R(uN 1 − uN 0 )) |y, x, R] = exp    1 2 (1 − R) exp − θN E [exp (θN 0 uR ) |uR < −xα, y, x] 2 0    1 2 +R exp − θN 1 E [exp (θN 1 uR ) |uR ≥ −xα, y, x] . 2

4

From Terza’s equation (8) and (9) we can reduce the expectation to  E [exp (uN 0 + R(uN 1 − uN 0 )) |y, x, R] = exp

which is equal to exp

1 2 2 σN



1 2 σ 2 N

 1 − Φ (θN 0 + xα) (1 − R) 1 − Φ (xα)  Φ (θN 1 + xα) +R Φ (xα)

ψ(θN 1 , θN 0 , α). Remove the exp

1 2 2 σN



term as this just shifts

the constant term in the estimation and we have derived the intended result. The conditional mean for the firm count equation is  E[N |y, x, R] = exp y1 δ1∗ + Ry2 δ2 ψ(θN 1 , θN 0 , α) Φ (θN 1 + xα) 1 − Φ (θN 0 + xα) ψ(θN 1 , θN 0 , α) = R + (1 − R) . Φ (xα) 1 − Φ (xα)

(1)

Terza’s methodology places an equality restriction on the covariances in equation (1) so that θN 0 = θN 1 . A contribution of our model is that we relax Terza’s restriction and allow the covariances to differ. Note that φ is the PDF from a standard normal distribution and Φ is the CDF. The parameter vector for the firm count equation is (α, δ ∗ , θN 1 , θN 0 ) where α are the parameters from the control function in the first stage, and δ ∗ are the parameters from the second stage firm count equation.3 The remaining parameters, θN i , represent the covariance between R∗ and N . The conditional mean for the price equation is E[pn |z, x, R] = z1 β1 + Rz2 β2 + Ψ(θp1 , θp0 , α) Ψ(θp1 , θp0 , α) = Rθp1

φ (xα) −φ (xα)) + (1 − R)θp0 . Φ (xα) (1 − Φ (xα))

(2)

Ψ(θp1 , θp0 , α) represents the usual selection control term. The parameter vector for the 3 ∗

δ differs from δ in equation (2) as the constant term in δ ∗ is shifted. See Terza (1998).

5

price equation is (α, β, θp1 , θp0 ) where β are the parameters from the second stage price equation and θpi represents the covariance between R∗ and pN . The ψ and Ψ functions in the firm count and price equations are the selection terms that control for any policy endogeneity associated with ISR authorization. For both the firm count and price equations we observe outcomes regardless of ISR authorization. Including these selection terms allows us to obtain consistent estimates of the parameters by accounting for selection into and out of ISR authorization. For the price equation, Ψ is the inverse Mill’s ratios included in standard selection models. For our firm count equation, ψ is a selection term applicable for count models similar to the inverse Mill’s ratios.

D

Estimating the Variance of β and θpi

Vectorize the price equation, equation (5), where p is a M × 1 vector, z1 is a M × k matrix, P PT and z2 is a M ×j matrix. With Nct carriers in country c during year t, M = C c=1 t=1 Nct . From the ISR authorization equation, equation (1), x is an expanded M × l matrix where xi is the ith row of x (1 × l) corresponding to the relevant country/year. The parameter vectors are β1 a k × 1, β2 a j × 1, and α a l × 1. Note that α are the parameter estimates from the ISR authorization equation. Define W1 and W0 as M × 1 vectors with row i as W1i =

φ(xi α) Φ(xi α)

and W0i =

−φ(xi α) 1−Φ(xi α) .

Also define IR as an M × M diagonal matrix with R on the diagonal, and I1−R = I − IR , an M × M diagonal matrix with 1 − R on the diagonal. Equation (5) is rewritten as

p = z1 β1 + IR z2 β2 + IR W1 θp1 + I1−R W0 θp0 + εp .

Let Z be a M × k + j + 2 matrix with Z = [z1 , IR z2 , IR W1 , I1−R W0 ], and let β be a k + j + 2 × 1 parameter vector with β = [β1 , β2 , θp1 , θp0 ]. The price equation is further

6

simplified to p = Zβ + εp . Connecting equation (5) with equation (3)

εp = IR εp1 + I1−R εp0

with εp1i = up1i − W1i θp1 and εp0i = up0i − W0i θp0 . εp is defined such that

E[εp1i |Ri = 1] = E[εp0i |Ri = 0] = 0.

Ri = 1 corresponds to uRi ≥ −xi α and from the moments of a truncated bivariate normal distribution (see (Greene, 1997; Maddala, 1983)) we have

E[up1i |Ri = 1] = W1i θp1 E[up0i |Ri = 0] = W0i θp0 2 2 V ar(up1i |Ri = 1) = σp1 − θp1 W1i (W1i + xi α) 2 2 V ar(up0i |Ri = 0) = σp0 − θp0 W0i (W0i + xi α).

From the E[upi |Ri ] and the V ar(upi |Ri ) above 2 2 E[u2p1i |Ri = 1] = σp1 − θp1 W1i (xi α)

2 2 E[u2p0i |Ri = 0] = σp0 − θp0 W0i (xi α).

The V ar(εp1i |Ri = 1) = V ar(up1i |Ri = 1) and the V ar(εp0i |Ri = 0) = V ar(up0i |Ri = 0) using the relationship between εpi and upi defined above. These conditions are used to 2 and σ 2 which are respectively V ar(u ) and V ar(u ). Similar to estimate both σp1 p1i p0i p0

7

Maddala (1983), our estimates of these variances are

2 σ ˆp1 =

M1   1 X 2 ˆ ˆ 1i + xi α) εˆ2p1i + θˆp1 W1i (W ˆ M1 i=1

2 σ ˆp0

M0   1 X 2 ˆ ˆ 0i + xi α) = εˆ2p0i + θˆp0 W0i (W ˆ M0 i=1

where M1 is the number of observations with Ri = 1, M0 observations have Ri = 0, and 2 and R = 0 for σ 2 . the summations are only over observations with Ri = 1 for σ ˆp1 ˆp0 i

ˆ and the derivation of our estimator is adapted from MadNow we estimate the V ar(β) dala (1983) (see the appendix of Chapter 8). From above, we rewrite the price equaˆ 1i )θp1 , and ˆ + ε˜p where ε˜p = IR ε˜p1 + I1−R ε˜p0 , ε˜p1i = εp1i + (W1i − W tion as p = Zβ ˆ 0i )θp0 . Expanding W ˆ i around α ε˜p0i = εp0i + (W0i − W ˆ = α, yields ˆ 1i ' W1i (W1i + xi α)xi (α W1i − W ˆ − α) ˆ 0i ' W0i (W0i + xi α)xi (α W0i − W ˆ − α).

To vectorize the above, define D1 as a M × M diagonal matrix with the ith diagonal term being W1i (W1i + xi α) and define D0 similarly with the ith diagonal term being W0i (W0i + xi α). We now have ˆ 1 ' D1 x(α W1 − W ˆ − α) ˆ 0 ' D0 x(α W0 − W ˆ − α)

and ε˜ = IR εp1 + IR D1 x(α ˆ − α)θp1 + I1−R εp0 + I1−R D0 x(α ˆ − α)θp0 .

8

The variance of the parameter vector is ˆ = (Z0 Z)−1 Z0 E[˜ V ar(β) εε˜0 ]Z(Z0 Z)−1 .

Some simplifications to point out are that E[IR εp1 εp0 0 I1−R ] = E[I1−R εp0 εp1 0 IR ] = 0 Cov(εp1 , α ˆ 0 ) = Cov(α, ˆ εp1 0 ) = 0 Cov(εp0 , α ˆ 0 ) = Cov(α, ˆ εp0 0 ) = 0.

The variance of the parameter vector simplifies to

2 ˆ = (Z0 Z)−1 Z0 IR V ar(εp1 ) + I1−R V ar(εp0 ) + θp1 V ar(β) IR D1 xV ar(α)x ˆ 0 D1 0 2 +θp0 I1−R D0 xV ar(α)x ˆ 0 D0 0 + θp1 θp0 IR D1 xV ar(α)x ˆ 0 D0 0 I1−R

 +θp1 θp0 I1−R D0 xV ar(α)x ˆ 0 D1 0 IR Z(Z0 Z)−1 .

The V ar(α) ˆ is the estimated variance of the parameter vector from the ISR authorization equation. The V ar(εp1 ) is an M × M diagonal matrix with V ar(εp1i |Ri = 1) as the ith diagonal term. Similarly V ar(εp0 ) is an M × M diagonal matrix with V ar(εp0i |Ri = 0) as the ith diagonal term.

E

Tests for Endogeneity

The potential endogeneity associated with ISR authorization becomes a concern when the residuals from the first stage, uRct , are not independent of the residuals from the second stage. The residuals in the second stage are determined conditional on ISR authorization

9

where in the firm count equation we have uN 1ct and uN 0ct . uN 1ct is the residual corresponding to when ISR is authorized and uN 0ct is the residual corresponding to when ISR is not authorized. In a similar fashion, we have up1nct and up0nct for the price equation. If the residuals uRct and uN 1ct are independent, then E[uRct uN 1ct ] = 0, and E[uRct uN 1ct ] = 0 if and only if θN 1 = 0 where θN 1 is an element of the covariance matrix ΣN . A check for endogeneity involves examining whether or not θN 1 , θN 0 , θp1 and θp0 are different from zero. Model coefficients are presented in Table 5 and using a standard t-test we cannot reject the null hypothesis that θN 1 = 0 and θN 0 = 0. In the price equation, both θp1 and θp0 are highly significant and with a probability greater than 0.99 we are able to distinguish these estimated values from zero. We also calculate the F-statistic for the joint hypothesis: θN 1 = 0 and θN 0 = 0 for the firm count equation. There are 1341 observations with 31 variables and two restrictions. The F-statistic is 5.3421 and the one percent critical value of F [2, 1311] is 4.63 so the joint hypothesis is rejected. In the price equation, there are 18633 observations with 32 variables and two restrictions. The F-statistic for the joint hypothesis that θp1 = 0 and θp0 = 0 is 88.9615. The one percent critical value is also approximately 4.63 so the joint hypothesis is rejected. These results are suggestive of an endogenous relationship between the price and ISR authorization, but the results are mixed for the firm count equation. The F-statistic for the firm count equation is supportive of an endogenous relationship between ISR authorization and firm counts, but the t-statistics are not.

10

F

Derivation of The Treatment Effects

Rewrite equations (2) and (3) as

N1 = exp (y1 δ1 + y2 δ2 + uN 1 )

N0 = exp (y1 δ1 + uN 0 )

pn1 = z1 β1 + z2 β2 + up1

pn0 = z1 β1 + up0

where a subscript of 1 indicates ISR authorization and 0 indicates that ISR is not authorized. From Heckman and Vytlacil (2007), the average treatment effects (ATEs) are defined as

AT EN (y) = E[N1 − N0 |y]

AT Epn (z) = E[pn1 − pn0 |z].

The ATE terms do not take the expectation conditional on R. While E[uN i |y] = 0, we find that E[uN i |y, R] 6= 0. The same is true for upi and both ATEs reduce to  AT EN (y) = exp y1 δ1∗ (exp (y2 δ2 ) − 1) AT Epn (z) = z2 β2

as reported in equations (6) and (7). The marginal average treatment effects (MATEs) show how a marginal change in either y2i or z2i result in a marginal change in the ATE. The number of MATEs is equal to the number of covariates included in y2 and z2 . Each MATE associated with the firm count equation is   ∂AT EN (y) ∗ = δ2i exp y1 δ1∗ exp (y2 δ2 ) + I{y2i ∈y1 } δ1i exp y1 δ1∗ (exp (y2 δ2 ) − 1) ∂y2i

11

∗ is the corresponding coefficient since y also appears in y . Each MATE where I{y2i ∈y1 } δ1i 2i 1

for the price equation is ∂AT Epn (z) = β2i . ∂z2i

G

Robustness Check to MNVO and MNP

Similar to ISR, mobile virtual network operator (MVNO) and mobile number portability (MNP) policies were intended to reduce entry costs and promote competition in the supply of mobile services. An MNVO lowers infrastructure costs by purchasing wholesale access from a network owner at bulk rates and then setting their own retail rates. MNP lowers switching costs by permitting consumers to keep their number when changing mobile providers. MNP was introduced in 27 countries between 1997 and 2004, and MVNO-related policies were introduced in six countries between 1999 and 2001 (Kim & Seol, 2007; Suehler, Dewenter, & Haucap, 2006). Our exclusion restrictions may be invalid if the introduction of MNP and MVNO are correlated with firm counts or IMTS prices and with the foreign country market structure variables we use as instruments. For robustness, we re-estimated the firm count and price equations with the variables MNP (equals one if the foreign country has MNP and zero otherwise) and MVNO (equals one if the foreign country has MVNO and zero otherwise) included as additional controls. Model results in Tables 1 and 2 below are very similar to the baseline results reported in Table 5.

12

Table 1: Model Coefficients for the Firm Count Equation Variable Constant R GDP POP TRADE LAGGED HHI LAGGED MARKUP RATEct COST AREA CAPACITY R × LAGGED MARKUP R × CAPACITY

Baseline 2.4928 (0.1014)*** 0.1054 (0.0794) 0.0438 (0.0144)*** 0.0204 (0.0081)** −0.0278 (0.0314) −0.2382 (0.0886)*** −0.0464 (0.0369) −0.1836 (0.1124) 0.0220 (0.0327) 0.0115 (0.0036)*** 0.0708 (0.1575) 0.0009 (0.0065) −0.0795 (0.0893)

MNVO

MNVO

MNP

2.4919 (0.1024)*** 0.1053 (0.0796) 0.0445 (0.0149)*** 0.0203 (0.0081)** −0.0278 (0.0313) −0.2382 (0.0887)*** −0.0463 (0.0368) −0.1833 (0.1119) 0.0222 (0.0329) 0.0115 (0.0036)*** 0.0705 (0.1569) 0.0008 (0.0064) −0.0788 (0.0881) −0.0091 (0.0246)

2.4925 (0.1025)*** 0.1053 (0.0798) 0.0440 (0.0149)*** 0.0204 (0.0081)** −0.0278 (0.0314) −0.2383 (0.0886)*** −0.0464 (0.0368) −0.1835 (0.1119) 0.0221 (0.0329) 0.0115 (0.0036)*** 0.0707 (0.1569) 0.0009 (0.0064) −0.0792 (0.0880)

MNP θ1 θ0

0.0133 (0.0251) −0.0425 (0.2566)

0.0130 (0.0246) −0.0421 (0.2571)

−0.0036 (0.0240) 0.0132 (0.0247) −0.0424 (0.2572)

Note: Standard deviations are in parenthesis. *, **, and *** indicatesignificance at the 10%, 5%, and 1% levels.

13

Table 2: Model Coefficients for the Price Equation Variable Constant R GDP POP TRADE LAGGED HHI LAGGED MARKUP COST AREA CAPACITY LAGGED SHARE RATEnct R × LAGGED MARKUP R × CAPACITY

Baseline −0.3829 (0.0717)*** −0.3314 (0.0644)*** −0.0018 (0.0188) 0.0160 (0.0070)** −0.2120 (0.0210)*** 0.2630 (0.0626)*** 0.0852 (0.0177)*** 0.0610 (0.0230)*** −0.0171 (0.0035)*** 0.2197 (0.0589)*** 0.0342 (0.0043)*** 0.3792 (0.0121)*** −0.0263 (0.0080)*** −0.1858 (0.0816)**

MNVO

MNVO

MNP

−0.3908 (0.0718)*** −0.3326 (0.0644)*** 0.0045 (0.0191) 0.0157 (0.0070)** −0.2122 (0.0210)*** 0.2636 (0.0626)*** 0.0858 (0.0177)*** 0.0636 (0.0230)*** −0.0173 (0.0035)*** 0.2175 (0.0589)*** 0.0342 (0.0043)*** 0.3795 (0.0121)*** −0.0272 (0.0080)*** −0.1785 (0.0817)** −0.0884 (0.0545)

−0.3934 (0.0718)*** −0.3342 (0.0644)*** 0.0063 (0.0191) 0.0154 (0.0070)** −0.2118 (0.0210)*** 0.2631 (0.0626)*** 0.0860 (0.0177)*** 0.0647 (0.0230)*** −0.0174 (0.0035)*** 0.2160 (0.0589)*** 0.0342 (0.0043)*** 0.3796 (0.0121)*** −0.0274 (0.0080)*** −0.1749 (0.0817)**

MNP θ1 θ0

0.2407 (0.0317)*** 0.3429 (0.0274)***

0.2381 (0.0318)*** 0.3465 (0.0274)***

−0.1153 (0.0532)** 0.2375 (0.0318)*** 0.3476 (0.0274)***

Note: Standard deviations are in parenthesis. *, **, and *** indicatesignificance at the 10%, 5%, and 1% levels.

14

H

Robustness Check to Cellular Service

Our sample period coincides with the dramatic growth in cellular phones in the United States and in many foreign countries. TeleGeography (2014) data show that the share of cellular in international terminated traffic increased from about 15 percent in 2000 to about 29 percent in 2004, which suggests that it is becoming a more viable substitute for traditional fixed-line IMTS. We control for growth in United States cellular markets with year fixed effects, however, it is possible that growth in foreign country cellular markets is correlated with IMTS prices and with the foreign country market structure variables we use as instruments. For robustness, when estimating the relationship between ISR and prices, we add two different cellular market variables to the price equation to eliminate this concern. In Table 3 below, Cell Subscriptions is the number of cellular subscribers (in millions) in each foreign country for each year and Cell Penetration is the number of Cell Subscriptions per 100 people in each foreign country for each year. The cellular data are from the World Bank (2015). Model results, in Table 3 below, are very similar to the baseline results reported in Table 5. As an additional robustness check, we include cell subscription observations with our MNVO and MNP variables and present the results in Table 4 below. The results are very similar to the baseline results suggesting that the baseline specification is robust.

15

Table 3: Model Coefficients for the Price Equation with Cell Market Variables Variable Constant R GDP POP TRADE LAGGED HHI LAGGED MARKUP COST AREA CAPACITY LAGGED SHARE RATEnct R × LAGGED MARKUP R × CAPACITY

Baseline

Cell Subscriptions

−0.3829 (0.0717)*** −0.3314 (0.0644)*** −0.0018 (0.0188) 0.0160 (0.0070)** −0.2120 (0.0210)*** 0.2630 (0.0626)*** 0.0852 (0.0177)*** 0.0610 (0.0230)*** −0.0171 (0.0035)*** 0.2197 (0.0589)*** 0.0342 (0.0043)*** 0.3792 (0.0121)*** −0.0263 (0.0080)*** −0.1858 (0.0816)**

CELL SUBSCRIPTIONS

−0.4191 (0.0728)*** −0.3288 (0.0646)*** 0.0508 (0.0252)** 0.0168 (0.0070)** −0.2129 (0.0210)*** 0.2622 (0.0635)*** 0.0955 (0.0187)*** 0.0709 (0.0232)*** −0.0159 (0.0035)*** 0.2165 (0.0595)*** 0.0346 (0.0044)*** 0.3790 (0.0122)*** −0.0275 (0.0080)*** −0.1826 (0.0818)** −0.0378 (0.0131)***

CELL PENETRATION θ1 θ0

0.2407 (0.0317)*** 0.3429 (0.0274)***

0.2391 (0.0318)*** 0.3488 (0.0275)***

Cell Penetration −0.4448 (0.0722)*** −0.2743 (0.0653)*** 0.0048 (0.0188) 0.0110 (0.0071) −0.1999 (0.0210)*** 0.2562 (0.0628)*** 0.0989 (0.0185)*** 0.0797 (0.0231)*** −0.0177 (0.0035)*** 0.1974 (0.0592)*** 0.0349 (0.0044)*** 0.3797 (0.0122)*** −0.0226 (0.0080)*** −0.2080 (0.0816)**

−0.2880 (0.0471)*** 0.2145 (0.0321)*** 0.3530 (0.0274)***

Note: Standard deviations are in parenthesis. *, **, and *** indicatesignificance at the 10%, 5%, and 1% levels.

16

Table 4: Model Coefficients for the Price Equation Variable Constant R GDP POP TRADE LAGGED HHI LAGGED MARKUP COST AREA CAPACITY LAGGED SHARE RATEnct R × LAGGED MARKUP R × CAPACITY

Baseline

Cell Subscriptions with MNVO and MNP

−0.3829 (0.0717)*** −0.3314 (0.0644)*** −0.0018 (0.0188) 0.0160 (0.0070)** −0.2120 (0.0210)*** 0.2630 (0.0626)*** 0.0852 (0.0177)*** 0.0610 (0.0230)*** −0.0171 (0.0035)*** 0.2197 (0.0589)*** 0.0342 (0.0043)*** 0.3792 (0.0121)*** −0.0263 (0.0080)*** −0.1858 (0.0816)**

CELL SUBSCRIPTIONS MNVO MNP θ1 θ0

0.2407 (0.0317)*** 0.3429 (0.0274)***

−0.4224 (0.0729)*** −0.3349 (0.0646)*** 0.0486 (0.0254)* 0.0159 (0.0070)** −0.2111 (0.0210)*** 0.2596 (0.0635)*** 0.0953 (0.0187)*** 0.0734 (0.0232)*** −0.0163 (0.0035)*** 0.2122 (0.0595)*** 0.0346 (0.0044)*** 0.3789 (0.0122)*** −0.0278 (0.0080)*** −0.1719 (0.0819)** −0.0325 (0.0132)** 0.4362 (0.2370)* −0.5151 (0.2321)** 0.2380 (0.0318)*** 0.3516 (0.0275)***

Note: Standard deviations are in parenthesis. *, **, and *** indicatesignificance at the 10%, 5%, and 1% levels.

17

I

Robustness to Market Size

As a robustness check, we include the share of a countrys minutes of IMTS compared to the minutes of IMTS worldwide. This allows us to see if ISR authorization affects small countries differently from larger countries as measured by their market size. This measure is lagged one year and included in the estimation of the model coefficients as the variable LAGGED MARKET SIZE. To determine the heterogeneous impact that market size has on ISR authorization we also include an interaction with ISR authorization, i.e., R × LAGGED MARKET SIZE. The results from this robustness check are reported in Tables 5 through 7 below. Table 6 indicates that larger markets have more firms, but that there is not a statistically significant heterogeneous impact of ISR authorization due to market size. With regards to the price, Table 7 also indicates the lack of a statistically significant heterogeneous impact of ISR authorization due to market size. While we would expect market size to have a heterogeneous impact on ISR authorization, our empirical evidence does not support this hypothesis.

18

Table 5: Model Coefficients for the ISR Equation Variable Constant GDP POP TRADE LAGGED HHI

Baseline 1.8088 (1.1210) 0.6917 (0.3260)** −0.4981 (0.2673)* −0.1893 (0.1995) −0.8005 (0.6815)

LAGGED MARKET SIZE LAGGED MARKUP RATEct COST AREA CAPACITY FHHI FCOMP

−0.2172 (0.3870) −11.6996 (1.3255)*** −0.0969 (0.3188) −0.0451 (0.0484) 0.2280 (0.5537) −0.7564 (0.4781) 0.5467 (0.1660)***

With Market Size 2.6084 (1.2030)** 0.4362 (0.3278) −0.2552 (0.2814) 0.9214 (0.5706) −0.5903 (0.6872) −0.2333 (0.1062)** −0.2865 (0.4022) −11.8404 (1.3221)*** −0.1473 (0.3231) −0.0717 (0.0511) 0.2116 (0.5625) −1.2574 (0.5550)** 0.5544 (0.1659)***

Note: Standard deviations are in parenthesis. *, **, and *** indicate significance at the 10%, 5%, and 1% levels.

19

Table 6: Model Coefficients for the Firm Count Equation Variable

Baseline

Constant

With Market Size

2.4928 (0.1014)*** 0.1054 (0.0794) 0.0438 (0.0144)*** 0.0204 (0.0081)** −0.0278 (0.0314) −0.2382 (0.0886)***

R GDP POP TRADE LAGGED HHI LAGGED MARKET SIZE

−0.0464 (0.0369) −0.1836 (0.1124) 0.0220 (0.0327) 0.0115 (0.0036)*** 0.0708 (0.1575) 0.0009 (0.0065) −0.0795 (0.0893)

LAGGED MARKUP RATEct COST AREA CAPACITY R × LAGGED MARKUP R × CAPACITY R × LAGGED MARKET SIZE θ1

0.0133 (0.0251) −0.0425 (0.2566)

θ0

Note: Standard deviations are in parenthesis. *, **, and *** indicate significance at the 10%, 5%, and 1% levels.

20

2.4766 (0.1085)*** 0.1133 (0.1372) 0.0613 (0.0131)*** 0.0099 (0.0114) −0.1085 (0.0334)*** −0.2461 (0.1225)** 0.0281 (0.0143)** −0.0418 (0.0397) −0.1796 (0.1673) 0.0268 (0.0394) 0.0125 (0.0032)*** 0.0678 (0.2732) 0.0010 (0.0097) −0.0803 (0.1263) −0.0118 (0.0164) 0.0103 (0.0381) −0.0266 (0.4917)

Table 7: Model Coefficients for the Price Equation Variable Constant R GDP POP TRADE LAGGED HHI

Baseline −0.3829 (0.0717)*** −0.3314 (0.0644)*** −0.0018 (0.0188) 0.0160 (0.0070)** −0.2120 (0.0210)*** 0.2630 (0.0626)***

LAGGED MARKET SIZE LAGGED MARKUP COST AREA CAPACITY LAGGED SHARE RATEnct R × LAGGED MARKUP R × CAPACITY

0.0852 (0.0177)*** 0.0610 (0.0230)*** −0.0171 (0.0035)*** 0.2197 (0.0589)*** 0.0342 (0.0043)*** 0.3792 (0.0121)*** −0.0263 (0.0080)*** −0.1858 (0.0816)**

R × LAGGED MARKET SIZE θ1 θ0

0.2407 (0.0317)*** 0.3429 (0.0274)***

With Market Size −0.3707 (0.0717)*** −0.3235 (0.0659)*** −0.0073 (0.0206) 0.0166 (0.0078)** −0.1874 (0.0460)*** 0.2599 (0.0627)*** −0.0012 (0.0085) 0.0850 (0.0177)*** 0.0578 (0.0230)** −0.0168 (0.0036)*** 0.2121 (0.0589)*** 0.0342 (0.0043)*** 0.3791 (0.0121)*** −0.0265 (0.0080)*** −0.1840 (0.0823)** −0.0066 (0.0072) 0.2345 (0.0318)*** 0.3567 (0.0278)***

Note: Standard deviations are in parenthesis. *, **, and *** indicatesignificance at the 10%, 5%, and 1% levels.

21

J

Quintiles of the Firm Count Equation

The quintiles are determined from the average number of firms in each country over time, where the first quintile includes the 20 percent of countries that on average have the fewest number of firms. Because the regression for each quintile contains multiple observations for a smaller number of countries we had to remove the country level covariates of GDP and TRADE from the quintile regressions. For the same reason, we also had to remove the COST variable, which is reported at the region level. The coefficients on R are of primary interest. The coefficient for the second quintile is the only one statistically significant at the ten percent level, and the coefficients for the first through third quintile have more significance than the last two. Ignoring statistical significance, it appears that ISR authorization has a non-monotonic affect on firm counts. Markets with low and high firm counts have even larger firm counts when ISR is authorized, but markets with moderate levels of firm counts experience a decrease in firm counts when ISR is authorized. An alternative interpretation that factors into the statistical significance of the results is that when ISR is authorized, firm counts increase in markets with lower firm counts and this effect dissipates for moderate to high levels of firm counts.

22

Table 8: Model Coefficients for the Firm Count Equation Variable Constant R POP LAGGED HHI LAGGED MARKUP RATEct 23

AREA CAPACITY θ1 θ0

1st Quintile 2.1075 (0.0961)*** 0.0459 (0.0769) 0.4264 (0.2273)* −0.0663 (0.0754) −0.0672 (0.0400)* −0.0052 (0.0940) −0.0023 (0.0155) 0.0084 (0.0715) 0.0657 (0.1075) −0.0622 (0.0609)

2nd Quintile 2.2422 (0.1007)*** 0.0837 (0.0495)* −0.0990 (0.1011) −0.0946 (0.0701) 0.0118 (0.0321) 0.0167 (0.0889) 0.0003 (0.0197) −0.0687 (0.0678) −0.0607 (0.0506) −0.0190 (0.1831)

3rd Quintile 2.4181 (0.0730)*** −0.0346 (0.0379) 0.0681 (0.0272)** −0.1157 (0.0594)* 0.0198 (0.0363) −0.1240 (0.0847) 0.0049 (0.0126) 0.0199 (0.0515) −0.0022 (0.0211) 0.0754 (0.0748)

4th Quintile

5th Quintile

2.6423 (0.0934)*** 0.0156 (0.1414) 0.0134 (0.0071)* 0.0637 (0.2840) −0.0269 (0.0782) −0.2233 (0.2260) 0.0048 (0.0029) 0.0126 (0.2463) −0.0137 (0.0281) −0.0231 (0.3794)

Note: Standard deviations are in parenthesis. *, **, and *** indicate significance at the 10%, 5%, and 1% levels. Quintiles are based on country average firm counts where the first quintile contains the 20% of counties with the lowest average number of firms.

2.7340 (0.1364)*** 0.0337 (0.1940) 0.0063 (0.0111) 0.1171 (0.4675) −0.0819 (0.1268) −0.3434 (0.3566) 0.0113 (0.0047)** 0.0053 (0.4142) 0.0239 (0.0467) −0.0353 (0.5650)

Table 9: Model Coefficients for the Firm Count Equation with ISR Interactions Variable Constant R POP LAGGED HHI LAGGED MARKUP 24

RATEct AREA CAPACITY R × LAGGED MARKUP R × CAPACITY θ1 θ0

1st Quintile 2.1150 (0.0973)*** 0.0756 (0.0859) 0.3797 (0.2276)* −0.0810 (0.0758) −0.0595 (0.0404) −0.0095 (0.0950) −0.0013 (0.0155) 0.0169 (0.0720) −0.0181 (0.0096)* −0.0742 (0.1604) 0.1357 (0.1259) −0.0623 (0.0612)

2nd Quintile 2.2412 (0.1003)*** 0.0663 (0.0643) −0.0989 (0.1012) −0.0960 (0.0712) 0.0116 (0.0324) 0.0187 (0.0905) 0.0002 (0.0197) −0.0672 (0.0709) 0.0070 (0.0094) −0.0011 (0.1045) −0.0576 (0.0480) −0.0190 (0.1806)

3rd Quintile 2.4386 (0.0744)*** −0.0744 (0.0617) 0.0644 (0.0276)** −0.1177 (0.0594)** −0.0059 (0.0429) −0.1272 (0.0849) 0.0051 (0.0124) 0.0175 (0.0604) 0.0106 (0.0061)* 0.0096 (0.0827) 0.0010 (0.0211) 0.0803 (0.0740)

4th Quintile 2.6209 (0.0852)*** 0.0785 (0.0683) 0.0131 (0.0077)* 0.0669 (0.2862) −0.0233 (0.1062) −0.2190 (0.2171) 0.0047 (0.0029) 0.0617 (0.3437) −0.0004 (0.0112) −0.0901 (0.2010) −0.0230 (0.0164) −0.0218 (0.3826)

Note: Standard deviations are in parenthesis. *, **, and *** indicate significance at the 10%, 5%, and 1% levels. Quintiles are based on country average firm counts where the first quintile contains the 20% of counties with the lowest average number of firms.

5th Quintile 2.6838 (0.1456)*** 0.2067 (0.0804)** 0.0036 (0.0156) 0.0681 (0.4225) −0.0802 (0.1851) −0.3061 (0.3187) 0.0122 (0.0056)** 0.1536 (0.6018) 0.0010 (0.0164) −0.2605 (0.3640) −0.0059 (0.0215) −0.0315 (0.5732)

References Greene, W. H. (1997). Econometric analysis (3rd ed.). Prentice Hall. Heckman, J. J. (1978). Dummy endogenous variables in a simultaneous equation system. Econometrica, 46 (4), 931–959. Heckman, J. J., & Vytlacil, E. J. (2007). Econometric evaluation of social programs, part I: Causal models, structural models and econometric policy evaluation. In J. J. Heckman & E. E. Leamer (Eds.), Handbook of econometrics (Vols. 6, Part B, pp. 4779 – 4874). Elsevier. Kim, B., & Seol, S. (2007). Economic analysis of the introduction of the mvno system and its major implications for optimal policy decisions in Korea. Telecommunications Policy, 31 , 290–304. Maddala, G. S.

(1983).

Limited-dependent and qualitative variables in econometrics

(Vol. 3). New York, NY: Cambridge University Press. Roy, A. D. (1951). Some thoughts on the distribution of earnings. Oxford Economic Papers, 3 (2), 135–146. Suehler, S., Dewenter, R., & Haucap, J. (2006). Mobile number portability in europe. Telecommunications Policy, 30 , 385–399. TeleGeography. (2014). Telegeography report. Washington, DC: PriMetrica Inc. Terza, J. V. (1998). Estimating count data models with endogenous switching: Sample selection and endogenous treatment effects. Journal of Econometrics, 84 (1), 129 – 154. World Bank. (2015). Mobile cellular subscriptions. Washington, DC: World Bank Open Data.

25