broadband internet, labor demand, and total factor ... · productivity evolves according to an...
TRANSCRIPT
Policy Research Working Paper 8318
Broadband Internet, Labor Demand, and Total Factor Productivity in Colombia
Carlos Ospino
Finance, Competitiveness and Innovation Global Practice GroupJanuary 2018
WPS8318P
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
edP
ublic
Dis
clos
ure
Aut
horiz
ed
Produced by the Research Support Team
Abstract
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.
Policy Research Working Paper 8318
This paper is a product of the Finance, Competitiveness and Innovation Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at [email protected].
This paper studies the relationship between information and communication technology, labor demand, and total factor productivity in Colombia. It estimates total factor productivity for the Colombian manufacturing sector using a method that assumes a law of motion where total factor productivity evolves according to an autoregressive pro-cess of order 1 as well as with the past use of broadband technologies. Using fixed effects models, the paper esti-mates the effect of broadband use on the labor demand of different workers, controlling for total factor productivity,
capital, and wages. To address the potential endogene-ity between broadband adoption and labor demand, the analysis uses state-industry level variation in broadband quality (speed) and intensity of use. The results show a pos-itive association of broadband adoption on labor demand, suggesting that adoption of information and communi-cation technology can offset the employment effects of technological growth. Attempts to identify causal effects using an instrumental variable approach were inconclusive.
BroadbandInternet,LaborDemand,
andTotalFactorProductivityinColombia
Carlos Ospino1
JEL codes: D24, L25, L60, O15
Keywords: Total factor productivity, broadband internet, labor demand, production function
estimation.
1This work has benefitted from funding by the World Bank’s Latin America and Caribbean’s Chief Economist Office, under the regional study “Digital technology adoption, skills, productivity and jobs in Latin America”. I appreciate the very helpful comments and suggestions from Joana Silva, Mark Dutz, Rita Almeida, Truman Packard, Giulia Lotti and the participants at the author’s workshop at the World Bank. Daniel Espinosa provided much appreciated research assistance to this project. Please send comments and suggestions to [email protected]
2
IntroductionThis section of the paper shows the international context of internet use in Colombia.
Figure 1 OECD Fixed broadband subscriptions per 100 inhabitants, by technology, June 2015
Source: OECD, Broadband Portal, http://www.oecd.org/sti/broadband/oecdbroadbandportal.htm, June 2015.
Figure 1 shows that as of June 2015, Colombia had a total number of internet subscriptions per 100
inhabitants similar to Mexico (11.2), but lower than Chile (14.6). The latter’s is half of the OECD
average of 28.8 subscriptions per 100 inhabitants. DSL and cable are the dominant technologies in
Colombia and both have the same number of subscriptions (5.2) per 100 inhabitants, despite an
ambitious plan to expand the optical fiber network to almost every municipality ((Ministerio de
Tecnologías de la Información y las Comunicaciones, 2017).
Figure 2 OECD fixed broadband penetration (per 100 inhabitants), percentage increase, June 2014‐June 2015, by country
Source: OECD, Broadband Portal, http://www.oecd.org/sti/broadband/oecdbroadbandportal.htm, June 2016.
00
05
10
15
20
25
30
35
Chile Mexico OECD Colombia
DSL Cable Fibre Satellite Fixed wireless Other
0.00
2.00
4.00
6.00
8.00
10.00
12.00
Chile Mexico OECD Colombia
3
Colombia showed a significant increase in fixed broadband penetration. The growth level was
almost three times the level of the OECD average and near one and a half times the growth level
displayed by Chile. This shows that at the household level, Colombia may still have room to expand
its broadband penetration. At the business level, the story is quite different. In 2003, 86% of small
and medium size enterprises (SME) and 96% of large ones had internet access (CINTEL) (2003). In
2015, 99% of firms in manufacturing and 99.7% in retail and services used internet (CRC) (2015).
This shows that internet access in the last decade has been relatively high in Colombia’s business
sector.
This high coverage of internet access poses the challenge of identifying the gains in productivity due
to the availability of internet and how it relates to labor demand; I therefore focus on a slightly
different question. In the next section, I explore the literature and conclude that the focus, rather
than being on internet access per se, should be on high‐speed broadband internet access for firms.
The second section discusses the relevant literature that is related to this paper. The third describes
the data used in this analysis. The forth section discusses the results of the ICT module in the Annual
Manufacturing Survey for the years 2008‐2014. The fifth section discusses the TFP estimation
procedure. The sixth section presents the main results of the paper in terms of the effect of TFP
gains due to ICT use on labor demand and wages. The sixth section shows robustness and
falsification exercises. The final section concludes with a discussion of the results.
LiteraturereviewThis paper is related to two different branches of the economics literature. The first branch studies
the impact of broadband internet and firm productivity ((Fabling, Grimes, & Grimes, 2016). The
second branch focuses on the estimation of production functions and total factor productivity. The
first contribution of the paper is adding to the scarce empirical evidence about the labor demand
predictions of task‐based models of technological change for heterogeneous firms in developing
countries, with a special focus on broadband internet. The second one is estimating total factor
productivity within the framework of proxy method models of production function estimation
proposed by De Loecker and Warzynski (2012) and De Loecker (2013). The innovation is modeling a
total factor productivity law of motion that explicitly considers the past adoption of high‐speed
broadband internet as a shifter in growth levels. I now discuss each branch of the literature and put
the paper’s contribution in context within each one.
The paper’s theoretical underpinnings follow Brambilla (2018). Her model predicts that because of
ICT adoption, firms will grow, pay higher wages, become specialized in complex tasks and more skill
intensive conditional on skilled labor and ICT being complements. Brambilla (2018) expands
traditional task‐based models of technical progress and labor markets to allow firm heterogeneity
and wage heterogeneity across firms. In this paper, ICT adoption will be modeled as the adoption of
high‐speed broadband internet. Akerman et al. (2015) provide evidence of the skill complementarity
of broadband internet adoption. Related to this point, there is also a growing literature about the
effect of broadband internet on firm productivity ((Bertschek, Cerquera, & Klein, 2013; Colombo,
4
Croce, & Grilli, 2013; Fabling et al., 2016; Grimes, Ren, & Stevens, 2012; Haller & Lyons, 2015). I now
briefly discuss this literature.
Akerman et al. (2015) use Norwegian data and a quasi‐experiment to show that ICTs are innovations
biased towards skilled workers. To identify the intent‐to‐treat effects, the authors exploit the
staggered roll‐out of broadband internet. To estimate the production function and thus
productivity, they use the Levinsohn and Petrin (2003) methodology. The authors find evidence that
the availability of broadband is associated with a substantial increase in the output elasticity of
skilled labor. They find the opposite effect for unskilled labor.
Fabling et al. (2016) study the effect of ultrafast broadband (UFB) internet on labor demand and
productivity in New Zealand between 2010 and 2012. UFB was defined by the government as
download speeds of up to 100Mbps and upload speeds of up to 50Mbps. The authors use the
proximity to schools as an instrument for broadband availability since the government targeted
schools and hospitals in their expansion plans of UFB to reach 80 percent of the population by 2022.
The technology through which UFB can be obtained and the authors focus on is (optic) fiber. These
authors do not find any statistically significant effects and suggest that most gains from UFB would
arise as a result of selection. Their results suggest that gains from UFB would be related to
complementary investments by firms.
Bertschek et al. (2013) do not find a statistically significant effect of broadband on labor productivity,
but find positive effects on firm innovation in Germany. Their analysis covers the period 2001 to
2003 when there was already a high penetration rate of broadband use by firms and focus on DSL
technology. The authors exploit broadband availability at the postal code level to identify the causal
effects.
Colombo et al. (2013) find negligible effects of adopting broadband applications in small and
medium enterprises in Italy during the period 1998‐2004. The authors base their analysis on services
which require broadband speeds and construct their variable of interest using principal component
analysis. Their results point to the complementarity of broadband with other investments at the
firm, those related to management practices.
The paper also relates to the literature about production functions and total factor productivity
(TFP) estimation ((D. A. Ackerberg, Caves, & Frazer, 2015; De Loecker, 2013; De Loecker &
Warzynski, 2012; Gandhi, Navarro, & Rivers, 2016; Levinsohn & Petrin, 2003; Olley & Pakes, 1996).
The paper contributes to this literature by explicitly modelling an endogenous total factor
productivity process which is governed by past productivity as in D. Ackerberg et al. (2006) and the
previous adoption of broadband through broadband technology, namely DSL, cable or optical Fiber.
A more detailed discussion about these methods is postponed to the section where the
methodology is explained.
DatadescriptionThe main data set is composed of an unbalanced panel of 14,479 plants in the manufacturing sector
in Colombia, spanning over 7 years (2008‐2014) of data from the Annual Manufacturing Survey
5
(EAM from its acronym in Spanish.). EAM is a census of manufacturing firms in Colombia that employ
at least 10 workers or have an output value about 73,000 USD.2 It gathers very detailed information
about employment levels and costs, asset levels and investment, output and intermediate materials
use for production. Starting in the year 2008 it also collects an ICT module where information about
internet use and technology can be found. In the sample, 6,999 (48.3%) of the plants have
information for every year and were used as a balanced sample to check the robustness of the
results. I also use broadband speed data at the state level kindly provided by the Communications
Regulation Commission for the years 2008‐2014. The sample used to estimate total factor
productivity was a subsample of these firms and it is described in detail in that section.
ICTuseintheColombianmanufacturingsectorIn this section I describe the results about ICT use in Colombian firms from the information collected
in EAM special modules for the years 2008 through 2014. The pooled sample summary statistics for
the five‐year period are discussed. In this section the analysis is broken down by firm size and I
report information collected in the ICT module about the use of computers at the firm, internet
access, and the type and speed of internet connections.
Table 1 Distribution of firms by total employment size
Source: EAM‐DANE 2008‐2014. Author’s calculations.
The top panel in Table 1 shows that 72% of firms in the sample hired up to 50 workers and are
considered by most definitions small firms; 18% of firms hired between 50 and 200, 8% between
200 and 1,000 and about 2% of firms hired more than 1,000 workers between 2008 and 2014. It is
clear that most firms in the sample are small units of production, and the reader should take this
into account when interpreting the results. The bottom panel restricts the sample to the initial year,
2008, where ICT information started to be collected. The main takeaway is the stability in the
distribution of firm sizes during the period of analysis.
2 137.2 million COP in the year 2013 and exchange rate of 1,867 COP per USD, which is the average exchange rate for that year.
Size Frequency Percent Cum.
[0,50] 91,861 72.07 72.07
(50,200] 22,776 17.87 89.94
(200,1000] 10,382 8.15 98.09
>1000 2,438 1.91 100
Total 127,457 100
Size Frequency Percent Cum.
[0,50] 6,708 69.21 69.21
(50,200] 1,884 19.44 88.65
(200,1000] 862 8.89 97.54
>1000 238 2.46 100
Total 9,692 100
Full sample period
Year 2008
6
Table 2 ICT and firm size
Source: EAM‐DANE 2008‐2014. Author’s calculations. Each column shows the average value of each variable for each bracket of firm
size.
The first column of Table 2 shows the average number of employees in each firm size category, to
put into context the variables described. The table also shows that the average number of
computers at each firm increases with size as does the share of workers that use computers. Column
four shows that the average use of internet is basically universal for firms of all sizes. The final
column shows that having a website tends to increase with firm size. In smaller firms about 1 in 3
workers uses a computer while in large firms this ratio is about 1 in 2. In the Colombian
manufacturing sector, less than a third of workers uses the internet for their work. This percentage
is higher for large firms (those hiring more than 1,000 workers) but is just about 42% of workers. In
small firms, just under 28% of workers use the internet in their work. This suggests that the impact
of internet use or availability on firms’ productivity would be more likely to be observed in larger
firms where internet exposition is greater for workers.
Table 3 Internet and firm size
Source: EAM‐DANE 2008‐2014. Author’s calculations. Each column shows the average value of each variable for each bracket of firm size.
The leading technology in the Colombian manufacturing sector is cable/fiber with just under 46% of
firms using this technology to access the internet. The second most popular technology is ADSL, and
this is particularly true for smaller firms. 25% for small firms use ADSL while only 5% of firms with
more than 1,000 workers use this technology. The most common contracted speed for internet use
is 1‐2 Mbps with just under 36% of firms. However, two‐thirds of large firms report internet speeds
over 2Mbps.
TFPestimationData Cleaning
SizeTotal
employment
Number of
computers
% of
workers
using
computers
Internet
use
% of
workers
use
internet
for work
Has a
website
[0,50] 14.3 3.1 32.2 96.6% 29.7 46.5%
(50,200] 91.1 16.6 35.7 99.9% 32.6 77.3%
(200,1000] 280.0 98.2 42.3 99.8% 36.3 90.2%
>1000 650.1 573.7 55.2 100.0% 47.0 93.7%
Total 63.8 30.2 34.5 97.7% 31.4 58.3%
Size Modem ISDN ADSL Cable/Fiber Wireless Mobile0‐256
Kbps
254‐1024
Kbps
1025‐2048
Kbps
>= 2049
Kbps
[0,50] 6.3% 0.8% 25.3% 39.0% 19.7% 6.5% 7.5% 26.9% 38.7% 24.9%
(50,200] 2.2% 0.5% 27.0% 51.2% 13.4% 5.6% 2.7% 18.6% 33.4% 45.2%
(200,1000] 1.0% 0.3% 13.1% 72.9% 7.7% 5.0% 1.2% 13.7% 24.7% 60.4%
>1000 0.2% 1.1% 5.2% 79.2% 7.0% 7.4% 1.2% 7.7% 16.0% 75.0%
Total 4.7% 0.7% 23.9% 45.8% 16.9% 6.2% 5.8% 23.4% 35.7% 33.8%
7
As discussed above, the sample for the estimation of the production function parameters from
which TFP is calculated was different from the main sample. The selection criteria were determined
using four items: Value added, prices and input consistency, industry size, and survival. I now discuss
each one in detail.
Value added: firms which reported zero or negative value added were deleted.
Wage bill and intermediate inputs: I required that both the value of the wage bill and the
costs of intermediate materials were strictly lower than the value of output. Firms that did
not fulfill this requirement were deleted.
Industry size: I required that that at least 10 firms were part of an industry at any moment
in time to include any firm in that industry.
Survival: For any firm to be included in the production function estimation sample, it had
to report information during four consecutive years.
Input selection
The production function estimation uses capital, labor and intermediate materials as inputs. All
monetary variables were deflated using the Producer Price Index which varies by industry class (two
digits in ISIC Rev. 3) and the base year is 2013. Capital is defined as the value of fixed capital stock
in each year. The stock of capital is constructed recursively using the change in investment levels
and assuming a depreciation value of 5% for all capital. ∗ 1 0.05 . Capital
includes three types of assets: Machinery, office equipment and transportation equipment. I have
decided to leave out buildings and land to make the analysis comparable across businesses and
sectors.3 Labor includes all personnel, regardless of their skill level or contract type. Finally,
intermediate materials include the value of raw inputs, materials and packaging used in the
production process.
Total factor productivity was estimated using the GMM methods proposed by De Loecker and
Warzynski (2012) and De Loecker (2013). I will refer to this methodology as DLW from now on. A
gross output production function was estimated to address the concerns by Gandhi, Navarro, &
Rivers (2016) that value‐added production functions may generate too much dispersion in the
estimated productivity.4 In this paper a firm’s broadband use at each moment was used to identify
the parameters of the Cobb‐Douglas production function in equation (1), which uses labor, capital
and intermediate materials ( , , , respectively). Where is the unobserved total factor
productivity and represents measurement error. The assumption is that the demand for inputs
differs significantly between firms that decide to use broadband internet and those that do not. All
inputs and gross output are expressed in logarithms.
3 Eslava et al. (2004) construct capital stock series which include equipment, machinery, buildings and structures, while excluding vehicles and land. 4 Gandhi et al. (2013) find that in Colombia, under a value‐added production function results imply that the 95th percentile firm would produce more than six times more output whereas under a gross output production function this firm would only produce twice as much.
8
1
As an innovation, the process that governs the evolution of productivity over time can be
determined by the firm’s use of broadband technologies at time t‐1, . The intuition follows De
Loecker's (2013) argument that if productivity evolves exogenously, as in D. A. Ackerberg et al.
(2015) then investment decisions such as exporting or the adoption of high‐speed internet would
have no effect on technological improvements by the firm. We therefore consider an explicit law of
motion which allows past decisions of adopting broadband internet to directly affect productivity
growth.
DLW propose a two‐stage estimation procedure. In the first stage, gross output is regressed on the
production function inputs and on variables that affect input demand. In our case the firm’s
broadband use is assumed to affect input demands differently for firms that have access to
broadband internet and those that do not. The function includes interactions of period t
broadband use, a third‐degree polynomial of all inputs, year and industry fixed effects.
, , , 2
The procedure assumes that (unobserved) total factor productivity, , is a monotonic function of
(observable) inputs and broadband use, , , , . Therefore, unlike other proxy
methods such as Levinsohn and Petrin (2003) none of the parameters is identified in the first stage.
As mentioned, I assume that productivity evolves as a function of the previous period productivity,
broadband use and a productivity shock, similar to De Loecker and Warzynski (2012).
; . Function is approximated using a third‐degree polynomial on lagged
productivity, lagged exporting status and a constant. The vector ≡ , , is used to
construct total factor productivity as:
, , ,
Finally, in the second stage the production function parameters are identified by imposing the
following moment conditions in a GMM estimation.
000
Capital is assumed to be predetermined at time , and therefore uncorrelated with the productivity
shock. Lagged variable inputs and are used in the moment conditions as current values
are likely to be correlated with the productivity shock (e.g. these are freely variable inputs). A single
production function is assumed and estimated for the whole manufacturing sector. In line with this
assumption, Casas and González (2016) showed that aggregating from sector‐specific productivity
provides similar productivity growth patterns to estimating a single production function.
9
This paper aims to assess the effect of TFP growth which can be explained by broadband use on
labor demand and wages. Since the estimation process imposes a structured autoregressive law of
motion on the evolution of TFP which is affected by past broadband use, this exercise can be
performed directly by regressing the variables of interest on the estimated TFP. De Loecker (2013)
points out that methods which do not consider an endogenous TFP process that depend on past
exporting behavior tend to underestimate the relationship between exporting and productivity.
Under the assumption that there are similarities in the decision‐making process between the
adoption of broadband technology and entering exporting markets, I conjecture that not
considering an endogenous TFP process which depends on broadband use is likely to bias the
relationship between ICT use and productivity. To assess the nature of the bias I will conduct
robustness exercises using the productivity estimator of D. A. Ackerberg et al. (2015) and the De
Loecker (2013) procedure which models the productivity law of motion as a function of past
exporting status.
Table 4 shows the estimated production function coefficients for each the methods used in this
paper. My preferred method’s coefficients for labor and capital fall within the range of values of OLS
and fixed effects estimations. Suggesting that both coefficients are biased upward under an OLS
estimation. On the contrary, the materials coefficient is higher than either OLS or fixed effects. Table
5 shows that the correlation of TFP across the three methods is very high.
Table 4 Production function estimated coefficients
Author’s estimations. Standard errors for DLW obtained via bootstrapping with 100 repetitions.
Table 5 Correlation of TFP estimates
Source: EAM‐DANE. Author’s calculations.
DLW estimation is performed under the assumptions that: 1‐Variable input labor demands are a
monotonic function of productivity. 2‐Input demand responds differently for firms that use
broadband internet and those that do not. 3‐TFP evolves as a function of the previous period
productivity and the previous period use of broadband internet. I now provide evidence that these
assumptions hold.
Method Labor Capital Materials
DLW (Broadband Use) 0.466 *** 0.126 *** 0.529 ***
ACF 0.458 *** 0.119 *** 0.539 ***
DLW (Exporting Status) 0.408 *** 0.112 *** 0.588 ***
OLS 0.492 *** 0.199 *** 0.381 ***
Fixed effects 0.373 *** 0.076 *** 0.418 ***
DLW (Broadband Use) ACF DLW (Exporting Status)
DLW (Broadband Use) 1
ACF 0.8257 1
DLW (Exporting Status) 0.9779 0.7849 1
10
Table 6 Input demand and TFP
Table 6 shows that there is a negative correlation, for all inputs, between TFP and input demand in
OLS estimations. This is true for variable inputs that enter the production function as well as for
energy which was not included but has been used by other authors in Colombia (Eslava,
Haltiwanger, Kugler, & Kugler (2004)). This same table shows that input demand is higher for firms
that have broadband internet and those that have broadband internet technologies. In this case
firms that have contracted broadband speeds through broadband technologies (DSL, cable/fiber,
wireless) showed a lower input demand than the rest of firms in the OLS estimation. Once I account
for firm fixed effects, the correlation between capital and productivity becomes positive and firms
that use broadband through broadband technologies have a higher variable input demand. This
interaction was chosen as the indicator for broadband use for the remainder of the paper, as it
captures the use of higher broadband internet speeds. It is an established fact that high‐speed
internet requires technologies such as DSL or optical fiber for its delivery, where regular broadband
tends to be more technology neutral (Calvo, 2012).
Table 7 shows the estimation of the TFP policy function through three different methods: Ordinary
least squares, fixed effects and dynamic panel models. All estimations show that TFP is a function
of past productivity. The coefficient on lagged TFP has the expected positive sign which points to
the persistence of TFP growth within firms. Lagged broadband use is not statistically significant and
lagged broadband technology is only statistically significant in the fixed effects estimation, which is
my preferred specification. In these specifications their interaction is not statistically significant.
(1) (2) (3) (4) (5) (6) (7) (8)
VARIABLES
TFP ‐0.254***‐0.317*** ‐0.052*** 0.026*** ‐1.566*** ‐1.669*** ‐0.156*** ‐0.253***
[0.008] [0.006] [0.013] [0.008] [0.012] [0.009] [0.013] [0.010]
1.broadband 0.847*** ‐0.008 1.372*** 0.005 1.353*** 0.049*** 1.184*** 0.032**
[0.031] [0.010] [0.049] [0.012] [0.047] [0.014] [0.049] [0.016]
1.BBTechnology 0.440*** ‐0.010 0.740*** ‐0.010 0.607*** 0.027*** 0.629*** 0.016
[0.021] [0.007] [0.034] [0.009] [0.032] [0.010] [0.034] [0.012]
1.broadband#1.BBTechnology ‐0.309*** 0.034*** ‐0.601*** 0.004 ‐0.469*** 0.033** ‐0.493*** ‐0.015
[0.032] [0.010] [0.051] [0.013] [0.049] [0.015] [0.051] [0.017]
Constant 4.250*** 4.859*** 12.972*** 12.783*** 21.034*** 21.339*** 12.281*** 12.593***
[0.044] [0.028] [0.069] [0.037] [0.066] [0.041] [0.069] [0.048]
Observations 55,098 55,098 55,098 55,098 55,098 55,098 55,013 55,013
R‐squared 0.103 0.065 0.142 0.030 0.346 0.443 0.170 0.015
Plant FE NO YES NO YES NO YES NO YES
Year FE YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES
model ols fe ols fe ols fe ols fe
Number of plants 9,917 9,917 9,917 9,910
Standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Total employment Capital Materials Energy
11
Column (4) shows that all terms of a more flexible, third‐degree polynomial of lagged TFP, are
statistically significant. More precisely, that TFP does respond to past broadband use.
Table 7 TFP policy function estimation
ICTuse,productivityandlabordemandIn this section I provide evidence of the relationship between productivity, broadband use and labor
demand. To do so I estimate longitudinal fixed effects linear models as specified by the following
equation:
Β Β Β Β
Where interest lies in coefficient which captures the elasticity of dependent variable to
changes in broadband use. The employment dependent variables are logs of: Total employment,
direct employment, outsourced employment, plant laborers, plant professionals, total plant
workers, managers, permanent, temporary employment, men and women. This rich range of labor
types attempts to identify whether broadband use is related to changes in the labor demand of
different workers. METRO, SECTOR and YEAR, are vectors of indicator variables for each
VARIABLES (1) (2) (3) (4)
Lagged TFP 0.904*** 0.310*** 0.637*** ‐1.352***
[0.002] [0.005] [0.017] [0.156]
Lagged TFP^2 0.257***
[0.023]
Lagged TFP^3 ‐0.012***
[0.001]
Lagged broadband ‐0.003 ‐0.003 0.002 ‐0.002
[0.007] [0.007] [0.010] [0.007]
Lagged BBTech 0.002 0.009* 0.010 0.01**
[0.005] [0.005] [0.006] [0.005]
Lagged broadbandXLagged BBTech ‐0.016** ‐0.004 ‐0.040*** ‐0.003
[0.008] [0.008] [0.010] [0.008]
Constant 0.475*** 3.182*** 1.315*** 6.530***
[0.010] [0.024] [0.221] [0.331]
Observations 44,714 44,714 44,714 44,714
R‐squared 0.831 0.094 0.098
Plant FE NO YES NO NO
Year FE YES YES YES YES
Sector FE YES YES YES YES
model ols fe gmm fe
Number of nordest . 9,363 9,363 9,363
Standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
12
metropolitan area, (two‐digit) industry of economic activity and year respectively. Finally, the vector
X captures variables which are theoretically related to labor demand. I have included the lagged
value of capital and the minimum wage (which is interacted with the share of unskilled labor in
industry total employment). All standard errors were two‐way clustered at the state and plant level.
The period of analysis is 2008‐2014 given that information on broadband use is only available for
this time frame. The estimation sample is restricted to firms that report a consistent industry code
during the sample period and therefore the vector of coefficients Β will not be identified separately
form the fixed effect.
The adoption of broadband services may be correlated with unobserved factors, such as the
availability of service providers at the plant’s location or the costs of internet services. Under this
assumption coefficient will be inconsistent to identify the effect of broadband adoption. I
instrument broadband use with an interaction of log broadband speeds available for businesses at
the state level and the industry average broadband use through broadband technology in the
previous year. Therefore, the estimation will pick up the effect of broadband use which can be
explained by the variation in broadband speed availability at the industry‐state level. Broadband
speed is defined as the download internet speeds weighted by the number of subscriptions each
internet service provider reported in each state from 2008 to 2014.
The identification of the effect of broadband use due to broadband service quality use on labor
demand relies on the relevance and exclusion restrictions being satisfied by the instrument. The
instrument was constructed as the product of two variables: 1‐The average use of broadband at the
(three‐digit) industry level in the previous year.5 2‐The (weighted6) average download speeds at the
state level. The first part of the instruments is relevant since it captures the degree of broadband
use in each industry. As Table 15 shows, there are statistically significant differences in labor market
variables depending on whether firms belong to an industry where broadband is more intensively
used. The second part of the instruments is relevant as it captures a precise definition of broadband
internet quality, it is download speeds in the state where the production unit is located.7
The exclusion restriction of the instrument is given by the fact that the average, three‐digit industry
broadband use is predetermined each period. The second part of the instrument exploits the fact
that internet quality to businesses is determined by the service provider infrastructure and should
only affect labor demand through its effects on broadband adoption and use and not directly. Figure
3 suggests that between 2008 and 2014 the average internet speed increased in Colombia. By 2008
only the darkest areas had internet speeds above 2mbs, while by 2014 some of the white areas had
broadband coverage. However, this increase has not radically changed the areas where the highest
broadband speeds were available. The Appendix shows the broadband speed information for every
5 Recall that in the analysis I include industry dummies at the two‐digit level, so this is a finer measure of industry that is still larger than the firm level. 6 The average is weighted by the number of subscriptions each Internet Service Provider has. 7 The state level may still be too broad a measure of broadband availability. However, the data provided to us for this research project by Colombia’s national statistics agency did not include a finer geographical identification variable due to data confidentiality concerns.
13
year. A violation of the exclusion restriction would suggest firms change labor demand as a response
to internet speeds in their state and not because it affects labor demand through its effect on
broadband adoption and use.
Figure 3
Table 8 shows some suggestive statistics for firms across two dimensions. The first is whether they
were classified as being part of an industry where broadband use was above the manufacturing
sector average in the year 2008. The second dimension captures whether the firm currently uses
broadband internet. Current use of broadband is related to higher unskilled labor demand, output
and productivity. However, there is not any additional correlation due to being part of an intensive
broadband use sector. These results suggest a certain degree of complementarity of broadband use
with the demand of unskilled labor in Colombia but not with skilled labor. Even in the presence of a
positive relationship with output and total factor productivity.
(847,2979](582,847](492,582][248,492]
Average Download Speed 2008
(4663,6717](3821,4663](2650,3821][590,2650]
Average Download Speed 2014
14
Table 8 Broadband use intensity and selected outcomes
I now turn to the main results. Table 9 shows that firms that report using broadband internet appear
to have a higher labor demand than those with slower internet access; about 2.2 log points.
However, there does not appear to be any relationship between broadband use and outsourced,
management and permanent labor demand, which suggests internet use is particularly related to
production workers’ labor demand. Temporary labor displays the highest correlation with
broadband use. Since TFP has been estimated from a structural model, under the set of assumptions
explained before, the results can be interpreted as a causal effect on labor demand.
These results show a negative relationship of productivity gains on labor demand for all types of
workers, holding everything else constant. Columns 2 to 9 compare pairs of possible labor
substitutes in the production process. Taken at face value, increases in productivity from past
adoption of broadband internet would reduce to a higher degree outsourced labor than direct labor,
low skilled than high skilled labor demand, production workers relative to managers, and temporary
workers relative to permanent ones. These results suggest that labor types which may be easier to
adjust, such as outsourced, unskilled and temporary would be the most likely to be substituted by
broadband adoption. Interestingly, the size of the effect is similar for men and women, suggesting
an absence of gender bias in the substitution between labor and productivity growth.
(1) (2) (3) (4) (5)
VARIABLES Laborers Professionals Value of production Output per worker TFP
Intensive & Broadband (1=Yes) ‐0.013 0.017 0.003 0.005 0.008
[0.016] [0.019] [0.013] [0.010] [0.006]
Broadband (1=Yes) 0.017*** 0.007 0.019** 0.003 0.034***
[0.006] [0.011] [0.008] [0.008] [0.004]
Observations 50,466 28,967 52,661 52,616 50,753
R‐squared 0.007 0.003 0.009 0.002 0.005
Number of plants 9,037 5,915 9,378 9,369 9,139
Plant FE YES YES YES YES YES
Year FE YES YES YES YES YES
Sector FE YES YES YES YES YES
City FE YES YES YES YES YES
Std. Errors
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Cluster Plant/State
15
Table 9 Labor demand and Broadband use
Table 10 shows the results of my instrumental variables approach. As discussed, I used the log
average download speeds at the state level interacted with the lagged average broadband use at
the industry level. While other potential instruments such as the number of broadband
subscriptions and the upload speed were available, the preferred specification uses download speed
given the results of statistical tests which can be seen in Table 15 and Table 16 in the Appendix.8
Except for outsourced and professional workers, the first‐stage statistics are well above the
conventional levels of statistical significance. This is also true about the weak instruments and weak
instrument robust inference statistics.
Correcting for the possible endogeneity in broadband adoption suggests strong positive effects on
labor demand of broadband use. The point estimate on total labor demand is around 22 log points,
but it is not statistically significant. Lack of significance may be due to the use of two‐way clustered
standard errors and the fact that variation of the instrument is at the state and industry level. The
fact that most point estimates are strictly greater than zero provides some evidence of a positive
effect, which nonetheless is not precisely identified.
Alternatively, given the model assumptions under which TFP was estimated, the underlying labor
demand should only be a function of input prices and TFP. Thus, under these assumptions, the effect
of broadband adoption should only be observed through its indirect effect on TFP. Therefore, lack
of statistical significance in the 2SLS estimation is consistent with this theoretical model.9
8 See Table 15 which shows the first‐stage regression and Table 16 which shows relevance, and weak inference statistics. 9 I thank the participants at the applied microeconomics seminar at Banco de la República in Cali, especially Juan Esteban Carranza, Camila Casas and Salvador Navarro for pointing out this alternative explanation.
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)VARIABLES Total Direct Outsourced Laborer Professional Production Management Permanent Temporary Women Men
Broadband (1=Yes) 0.022*** 0.023*** ‐0.007 0.021*** 0.015** 0.023*** 0.006 0.008 0.037** 0.018*** 0.021***
[0.005] [0.006] [0.018] [0.005] [0.007] [0.005] [0.006] [0.007] [0.017] [0.005] [0.006]
TFP ‐0.332*** ‐0.291*** ‐0.745** ‐0.341*** ‐0.099** ‐0.304*** ‐0.195*** ‐0.204*** ‐0.328*** ‐0.253*** ‐0.277***
[0.056] [0.049] [0.324] [0.078] [0.049] [0.055] [0.037] [0.039] [0.083] [0.041] [0.051]
Observations 50,753 50,255 13,479 48,805 27,910 49,823 48,020 44,098 33,711 49,050 50,186
R‐squared 0.071 0.042 0.027 0.046 0.004 0.047 0.021 0.015 0.013 0.029 0.040
Number of plants 9,139 9,082 2,716 8,837 5,759 8,989 8,809 8,196 6,593 8,920 9,064
Plant FE YES YES YES YES YES YES YES YES YES YES YES
Year FE YES YES YES YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES YES YES YES
Std. Errors
model
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Cluster Plant/State
fe
16
Table 10 2SLS estimation Labor demand and Broadband use
RobustnessexerciseIn this section I use a balanced panel sample to assess the degree to which sample selection could
change the previous results. The balanced panel has a slightly higher share of large firms than the
unbalanced panel in the main exercise, but the size distributions are not too different from one
another.
Table 11 Size distribution in the balanced panel.
Using a balanced panel does not change the main results, so there should be little concern about
the effect of sample selection in the OLS regression. The fact that the balanced sample is comprised
of about 70% of plants in the period is also reassuring. When I instrumented broadband on the
balanced panel, the instrument loses its relevance, making the comparison between the two
samples difficult. This suggests, however, that entry and exit dynamics could be related to
broadband quality in a way that is systematically different for firms that remain in the sample since
2008 and those that enter or leave. Therefore, the 2SLS results in the main section should be
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
VARIABLES Total Direct Outsourced Laborer Professional Production Management Permanent Temporary Women Men
Broadband (1=Yes) 0.217 0.017 1.672* 0.182 ‐1.217 0.166 ‐0.100 ‐0.011 0.607 ‐0.047 0.361
[0.293] [0.396] [1.001] [0.393] [2.076] [0.393] [0.409] [0.415] [0.812] [0.285] [0.339]
TFP ‐0.304*** ‐0.250*** ‐0.698* ‐0.316*** ‐0.001 ‐0.280*** ‐0.157*** ‐0.174*** ‐0.316*** ‐0.213*** ‐0.267***
[0.060] [0.060] [0.360] [0.093] [0.133] [0.068] [0.050] [0.053] [0.080] [0.048] [0.056]
Observations 44,816 44,396 11,597 43,030 24,517 43,970 42,347 38,880 29,570 43,266 44,298
R‐squared 0.009 0.035 ‐0.668 0.012 ‐0.599 0.017 0.009 0.012 ‐0.077 0.021 ‐0.082
Number of plants 8,853 8,802 2,574 8,549 5,503 8,703 8,522 7,909 6,354 8,629 8,777
Plant FE YES YES YES YES YES YES YES YES YES YES YES
Year FE YES YES YES YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES YES YES YES
Errors
Instrument
Weak identification test (Kleibergen‐
Paap rk Wald F statistic) 12.73 12.01 4.126 13.42 2.841 12.39 12.77 11.02 9.504 13.96 13.25
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Cluster Plant/State
Download speed at state level
Size FrequencyPercent Cum.
[0,50] 42,241 70.19 70.19
(50,200] 10,868 18.06 88.25
(200,1000] 5,509 9.15 97.41
>1000 1,561 2.59 100
Total 60,179 100
Size FrequencyPercent Cum.
[0,50] 5,863 68.2 68.2
(50,200] 1,699 19.76 87.96
(200,1000] 802 9.33 97.29
>1000 233 2.71 100
Total 8,597 100
Full sample period
Year 2008
17
interpreted as being driven mainly by entry and exit dynamics in the manufacturing sector. Further
work to account for these dynamics would be an interesting extension to the current results.
Table 12 Labor demand and Broadband use
Table 13 14 2SLS estimation Labor demand and Broadband use
DiscussionColombia is a country where businesses have a relatively high level of internet access. However,
there is a great amount of heterogeneity regarding internet speed. Considering that the
identification strategy proposed for the paper relies on the geographical nature of internet rollout
to identify the causal effect of ICT on labor demand which is mediated through productivity gains, a
finer measure of internet access was considered. The measure combines broadband speeds (greater
than 1Mbps) and broadband technology use (DSL, cable/fiber, wireless). Figure 4 and Figure 5
showed that this measure displayed both time and industry variation, making it a good candidate to
model productivity gains due to broadband use. Since broadband download speed might be a key
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)VARIABLES Total Direct Outsourced Laborer Professional Production Management Permanent Temporary Women Men
Broadband (1=Yes) 0.019*** 0.021*** ‐0.012 0.017** 0.018* 0.022*** 0.004 0.011 0.020 0.015*** 0.018***
[0.006] [0.007] [0.022] [0.007] [0.009] [0.007] [0.007] [0.008] [0.019] [0.005] [0.006]
TFP ‐0.276*** ‐0.238*** ‐0.580** ‐0.305*** ‐0.076* ‐0.269*** ‐0.176*** ‐0.176*** ‐0.289*** ‐0.220*** ‐0.230***
[0.052] [0.045] [0.272] [0.084] [0.040] [0.060] [0.031] [0.039] [0.076] [0.035] [0.046]
Observations 34,602 34,232 10,511 33,236 19,944 33,944 32,839 30,384 24,095 33,488 34,282
R‐squared 0.060 0.033 0.019 0.040 0.003 0.042 0.019 0.014 0.010 0.025 0.032
Number of plants 6,363 6,322 2,248 6,150 4,180 6,256 6,160 5,751 4,830 6,225 6,323
Plant FE NO NO NO NO NO NO NO NO NO NO NO
Year FE YES YES YES YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES YES YES YES
Std. Errors
model
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Cluster Plant/State
fe
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
VARIABLES Total Direct Outsourced Laborer Professional Production Management Permanent Temporary Women Men
Broadband (1=Yes) ‐0.121 ‐0.399 1.296 ‐0.086 ‐2.233 ‐0.095 ‐0.232 ‐0.572 0.601 ‐0.319 ‐0.045
[0.247] [0.417] [1.082] [0.365] [1.926] [0.371] [0.382] [0.530] [0.816] [0.356] [0.320]
TFP ‐0.263*** ‐0.199*** ‐0.637** ‐0.293*** 0.060 ‐0.257*** ‐0.155*** ‐0.127** ‐0.322*** ‐0.190*** ‐0.224***
[0.060] [0.063] [0.316] [0.102] [0.128] [0.075] [0.054] [0.065] [0.091] [0.056] [0.052]
Observations 34,602 34,232 10,511 33,236 19,944 33,944 32,839 30,384 24,095 33,488 34,282
R‐squared 0.033 ‐0.143 ‐0.407 0.029 ‐2.067 0.026 ‐0.027 ‐0.205 ‐0.082 ‐0.069 0.028
Number of plants 6,363 6,322 2,248 6,150 4,180 6,256 6,160 5,751 4,830 6,225 6,323
Plant FE NO NO NO NO NO NO NO NO NO NO NO
Year FE YES YES YES YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES YES YES YES
Errors
Instrument
Weak identification test (Kleibergen‐
Paap rk Wald F statistic) 6.114 5.409 2.662 6.852 4.456 6.803 5.672 6.131 5.684 5.201 7.062
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Cluster Plant/State
Download speed at state level
18
factor explaining the adoption of broadband internet, it was considered a suitable candidate for an
instrument.
In this paper, I proposed and estimated total factor productivity under the assumption that it
evolves according to a process which depends on past broadband adoption. I estimated the effect
of broadband adoption on labor demand for different worker types. I find positive correlations
between broadband adoption and labor demand of most worker types of about 2%. The
instrumental variable estimation showed a positive but statistically insignificant effect of broadband
use on total labor demand. The fact that I used two‐way clustered standard errors could have
affected the statistical significance of the test. Alternatively, lack of statistical significance could
come from the fact that broadband should only affect labor demand through its direct effect on TFP
growth. In Table 8, I showed a positive relationship between broadband use and total factor
productivity, which suggests a positive complementarity between broadband use and productivity.
To further asses the possible skill complementarity of broadband, I interacted total factor
productivity and broadband adoption. Table 17 in the Appendix shows that labor demand for
unskilled labor is negatively correlated to productivity increases even further for firms that use
broadband, regardless of the method used to estimate TFP. Table 18 shows that for exporting firms
although unskilled labor demand is negatively related to TFP growth, the value of production and
output per worker are positively related to TFP growth.
The paper contributes by providing evidence about the effect of ICT adoption which is associated
with productivity growth on labor demand. The lack of significance in the instrumental variable
estimations suggests that other identification strategies are needed to capture the causal effect of
broadband use on labor demand. Particularly, since the instrument appears to be highly sensitive
to entry and exit dynamics. Although several possible instruments were available, download speed
appears to be a strong predictor of broadband adoption to a greater extent than the number of
subscriptions or upload speed, but more work should be conducted in determining the appropriate
instruments to pin down the causal effects of broadband adoption on labor demand and
productivity.
References(CINTEL), C. de I. de las T. (2003). Análisis del Mercado Servicios de Banda Ancha en Colombia.
Bogotá.
(CRC), C. de R. de C. (2015). IV Reporte de industria del sector TIC. Bogotá.
Ackerberg, D. A., Caves, K., & Frazer, G. (2015). Identification Properties of Recent Production Function Estimators. Econometrica, 83(6), 2411–2451. https://doi.org/10.3982/ECTA13408
Ackerberg, D., Caves, K., & Frazer, G. (2006). Structural Identification of Production Functions. MPRA Paper No. 38349, 12(38349), 1–35. Retrieved from http://folk.uio.no/rnymoen/Ackerberg_Caves_Frazer.pdf
Akerman, A., Gaarder, I., & Mogstad, M. (2015). The skill complementarity of broadband internet. Quarterly Journal of Economics, 130(4), 1781–1824. https://doi.org/10.1093/qje/qjv028
19
Bertschek, I., Cerquera, D., & Klein, G. J. (2013). More bits – more bucks? Measuring the impact of broadband internet on firm performance. Information Economics and Policy, 25(3), 190–203. https://doi.org/10.1016/j.infoecopol.2012.11.002
Brambilla, I. (2018). Digital Technology Adoption and Jobs: A Model of Firm Heterogeneity.
Calvo, A. G. (2012). Universal Service Policies in the Context of National Broadband Plans. OECD Digital Economy Papers, (203). https://doi.org/10.1787/20716826
Casas, C., & González, A. (2016). Productivity Measures for the Colombian Manufacturing Industry * (Borradores de economía No. 947). Cali.
Colombo, M. G., Croce, A., & Grilli, L. (2013). ICT services and small businesses’ productivity gains: An analysis of the adoption of broadband Internet technology. Information Economics and Policy, 25(3), 171–189. https://doi.org/10.1016/j.infoecopol.2012.11.001
De Loecker, J. (2013). Detecting learning by exporting. American Economic Journal: Microeconomics, 5(3), 1–21. https://doi.org/10.1257/mic.5.3.1
De Loecker, J., & Warzynski, F. (2012). Markups and firm‐level export status. American Economic Review, 102(6), 2437–2471. https://doi.org/10.1257/aer.102.6.2437
Eslava, M., Haltiwanger, J., Kugler, A., & Kugler, M. (2004). The effects of structural reforms on productivity and profitability enhancing reallocation: Evidence from Colombia. Journal of Development Economics, 75(2 SPEC. ISS.), 333–371. https://doi.org/10.1016/j.jdeveco.2004.06.002
Fabling, R., Grimes, A., & Grimes, A. (2016). Picking up speed : Does ultrafast broadband increase firm productivity ?, (November).
Gandhi, A., Navarro, S., & Rivers, D. (2016). On the Identification of Production Functions: How Heterogeneous is Productivity? Mimeo, (May 2006).
Grimes, A., Ren, C., & Stevens, P. (2012). The need for speed: impacts of internet connectivity on firm productivity. Journal of Productivity Analysis, 37(2), 187–201. https://doi.org/10.1007/s11123‐011‐0237‐z
Haller, S. A., & Lyons, S. (2015). Broadband adoption and firm productivity: Evidence from Irish manufacturing firms. Telecommunications Policy, 39(1), 1–13. https://doi.org/10.1016/j.telpol.2014.10.003
Levinsohn, J., & Petrin, A. (2003). Estimating Production Functions Using Inputs to Control for Unobservables. Review of Economic Studies, 70(2), 317–341. https://doi.org/10.1111/1467‐937X.00246
Ministerio de Tecnologías de la Información y las Comunicaciones. (2017). Proyecto Nacional de Fibra Óptica. Retrieved February 12, 2017, from http://www.mintic.gov.co/portal/vivedigital/612/w3‐propertyvalue‐647.html
Olley, G. S., & Pakes, A. (1996). The Dynamics of Productivity in the Telecommunications Equipment Industry. Econometrica, 64(6), 1263–1297.
20
AppendixTable 15 First Stage Regression for 2SLS Estimation
Note: Each column shows the sample over which the 2SLS estimation took place. The dependent variable is broadband use.
Table 16 First Stage test statistics
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Sample Total Direct Outsourced Laborer Professional Production Management Permanent Temporary Women Men
TFP 0.090*** 0.090*** 0.051 0.109*** 0.060*** 0.095*** 0.085*** 0.079*** 0.064*** 0.088*** 0.091***
[0.019] [0.018] [0.044] [0.032] [0.016] [0.022] [0.018] [0.018] [0.019] [0.019] [0.019]
Capital (Lagged) ‐0.001 ‐0.001 ‐0.022** ‐0.001 ‐0.012* ‐0.002 ‐0.002 0.002 ‐0.010 ‐0.001 ‐0.002
[0.008] [0.008] [0.009] [0.008] [0.007] [0.007] [0.007] [0.008] [0.010] [0.008] [0.008]
Minimum Wage X Share of Laborers in Industry 0.000* 0.000* ‐0.000 0.000* 0.000 0.000* 0.000 0.000 0.000** 0.000* 0.000*
[0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000]
Excluded instrument from structural equation 0.060*** 0.057*** 0.042** 0.061*** 0.030* 0.061*** 0.062*** 0.050*** 0.054*** 0.061*** 0.061***
[0.017] [0.016] [0.021] [0.017] [0.018] [0.017] [0.017] [0.015] [0.018] [0.016] [0.017]
Observations 44,816 44,396 11,597 43,030 24,517 43,970 42,347 38,880 29,570 43,266 44,298
Number of plants 8,853 8,802 2,574 8,549 5,503 8,703 8,522 7,909 6,354 8,629 8,777
Plant FE NO NO NO NO NO NO NO NO NO NO NO
Year FE YES YES YES YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES YES YES YES
Errors
Instrument
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Download speed at state level
Cluster Plant/State
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Statistic Total Direct Outsourced Laborer Professional Production Management Permanent Temporary Women Men
1St Stage F Test 12.73 12.01 4.13 13.42 2.84 12.39 12.77 11.02 9.5 13.96 13.25
Prob> F 0.0014 0.0018 0.0545 0.0011 0.1034 0.0016 0.0014 0.002 0.0048 0.0009 0.0011
Weak identification test (Cragg‐Donald
Wald F statistic) 19.27 17.22 3.03 18.99 3.08 19.65 19.33 11.98 11.27 19.18 20.05
Weak‐instrument‐robust inference
(Anderson‐Rubin Wald test) 0.6 0 5.62 0.24 0.58 0.19 0.06 0 0.75 0.03 1.38
P‐Value 0.4452 0.9672 0.0269 0.6303 0.4541 0.6646 0.8126 0.9801 0.394 0.8727 0.2505
Note: The Stock‐Yogo weak ID test critical values: 10% maximal IV size is 16.38.
21
HeterogeneouseffectsofbroadbanduseTable 17
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
VARIABLES
TFP#1.broadband ‐0.087** 0.001 ‐0.012 0.015
[0.042] [0.023] [0.025] [0.013]
Broadband (1=Yes) 0.421** 0.035 ‐0.304*** 0.012 ‐0.008 ‐0.190 0.080 ‐0.082 ‐0.266* ‐0.068 ‐0.032 0.041
[0.194] [0.061] [0.112] [0.106] [0.129] [0.209] [0.109] [0.106] [0.149] [0.064] [0.054] [0.329]
TFP ‐0.293*** ‐0.099** ‐0.253*** 0.061***
[0.068] [0.040] [0.051] [0.017]
Lagged Capital 0.081*** 0.075*** 0.066***0.074***0.072***0.069*** 0.091*** 0.084*** 0.097*** 0.003 0.004 0.029***
[0.006] [0.007] [0.007] [0.007] [0.007] [0.007] [0.007] [0.009] [0.010] [0.007] [0.007] [0.008]
Minimum Wage 0.000 0.000 0.000 ‐0.000* ‐0.000** ‐0.000** 0.000*** 0.000*** 0.000***0.000***0.000***0.000***
[0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000]
TFP#1.broadband ‐0.007 0.006 0.025 0.010
[0.016] [0.034] [0.029] [0.014]
TFP(DLW) ‐0.232*** ‐0.051 ‐0.236*** ‐0.004
[0.052] [0.038] [0.043] [0.018]
TFP#1.broadband 0.055*** 0.035 0.048* ‐0.007
[0.020] [0.036] [0.025] [0.056]
TFP(ACF) ‐0.566*** ‐0.186*** 0.307*** 0.939***
[0.100] [0.041] [0.070] [0.143]
Observations 48,805 48,573 48,573 27,910 27,834 27,834 50,747 50,442 50,442 50,747 50,442 50,442
R‐squared 0.047 0.021 0.055 0.004 0.003 0.005 0.032 0.021 0.030 0.003 0.001 0.183
Number of plants 8,837 8,737 8,737 5,759 5,704 5,704 9,138 9,030 9,030 9,138 9,030 9,030
Plant FE NO NO NO NO NO NO NO NO NO NO NO NO
Year FE YES YES YES YES YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES YES YES YES YES
Std. Errors
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Laborers Professionals Value of production Output per worker
Cluster Plant/Depto
22
Table 18
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)
VARIABLES
TFP#.1exports ‐0.335** ‐0.008 0.059*** 0.094*
[0.137] [0.032] [0.022] [0.056]
Exports (1=Yes) 1.550** 0.936* 0.307 0.051 0.130 0.033 ‐0.227** ‐0.418*** ‐0.184 ‐0.406 ‐0.428* 0.003
[0.620] [0.553] [0.259] [0.132] [0.202] [0.231] [0.097] [0.091] [0.338] [0.263] [0.244] [0.696]
TFP ‐0.321*** ‐0.096** ‐0.266*** 0.058***
[0.074] [0.044] [0.047] [0.014]
Lagged Capital 0.081*** 0.086*** 0.077*** 0.074*** 0.097*** 0.094*** 0.091*** 0.117*** 0.136*** 0.003 0.019*** 0.050***
[0.006] [0.011] [0.010] [0.007] [0.009] [0.009] [0.007] [0.011] [0.009] [0.007] [0.005] [0.007]
Minimum Wage 0.000* ‐0.000 ‐0.000 ‐0.000* ‐0.000*** ‐0.000** 0.000*** 0.000*** 0.000*** 0.000*** 0.000*** 0.000***
[0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000] [0.000]
TFP#.1exports ‐0.240 ‐0.024 0.131*** 0.122*
[0.148] [0.054] [0.024] [0.064]
TFP(DLW) ‐0.345*** ‐0.090* ‐0.367*** ‐0.035**
[0.077] [0.046] [0.064] [0.016]
TFP#.1exports ‐0.046 0.001 0.044 0.004
[0.045] [0.040] [0.058] [0.120]
TFP(ACF) ‐0.664*** ‐0.206*** 0.365*** 1.101***
[0.093] [0.052] [0.062] [0.131]
Observations 48,805 72,460 72,460 27,910 41,671 41,671 50,747 75,077 75,077 50,747 75,077 75,077
R‐squared 0.050 0.038 0.073 0.004 0.008 0.010 0.032 0.043 0.040 0.004 0.006 0.216
Number of Plants 8,837 10,033 10,033 5,759 6,643 6,643 9,138 10,334 10,334 9,138 10,334 10,334
Plant FE YES YES YES YES YES YES YES YES YES YES YES YES
Year FE YES YES YES YES YES YES YES YES YES YES YES YES
Sector FE YES YES YES YES YES YES YES YES YES YES YES YES
City FE YES YES YES YES YES YES YES YES YES YES YES YES
Std. Errors
Robust standard errors in brackets
*** p<0.01, ** p<0.05, * p<0.1
Laborers Professionals Value of production Output per worker
Cluster Plant/State
23
Figure 4 Broadband use through broadband technology by industry 2008‐2014.
Figure 5 Broadband use through broadband technology by industry 2008‐2014.
.5.6
.7.8
.9.5
.6.7
.8.9
.5.6
.7.8
.9
2008 2010 2012 2014 2008 2010 2012 2014 2008 2010 2012 2014 2008 2010 2012 2014
15 16 17 18
19 20 21 22
23 24 25 26
95% CI lpoly smooth: BBTechXbroadband
lpoly smoothing grid
Graphs by ciiu2_r3
.5.6
.7.8
.9.5
.6.7
.8.9
.5.6
.7.8
.9
2008 2010 2012 2014 2008 2010 2012 2014
2008 2010 2012 2014 2008 2010 2012 2014
27 28 29 30
31 32 33 34
35 36
95% CI lpoly smooth: BBTechXbroadband
lpoly smoothing grid
Graphs by ciiu2_r3
24
Figure 6 Average download and upload broadband speed by state 2008‐2014. Source: Comisión de Regulación de Comunicaciones.
(847,2979](582,847](492,582][248,492]
Average Download Speed 2008
(439,1501](308,439](253,308][176,253]
Average Upload Speed 2008
(1227,2766](960,1227](765,960][211,765]
Average Download Speed 2009
(571,924](469,571](396,469][134,396]
Average Upload Speed 2009
25
(1822,3704](1702,1822](1435,1702][270,1435]
Average Download Speed 2010
(897,1148](825,897](720,825][188,720]
Average Upload Speed 2010
(2442,4198](2197,2442](1773,2197][337,1773]
Average Download Speed 2011
(1128,1376](1042,1128](878,1042][198,878]
Average Upload Speed 2011
26
(3068,5091](2643,3068](1879,2643][593,1879]
Average Download Speed 2012
(1280,1699](1200,1280](936,1200][416,936]
Average Upload Speed 2012
(3699,5643](3240,3699](2376,3240][559,2376]
Average Download Speed 2013
(1422,2075](1300,1422](1056,1300][443,1056]
Average Upload Speed 2013
27
(4663,6717](3821,4663](2650,3821][590,2650]
Average Download Speed 2014
(1368,2151](1245,1368](1050,1245][517,1050]
Average Upload Speed 2014