Learning-by-doing, learning-by-exporting, and productivity: evidence - - PowerPoint PPT Presentation

Learning-by-doing, learning-by-exporting, and productivity: evidence from Colombia

Ana M. Fernandes Alberto E. Isgut The World Bank Wesleyan University

LACEA Conference Paris, October 27-29, 2005

Does participation in export markets increase plant-level productivity?

Appealing idea but inconclusive empirical evidence so far What is the direction of causality? Learning-by-exporting hypothesis: Export participation => Improvement in plant-level productivity Or self-selection hypothesis: Higher plant productivity => Export participation?

What have other researchers found?

Method 1: Difference-in-differences specifications

Sample partitioned into a treatment group and a control group, e.g. entrants into export markets vs. nonexporters YiT measure of plant performance, e.g. TFP Di0=1 if plant i belongs to treatment group Zi0 controls, e.g. plant size, industry, year

( )

it i i i iT iT

Z D Y Y T Y ε β β β + + + = − ≡ ∆

2 1

' ln ln 1 ln

No OLS Chile Alvarez and Lopez (2004) Yes Matched samples Slovenia De Loecker (2005) No Matched samples Germany Arnold and Hussinger (2004) Yes OLS Korea Hahn (2004) Yes Matched samples UK Girma et al. (2004) Yes OLS Canada Baldwin and Gu (2003) No IV, c-s (time of 1st exports) Morocco Fafchamps et al. (2002) Yes (except Korea) OLS, c-s (born exporters) Indonesia, Korea, Malaysia, Philippines, Thailand Hallward-Driemeier et al. (2002) Yes OLS (export intensity) Italy Castellani (2002) Some (young) Nonparametric Spain Delgado et al. (2002) No OLS Colombia Isgut (2001) Some (Taiwan) OLS Korea, Taiwan Aw et al. (2000) No OLS USA Bernard and Jensen (1999) No OLS Germany Bernard and Wagner (1997) LBE?

• Estim. method

Country Study

Method 2: Dynamic specifications

Yit plant performance measure Dit-1= 1 if plant i exported in t-1 Xit controls (e.g., industry, year) ? it plant-specific effect unobserved by the econometrician

it it it it it it

X Y D Y ε ω β β β β + + + + + =

− −

' ln ln

2 1 2 1 1

Yes SYS-GMM Cameroon, Cote d’Ivoire, Ghana, Kenya, Tanzania, Zambia, Zimbabwe Van Biesebroeck (2005) Yes Various (current exports) Indonesia Blalock and Gertler (2004) Yes FIML w/ export participation equation Yes Olley and Pakes (1996) Yes FIML nonparametric errors No FIML w/ export participation equation Cameroon, Ghana, Kenya, Zimbabwe Bigsten et al. (2004) Yes IV (lagged export intensity) China Kraay (1999) Some (Morocco) GMM No FIML w/export participation equation Colombia, Morocco Clerides et al. (1998) LBE?

• Estim. method

Country Study

What previous studies have in common?

• Most measure export involvement using an export

participation dummy variable

But shouldn’t learning-by-exporting depend on the extent of exposure to export markets?

• Most use the whole sample of manufacturing plants

But are all manufacturing plants equally likely to learn?

What is learning by exporting?

• Define learning-by-exporting drawing on Arrow (RES,

1962) characterization of learning-by-doing:

• “Learning is the product of experience. Learning can only take

place through the attempt to solve a problem and therefore only takes place during activity” (p. 155).

• “Learning associated with repetition of essentially the same

problem is subject to sharply diminishing returns… To have steadily increasing performance, then, implies that the stimulus situations must themselves be steadily evolving rather than merely repeating” (pp. 155-6).

What do firms learn from exporting?

• To adopt stringent technical standards demanded by

sophisticated consumers

• To use new, more efficient equipment that might need to

be introduced for export production

• To meet orders in a timely fashion and ensure the quality
• f their products

If learning is the product of experience, then learning-by-exporting should depend on measures of export experience rather than export participation If learning is subject to sharply diminishing returns, then learning-by-exporting should not be observed in established exporters => focus on young exporters Firms might learn both by exporting and “by doing” => we control for learning-by-doing effects in our estimation of learning-by-exporting effects

Implications for measurement

Experience measures

• Output experience:
• Plant age
• Index of cumulative production up to t-1:
• In the regressions we enter these variables in reciprocal form
• Export experience:
• Number of years plant exported up to t-1
• Index of cumulative exports up to t-1:

i i

FY i t FY i it

Y Y YE

, 1

/

∑

− =

=

τ τ

i i

FE i t FE i it

E E EE

, 1

/

∑

− =

=

τ τ

Empirical specification

• Production function:

Lit labor Mit materials Kit capital Qit vector of factor quality measures, including skill intensity Sit, wage premium Wit, and capital vintage Vit Ait total factor productivity

( )

( )

it it it EE it YE it it it q it it it it

EE YE A A Q K M L Y

k m l

ε ω β β β

β β β

+ + + = = exp ' exp

Estimation

• Two-step approach

1) Olley-Pakes and Levinsohn-Petrin assumptions:

a. ? it follows a Markov process

• b. Manager chooses variable inputs based on the

expected value of ? it and the other state variable, capital

2) To control for differences in input quality, we add an additional state variable, capital vintage, and two additional inputs, skill ratio and wage premium 3) For now we assume that the choice of inputs is uncorrelated with output and export experience

Estimation (Cont’d)

4) In the first step we estimate using Levinsohn and Petrin 5) Our measure of TFP is 6) In the second step we estimate

• Notice that we add a fixed plant effect

it it it

a ε ω ˆ ˆ ˆ + =

it i it EE it YE it

u f EE YE a + + + + = β β β0 ˆ

( )

( )

it it it q it it it it

Q K M L Y

k m l

ε ω β

β β β

+ = exp ' exp

Estimation (Cont’d)

7) Because ? it follows a Markov process, the error term uit is likely to be autocorrelated. We model it as

If ? ? 0, the estimating equation can be written as Taking first differences to get rid of the fixed effect we get

( ) ( ) ( ) ( )

it i it it EE it it YE it it

f EE EE YE YE a a ν ρ ρ β ρ β ρ β ρ + − + − + − + + − =

− − −

1 ˆ 1 ˆ

1 1 1

) , .( . . ~ , 1 ,

1 ν

σ ν ρ ν ρ d i i u u

it it it it

≤ + =

−

it it EE it EE it YE it YE it it

EE EE YE YE a a ν ρβ β ρβ β ρ ∆ + ∆ − ∆ + ∆ − ∆ + ∆ = ∆

− − − 1 1 1

ˆ ˆ

Estimation (Cont’d)

8) As we know from the literature on dynamic panel data models, . Also, using the cumulative production index as output experience, , which is also correlated with the error term. 9) According to the self-selection into export markets hypothesis, past and current favorable productivity shocks can facilitate a firm’s access to export markets. We assume, following Kraay (1999) that future productivity shocks are uncorrelated with export participation: . Therefore, the first differences of our measures of export experience are correlated with the error term.

) , ˆ (

1

≠ ∆ ∆

− it it

a Cov ν

i

FY i it it

Y Y YE

, 1 / −

∆ = ∆

, ) , ( > =

+

s E Cov

s it it ν

Estimation (Cont’d)

10) Conclusion: we need to use an IV method. In the paper we use Blundell and Bond’s (1998) system- GMM estimator. 11) Notice that if ? = 1, the model simplifies to In this case we do not need to instrument for the differences in experience variables because they are transformations of lagged output and exports, which are uncorrelated with the error term at time t.

it it EE it YE it

EE YE a ν β β + ∆ + ∆ = ∆ˆ

Data

• Plant-level data from manufacturing census of

Colombia 1981-1991 (DANE)

• Output and intermediates deflated by industry-level

price indexes, accounting for exports and imported materials

• Main sample includes plants born in or after 1981
• No export data before 1981
• Exclude plants with three or less consecutive observations, in

industries with less than 100 observations, first year of each plant, and outliers

• 15,457 plant-year observations (3,091 plants)

Effect of export experience on plant TFP: System-GMM estimation – Assumption: ? ? 0

• 0.0010

(0.0024) Lagged cumulative export index 0.0013 (0.0018) Cumulative export index

• 0.097**

(0.043) Lagged number of years plant exported 0.092** (0.037) Number of years plant exported 1 2 0.999*** (0.006) 0.995*** (0.008) Lagged productivity

Sample of plants with 8 or more observations; robust standard errors in parentheses; regressions include year dummies

• Our estimates of ? are not significantly different from

1, with p-values of 0.25 and 0.47.

• Unit root tests on confirm this result.
• Therefore, we feel we are justified to estimate the

model with the simpler first difference specification.

• We show results for different subsamples
• In all cases we control for current exports by adding a

current exports dummy (in first differences) to the estimating equation.

it

a ˆ

Effect of export experience on plant TFP: ? = 1

0.049*** (0.007)

Full sample (n=3091) First differences (OLS)

0.041*** (0.011)

Exporters (n=465)

0.073*** (0.025)

Born exporters (n=130)

Coefficients of number of years exported, controlling for current exports

Regressions include year dummies

Note: In current draft we estimated the model using long differences and the within estimator. The rationale is

• this. The model with ? = 1 can be expressed as

Taking long differences, we get While it seems like this model can be estimated by WLS to account for heteroscedasticity, the problem is that the error term is correlated with both experience variables.

∑

=

+ + + + =

t F i i it EE it YE it

i

f EE YE a

τ τ

ν β β β0 ˆ

( ) ( ) ∑

+ =

+ − + − = −

i i i i i i i i

L F i iF iL EE iF iL YE iF iL

EE EE YE YE a a

1

ˆ ˆ

τ τ

ν β β

The same problem occurs if we estimate the model using the within estimator. Letting , the within estimator is which also has the error term correlated with the experience variables. Nevertheless, in the next table we compare the three estimators (differences, long differences, within) to assess how serious is the endogeneity bias.

( ) ( )

. . . .

ˆ ˆ

i it i it EE i it YE i it

EE EE YE YE a a Ω − Ω + − + − = − β β

∑

=

= Ω

t F i it

i

τ τ

ν

Effect of export experience on plant TFP: ? = 1

0.040*** (0.005)

Long differences (WLS)

0.044*** (0.003) 0.049*** (0.007)

Full sample (n=3091) Within estimator (WLS) First differences (OLS)

0.036*** (0.008) 0.039*** (0.004) 0.041*** (0.011)

Exporters (n=465)

0.074*** (0.013) 0.078*** (0.008) 0.073*** (0.025)

Born exporters (n=130)

Comparison of estimators

Regressions include year dummies

Variant: Taking advantage of our TFP estimates, we replicate studies of learning-by-exporting based on differences in differences and matched samples.

Replication: Growth rate of plant TFP after entry into export markets

0.015 (0.014) 1 year

Diff.-in-diff. (OLS)1 Time horizon

0.030*** (0.007) 5 years 0.026*** (0.008) 3 years

1. Initial wage, initial skill, initial size (labor), initial capital intensity, and year, industry, and region dummies used as controls. 2. Nonexporters matched with exporters in the same year and industry.

0.033*** (0.010) 0.041*** (0.014) 0.036* (0.021)

Propensity score matching2

Do entrants into export markets become more productive in their respective industries?

1.3 (1.5) 1 year

Diff.-in-diff. (OLS) Time horizon

12.8*** (4.6) 5 years 5.1** (2.6) 3 years 11.8** (4.6) 9.2** (4.0) 1.8 (2.2)

Propensity score matching

Dependent variable: Change in percentile of plant in industry-year distribution of TFP

One-step approach

• Estimate link between export experience and plant TFP

directly from the production function

• Now we assume that the manager chooses variable

inputs based on the expected value of ? it, capital, and two other state variables, output experience and export experience

• Here we allow for the possibility that the choice of

inputs is correlated with experience

One-step approach

• In a variant of the estimation we include a dummy for

current exports, which we model as a choice variable.

• In another variant we include a dummy for exporters,

which we model as another state variable.

• We estimate this model separately for the five largest

manufacturing industries in Colombia.

Learning-by-exporting: One-step regressions

0.036* Textiles (n=997) 0.092*** Apparel (n=3045) 0.038** Plastics (n=914) Metal products (n=1218) Food products (n=1937) 0.060** 0.045**

Basic specification

0.028* 0.041* 0.026* 0.048* 0.024

Specific. w/export dummy Coefficients of number of years exported

***, ** and * indicate significance at the 1%, 5% and 10% confidence level, respectively

0.059** 0.065** 0.034** 0.046** 0.030*

Specific. w/exporter dummy

Extension 1

• Do old plants learn from exporting as much as young

plants?

• To investigate this question, we expand the sample by

including “old” plants, those that started operations prior to 1981

46,574 plant-year observations (6,171 plants) Assumption: old plants showing at least three years with zero exports before exporting for the first time during the 1981-1991 period are new entrants into export markets

Learning-by-exporting: Young vs. Old

Old Young 0.064*** (0.004) 0.066*** (0.003) 0.018*** (0.003) 0.022*** (0.002)

Exporters Full sample

Coefficients of number of years exported, controlling for current exports, within (WLS) Note: Provisional – Estimation to be revised

Regressions include year dummies

Extension 2

• What causes differences in learning-by-exporting across

industries?

H1: High income countries’ consumers are more discriminating about the quality of their imports => learning-by-exporting is positively related to the percent of industry exports delivered to high income countries H2: Industries that export more have better developed channels of distribution, facilitating access of newcomers to export markets => learning-by-exporting is positively related to the value of industry exports

Learning-by-exporting: Explaining differences across industries Note: Provisional – Estimation to be revised

0.063*** (0.014) Number of years plant exported interacted with share of industry exports to high income countries 0.068*** (0.011) Current exports dummy 0.012 (0.008) Number of years plant exported 1 0.068*** (0.011)

• 0.140***

(0.031) 2 0.016*** (0.003) Number of years plant exported interacted with log of value of industry exports

Fixed effects regressions with year dummies; robust standard errors in parentheses

Conclusions

• Strong evidence of learning-by-exporting for young plants in

Colombia

• TFP increases, on average, 4%-5% for each additional

year a plant has exported, controlling for current exports

• Results are robust to the use of different estimation

methods and different subsamples of the data

• TFP of entrants into export markets grows 3%-4% faster

than that of non-exporters

• 1-step approach confirms findings of learning-by-

exporting across industries

• Young plants’ TFP benefits more from export

experience than old plants’ TFP

• There is more learning-by-exporting in industries:
• that deliver a larger percent of their exports to high-

income countries

• with larger export values

• Some tentative policy implications based on the results:
• Exporting increases plant productivity => avoid

policies that discourage access of domestic firms to export markets

• Young plants have potential for learning => facilitate

the process of creative-destruction by which failing plants give rise to new plants