Import penetration and returns to tasks: recent evidence from the Peruvian labour market

Abstract

This paper provides original evidence on the impact of import penetration on wages of individuals performing manual/cognitive task-intensive jobs in the Peruvian labour market. Matching labour force surveys with task indicators from the us O*Net database and with information on industry- and occupation-specific import exposure, we build a continuous measure of manual intensity to uncover the heterogeneous effect of import penetration on workers’ wages. In order to tackle the endogeneity hampering the consistent estimation of our effects of interest, we combine an identification strategy based on heteroskedasticity with the traditional instrumental variable approach. We find that workers employed in highly cognitive/less manual-intensive jobs in the Peruvian manufacturing sectors are positively affected by industry-specific import penetration. This evidence is confirmed and magnified for the whole economy when the effects of occupation-specific import exposure are addressed.

Appendices

A Additional material for the data section

1.1 A.1 Additional figures and tables

See Table 8, Figs. 6 and 7.

Table 8 Correlation coefficients between selected individual variables: EPE, 2004–2009.

Fig. 6

Source: EY, 2014

Distribution of economic activities in Peru.

Fig. 7

Source: WITS-COMTRADE 2003–2009. Authors’ own elaborations

Peruvian imports—shares by market of origin.

1.2 A.2 Description of task indicators

Matchingo*netand EPE We collected information on tasks from the o*net, the Occupational Information Network of the us Department of Labour, which includes, among other information, an abilities survey providing data on 52 abilities required for each o*net-soc2010 occupation. As occupations from the Peruvian Labour Force Surveys are classified according to cno-95 at the three-digit level, it was necessary to find a suitable crosswalk between cno-95 and o*net-soc2010. Using readily available crosswalks, we converted the o*net-soc2010 occupation coding to the isco-88 occupation coding. However, since there was no table of correspondence available, we built a table of correspondence between the cno-95 and the isco-88 based on the description of occupations provided in the two classification schemes (see Table 9 below). This procedure allowed the o*net survey to be linked to the isco-88 coding and then to the Peruvian labour force survey.

Table 9 Table of correspondence between CNO-95 and ISCO-88

Definition of manual and cognitive tasks and comparison with task indicators from extant literature The manual task indicator takes after Peri and Sparber (2009) and, although it only rests on the subset of o*net abilities labelled as “Physical” and “Psychomotor”, it well describes the physical conditions required by the work context. Our proxy turns to be highly correlated (0.92) with the average intensity of the subset of o*net items labelled as “Physical Work Conditions”. Therefore, our definition of manual is exhaustive in terms of the relevant physical and psychomotor abilities required to perform a job. To capture the cognitive nature of a job, instead, we follow Bacolod and Blum (2010) and select a subset of thirteen items from the o*net “Cognitive” abilities section and calculate their average. The resulting proxy fully reflects the information included in the whole section, as its correlation with the mean value of all cognitive abilities included in the section is 0.98. Also, our proxy is highly correlated (0.92) with the mean value of items in the “Mental Work Activities” section; therefore, it wholly captures the extent of mental activities required by a job. Moreover, we compare our task measures with existing ones in the related literature. In particular, we consider the split by Autor and Dorn (2013) who build abstract, non-routine cognitive, routine cognitive, non-routine manual and routine manual task measures. It is worth mentioning that the comparison between our task measures and theirs is not straightforward, due to the fact that the authors hinge on the dot, while we use the o*net database. However, we have built from the o*net database task measures analogous to theirs, and we show corresponding pairwise correlations between them in Table 10. As shown in the table, Manual captures information on Autor and Dorn ’s manual dimension of Non-routine and Routine. Cognitive, instead, is very close to the analytical and interpersonal dimension of Autor and Dorn ’s Abstract measure. Worthy of note is the correlation between our manual and cognitive task indicators which is equal to $-\,0.52$. This suggests that the two measures give opposite, but not overlapping, indications on the nature of a job.

Table 10 Correlations between MS, manual, cognitive and other literature-based task measures.

Fig. 8

Source: Authors’ own elaboration based on results of Table 5 and 6. The marginal effect in a is computed using the ols estimation results reported in Column [2] of Table 5, and confidence bands are based on cluster-robust standard errors. The marginal effect in b is computed using the kv-iv estimation results reported in Column [6] of 6, and confidence bands are based on cluster-bootstrapped standard errors

Marginal effects MS, a industry-specific and b occupation-specific baseline models, epe 2004–2009.

Marginal effects of MS Figure 8 shows the marginal effects of MS according to the percentiles of industry- and occupation-specific import penetration measures. In the sector-level model a positive effect of the manual intensity is only recorded on wages of workers employed in sectors that are only modestly exposed to import penetration, that is Peruvian’s comparative advantage sectors. As a matter of fact, in our data the average import exposure of the Food Sector is 8%, of the Textile sector is 15%, of the Apparel sector is 6%, while in comparative disadvantage sectors like Machinery and Transport Equipment the import exposure is about 48% and 43%, respectively. Turning to the effect of manual intensity on wages according to the percentiles of the occupation-specific import penetration measure, the effect is negative and declining as the import exposure of an occupation increases. As an example, manual intensity starts exerting a negative effect on wages for blacksmiths and metal workers whose import exposure is about 15% and 18%, respectively. Its effect worsens for chemical-processing plant operators whose import penetration measure is about 27% and reaches the lowest value for fibre preparers and shoe-makers whose import penetration measures are about 50% and 58%, respectively.

B Complete baseline estimation results

See Table 11, 12 and 13.

Table 11 Wage equation—industry-specific IP baseline model, epe 2004–2009

Table 12 Wage equation—occupation-specific IP baseline model, epe 2004–2009

Table 13 Wage equation—first-stage estimation results: manual intensity (MS) and import penetration (IP), epe 2004–2009

C Robustness checks

1.1 C.1 Instrumental variables

The first part of this exercise elaborates on the choice of our instrumental variable for the identification of import penetration effects. In particular, our concern is on the potentially improper control for common product demand shocks between Peru and the remaining low- and middle-income economies outside lac, which may threaten the exogeneity of ${ Exp}^{US}_{lt}$. Following Autor et al. (2013), we adopt a gravity-based approach where we instrument import penetration in Peru with the inferred relative change of us comparative advantage and market access vis-à-vis that of Peru. In particular, a standard gravity model of trade (Feenstra 2004) sees exports from country A to country B as dependent on the export capability—e.g. sectoral productivity—of country A and on transport costs between the two economies. Therefore, we estimated a gravity model by regressing the log difference of exports from the usa and Peru to the world on industry and partner country dummies and took the residual which should measure changes in the relative export performance of usa to Peru due to changes in the former’s comparative advantages and relative market access. The gravity equation we estimate is:

$$\begin{aligned} lnExp^{US}_{jkt}-lnExp^{Peru}_{jkt}= \alpha _{j} + \alpha _{k} +\epsilon _{jkt}, \end{aligned}$$

(13)

where ${ Exp}^{US}_{jkt}$ and ${ Exp}^{Peru}_{jkt}$ are us and Peruvian sector j exports to country k and $\alpha _{j}$, respectively, and $\alpha _{k}$ are sector and country fixed-effects, respectively. The latter are meant to purge from any possible demand conditions in the foreign country k, while sector fixed-effects are meant to capture initial us comparative advantages relative to Peru. As a consequence, relying on a standard theoretical derivation of the gravity model (Feenstra 2004), residuals from the above equation are expected to capture differential time-varying comparative advantage and trade costs of the usa relative to Peru for sector j. Then, we estimate gravity model 13 for us and Peru 3-digit ISIC exports to the world over the period 2004–2009 and take the mean change in the residuals for industry j across destination markets k between year t and t − 1. When the change in residual is multiplied by Peruvian imports from the world in industry j in the pre-sample year 2003, we obtain the change in Peruvian imports predicted by Peruvian major partner’s changing comparative advantage and falling trade costs. The results of the estimated kv-iv model with the gravity-based instrument are presented in Columns [1] and [4] of Table 14. Notice that, even though the gravity-based instrument does not exhibit strong explanatory power in predicting import penetration, our main findings hold for both the industry- and occupation-specific models.

Table 14 Wage equation—robustness to alternative iv, epe 2004–2009

In the same line, we further investigate this issue by adopting an alternative instrumental variable approach based on transport costs. We retrieve us import flows from Peter Schott’s web page³⁵, and we calculate transport costs tc as the ratio of import charges—the costs of shipping a good from the source to the home market—over import value and we run the following regression:

$$\begin{aligned} tc_{us, g, k}= & {} \alpha + \beta _{1}\,\frac{import\ weight_{us,g,k}}{import\ value_{us, g,k}} + \beta _{2}\, distance_{us, k} \nonumber \\&+ \beta _{3}\, oil_{t} + \beta _{4} distance_{us, k}* oil_{t} \nonumber \\&+ \beta _{5} \frac{import\, weight_{us,g,k}}{import\, value_{us, g, k}}*distance_{us, k}* oil_{t} + \epsilon _{g,k,t} \end{aligned}$$

(14)

where $import\, weight_{us, g, k}$ measures the weight of us imports of HS2002 6-digit good g from country k, $distance_{us, k}$ measures geodesic distance between the usa and country k (available from CEPII), and $oil_{t}$ measures oil price at time t. Once we estimate this model on us imports, we take $\beta$ coefficient estimates and calculate the model predictions of transport costs on the bases of Peruvian right-hand side regressors based on Peruvian HS2002 6-digit import flows, quantities and source countries available from the CEPII BACI database. Then, we aggregate transport costs at the ISIC 3-digit level across goods and source countries weighting the latter by their share in the ISIC sector j in the pre-sample year 2003.

Corresponding results are presented in Columns [2] and [5] of Table 14. Notice that our main results still remain unaffected by the choice of an alternative instrumental variable, even though there is a statistically significant correlation between transport cost instrument and our sector-level import penetration measure.

These results provide two main takeaways. First, ${ Exp}^{US}_{lt}$ does not seem to carry residual correlation with common demand shocks between Peru and non-lac low- and middle-income countries, since the kv-iv strategy based on ${ Exp}^{US}_{lt}$ yields very similar results to those obtained with the gravity and transport costs-based instruments. Second, even though these instrumental variables may reveal to be rather weak, especially in predicting the occupation-level import penetration measure, the combination of the iv strategy with identification through heteroskedasticity provides stable and reliable results. This last finding is corroborated by the results in Columns [3] and [6] in Table 14, where we drop the instrumental variable and rely only on heteroskedasticity for identification.

1.2 C.2 Model misspecification

The second part of our robustness analysis looks into model misspecifications possibly arising from further uncontrolled heterogeneity in the sample and from the choice of the functional form of the wage regression equation.

With respect to the choice of our sample, we first look at the period between 2004 and 2008. Despite Peru’s exceptional growth during the whole period of analysis, the worldwide economic crisis has hit the country hard in 2009. We consider this feature in our baseline specification by including the interaction of sector or occupation fixed-effects with a dummy for 2009. Nevertheless, for robustness, we also limit our analysis to 2008 so that we eliminate any source of negative exogenous shock that may affect our findings. Results reported in Columns [1] and [5] of Table 15 confirm our main results.

Furthermore, we restrict our sample by considering only the first time each worker is interviewed. In this case our sample is a pure cross section consisting of individuals observed only once. As individuals appear up to four times in the 2004–2006 waves and up to two times in the 2007–2009 waves (see Sect. 4), we want to check that our inference is not inflated by repeated measurements and/or affected by a different survey design across the years. In Columns [2] and [6] of Table 15 we instead report the estimation results for the sample of cross sections. At least for the occupation-level import penetration measure, which is the specification that provides a significant effect of import penetration on wages according to the occupation manual intensity, the results confirm our main findings.

As for the choice of the functional form, we look into possible misspecification due to omitted quadratic terms. Balli and Sorensen (2013) argue that interaction terms are a partial specification of a multiple regression where explanatory variables represent the terms of a Taylor expansion approximating the functional relationship between the dependent and the independent variables. As such, the set of covariates should include at least quadratic terms. We show the result of this exercise in Columns [3] and [7] where we add the quadratic terms for IP and MS. We report only the results obtained by ols since managing endogeneity of the quadratic terms as well proved computationally very cumbersome in our context. We show that the statistical significance of our main variable of interest, i.e. the interaction term IP$\times$MS, is not the result of the omitted quadratic terms.

Last, we elaborate on our choice to deal with sample selection by means of a linear control function approach based on Heckman and Hotz (1989). The standard approach would be to embed Heckman ’s (1979) two-step strategy in the estimation framework laid out so far. The latter amounts to specifying an equation for the binary participation $s_i$, for $i=1, \ldots , n$ as

$$\begin{aligned} s_i = 1\left\{ w^*_i > 0 \right\} \qquad w^*_i = \varvec{S}_i' \varvec{\vartheta } + \zeta _i \quad \end{aligned}$$

(15)

where $1\{ \cdot \}$ is an indicator function and $s_i$ takes value 1 if worker i is employed and zero otherwise, $w_i^*$ is worker i’s reservation wage, and $\varvec{S}_i$ is a set of covariates including $X_i$ and possibly exclusion restrictions. Under the assumption that $\zeta _i \sim N(0,1)$, the error term $\varepsilon _{ilt}$ in Eq. (1) is assumed to have expected value conditional on participation

$$\begin{aligned} \mathrm {E}\left( \varepsilon _{ilt}| s_i = 1 \right) = \lambda \left( \varvec{S}_i' \varvec{\vartheta } \right) \end{aligned}$$

(16)

where $\lambda (\cdot )$ is the inverse Mills ratio $\varphi (\cdot )/\Phi (\cdot )$, with $\varphi (\cdot )$ and $\Phi (\cdot )$ being the standard normal density and distribution function, respectively. Under Eq. (15) and (16) estimation then follows a standard two-step approach.

As argued in Sect. 3.2.2, this sort of strategy poses an additional problem in this context. Standard errors for the wage equation have to be clustered at either the sector or the occupation–year level to correct for the possible misleading inference arising from repeated measurements of MS and IP across workers. When we reject the null hypotheses of exogeneity of MS and IP, standard errors have to be cluster-bootstrapped. In this context, the two-step approach would require us to resample clusters of sectors or occupations for the non-employed, which are obviously undefined.

In order to overcome this issue, we decided to rely on the assumption that workers’ employment participation follows the mechanism of selection on observables, that is the dependence between earnings and participating in the labour market is approximated by a set of observable covariates describing the family composition (see Sect. 3.2.2), that we include in $\varvec{X}_i$. This amounts to assuming that

$$\begin{aligned} \mathrm {E}\left( \varepsilon _{ilt}| s_i = 1 \right) = \mathrm {E}\left( \varepsilon _{ilt}| \varvec{X}_i \right) = \varvec{X}_i' \varvec{\beta } \end{aligned}$$

(17)

where we also specify the dependence between $\varepsilon _i$ and $\varvec{X}_i$ as linear. This approach was first put forward by Heckman and Hotz (1989) to deal with selection on observables. Notice that, in the context of our economic application, we must assume that neither ${ MS}_k$ nor ${ IP}_{lt}$, with $l=j$ or $l=k$, impact the participation decision.

Table 15 Wage equation—robustness to misspecification

In Columns [4] and [8] of Table 15, we compare our results with those based on Eq. (16). Worthy of note is that the results in the table are based on uncorrected standard errors. Our results are robust and prove that the coefficients of interest are not sensitive to our approach based on the linear control function estimator.

Table 16 Wage equation—robustness to omitted variables

1.3 C.3 Omitted variables

This section presents a sensitivity analysis of our main model to the potential omission of relevant confounding factors that might affect productivity levels and/or wages according to workers’ degree of manual intensity.

First, we account for the fact that the equivalent effects of import penetration and technological change (see Sect. 3.2.2) may vary with a job’s manual intensity. As technological improvements, in the form of more automated production processes, complement low manual-intensive workers but substitute high manual-intensive workers, our concern is that this unaccounted effect may influence our coefficient of main interest, ${ IP}_{lt}\times { MS}_k$, $l=j,k$. To this purpose, we add to our main specification the interaction term Mean wages$\times { MS}_k$ which should capture the heterogeneous effect of technology and informal labour on workers with different levels of manual intensity (see Columns [1] and [5] in Table 16).

As an alternative measure of technological progress, we add to our main set of covariates a measure of ict capital, at the sector or occupation level depending on the specification, to directly control for the evolution of technology in Peru. As previously stated, we unfortunately lack data on technology indicators for Peru and for this reason we decided to use information for the world technology leader, the usa. For this purpose, we retrieved industry-level data from the Bureau of Economic Analysis (bea) of the us Department of Commerce on ict capital equipment stocks.³⁶ We take the 2003 pre-sample value of $ICT_{{ Kap}}$ and multiply it by a trend to match to our data as a proxy of potential technological change in Peru. Taking the pre-sample value of $ICT_{{ Kap}}$, with a time variation guaranteed by an exogenous trend, avoids potential further endogeneity issues. Also, we allow $ICT_{{ Kap}}$ to have a different effect on wages according to the workers’ manual intensity by augmenting the model with the interaction term $ICT_{{ Kap}}\times$MS. Although the data are for the usa, the fact that ict technology has known a remarkably rapid diffusion across the world makes us confident that this variable is a fair proxy of the time evolution of technical progress in Peru. This robustness check is presented in Columns [2] and [6] of Table 16.

Furthermore, in order to account for potential wage changes induced by trade through channels other than import exposure, we add to our main set of covariates Exports calculated as the sector or occupation level, depending on the specification, change in Peru’s export share against the world as of 2003 and multiplied by a trend. As above, taking the pre-sample value of Exports, with a time variation guaranteed by an exogenous trend, avoids potential further endogeneity issues arising from the joint evolution of Exports and wages (Colantone and Crinó 2014). In our framework, Exports should capture any exogenous shock that may directly influence the competitiveness of the Peruvian economy and indirectly workers’ wages. We also add to our main set of covariates Exports and its interaction with ${ MS}_{k}$ to account for the possible moderating effect of the manual intensity on the impact of exports on wages. Columns [3] and [7] of Table 16 show that, even in this case, there are no significant changes to our main estimation results.

Fig. 9

Source: Authors’ own elaboration based on results of Table 17. The marginal effect in a is computed using the ols estimation results reported in Column [1], and confidence bands are based on cluster-robust standard errors; The marginal effect in b is computed using the kv-iv estimation results reported in Column [5], and confidence bands are based on cluster-bootstrapped standard errors

Marginal effects IP, a industry-specific and b occupation-specific baseline models, epe 2004–2009.

Table 17 Wage equation—alternative specification based on relative wages

Finally, a further identification issue comes from the widespread use of informal labour in the Peruvian economy (see Sect. 3.2.2). Import competition may affect an individual’s probability of switching from a formal to an informal job. Although informality is a persistent feature of the Peruvian economy and sector dummies could be enough to control for this, the incidence of informality may still heterogeneously vary through time across sectors and this can affect the identification of our import competition measure. In this robustness check we account for the share of informal workers by sector and occupation from the ENAHO survey. Notice that this information is only available for the years 2007–2009. Looking at the results in Columns [4] and [8] in Table 16, our results are confirmed even in the presence of this control which only turns positive and significant in the occupation-specific specification. The positive association between informality and wages could stem from lower wage earners being more likely to be informal or, in the presence of market segmentation between formal and informal workers, from the lower availability of formal workers in a sector driving up formal wages.

1.4 C.4 Alternative specification based on relative wages

This section presents a final sensitivity check based on an alternative definition of our dependent variable. As previously discussed (see Sect. 3.2.2), sector- or occupation-level proxies for technological change are unavailable for Peru. Hence, we have included the average wage by industry, occupation and year as an explanatory variable in $\varvec{X}_i$ in the wage equation. Since average wages are likely to be endogenous as well, the identification of import penetration effects is at risk if, beyond the inclusion of the baseline controls and industry-/occupation-specific fixed-effects, a residual correlation exists between mean wages and our IP variables. Therefore, in order to ascertain whether the potential endogeneity of mean wages affects our baseline evidence, we remove this variable from the set of regressors and redefine our dependent variable as the log ratio of individual wages over mean wages at the sector–occupation–year level. This allows us to retain the same framework while treating mean wages as endogenous. Corresponding results are shown in columns [1]–[2] and [4]–[5] of Table 17, and marginal effects are displayed in Fig. 9. Our baseline evidence is corroborated for both the sector- and occupation-specific models, thus confirming the high similarity of our baseline models with the empirical models for relative wages explored in this Appendix. Finally, in columns [3] and [6] of Table 17 we show that our findings are robust to the inclusion of the ict capital proxy used above in “Appendix C.3”.