Predictamen de la Comisión de Salud y Población

In devising an estimation strategy for Hypotheses H4b to H12b (H4a has been excluded from consideration) it is useful to first briefly summarise some key points in the related empirical literature:

1. Estimator choice: all the related literature uses OLS, with Lederman et al. (2010) also accounting for selection and instrumental variable estimation. The choice for OLS in a cross-sectional setting follows from this literature.

2. Multicollinearity: how the independent variables of interest are assessed with this OLS framework depends on the degree to which multicollinearity is an issue. One study adopts a hierarchical addition approach to account for multicollinearity (Wilkinson et al., 2011) and another utilises separate regressions for the activities' budgets to account for this multicollinearity issue (Morisset et al., 2003). However, most studies regress all variables simultaneously (Lederman et al., 2010; Wilkinson & Brouthers, 2000; Peng & York, 2001; Martin, 2003; Kotabe, 1993). In this study, the choice of hierarchical versus separate versus simultaneous addition of the variables depends on the variance inflation factor. If this lies below the acceptable threshold of 10 (O’brien, 2007) the estimation does not suffer from multicollinearity.

3. Variable interactions: the empirical literature offers a guide to the assessment of the offices' activities within the context of the institutional environment. Wilkinson et al. (2011) do so by adding an interaction term between activities and entrepreneurial climate. Similarly, Harding and Javorcik (2011) interact their variable of interest with two sets of institutional environment variables to identify whether sector targeting is more effective when information availability is lower and transaction costs are higher. For the former they use proxies for linguistic and cultural distance, and for the latter they use indicators for the ease of doing business and the (in)effectiveness of bureaucracy. They interact both elements in separate regressions. Interactions terms are also utilised in Chapter 5 to identify the effect of the presence of foreign missions

160 under varying institutional circumstances, and in the survey data analysis they are relevant too for Hypotheses H11a to H12b.

When assessing which of the resource and activity variables work to increase exports, the analysis starts with a standard gravity model of trade which includes exports as the dependent variable, and typical gravity model variables as the independent variables. So, it follows that:

(7.1) ln(Xij) = β0 + β1ln(Distij) + β2ln(GDPcapi) + β3ln(GDPcapj) + β4ln(Popi) +

β5ln(Popj) + εij

Where i is the export flow's origin country, j is the export flow's destination country,

Xij stands for exports from i to j, Distij is the geographic distance between i and j,

GDPcap are the GDP per capita terms, Pop are the population size terms, and ε is the

error term. This model is the first to be estimated and utilises total exports and industry- adjusted total exports in two separate estimations to identify whether the general gravity model holds when utilising industries the offices operate in.

The effect of the offices' resources and activities on trade can be determined as follows: (7.2) ln(Xij) = β0 + βS[Surveyij] + βZZ + εij

Where Z is the vector of control variables in equation (7.1). The term [Surveyij] is the

set of survey variables from Table 7.1. To identify the consistency of the survey variables in the face of the standard gravity model variables, the estimation takes place twice: once without the Z vector, and once with. In addition, due to multicollinearity issues when adopting all relevant survey variables into one model (i.e. a Variance Inflation Factor (VIF) larger than 10) this procedure takes place separately for each of the four groups in the instrument design in Chapter 6. In doing so, the relational network aspects65 _{are excluded because they are relevant only when it comes to}

information asymmetry alleviation, which is discussed later. The survey variables consist of: office characteristics (ofsize, locrat, ageoff, oftype), the commercial diplomat's characteristics (proac, exp1, exp3, exp3, pubrat, lang1, lang2), activity characteristics (netwact, intelact, imagact, fdivact, bsupact, numind, smeo, smed), and structural network characteristics (nsizo, nsizd, precon).

65 _{Tie strength, network communication, and the three network composition variables that outline the} importance of a specific group of contacts (public officials, business contacts, and innovation contacts) in the office's network.

161 Estimation of the above equations proceeds in three steps. First is an analysis of the control variables' behaviour on total exports, and value of total exports – adjusted for the industries that an office operates in – to identify whether the industry-adjusted exports are consistent with the gravity model framework. This is done with equation (7.1). Second is estimating the effect of the survey variables on exports. This is done for both total exports and industry-adjusted exports as per Hypothesis H8a. In addition, to account for potential aggregation issues, the four groups of variables are estimated separately. This takes place with and without the control variables, and based on equation (7.2). The third step analyses whether the offices' activities and networks alleviate information asymmetry using interactions between the relevant commercial diplomacy office variables and the trade barrier variables.

To assess the interaction estimation required for Hypotheses H11a to H12b, a modification to equation (7.2) is required. The relationships of interest are between the information asymmetry variables and the relevant survey variables: activity performance, and the relational network variables. Determining this follows the analysis in Chapter 5, Wilkinson et al. (2011) and Harding and Javorcik (2011). The model includes interaction terms between relevant variables. The inclusion of interaction effects in the estimating equation is as follows:

(7.3) ln(Xij) = β0 + βSI([Surveyij] x [IA]) + βS[Surveyij] + βIA[IA] + βZZ + εij

Here, the term [Surveyij] x [IA] denotes an interaction between the survey element of

interest and the information asymmetry variable of interest. The variables within this term also enter separately into the equation and appear before the vector of control variables Z. Their inclusion is necessary for the correct interpretation of the interaction term (Jaccard, Wan, & Turrisi, 1990).

The four information asymmetry variables that interact are linguistic distance, religious distance, institutional quality, and the level of internet use in the host country. As in Harding and Javorcik (2011) and the economic literature in general (e.g. Moenius & Berkowitz, 2011; Borrmann et al., 2006), each interaction term requires a separate estimation of equation (7.3).

Based on the relevant hypotheses, the survey variables of interest are the five activity group variables (netwact, intelact, imagact, fdivact, bsupact), and the variables for the relational network characteristics (tiestr, netcom, tiepub, tiebus, tieinno). As such, the ten survey variables will each interact with four information asymmetry variables in

162 separate estimations to alleviate the multicollinearity concerns that interacted variables face. As a robustness check, additional estimations are done using alternative variables for institutional quality and internet use, and the survey variable instdist will be used as a robustness check.