Balance of Payments Data

Balance of payments data set out to measure all capital transactions between the MNC parent and the foreign affiliate. Capital transactions are classified into three categories: (1) equity investment, i.e. the purchase of shares and other capital transfers such as the provision of machinery; (2) reinvested earning, i.e. the money earned by the foreign affiliate but not remitted to the parent MNC; and (3) other capital transactions, mainly inter-company debt flows.Footnote ¹³ The sum of these transactions is then netted against repatriated earnings and debt repayment back to the parent company. Repatriated earnings and reverse debt flow occasionally outweigh other capital transactions, resulting in negative FDI inflows. Two types of these data are typically reported and used in the literature: flow data and stock data. Flow data capture the net value of the above-described capital transactions that take place in a given year. Stock data capture the accumulation of flows over time in the host country. Much of these data are gathered by central banks as a byproduct of tracking the balance of payments. These methods capture equity investments, debt flow and repatriated income that cross borders between parent MNCs and their foreign affiliates. Reinvested earnings do not cross a border and data collected by central banks do not pick it up.Footnote ¹⁴ These data are gathered through surveys of MNCs and their affiliates and added to the data collected by the central bank.

To our knowledge, balance of payment data are the only data available for a large set of countries over long time-periods, leaving an analyst intent on employing data from a comprehensive number of source and/or host countries over multiple decades with little choice but to use them. There are, however, some problems associated with these data. Some of these problems are tangential to the issue of liquid versus illiquid assets that is the focus of this article, so we touch on them only briefly. We stress that this brief treatment is not motivated by a belief that these problems with BOP data are trivial. To the contrary, we believe that these problems alone strongly caution against the use of BOP data for testing theories of political risk.

The US Bureau of Economic Analysis Data

The above noted problems can be avoided by using data collected by the US Bureau of Economic Analysis (BEA) in their annual and benchmark (conducted every five years) surveys of American MNCs. These surveys provide as comprehensive a view of MNC activities as exists. Firms are required to report on the operations of the parent company and all overseas affiliates with sufficient assets.Footnote ²⁵ The firm-level data is kept confidential (and is therefore considered reliable), but several aggregates are publicly available. These data provide unique insights into the operations of foreign direct investors, though it does require a singular focus on the operations of affiliates of US based MNCs. The firms themselves report the data according to a single standard set by the BEA, which avoids the problems of non-random missing data on reinvested earnings or other non-random irregularities in reporting standards and methods. By focusing on the assets and activities of the affiliates, rather than on transactions between affiliate and parent, these data sidestep problems related to holding companies and local financing.

Importantly for our purposes, the fine-grained nature of the data allows us to observe and separate liquid and illiquid assets better. Figure 1 shows the aggregate balance sheet for foreign affiliates (excluding financial industry firms and holding companies) of US MNCs based on the 2004 benchmark survey.Footnote ²⁶ The striking thing about these data is that 43 per cent of US MNCs’ balance sheets are accounted for by current assets – i.e. highly liquid assets that are expected to be converted into cash within a year under normal operating conditions; 6 per cent of assets are held in cash; another 24 per cent of assets are held in receivables, which are highly liquid and can typically be converted into cash via factoring. ‘Other current assets’, a category which includes liquid marketable securities and pre-paid expenses makes up another 6 per cent; inventories make up another 7 per cent of assets, and are often, but not always, fairly liquid.

Fig. 1

Balance sheet of non-bank affiliates of US MNCs Source: US Bureau of Economic Analysis, 2004

Non-current assets – i.e. assets not intended to be liquidated in the coming year under normal business operations – make up 57 per cent of foreign affiliate assets; of these, 9 per cent are equity investments in other foreign affiliates owned by the parent company; 24 per cent are ‘other non-current assets’, a category which includes the value of intangible assets – such as brand names, trademarks or patents – and securities not intended to be liquidated within the year. Plant, Property and Equipment, net of depreciation and depletion (net PPE), which includes the value of physical structures, land, machinery, equipment and the book value of land, timber, mineral and similar rights owned by the foreign affiliate, make up 24 per cent of assets.

The lesson we draw from the data in Figure 1 is that a substantial portion of the assets held by the foreign affiliates of US MNCs – and certainly assets held in cash, receivables and investments in marketable securities – are at least as liquid as typical forms of portfolio investment. Their accumulation is unlikely to affect the level of political risk that foreign direct investors face. A very large portion of the average foreign affiliate's assets are irrelevant to questions of political risk if, as is in keeping with the logic of the obsolescing bargain, such risk is primarily a function of asset illiquidity.

At a minimum, the prevalence of relatively liquid capital on the balance sheets of foreign affiliates of US MNCs introduces some noise into estimations of the relationship between political risk and MNC behaviour. However, some projects’ assets are far more concentrated in fixed capital than others and this has the capacity to introduce bias. Figure 2 shows the same balance sheet data as Figure 1, but separates affiliates by several (broad) industrial aggregates. Net PPE constitutes over 50 per cent of assets in the mining and utilities industries, is closer to 20 per cent of assets in the information and manufacturing industries, and less than 10 per cent in services and wholesale trade. The combined value of cash, receivables and other non-current assets varies similarly. The gap between what is being measured in BOP data and illiquid capital will be largest in countries that are dominated by service sector and manufacturing investments and smallest in countries that are dominated by investments in resources extraction. Industrial makeup is in practice correlated with many of the same political institutions used to explain FDI allocation. This has the capacity to introduce bias in pooled estimates of political institutions’ effects on FDI, though country fixed effects are likely to diminish this bias to the extent that industrial composition stays relatively constant within countries over the duration of the observation period.

Fig. 2

Balance sheet of non-bank affiliates of US MNCs by industry Source: US Bureau of Economic Analysis, 2004. Note: Services refers to ‘Professional, scientific, and technical services’.

In sum, there are reasons to be wary of using BOP- based flow and stock data in analyses of FDI allocation, particularly in analyses that relate the deployment of illiquid capital to levels of political risk embodied in political institutions. Such analyses will unavoidably be riddled with excessive noise and possibly bias. The vast majority of work relating BITs to FDI relies on BOP-based flow and stock data. Whether and how BITs impact MNC behaviour therefore remains an open question that we address in the following section.

Research Design and Dependent Variables

We test BITs’ effect on MNC behaviour through a series of regression analyses relating different measures of US MNC activities to the presence or absence of a US BIT with the host-country. The measures of MNC activity that we use as dependent variables are: (1) The log of the value of total assets held by foreign affiliates of US MNCs (lnTotal Assets); (2) the log of the value of fixed capital held by US MNCs (lnPPE); (3) the log of the value on non-fixed capital assets held by US MNCs (lnnon-PPE); and (4) plant, property and equipment as a percentage of total assets (PPE/Total Assets).Footnote ²⁷

All of these data are measured at the country-year level and are taken from the publicly available aggregates of BEA's annual surveys of US MNCs. These data exclude investments in financial sector firms and holding companies. All measures are expressed in millions of constant US dollars. We expect that BITs will be positively and most clearly related to lnPPE. We also expect to see a positive relationship between BITs and PPE/Total Assets, indicating a higher fixed capital intensity where BITs are in place. We are agnostic about BITs’ effect on the other variables.Footnote ²⁸

Our estimating equation is as follows:

$$\ln \!FD{{{\rm{I}}}_{{\rm{it}}}}\,{\rm{ = }}\,\alpha \:{\rm{ + }}\:{{\beta }_{\rm{1}}}\,{\rm{\ast}}\,USBI{{T}_{it\,{\rm{ - }}\,{\rm{1}}}}\:{\rm{ + }}\:{{\beta }_{\rm{2}}}\,{\rm{\ast}}\,{{Z}_{it\,{\rm{ - }}\,{\rm{1}}}}\:{\rm{ + }}\:{{Y}_t}\:{\rm{ + }}\:{{V}_i}\,{\rm{ + }}\,{{{\epsilon}}_{it}}{\rm{,}}$$

where i indexes the host country, t indexes the year, USBIT_{it −1}indicates the existence of BIT with the United States at time t−1 and Z_{it −1} represents a vector of country-level control variables, Y_t represents year fixed effects and V_i represents country fixed effects. All models are estimated with Driscoll–Kraay standard errors to correct for autocorrelation, heteroscedasticity and correlations across panels.Footnote ²⁹ All results are robust to estimation with panel corrected standard errors.Footnote ³⁰

The Sample

Our sample includes every developing country-year observation for which data is publicly available from the BEA. We define developing countries as all non-OECD countries as of 1995, plus Mexico and Turkey.Footnote ³¹ This results in samples of between ninety-four and ninety-nine countries, depending on the dependent variable being used. A list of the countries included in the most expansive model is available in the Appendix. Our data includes observations from 1997 to 2007, which is determined by availability of BEA data.

All of the FDI figures that the BEA makes available contain a considerable amount of missing data. Some instances of missing data occur because the amount of money is so small (between −$500,000 and $500,000) that the BEA does not record it (but notes the reason for its being missing). In these cases, we replace the missing value with 0. In the remaining cases, we count missing data as ‘missing’.Footnote ³²

Our use of panel fixed-effects in a short sample means that the coefficient on USBIT reflects the impact of the nineteen investment treaties that were ratified by the United States over the ten-year window and are included in this sample. Relative to the larger universe of US BITs in force, our estimates are disproportionately reflective of data from Latin America and under-represent data from Africa. European BITs and Asian BITs are represented in our sample in rough proportion to their representation in the larger universe of US BITs. One result of this is that our estimates rely on a sample of BITs with countries that are, among other things, somewhat richer and more democratic than is true of the universe of countries with which the United States has BITs. To the extent that BITs work better (worse) in richer or more democratic countries, our estimates overstate (understate) their effectiveness. We do not see any obvious reasons why this sample should bias us towards or away from finding that BITs are more closely correlated with investments in fixed capital than other forms of FDI. Nonetheless, we readily point out the potential threat to external validity presented by this sample.

Independent Variables

Our measure of BITs is taken primarily from UNCTAD and is coded 1 if a country has a BIT in force with the United States in year t and 0 otherwise. We code countries that have free trade agreements in force with the United States that include an investment chapter as having a BIT.

We do not have any theoretical expectations over the time lag through which a BIT is likely to affect MNC behaviour. On one hand, fixed capital cannot be accumulated instantaneously, which suggests a lagged effect. On the other hand, US BITs cover pre-existing investments and, given the delay between the signing of a BIT, its ratification and entry into force, an impending BIT should be easy for firms to anticipate.Footnote ³³ There may be little incentive for firms to postpone planned investments in order to wait for the BIT to come into force officially, suggesting a lead effect. In practice we found that including our BIT variable at a one-year forward lag provides the best model fit (which we adjudicated by comparing AIC statistics), though specifications that measure BITs with the United States at a one-year lag go farther to mitigate reverse causality and are more in keeping with the extant literature. We report models using the traditional one-year lag of the BIT variable, the results of which are generally consistent with models estimated using alternative lag lengths.Footnote ³⁴

BITs are clearly not randomly assigned and are almost certainly more likely to arise between countries that are structurally predisposed to having FDI flowing from one country to the other, either by virtue of proximity, size, history, commercial relationships or some other feature. Our use of country fixed effects in combination with time varying controls such as GDP, GDP per capita and trade flows is likely to reduce the degree of this endogeneity substantially, particularly over a panel period as short as this. Some endogeneity may remain to the extent that the timing of BITs’ entry into force in a given country reflects lobbying by investors who were already planning on future investments. The BIT and the investment may show up at roughly the same time, but both are being caused by investors’ otherwise determined intentions. Measuring BITs at a one-year lag mitigates this endogeneity to some degree, but not entirely.

Ideally, we would address this endogeneity with an instrument that is correlated with BITs but otherwise unrelated to FDI. Such an instrument would have to vary within countries over time in order to be effective in our fixed effects specification. We do not have such an instrument at our disposal and, therefore, do not offer any methodological innovation to deal with endogeneity beyond what is already available in the extant literature. However, the basic goal of this empirical exercise – determining whether or not BITs are more relevant for investments in fixed capital than for other forms of investment – is not greatly undermined by this form of endogeneity. Even if FDI and BITs are in part jointly determined by investors’ future intentions, it remains to be seen whether the entangled relationship exists more strongly between BITs and investments in fixed capital than between BITs and investments in liquid capital or between BITs and total FDI. While we would naturally prefer a research design that precisely establishes the direction of causality, the critical point for our purposes is to establish that the correlation exists differently across different types of investment and our research design is capable of doing that.

Control Variables

Our models include standard country-level variables and largely mirror those used elsewhere in the literature. These are: the log of GDP, the log of GDP per capita, host country trade as a percentage of GDP, host country skill levels (which we proxy with the tertiary school enrolment share), the log of inflation, the presence of a preferential trade agreement (PTA) with the United States, the number of veto players (measured using Henisz's POLCON dataset) and the level of democracy (taken from the Polity IV dataset). Unless otherwise noted, all of the data is taken from the World Development Indicators. As noted, we include year and country fixed effects in all models. Descriptive statistics for all variables can be found in the appendix.