A Data construction and “HS 8-plus” level of product detail

Our main analysis relies on a merged dataset built by a firm-by-firm match of TIPS and Prowess data, excluding wholesalers for the sake of focusing on the trading behavior of production firms. TIPS data required considerable preparation for the merge, over and above simply aggregating its daily data to a fiscal year basis. A final merge with CEPII data brought destination market characteristics into the data set. The entire data construction process is described in detail in the appendix to Anderson et al. (2016).

As noted in the text we use an “HS 8-plus” product definition in the paper. Nearly all of the HS 8 tariff lines in TIPS have product definitions recorded by Indian Customs, and moreover those definitions vary within 8-digit lines. For example, a single HS 8 line might contain multiple types of textiles, and another might include multiple types of cooking spices, all with product definitions that define more precisely than the HS 8 code what was actually exported. Our challenge was to create usable information from these notes recorded by individual customs agents, often with different word orders or different spellings, across thousands of tariff lines over four years. We used two fuzzy-logic computer algorithms to identify common product definitions within each HS 8 line (Levenshtein distance and bigram comparisons), and added this information to the HS 8 codes to create the product definitions in our dataset.

Inside of an HS 8 or HS 8-plus code the quantity units can vary widely. The dependent variable in our empirical work is the export product price, defined as an export unit value and calculated as the relevant total value of exports divided by quantity. So, for instance, a firm's average price for selling a particular product to the United States in any given year would be the value of sales divided by, say, the metric tons sold. But in many of the single firm-product-destination categories, export values are reported in several different units, such as “buckles,” kilos, pounds, and boxes, the sum of which yields the total value of exports for that firm-product-destination observation. It is not possible to make meaningful unit value comparisons, or aggregations, across different units in these instances.

We therefore drop all observations that are measured in units that are not official units recognized by Indian Customs (see http://www.cybex.in/International-Trade-Resources/Unit_Quantity_Code.aspx). Further, we aggregate and “harmonize” the remaining values where there are well-established conversion factors for the units. Therefore in many instances we convert pounds to kilos, and tons to metric tons, and so on, prior to calculating unit values. However, there remain thousands of lines of data where the conversion factors are unknown, or for which the reporting of separate lines based on different quantity measures strongly suggests that there are in fact underlying differences between the goods reported in those lines (even when they are in the same 8-digit HS category). Accordingly, for the analysis reported here we keep only the top three units in each HS line, by value, and drop the others, a trim which costs approximately 2.5% of all observations.

The sigma indices of the quality of imported inputs are constructed from TIPS import data (aggregated and harmonized as described above). We calculate an annual z-score for the price a firm pays for each imported input (relative to the universe of prices paid for that good by all importing firms in TIPS), and then calculate an annual import value-weighted mean z-score across the firm's import bundle each year. The higher the sigma, the higher the relative price (and, by implication, the quality) of the firm's imported inputs. The sigma for capital goods imports uses the HS-based definition of capital goods developed in Bakht, Yunus, and Salimullah (2002).

The final data set contains 25,962 individual firm-product-destination-year observations over fiscal 2000–2003 for exporters including matched information on the prices these firms paid for imports (if any), drawn from 1,098 unique firms. Although by name alone we are able to match more firms than this, many observations are lost because they are not manufacturing firms (e. g., wholesalers), have incomplete information (e. g., missing input information in Prowess), or do not survive our procedures to clean the data.

We calculate TFP using the Stata implementation of the Levinsohn and Petrin (2003) technique, following Topalova and Khandelwal (2011) approach (pp. 998–999) to put each firm's productivity into index form (which itself depends on Aw, Chen, and Roberts 2001), which allows productivity comparisons within and between industries. We measure firm output with value-added (Topalova and Khandelwal 2011, use sales). Capital is measured as the size of each firm's gross fixed assets, and labor is proxied by the wage and salary bill (the number of employees is not included in Prowess). Note that this is the measure of labor used both in the TFP calculation and directly (in log form, “ln_labor”) on the right hand side of our regressions reported in Table 5 and Table 6 as our proxy for firm size; we also calculate the capital/labor ratio used in the regressions (“ln_klabor”) from these capital and labor variables.

We estimate TFP at the 4-digit National Industrial Classification (NIC) code level where possible, and at the 3-digit level when necessary due to a small number of firms at the 4-digit level (less than 20). We use Prowess data on firms’ spending on raw materials and electric power as the proxy for productivity shocks. All variables are expressed in real terms: output is deflated by two-digit industry-level wholesale prices indices from Ahsan (2013); capital expenditures are deflated by a capital goods wholesale price index we construct from several sub-industry wholesale price indices (including machine tools, electric machinery, and other capital goods); materials and power are likewise deflated with separate materials and power wholesale price indices we construct; and finally the wage and salary bill is deflated by the Economist Intelligence Unit's Indian labor cost index.

We calculate remoteness as in Harrigan, Ma, and Shlychkov (2015): the GDP-weighted distance of an export partner from all other export partners. So, for example, when we observe a transaction with the Philippines we sum the GDP-weighted distances between the Philippines and India's other export partners. Therefore Rd=∑Yodistod−1−1, where R_d is the remoteness of export partner d, Y_o is the GDP of country o, a member of the set of India's trading partners, and dist_od is the distance between d and a given country o.

B Controlling for Destination Market and Firm Characteristics

As in Harrigan, Ma, and Shlychkov (2015), we consider the possibility of selection bias because firm prices are only observed if firms choose to export to particular destinations, and we implement their three-stage estimator, itself an extension of Wooldridge (1995). The first stage is a Probit of entry (of a firm in a particular destination in a particular year) on all exogenous export-market characteristics (X_d), firm characteristics (X_f), and a year-specific intercept α. Omitting time subscripts we have:

where Y_fpd indicates firm f's revenues from exports of product p to destination d.  Equation (1) is estimated over an expanded sample of all possible firm-destination-year combinations; that is, it is applied to a “rectangularized” data set with zeros added. The inverse Mills ratio λˆfpd is then included in the second stage which explains observed (i. e., positive) firm-product-destination revenue based upon export-market and firm characteristics and product fixed effects (αp):

Quasi-residuals, formed as the actual residuals plus the estimated term for the inverse Mills ratio, ηˆfpd= uˆfpd+γλˆfpd, from this second stage are then entered as a selection control in the price regression:

This approach is more flexible than the two-step Tobit approach proposed by Wooldridge (1995) in that the estimated effects on entry, the δ’s in eq. (1), are allowed to differ from the effects on export intensity, the ζ’s in equation (2).^[22]