# The Acquisition Decision – Estimation Strategy

III. The Acquisition Decision A. Estimation Strategy The first set of predictions arising from the model reveals which domestic firmsare likely to be the targets…

III. The Acquisition Decision A. Estimation Strategy

The first set of predictions arising from the model reveals which domestic firmsare likely to be the targets of foreign acquisitions. When there is a complementaritybetween a firm’s initial productivity level and the amount of innovation, foreignfirms acquire the most productive firms in the economy (those with higher φ i ), sothat there is positive selection. In an alternative scenario, in which foreign firmstransfer their own productivity level to the domestic firm regardless of which firmsthey buy, negative selection emerges: foreign firms acquire the least productive firms(those with lower φ i ).We estimate the type of selection at work in the data in the following way:equation(4) says that the share of the surplus generated by the acquisition goingto the acquiring firm is equal to K for all acquired firms, and the free entry conditionfor foreign firms in the acquisition market implies that acquisition takesplace whenever ( V i* F − V i* D ) ≥ K. Rearranging this inequality, we denoteF it = ( V i F − V i* D ) − K. The binary outcome of the acquisition decision Fi t canbe seen as reflecting a threshold rule for the underlying latent variable F i*t , so thatFi t = 1 (foreign ownership) if F *it ≥ 0, and Fi t = 0 (domestic ownership) if F i*t < 0.We also allow the average probability of acquisition to vary by year and industry byincluding year ( d t ) and industry ( d s ) dummies. Given these assumptions, the probabilitythat a given firm i in industry s is acquired in year t can be estimated usingthe following linear approximation:(5) Fi t = α + β φ i t−1 + dt + ds + ν it .We first measure the productivity of firm i, φ i0 , in the base year (the first year the firmappears in the data, which we subsequently exclude from the analysis) and examinethe probability in the data that a firm will ever be acquired (such that we use one observationper firm). We then allow for a time-varying measure of lagged productivity,φ it−1 , to examine the probability of being acquired in any given year, conditional onbeing domestically owned the year before. Empirically, lagged and initial productivityare highly positively correlated, but the ordering of firms based on lagged productivitymay better reflect the attractiveness of any one firm at the time of potential purchase.Under the assumptions of the model, β is predicted to be positive. In contrast,with negative selection, β is expected to be negative. Hence, the observed selectioneffect offers insight into the actual nature of the potential technology transfer frommultinational parents to domestic subsidiaries.B. Foreign Firms Select the Most Productive Domestic FirmsBefore turning to the analysis, we use our dataset to explore the patterns of selectiongraphically. Figure 2 plots the distribution of initial productivity (as measuredby ln sales) for two groups of firms: those that are acquired by a foreign firm fouryears after our baseline productivity is computed and those that remain domestic.One can clearly see that the distribution of acquired firms (solid line) lies to the right3606 THE AMERICAN ECONOMIC REVIEW December 2012of those that remain domestic.Since our measure of productivity is demeaned relativeto the industry, this does not reflect differences in firm size by industry. Figure3 reproduces Figure 2 by industry. A striking pattern emerges: Positive selection ispresent in every industry. These two figures provide prima facie evidence that thepositive selection predicted in our model dominates in the Spanish data.

productivity. These are the logarithm of real firm sales (columns 1 to 3) and thelogarithm of labor productivity (columns 4 to 6), each relative to its industry mean.The regressions in panel A use baseline (initial) productivity measured by thesetwo variables and one observation per firm to estimate the probability of ever beingforeignacquired (within the sample). Panel B uses (time-varying) lagged productivityas an independent variable to estimate the probability of being acquired in anygiven year, conditional on being domestic the year before. All regressions includeindustry dummies. Additionally, panel B includes year dummies and industry trends,so the results can be interpreted as within-industry differences in the probability ofacquisitionas a function of initial productivity, controlling for possible differentialtrends in acquisitions by industry.Regardless of the productivity measure used, we find that more productive firmsare more likely to become foreign owned. For example, the coefficient in column 1a(0.0351) implies that, conditional on being domestic the year the firm enters thesample, a one standard deviation increase in initial productivity makes a firm6 percentage points more likely to be acquired by the end of the sample. The sameincrease in lagged productivity is associated with a 1 percent higher yearly probabilityof being acquired (column 1b).23Columns 2 and 5 replace the productivity variable with indicator variables foreach productivity quartile. For example, in column 2a, being in the second salesquartile increases the probability of becoming foreign owned during the sampleyears by 3.3 percentage points relative to firms in the first quartile (correspondingto a yearly figure of 0.4 in column 2b); being in the third quartile by 4.8 percentagepoints (1.1 yearly, in column 2b); and being in the highest productivity quartileby as much as 14.8 percentage points (2.7 yearly, in column 2b). A similar patternemerges when using labor productivity as the productivity measure. Therefore,firms at the upper end of the productivity distribution are substantially more likely tobecome foreign owned, and the effect is increasing in firm productivity, with firmsin the upper quartile having a much higher probability of acquisition.Finally, columns 3 and 6 explore the possibility that foreign firms are selectingexporters (because, for example, exporting firms have less exchange rate exposure),and exporting is positively correlated with initial productivity. We introducea dummy variable for exporting status and interact it with initial productivity. Initialproductivity always remains positively related to the probability of being acquired,regardless of export status. There is also some evidence that multinationals aremore likely to target exporters, but we find no systematic evidence of differentialpositive selection among exporters. So, overall, even though some firms may beacquired because of their exporter status, positive selection persists, and multinationalsare more likely to acquire the most productive firms among both exportersand nonexporters.Table 2, therefore, reinforces the results from Figures 2 and 3 and shows that,within our cross-section of firms, the more productive domestic firms are morelikely to become foreign owned—evidence of positive selection and the presence of“cherry-picking.” These selection patterns are inconsistent with a model in which23 Table 1 shows that 3.5 percent of our observations are firms under foreign ownership. This corresponds to165 firms (or 4.6 percent) being acquired during the sample.3610 THE AMERICAN ECONOMIC REVIEW December 2012foreign firms always find it optimal to transfer their superior technology acrossinternational (or firm) borders to any domestic firm, as is often assumed.While the results in a number of papers point to the presence of positive selectionby foreign firms in other countries (e.g., for Chile, Ramondo 2009; for Indonesia,Arnold and Javorcik 2009; for the UK, Criscuolo and Martin 2009), to the bestof our knowledge, no prior studies have explained this empirical regularity. Whenviewed within the context of our model, our finding suggests that acquisition patternsreflect an underlying complementarity between the initial productivity of theacquired firm and the extent of innovation post-acquisition. As we will see later, thisfinding has significant implications for the relationship between multinational activityin a country and the evolution of the productivity distribution.IV. The Innovation DecisionA. Estimation Strategy: Fixed Effects and Propensity ScoreHaving established that foreign firms positively select domestic firms as targets,we now test the set of predictions relating productivity-enhancing investments toacquisition—namely, that upon being acquired, foreign subsidiaries invest more ininnovation than they would have done had they remained domestic. Our model suggeststhat acquired firms undertake more investment activity, controlling for the initialproductivity of the acquired firm. This can be seen in Figure 1 as the differencebetween λf and λ D .The optimal level of investment under each ownership structure is determined bythe first-order condition given in equation (2). In this case, innovation can increaseupon acquisition for several reasons. The foreign firm could provide access to alarger market and/or bring with it lower innovation costs, such that ( A i_ b i ) F > ( A i_ b i ) D .Our innovation variables are based on the firm-level responses to the questionsabout whether the firm made specific types of innovation in that year, which weinterpret as improvements to firm technology. We are interested in how the firm’stechnology, which is the result of successive innovations, changes with foreign ownership.Since, at any point in time, the firm’s technology can be characterized as thesum of innovations made up to that point, we use the yearly variables on firm-levelinnovation described in Section II to measure the firm’s technology at time t as:I it = Σ j= t 0 tI ij , where t 0 is the year the firm entered the data.24 Any differences intechnology across firms in the year they enter the data will be captured by the firmfixed effects in our empirical specifications.25 As a result, all the variation in a firm’sinnovative activity—and the resulting technology level—that we relate to changesin the firm’s ownership structure occurs within the sample.2624 We have allowed the stock of innovation to depreciate at different rates over time. The results are qualitativelysimilar to the ones presented with this—the simplest—specification.25 First differences specifications of the estimations with three different measures of the innovation stock (processinnovation; product innovation; and process innovation that includes both new machines and new organizationalpractices) are presented as a robustness test in Appendix Table A1.26 Online Appendix Table S3 shows that each measure of the stock of innovation I it , enters the production functionas a significant shifter of productivity.VOL . 102 NO. 7 guada lupe et al.: innovati on and foreign ownership 3611Empirically, we first estimate the effect of acquisition on technology using thepanel structure of the dataset and including year fixed effects as follows:(6) I it = α + γ Fi t−1 + Σ jβ j X it−2 j

d t + η i + ϵ it ,where I it is a proxy for productivity-enhancing innovation. The fact that the initiallevel of productivity affects investment directly for foreign-owned firms isabsorbed by the firm fixed effects, η i , along with any other permanent unobservedcharacteristicsof firms. Including firm fixed effects implies that the estimatedparameter γ is a measure of the change in investment after being acquired, controllingfor the fact that foreign firms choose to acquire higher initial-productivity firms,and this is predicted to be positive.The fixed effects specification controls for selection based on time-invariant firmcharacteristics (e.g., initial productivity). However, it is important in the context ofour 17-year panel to acknowledge that firm characteristics may evolve differently overtime (for reasons outside the model) and impact multinational selection decisions differentially.In particular, selection may be driven by lagged firm characteristics anddecisions that could be correlated with future innovation. To address this and ensurethat the estimates of the parameter γ reflect changes in innovation activity associatedwith acquisition, we use three different strategies. First, we include X it−2 j, a set of jfirm-level characteristics, lagged relative to the acquisition decision, that control forselection on time-varying observables.27 Second, we include an indicator in equation(6) for whether the firm is acquired in the current period (F i t ) and in the followingperiod ( Fi t+1 ). This allows us to test directly in the data whether there was a change inthe dependent variable that was already taking place prior to the acquisition (in whichcase, the coefficient on the lead variable should be different from zero).Third, we use a propensity score estimator to reweight firms in equation (6) toreflect differences in the probability of being acquired based on prior characteristics.We calculate the propensity score for each firm in the following way. For eachyear, we consider firms acquired in that year as treated observations and firms thatare never acquired as control observations. We pool treated and control observationsacross all years to estimate the probability that a firm is acquired as a function of anumber of characteristics (see Lechner 1999). This estimated probability is the propensityscore, p  . The characteristics used to obtain the propensity score are laggedproductivity (measured by both log firm sales and log labor productivity), lagged logsales growth, lagged export status, lagged average wage, lag of the process innovationstock, innovation activity in the previous year, lagged log capital per employee,lagged log capital, and a year trend. We also allow for this relationship to vary acrossindustries by estimating the propensity score separately for each industry.2827 The variables that may be correlated both with being acquired and with subsequent innovation activity that areincluded as controls are: (i) the log of the level of firm sales; (ii) the log of labor productivity (to control for timevaryingselection on firm size and productivity); (iii) the log of sales growth (to control for time-varying selectionon productivity growth); (iv) export status (to control for time-varying selection on the international presence ofthese firms and potentially related productivity effects not captured by other variables); (v) average wage (to controlfor potential selection on changes in the skill mix of firms); (vi) log capital per employee; and (vii) log capital (tocontrol for potential selection on the evolving level of capital and capital intensity of firms).28 We also performed the standard tests to check that the balancing hypothesis holds within each industry. Wefound that all covariates are balanced between treated and control observations for all blocks in all industries. The3612 THE AMERICAN ECONOMIC REVIEW December 2012One can transform the propensity score estimates into weights such that thepropensity score reweighted regression yields a consistent estimate of a parameterof interest (Dehejia and Wahba 1999; Busso, DiNardo, and McCrary 2009).Specifically, weighting each treated firm by 1/ p , and weighting each control firm by1/(1 − p ), provides an estimate of the Average Treatment Effect (ATE) of acquisitionon innovation in a specification like equation (6).29 We restrict the analysis to firmsthat fall within the common support. Busso, DiNardo, and McCrary (2009) showthat the finite sample properties of this propensity score reweighting estimator aresuperior to the propensity score matching techniques (where each treated firm ismatched to one or several controls).The propensity score reweighting estimator obtained by reweighting observationsin equation (6) allows us to control not only for selection into being acquiredon time-invariant characteristics of firms (just like the equal-weighted fixed effectsregression), but also for time-varying characteristics through the propensity score.The underlying assumption in the estimation is that, conditional on observable timevaryingand any time-invariant characteristics that affect selection, treatment is random.Hence, outcomes for treated firms are attributable only to treatment status (thisis typically referred to as the ignorability assumption, or selection on observables).B. Acquired Firms Undertake More InnovationSince we have detailed information on the types of innovation domestic firmsundertake upon foreign acquisition, our data allow us to shed light on the actualprocess of technology adoption by domestic firms, and on precisely what types ofinnovations are more likely to be adopted/transferred.Our measures of innovation are specific actions related to the implementation ofproduct and process innovation, as well as the assimilation of foreign technologies.All the columns in Table 3 reflect regressions of an innovation variable on the laggedforeign ownership variable. As we will see, we observe empirically that innovationstake place mainly with a one-year lag, reflecting the fact that it takes some time forfirm strategies to change after acquisition. Lagging this independent variable alsoreduces possible concerns about reverse causality.In Table 3, we report the results for each investment variable: process innovation(panel A), product innovation (panel B), and assimilation of foreign technologies(panel C). The first column in each panel includes only firm fixed effects; the secondalso includes industry-specific time trends; the third adds a large set of laggedcontrols (to control for possible differences in innovation related to previous firmcharacteristics); the fourth column also adds the lead and contemporaneous indicatorsof acquisition; the fifth column presents the propensity score reweighted estimates.30relationships between each of these variables and the probability of being acquired are shown in online AppendixTable S4. Lagged ln firm sales is the most significant predictor of acquisition, consistent with our model.29 Since never-acquired firms may be used as controls more than once, we sum the control weights by firmto obtain the weight for the control firm (Lechner 1999). We also winsorize the weights at 1 percent because ofextreme outliers in the weights. The results are not sensitive to the exact outlier cutoff.30 As we will see, the number of observations changes with the specification used, because of changes in the numberof nonmissing observations as we include more variables and their lags. Online Appendix Tables S6 and S7 repeatall the analysis that follows, restricting the sample to only the nonmissing observations of the most restrictive sample.The results are similar, so we chose to provide the estimates on the unrestricted sample in the main body of the article.