scispace - formally typeset
SciSpace - Your AI assistant to discover and understand research papers | Product Hunt

Journal ArticleDOI

25 years of European merger control

01 May 2021-International Journal of Industrial Organization (North-Holland)-Vol. 76, pp 102720

AbstractWe study the determinants of common European merger policy over its first 25 years, from 1990 to 2014. Using a novel dataset at the level of the relevant antitrust markets and containing all relevant merger cases notified to the European Commission, we evaluate how consistently arguments related to structural market parameters – dominance, rising concentration, barriers to entry, and foreclosure – were applied over time and across different geographic market definitions. On average, linear probability models overestimate the effects of structural indicators. Using non-parametric machine learning techniques, we find that dominance is positively correlated with competitive concerns, especially in markets with a substantial increase in post-merger concentration and in complex mergers. Yet, its importance decreased following the 2004 merger policy reform. Competitive concerns are also correlated with rising concentration, especially if entry barriers and foreclosure are of concern. The impact of these structural indicators in explaining competitive concerns is independent of the geographic market definition and does not change over time.

Summary (6 min read)

1 Introduction

  • Competition policy, that is, the design and enforcement of competition rules, is a cornerstone of the European Union (EU)’s program to enhance the European single market and foster growth.
  • The three overturned prohibitions by the Court of First Instance at the beginning of the 2000s marked the peak of this process.
  • Thus, instead of only looking at the determinants of a merger decision in the aggregate, the authors also investigate the factors that caused competitive concerns in specific sub-markets and how they have changed over time.
  • The authors find that the existence of barriers to entry, the increase of concentration measures and, in particular, the share of product markets with competitive concerns are positively associated with the likelihood of an intervention.
  • After this static investigation, the authors then study the dynamics of the impact of a number of key determinants over time.

2.1 Institutional Details

  • The European Communities Merger Regulation (ECMR) was passed in 1989 and came into force in September 1990.2.
  • Following this second investigation phase, the EC can again unconditionally clear the merger (phase-2 clearance), clear the merger subject to commitments by the merging parties (phase-2 remedy) or prohibit the merger (phase-2 prohibition).
  • Significant changes to European merger control were introduced in 2004 through an amendment to ECMR with the aim of bringing merger control closer to economic principles: the concept of an efficiency defense was introduced, a chief economist was appointed, the timetable for remedies was improved and horizontal merger guidelines were issued.
  • After the 2004 reform, the test used by the European Commission can be most accurately described as a significant impediment of effective competition (SIEC) test, which is more closely aligned with US practice (Bergman et al., 2007; Szücs, 2012).

2.2 Previous Literature

  • Mergers are studied extensively, with a large body of both theoretical and empirical literature on questions such as firms’ incentives to merge and merger policy effectiveness.
  • Post-reform, mergers between US firms, full mergers, and cross-border mergers, decrease the probability of intervention while conglomerate mergers are more likely to be challenged.
  • While the authors use control variables measuring relative market size and market concentration, both HHI as well as market size are based on European-wide industry sales data6 rather than on the market shares of merging parties and competitors as reported in the case documents.
  • He then estimates probit models at this concerned market level for horizontal overlap markets, interacting all explanatory variables with a post-reform indicator variable.
  • Thus, Mini (2018) is the only paper that studies the determinants of merger policy interventions at the relevant product and geographic market level based on the population of European merger decisions as the authors do.

3 Data and Descriptives

  • The data contain almost the entire population of DG Comp’s merger decisions, both in the dimension of time and with regard to the scope of the decisions encompassed.
  • The data set contains information on the name and country of the merging parties (acquirer and target), the date of the notification, the date of the decision9 and the type of decision eventually taken by DG Comp (clearance, remedy, and prohibition) or whether the proposing parties withdrew the notification.
  • The dataset further contains information on the nature of mergers.
  • If for example the market share range indicated is [0-10] percent, the authors record a market share of Table 3 shows summary statistics for the market share related variables.
  • The variable Post-merger HHI (low) is a lower bound of the post-merger HHI: it is calculated as the square of the merging parties’ joint market share plus the sum of squared market shares of competitors, whenever information on competitors’ market shares is available.

4 Linear probability model

  • The authors explore the association between merger characteristics and the intervention decision by DG Comp within a parametric approach.
  • The authors first replicate the results of the existing literature, which explain a competition authority’s decision as a function of merger characteristics at the merger level.
  • In contrast to previous studies, the authors explicitly estimate different models in various sub-samples to assess the issue of sample selection, which could arise because some important indicators – prominently market share and concentration measures – are only observable for ca. 60% of the mergers.
  • Second, as a merger often affects many different markets, while its characteristics and effects on competition can be heterogeneous across these affected markets, the authors investigate in a second step the correlation between merger characteristics and DG Comp’s intervention decision at the market level.
  • Lastly, in order to allow for heterogeneity in the correlation between merger characteristics and intervention decisions, the authors look at the evolution of these relationships over time.

4.1 Methodology

  • The authors employ a linear probability model to estimate the relationship between merger characteristics and the intervention decisions of DG Comp.12.
  • The authors define the indicator variable intervention to be equal to one if DG Comp prohibited the merger, cleared the merger subject to remedies in phase-1, cleared the merger subject to remedies in phase-2, or the merging parties withdrew the merger proposal in phase-2.
  • Thus, in a second step, the authors estimate the correlation between market and merger characteristics and DG Comp’s assessment at the level of the concerned product/geographic market.
  • In each of the year-specific OLS regressions, the authors include industry fixed effects.
  • 14We also run models where the authors use the level of the market shares rather than the dummy variable for high market shares.the authors.

4.2.1 Determinants of Intervention - Merger Level

  • The authors present four specifications run at both the merger and market levels.
  • Hence, this specification basically includes all mergers decided by DG Comp.
  • Hence, specifications 2 and 3 present the results for the same specification as 1 split into those cases without information on market shares (specification 2) and those with information on market shares (specification 3).
  • Neither merger characteristics (full mergers and joint ventures) nor the variables indicating alternative theories of harm (foreclosure concerns, vertical mergers, conglomerate mergers) significantly affect the Commission’s decisions.
  • Finally, in the sample including market share information (column 4), the indicator for a joint market share above 50% has no effect whereas the indicator pertaining to HHIs strongly and significantly increases the probability of challenge.

4.2.2 Determinants of Concern - Market Level

  • Table 6 contains the same sets of regressions at the concerned market level.
  • In general, more covariates appear to be significantly associated with competitive concerns at the market level than what is observed at the merger level.
  • While this might be a statistical results due to the larger number of observations in these regressions, it is likely that the aggregation to the merger level hides some of the EC’s more fine-grained considerations concerning specific markets.
  • In addition, the risk of foreclosure also has a positive and significant, though smaller, effect.
  • Market size now plays a more decisive role, with national markets increasing the probability of concerns in all specifications except (2).

4.2.3 Determinants of Concern - Market Level - Split Sample over Time

  • The authors explore the heterogeneity in the correlation between merger characteristics and competitive concerns by DG Comp over time by running separate OLS regressions splitting the market-level dataset over years (regrouping notification years 1990-1994).
  • The authors consider market share and concentration to be important determinants of merger decisions, thus these are included in the analysis.
  • As discussed in the previous section, while the estimated coefficients might differ across samples, the relevant determinants of intervention or competitive concerns are the same across the different subsamples.
  • Thus, in the last six years of the data, 2008 - 2013, high concentration was not a significant determinant of competitive concerns.
  • OLS regression results, as well as coefficient plots equivalent to the ones shown here.

5 Machine Learning/Causal Forests

  • In Section 4, the authors explore the association between concentration, market shares, entry barriers, and the risk of foreclosure with the intervention decision by DG Comp parametrically.
  • Causal forests are a flexible tool to uncover heterogeneous effects, in particular when there are many covariates and potentially complex interactions between them.
  • First, this approach allows a much better modelling of the process that leads to a particular decision by taking into account the specificities of each merger.
  • While the authors still should be careful to interpret the coefficient estimates in a causal way, the potential bias in the coefficient estimates should be reduced.
  • Therefore, some of their key concepts are measured by means of simple dichotomous dummy variables rather than more complex metrics.

5.1.1 Background on Heterogeneous Treatment Effects

  • This question relates to the literature on heterogeneous treatment effects, where one major problem is the fear that researchers might iteratively search for subgroups with high treatment effects and only report results for these subgroups.
  • The reported heterogeneity in treatment effects might then be purely spurious.
  • Yij is the outcome variable (binary in the present case) for market i in merger j, Wij is a binary treatment variable (i.e. their structural indicators), τ(Xij) is the effect of Wij on Yij at point Xij in covariate space, and eij is an error term that may be correlated with Wij.
  • In these instances, methods such as nearest-neighbor matching or other local methods allow for consistently estimating τ(x).

5.1.2 Estimation using Causal Forests

  • The authors use the causal forest algorithm by Athey et al. (2017) implemented in the generalized random forest (grf) package in R to investigate how the correlation between the treatment variables and DG Comp’s intervention decision varies with merger characteristics.
  • Causal forests are based on the random forest methodology by Breiman (2001).
  • The outcome Yij for observation ij is then predicted by identifying the leaf containing observation ij based on its characteristics Xij and setting the prediction to the mean outcome within that leaf.
  • Minimizing the expected mean squared error of predicted treatment effects (rather than the infeasible mean squared error), is shown to be equivalent to maximizing the variance of the predicted treatment effects across leaves with a penalty for within-leaf variance (variance of treatment and control group mean outcomes within leaves).
  • These are the same four indicator variables as those used in the previous regressions: high post-merger concentration, joint market share above 50%, barriers to entry, and risk of foreclosure.

5.2 Estimation Results

  • The authors present the results of the correlation analysis between the four main variables of interest and the competitive concerns by DG Comp using causal forests.
  • The authors set all the other covariates included in X to their mean respectively median sample value.
  • The authors then predict the treatment effects at the data points of this prediction dataset using the causal forest grown and plot the treatment effect along with the point-wise 95% confidence intervals.
  • The estimated conditional average treatment effect did not change much using these different node sizes.
  • 20Rather than taking the mean merger over the entire sample, the authors also created a prediction dataset based on the mean merger for which they have information on the market shares and concentration variables.

5.2.1 Treatment - High Concentration

  • Figure 6 shows the predicted correlation between the high concentration indicator variable and competitive concerns of DG Comp over time setting all other covariates to their mean (dark blue), respectively median (light blue), value.
  • The conditional average treatment effect predicted by the causal forest is 0.14, which is slightly higher than the coefficient on the high concentration indicator in specification 4 in Table 6.
  • This indicates that, once the authors use a richer model that better describes the process behind DG Comp’s decisions, the impact of this structural indicator is less volatile and much more consistent over time.
  • Nonetheless, the importance of concentration appears to follow a downward trend over the years.
  • For the predicted correlation setting all other covariates to median rather than mean values, the drop in correlation in 2001/2002 is even more pronounced and insignificant as of 2001.

5.2.2 Treatment - Joint Market Share above 50%

  • Figure 7 shows the predicted correlation between the indicator variable for merging parties’ market shares above 50% and competitive concerns of DG Comp over time, as before setting all other covariates to their mean (dark blue), respectively median (light blue), value.
  • While the predicted correlation is positive and significant up until 2010 (at least setting all other covariates to their mean), market shares seem to become a less important intervention decision criterion since the early 2000s and even become insignificant as of 2011.
  • Notice again that, as for concentration, the correlations estimated by means of the causal forest seem to be much less volatile and more consistent over time than those estimated based on the simple linear probability model.
  • Putting the developments of the correlation between concentration and market share measures with the intervention decision by DG Comp together highlights the shift away from evaluating mergers based on structural indicators towards a more economics based approach.
  • By much, the authors only report the predictions based on the mean merger over the entire sample.

5.2.3 Treatment - Barriers to Entry

  • Figure 8 shows the predicted correlation between the presence of entry barriers in the concerned market and competitive concerns of DG Comp over time, again setting all other covariates to their mean (dark blue), respectively median (light blue), value.
  • The conditional average treatment effect predicted by the causal forest is 0.46, which is higher than the coefficient on the entry barrier indicator in any specification in Table 6.
  • Furthermore, there is considerable heterogeneity in the predicted correlation between the existence of entry barriers and competitive concerns over time.
  • While the predicted correlation with concerns was essentially zero up to 1997, it becomes positive, significant, and of increasing importance since 1998.
  • This development is also in line with the shift of DG Comp’s merger policy toward a more economics based approach.

5.2.4 Treatment - Risk of Foreclosure

  • Lastly, Figure 9 shows the predicted correlation between the indicator variable for risk of foreclosure in the concerned market and competitive concerns of DG Comp over time, setting all other covariates to their mean (dark blue), respectively median (light blue), value.
  • The conditional average treatment effect predicted by the causal forest is 0.51, which is more than the double of the coefficient on the foreclosure indicator in the specifications in Table 6.
  • The confidence intervals for the predicted correlation are very wide, especially in the early years with fewer merger cases, and no clear pattern for the relationship between risk of foreclosure and competitive concerns emerges.
  • There is a positive and mostly significant correlation that, if anything, seems to become more important over time.

6 Conclusion

  • The authors study the time-dynamics of the EC’s merger decision procedure over the first 25 years of European merger control using a new dataset containing all merger cases with an official decision documented by DG Comp (more than 5000 individual decisions).
  • The authors find that the existence of barriers to entry, the increase of concentration measures and, in particular, the share of product markets with competitive concerns increase the likelihood of an intervention.
  • In order to obtain a more fine-grained picture of the decision determinants, the authors extend their analysis to the specific product and geographic markets concerned by a merger.
  • The parametric estimations are quite volatile and do not allow for uncovering clear patterns over time.
  • In particular, the authors find that concentration as well as the merging parties’ market shares have become less important decision determinants over time and are even insignificant in most recent years.

Did you find this useful? Give us your feedback

...read more

Content maybe subject to copyright    Report

Discussion
Papers
25 Years of European Merger
Control
Pauline Affeldt, Tomaso Duso and Florian Szücs
1797
Deutsches Institut für Wirtschaftsforschung 2019

Opinions expressed in this paper are those of the author(s) and do not necessarily reflect views of the institute.
IMPRESSUM
© DIW Berlin, 2019
DIW Be
rlin
German Institute for Economic Research
Mohrenstr. 58
10117 Berlin
Tel.
+49 (30) 897 89-0
Fax +49 (30) 897 89-200
http://www.diw.de
ISSN el
ectronic edition 1619-4535
Papers can be downloaded free of charge from the DIW Berlin website:
http://www.diw.de/discussionpapers
Discus
sion Papers of DIW Berlin are indexed in RePEc and SSRN:
http://ideas.repec.org/s/diw/diwwpp.html
http://www.ssrn.com/link/DIW-Berlin-German-Inst-Econ-Res.html

25 Years of European Merger Control
Pauline Affeldt
, Tomaso Duso
and Florian Szücs
§
April 2, 2019
Abstract
We study the evolution of the EC’s merger decision procedure over the first 25 years of European com-
petition policy. Using a novel dataset constructed at the level of the relevant markets and containing all
merger cases over the 1990-2014 period, we evaluate how consistently arguments related to structural market
parameters were applied over time. Using non-parametric machine learning techniques, we find that the
importance of market shares and concentration measures has declined while the importance of barriers to
entry and the risk of foreclosure has increased in the EC’s merger assessment following the 2004 merger
policy reform.
JEL Classification: K21; L40
Keywords: Merger policy; DG Competition; causal forests
We thank Ivan Mitkov, Fabian Braesemann, David Heine, Juri Simons and Isabel Stockton for their help with data collection.
Deutsches Institut für Wirtschaftsforschung (DIW Berlin) & Technische Universität (TU) Berlin
Deutsches Institut für Wirtschaftsforschung (DIW Berlin), Technische Universität (TU) Berlin, Centre for Economic and Policy Re-
search (CEPR) & CESifo
§
Wirtschaftsuniversität Wien
1

1 Introduction
Competition policy, that is, the design and enforcement of competition rules, is a cornerstone of the European
Union (EU)’s program to enhance the European single market and foster growth.
1
The European Commis-
sion’s (EC) Directorate General for Competition (DG Comp) ensures the application of EU competition rules
and retains jurisdiction over community-wide competition matters, representing the lead antitrust agency
in the European context. Competition policy covers several areas ranging from monitoring and blocking
anticompetitive agreements in particular hardcore cartels to abuses by dominant firms, to mergers and
acquisitions as well as to state aid. Among these areas of antitrust enforcement, merger control plays a pe-
culiar role. First, it is the only area where there is ex-ante enforcement. Second, it has important implications
for the other areas of antitrust: if anticompetitive mergers that reduce competition and strengthen the domi-
nant position of the merging firms are not prevented, it might make the ex-post control of abusive behaviors
more difficult. Finally, mergers are the area of antitrust where the largest consensus on best practices exists.
Therefore, among competition policy tools, it is an area that attracted much policy interest and economic
research.
The European Communities Merger Regulation (ECMR), the legal basis for common European merger
control, came into force in 1990. Over the course of the next 25 years, European merger control saw sig-
nificant changes. While in the early 1990s there were approximately 50 notified cases per year, the annual
workload increased significantly in the late 1990s and has averaged around 280 cases in the 2000s. DG
Comp’s enforcement activity reflects these changes. Procedurally, many novelties were implemented in the
2004 amendment to the ECMR: not only were new horizontal merger guidelines and the office of the chief
economist introduced, but also, more importantly, a new substantive test, the so called "significant imped-
iment of effective competition" (SIEC) test and an efficiency defense were introduced. These amendments
marked a substantial change in the legal basis for merger control enforcement in Europe. Yet, the pressure
for these changes began much earlier with the increasing belief that a mere form-based assessment of mergers
could often result in wrong decisions. The three overturned prohibitions by the Court of First Instance at the
beginning of the 2000s marked the peak of this process.
In this paper, we employ a new dataset containing all merger cases with an official decision documented
by DG Comp (more than 5000 individual decisions) to evaluate the time dynamics of the EC’s decision
procedures (see Affeldt et al. (2018)). Specifically, we assess how consistently different arguments related to
the so called structural market parameters market shares, concentration, likelihood of entry, and foreclosure
put forward to motivate a particular decision were applied over time. In order to obtain a more fine-grained
picture of the decision determinants, we extend our analysis to the specific relevant product and geographic
markets concerned by a merger. Thus, instead of only looking at the determinants of a merger decision in the
aggregate, we also investigate the factors that caused competitive concerns in specific sub-markets and how
they have changed over time. This step is particularly important because larger mergers typically affect many
different product markets in many different geographic regions. For example, the mergers in our data affect
an average of six markets. Therefore, by analyzing individual markets, thus conducting a more disaggregate
analysis, we better model the process that lead to a specific merger decision. Thus, the scope and depth of
1
Gutiérrez and Philippon (2018) claim that since the 1990s, European markets have become more competitive than their US counter-
parts because of the increased economic integration and the enactment of the European single market. They attribute a key role in this
process to the tough enforcement of competition policy rules.
2

our data allow us to go beyond the existing literature by i) not relying on a sample of decisions but instead
reporting patterns for the whole population of merger cases examined by DG Comp; and ii) allowing for
heterogeneity within merger cases by examining the individual product and geographic markets concerned.
In a first step, and in line with the existing literature, we start by estimating the probability of intervention
as a function of merger characteristics at the merger level. We find that the existence of barriers to entry, the
increase of concentration measures and, in particular, the share of product markets with competitive concerns
are positively associated with the likelihood of an intervention. This approach naturally extends to the level
of the individual markets: instead of estimating the overall probability of an intervention, we estimate the
likelihood that competitive concerns are found in that specific product/geographical market under consider-
ation. We find that, again, barriers to entry, but also the risk of foreclosure play a role. While tightly defined
(national) markets increase the probability of concerns, the number of active competitors decreases it. Struc-
tural indicators of market shares and concentration show the expected positive and significant correlation
with the likelihood of competitive concerns. After this static investigation, we then study the dynamics of the
impact of a number of key determinants over time. We find that the importance of ’structural’ indicators of
market power has declined over the years, though we observe a large volatility in the estimates over time.
In a second step, we bring well-developed non-parametric prediction methods to the analysis of com-
petition policy outcomes: supervised machine learning techniques. In particular, we implement the causal
forest algorithm proposed by Athey and Imbens (2016). This step allows a more flexible approach to model
the heterogeneity in merger control decisions. Specifically, the association between structural indicators and
the Commission’s decisions is made a function of all other covariates. Especially after the reform of 2004,
a so-called effects-based approach centered on a clearly stated theory of harm was made a cornerstone of
EU merger control. In such an approach, the reliance on structural parameters was expected to decrease,
leaving space for the use of counterfactual analysis where the interactions of different elements might play a
crucial role to substantiate the theory of harm. Using this model, we find that the importance of market share
and concentration measures has declined over time while the importance of barriers to entry and the risk of
foreclosure has increased in DG Comp’s decision making. Yet, the impact of structural indicators appears
to be much less volatile than in the simple linear probability model. Thus, the arguments put forward by
the EC to substantiate its decisions appear to be more consistently applied once the process underlying these
decisions is modelled in a flexible way.
The remainder of the study is structured as follows. In Section 2, we discuss the institutional details of
European merger control and review recent studies that empirically investigate the determinants of merger
intervention. In Section 3, we describe the data set used in estimation. We present the parametric model
as well as estimation results for the determinants of EC merger interventions in Section 4, while Section 5
presents the model and results for non-parametric estimation of heterogeneous correlations between merger
characteristics and intervention by the EC. We conclude in Section 6.
3

Citations
More filters

Journal ArticleDOI
TL;DR: Using the highly flexible, non-parametric random forest algorithm to predict DG Comp’s assessment of competitive concerns in markets affected by a merger, it is found that the predictive performance of the random forests is much better than the performance of simple linear models.
Abstract: I study the predictability of the EC’s merger decision procedure before and after the 2004 merger policy reform based on a dataset covering all affected markets of mergers with an official decision documented by DG Comp between 1990 and 2014. Using the highly flexible, non-parametric random forest algorithm to predict DG Comp’s assessment of competitive concerns in markets affected by a merger, I find that the predictive performance of the random forests is much better than the performance of simple linear models. In particular, the random forests do much better in predicting the rare event of competitive concerns. Secondly, postreform, DG Comp seems to base its assessment on a more complex interaction of merger and market characteristics than pre-reform. The highly flexible random forest algorithm is able to detect these potentially complex interactions and, therefore, still allows for high prediction precision.

2 citations


References
More filters

Journal ArticleDOI
01 Oct 2001
TL;DR: Internal estimates monitor error, strength, and correlation and these are used to show the response to increasing the number of features used in the forest, and are also applicable to regression.
Abstract: Random forests are a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest. The generalization error for forests converges a.s. to a limit as the number of trees in the forest becomes large. The generalization error of a forest of tree classifiers depends on the strength of the individual trees in the forest and the correlation between them. Using a random selection of features to split each node yields error rates that compare favorably to Adaboost (Y. Freund & R. Schapire, Machine Learning: Proceedings of the Thirteenth International conference, aaa, 148–156), but are more robust with respect to noise. Internal estimates monitor error, strength, and correlation and these are used to show the response to increasing the number of features used in the splitting. Internal estimates are also used to measure variable importance. These ideas are also applicable to regression.

58,232 citations


Journal ArticleDOI
Abstract: (2004). The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Journal of the American Statistical Association: Vol. 99, No. 466, pp. 567-567.

7,980 citations


Journal ArticleDOI
Abstract: A discussion of matching, randomization, random sampling, and other methods of controlling extraneous variation is presented. The objective is to specify the benefits of randomization in estimating causal effects of treatments. The basic conclusion is that randomization should be employed whenever possible but that the use of carefully controlled nonrandomized data to estimate causal effects is a reasonable and necessary procedure in many cases. Recent psychological and educational literature has included extensive criticism of the use of nonrandomized studies to estimate causal effects of treatments (e.g., Campbell & Erlebacher, 1970). The implication in much of this literature is that only properly randomized experiments can lead to useful estimates of causal effects. If taken as applying to all fields of study, this position is untenable. Since the extensive use of randomized experiments is limited to the last half century,8 and in fact is not used in much scientific investigation today,4 one is led to the conclusion that most scientific "truths" have been established without using randomized experiments. In addition, most of us successfully determine the causal effects of many of our everyday actions, even interpersonal behaviors, without the benefit of randomization. Even if the position that causal effects of treatments can only be well established from randomized experiments is taken as applying only to the social sciences in which

7,232 citations


Posted Content
TL;DR: This is the first set of results that allows any type of random forest, including classification and regression forests, to be used for provably valid statistical inference and is found to be substantially more powerful than classical methods based on nearest-neighbor matching.
Abstract: Many scientific and engineering challenges -- ranging from personalized medicine to customized marketing recommendations -- require an understanding of treatment effect heterogeneity. In this paper, we develop a non-parametric causal forest for estimating heterogeneous treatment effects that extends Breiman's widely used random forest algorithm. In the potential outcomes framework with unconfoundedness, we show that causal forests are pointwise consistent for the true treatment effect, and have an asymptotically Gaussian and centered sampling distribution. We also discuss a practical method for constructing asymptotic confidence intervals for the true treatment effect that are centered at the causal forest estimates. Our theoretical results rely on a generic Gaussian theory for a large family of random forest algorithms. To our knowledge, this is the first set of results that allows any type of random forest, including classification and regression forests, to be used for provably valid statistical inference. In experiments, we find causal forests to be substantially more powerful than classical methods based on nearest-neighbor matching, especially in the presence of irrelevant covariates.

816 citations


Journal ArticleDOI
Abstract: Many scientific and engineering challenges—ranging from personalized medicine to customized marketing recommendations—require an understanding of treatment effect heterogeneity. In this paper, we develop a non-parametric causal forest for estimating heterogeneous treatment effects that extends Breiman's widely used random forest algorithm. In the potential outcomes framework with unconfoundedness, we show that causal forests are pointwise consistent for the true treatment effect, and have an asymptotically Gaussian and centered sampling distribution. We also discuss a practical method for constructing asymptotic confidence intervals for the true treatment effect that are centered at the causal forest estimates. Our theoretical results rely on a generic Gaussian theory for a large family of random forest algorithms. To our knowledge, this is the first set of results that allows any type of random forest, including classification and regression forests, to be used for provably valid statistical infe...

789 citations


Frequently Asked Questions (2)
Q1. What are the contributions in this paper?

The authors study the evolution of the EC ’ s merger decision procedure over the first 25 years of European competition policy. Using non-parametric machine learning techniques, the authors find that the importance of market shares and concentration measures has declined while the importance of barriers to entry and the risk of foreclosure has increased in the EC ’ s merger assessment following the 2004 merger policy reform. 

In order to obtain a more fine-grained picture of the decision determinants, the authors extend their analysis to the specific product and geographic markets concerned by a merger. The authors find that more determinants significantly affect the Commission ’ s competitive concerns at the market level than seen at the merger level. In particular, the authors find that concentration as well as the merging parties ’ market shares have become less important decision determinants over time and are even insignificant in most recent years.