Heterogeneity Effect of Corporate Financialization on Total Factor Productivity

Abstract

As corporate financialization becomes an important stylized fact, policymakers and economists are concerned that corporate financialization may harm firms’ productivity by reducing operation investment. However, no previous literature has assessed the heterogeneous effect of corporate financialization on total factor productivity induced by self-selection. In this study, we do so by leveraging the framework of marginal treatment effects (MTE) using samples of Chinese listed non-financial companies during the period from 2007 to 2018. We find that firms with high resistance to financialization can result in gains in productivity by choosing financialization. In contrast, firms with low resistance to financialization have a significant reduction in total factor productivity (TFP) when choosing financialization. Meanwhile, based on our counterfactual analysis, a stable fluctuation of housing growth is shown to have a significant impact on total factor productivity improvement, and the evidence is supportive of the TFP loss being mitigated by a stable housing market. Moreover, the mechanism analysis confirms a strong negative impact of financialization on innovation, which is the potential channel to depress aggregate total factor productivity. The detected heterogeneity effect highlights the importance of housing price in the self-selection of financialization choices to affect the TFP. Policymakers may wish to focus on corporate financialization, which is accompanied by inhibition in TFP growth, keeping the housing market stable in order to minimize the side effects of financialization on productivity.

1. Introduction

The most important stylized fact about the Chinese economy is the tendency toward financialization, which is referred to as a key feature of capital extension following the reversal of the real economy. Corporate financialization can be defined as a firm making substantial investments in financial assets [1]. Some empirical research provides evidence that corporate financialization can stimulate the short-term performance of firms and reach higher production efficiency yields [2,3], while over-financialization may hinder economic growth by extracting additional profits from the economy into the financial sector, thereby reducing production efficiency [4,5].

Several research studies [6,7] have empirically established that financialization has a negative effect on capital accumulation when resource and production input factors are established. The capital misallocation causes poor allocation of resources, leading to a negative effect on aggregate total factor productivity (TFP). Unlike these papers, recent works [8,9,10] highlight a positive effect. These papers suggest that non-financial companies could capture higher value through financialization during market booms. Both of these arguments are true to some extent. Previous literature states that corporate financialization is mainly driven by profit-seeking and risk aversion [11,12]. Financialization in high-yielding firms is more likely to capture the value of surplus liquidity and achieve optimal capital allocation [13], while financialization in low-yielding firms shows a tendency to pursue a higher return as its survival is threatened [14]. The extent to which TFP is affected by financialization depends on how corporate financialization decisions are made. More specifically, if motivated by speculation, firms will hold more financial assets considering the balance between risk and return. We can infer that the impact of corporate financialization on firms’ TFP will most likely differ across firms. The natural question then is: When does corporate financialization stimulate TFP, and under what conditions does financialization hinder the growth of TFP? Responses to such causal effects will most likely differ across firms. Thus, it is interesting to contrast potential heterogeneity in TFP due to a firm’s own financialization choice.

Corporate investment choices are not calculated at random, but rather are made intentionally by firms or managers based on their preferences [15]. Corporate investment behavior analyses have shown that the process of making investment decisions often includes subjective judgment of the investor’s and manager’s behaviors [16,17]. In many cases, corporate investment decisions face a trade-off between short-term profit maximization and long-term business strategy. It can be seen that there is a selection effect along with corporate financialization, which is influenced by both observable and unobservable factors. Existing literature has established that corporate financialization is associated with a statistically significant effect on aggregate productivity, explained by firm-specific determinants. Firm-specific determinants (size, profitability, growth, etc.) are the most important observable factors in choosing financialization. Measuring unobservable factors is not an easy task, and yet, little has been achieved regarding unobserved factors to assess the effect of corporate financialization on TFP.

In summary, the existing literature points to the impacts of corporate financialization on TFP. Nevertheless, to the best of our knowledge, no previous research has assessed the heterogeneous effect of corporate financialization on TFP induced by self-selection. In this study, we do so under the framework of marginal treatment effects (MTE) using samples of Chinese listed non-financial companies during the period from 2007 to 2018. Translated into our research question, the MTE is the effect of financialization on outcomes for firms at the margin of choosing financialization. The MTE framework also allows for depicting the marginal effects along with the probability of choosing financialization (which is the local average treatment effect). Accordingly, as an empirical matter, we find that the use of housing prices as the instrument affects financialization exactly as expected. In particular, we further conduct a counterfactual analysis to reconstruct house price fluctuations and corporate financialization decisions from a foresight perspective. We then assess the level of TFP according to a given counterfactual environment. We also try to substantiate our results by looking at the potential mechanism of the average effects from the view of innovation. We show that in companies with low financialization propensity, there is a positive effect of corporate financialization on TFP, whereas firms with high financialization propensity and low investment resistance experience a suppressive effect pf TFP.

Our paper makes several contributions. First, the effect of firm financialization behavior is treated homogeneously in existing studies, while unobservable characteristics of the firm are not considered. This paper quantifies the heterogeneous effect of corporate financialization on TFP and reveals some information on self-selection based on unobserved factors. We shed light on this surprising finding by estimating the distribution of treatment effects with respect to TFP. Along this line, this outcome complements existing evidence in identifying the economic consequences of corporate financialization by the potential heterogeneity of corporate in TFP due to a firm’s own financialization choice. Second, as an empirical matter, we find that the set of housing prices as the instrument affected financialization exactly as expected. Since the trend of housing prices is a key driver of corporate financialization decisions, understanding the effects of corporate financialization on productivity is particularly relevant in the housing boom. Comparing the effects of housing price changes on corporate financialization decisions in the counterfactual and factual cases, they are supportive of a stable fluctuation of housing growth that moderates the TFP loss, and steady fluctuations in housing growth are conducive to improved production efficiency. To the extent that productivity changes across firms’ financialization choices during the housing boom, we can use housing policy to mitigate firm-level TFP loss. In this respect, our work provides useful empirical input to a current policy issue.

The remainder of this paper is organized as follows: The literature review is in Section 2. Section 3 outlines the MTE conceptual framework of our analysis and describes the exogenous variation we exploit and presents the collected data. Section 4 reports the regression results and discusses policy counterfactuals. Section 5 is the part of the mechanism analysis, wherein we try to assess whether the change of innovation could explain our result. We offer some conclusions in Section 6.

2. Literature Review

A fundamental question with investment decisions is whether to invest, as well as the allocation of capital to the right investment projects, such as that modeled by Dobrowolski and Drozdowski [18] and Drozdowski [19]. These papers have contributed to our understanding of how the net present value (NPV) can be used in capital budgeting and investment planning. Some argue that the observed TFP reflects differences in the composition and quality of labor and other productive factors [20,21]. In particular, Männasoo et al. [22] show that human capital endowment positively affected TFP growth. The quality of human capital may depend on some unobservable factors such as personal education, training, new qualifications, and skills [23]. Capturing this sentiment, as the TFP is typically measured relative to the labor productivity, the input quality is determined by the capital intensity and average ability of the composition of the labor force, which can be difficult to quantify in practice. Others contend that investment periods vary across firms with information asymmetries. The “Holdup Losses” theory suggests that the manager prefers long-term projects to short-term projects. However, Narayanan [24] asserts that the manager selects short-term projects in an attempt to inflate the labor market’s perception of his value. Barnett et al. [25] suggest that firms tend to shift their sustainability strategy in long-term innovative ways that enable them to maximize firm value. Undoubtedly, both lines of explanation play a role, and it is important to disentangle their relative importance between short-term profit maximization and long-term business strategy.

Corporate financialization can be defined as a firm making substantial investments in financial assets [26]. To date, there has been valuable theoretical and empirical research on corporate financialization, and most of the studies tend to focus on the impact of corporate financialization. Two broad lines of corporate financialization consequences have been brought forward. Some empirical research provides evidence that corporate financialization can stimulate the short-term performance of firms [2,3], while in the long term, with a high level of financialization, corporate firms tend to pay less attention to the core business, resulting in lower physical investments and productivity loss. The literature has established that corporate financialization could crowd out operating cash flows that promote innovation [27,28]. Motivated by the conflicting empirical evidence on the subject, there is no clear theoretical consensus as to the most favorable forms of corporate investment, while a sizeable body of literature has engaged in documenting and analyzing the outcomes of financialization decisions and assessing whether the choice is an optimal investment.

The extant literature inspired by Cortty [29] calls for research that focuses on the observable causes of corporate financialization such as capital expenditures, liquidity, and financial constraints, whereas we are focused on unobservable factors changes in corporate investment behavior. We also highlight the choice of corporate financialization in affecting TFP. Undoubtedly, both capital surplus and managers’ preferences can exert a selection effect on financialization decisions. Furthermore, investment decisions primarily reflect the subjective judgment of the investor’s and manager’s behaviors [16,17]. Hence, we infer that unobservable factors may play a role in the effect of corporate financialization. The MTE framework allows us to assess its effects on firm efficiency to an extent not found in earlier studies.

Evidence on the effect of corporate financialization on the TFP is mixed, with effects ranging from negative to positive. Responses to such causal effects will most likely differ across firms and choice margins. Our goal is to better understand when firms benefit most from financialization and whether there exists treatment effect heterogeneity due to a firm’s choice probability of financialization. Much of the applied work continued to assume homogeneous treatment effects between corporate financialization and the TFP, which ignored the problem of endogeneity caused by self-selection into treatment based on unobserved characteristics. The MTE is informative about the nature of selection into treatment, which can clarify the connection of the switching regime self-selection. Moreover, the MTE allows heterogeneous treatment effects along the choice margin, which can reconcile the mixed empirical evidence on the relationship between corporate financialization and productivity.

3. Empirical Strategy and Research Design

Our estimation framework builds on Heckman and Vytlacil [30]. The model is stylized and simple, yet provides interesting insights based on treatment effect heterogeneity. The MTE method consists of two parts. In the first part, an effective exclusionary instrumental variable needs to be constructed to deal with the self-selection and heterogeneity characteristics of the sample. The second part is to estimate the marginal treatment effect by utilizing the estimated propensity score in the first step.

3.1. Framework of Analysis and Definition of Treatment Effects

MTE approach has two cornerstones: (i) A choice-theoretic structure based on the Roy model that defines each individual’s margin of indifference regarding treatment selection, and (ii) LIV estimation of MTE.

We assume that treatment is a binary variable denoted by D_i, D_i as the corporate behavioral decision is determined by various financial factors and operating conditions, such as operating conditions, industry profitability, and capital liquidity, and is also influenced by the firm’s unobservable characteristics, and is not a random choice.

Let Y_1,t and Y_0,t be the potential outcomes with and without treatment. For each sample corporate i, the effect on the TFP is related to the binary choice D_i. Specifically, the observed effects Y either equals Y_1,t, in the case that a firm received a treatment, which chooses financialization here (D_i = 1), or Y_0,t is the effect under the hypothetical scenario that the firm is not treated (D_i = 0). Consider the following discrete choice model for selection into treatment, which forms the basis for the MTE approach:

$$Y_{i,t}={\beta }_i{D}_i+{\alpha }_k{\displaystyle \sum _k{X}_{ik}}+U_i$$

(1)

$$D_i{}^{*}={\mu }_0(X_i,Z_i)-V_i$$

(2)

$D_i=1$ if $D_i{}^{*}\ge 0$, otherwise $D_i=0$.

The levels of potential outcomes of Y_i,t depend on firm characteristics, such as corporate size, profit, sales, etc., all contained in vector X_i,t, the impact of firm characteristics on the levels of potential outcomes may vary depending on treatment status, and the vector β_i captures treatment effect heterogeneity in terms of firm characteristics. To overcome the estimation bias caused by heterogeneity and selectivity bias, it is necessary to introduce the exclusive instrumental variable Z_i. Since Z_i is exogenous, it is uncorrelated with U_i and is, therefore, a valid instrument for treatment. U_i captures unobservable determinants of Y_i,t. It is likely that the level of potential outcomes will differ between treated and untreated firms since both U₁ and U₀ depend on a set of unobservable components. $D_i^{*}$ is the latent desire to choose financialization that depends on observed variables ${\mu }_0$($X_i,Z_i$) and unobservable $V_i$, where ${\mu }_0$($X_i,Z_i$) includes all variables in X_i plus the instrument Z_i.

U₁, U₀, $V_i$ are potentially correlated, inducing the endogeneity problem. It should be noted that the two potential outcomes, Y_1,t and Y_0,t, are never jointly observed for the same firm. To separately identify the right-hand side of Equation (1), unconditional independence is required: ($U_0$, $U_1,{ V}_i$) $\perp Z_i$, as we observe Y_i,t depending on the treatment status, the above model can be also represented as:

$$Y_{i,t}=(1-D_i)Y_{i,0}+D_i{Y}_{i,1}=Y_{i,0}+D_i(Y_{i,1}-Y_{i,0})$$

The MTE can be estimated from the model for the observed outcome:

$$Y_i={\mu }_0(X_i,Z_i)+D_i[{\mu }_1(X_i,Z_i)-{\mu }_0(X_i,Z_i)+U_{i,1}-U_{i,0}]+U_{i,0}$$

Given our models of potential outcomes and treatment, the treatment effect is:

$${\Delta }_i=Y_{i,1}-Y_{i,0}={\mu }_1(X_i,Z_i)-{\mu }_0(X_i,Z_i)+U_{i,1}-U_{i,0}$$

where ${\mathsf{\Delta }}_i={\beta }_i$, in which the coefficient on the treatment dummy varies across firms and is equal to:

$${\beta }_i=({\mu }_1-{\mu }_0)(X_i,Z_i)+U_{i,1}-U_{i,0}$$

where ${\mu }_1\left(X_i,Z_i\right)-{\mu }_0\left(X_i,Z_i\right)$ reflects treatment effects with given observed characteristics, and an idiosyncratic individual-specific gain $U_{i,1}-U_{i,0}$.

The average treatment effect (ATE) consists of both the average treatment effect on the treated (ATT) and the average treatment effect on the untreated (ATU). By averaging these parameters over the appropriate distribution of $=x$, they can also be defined conditionally:

$$ATE(x)=E({\Delta }_i|X_i=x)={\mu }_1(x)-{\mu }_0(x)$$

$$ATT(x)=E({\Delta }_i|X_i=x,D_j=1)={\mu }_1(x)-{\mu }_0(x)+E(U_{i,1}-U_{i,0})|X_i=x,D_j=1)$$

$$ATU(x)=E({\Delta }_i|X_i=x,D_j=0)={\mu }_1(x)-{\mu }_0(x)+E(U_{i,1}-U_{i,0}|X_i=x,D_j=0)$$

In this case, the common treatment parameters ATE, ATT, and LATE do not coincide. Following Heckman et al. [31], translated into our research question, the marginal treatment effect, MTE, is the effect of TFP indifferent to financialization and non-financialization, which is the next best alternative. P(Z) denotes the propensity score—a firm’s probability of choosing financialization, that is P(Z) = Pr (D = 1|Z). For this, the following transformation of the selection rule is:

$${\mu }_0(X_i,Z_i)-V_i\ge 0\Rightarrow {\mu }_0(X_i,Z_i)>V\Rightarrow \mathsf{\Theta }[{\mu }_0(X_i,Z_i)]\ge \mathsf{\Theta }(V_i)$$

More specifically, we specify the MTE as:

$$MTE(X=x,U_D={\mu }_D)=E(Y_1-Y_0|X=x,U_D={\mu }_D)$$

$$\begin{array}{l}MTE(x,{\mu }_D)=E(Y_1-Y_0|X=x,U_D={\mu }_D)\\ =\underset{\mathrm{observable}}{\underbrace{x({\beta }_1-{\beta }_0)}}+\underset{\mathrm{unobservable}}{\underbrace{E(U_1-U_0|U_D={\mu }_D)}}\end{array}$$

(3)

The MTE varies along the line of U_D in the case of heterogeneous treatment effects that arise if firms self-select into the treatment based on their expected idiosyncratic gains.

The MTE framework allows one to explore the policy implications of the treatment effects. Suppose D_i is the treatment choice under the baseline policy, and D_i’ is the treatment choice under the alternative policy. The policy-relevant treatment effect (PRTE) conditional on X_i = x is defined as follows:

$$\begin{array}{c}PRTE=\frac{E[{\mu }_1(X_i)-{\mu }_1(X_0)|D^{\prime }{}_i=1]E[D_i^{\prime }]-E[({\mu }_1(X_i)-{\mu }_1(X_0)|D_i=1]E[D_i]}{E[D^{\prime }{}_i]-E[D_i]}\\ +\frac{E[U_{1,i}-U_{0,i}|D^{\prime }{}_i=1]E[D^{\prime }{}_i]-E[U_{1,i}-U_{0,i}|D_i=1]E[D_i]}{E[D^{\prime }{}_i]-E[D_i]}\end{array}$$

(4)

The PRTE is the net effect of switching from a baseline policy to an alternative policy. Moreover, it represents the weighted difference between the ATT of the two scenarios. In this paper, the counterfactual refers to the simulated growth rates in housing prices during the sample period. Accordingly, PRTE measures how firms’ financialization decisions change when housing prices differ from reality and how the results impact the level of TFP.

3.2. Data

The data sources used in this paper include the China Stock Market and Accounting Research Database (CSMAR) and the China Entrepreneur Investment Club (CEIC), which provide detailed financial information and enterprise characteristics of listed companies, while the latter supplies the yearly city-level average housing prices of each city, as well as the macro environment indicators during the study period.

Our sample of non-financial listed companies covers the period from 2007 to 2018. We match companies’ data with city-level house prices based on the cities where the headquarters of listed companies are located. We exclude the sample of companies with missing main variables, special treatment (ST)/*ST firms, and select firms that have at least three consecutive observations for the dependent variable, which is also required for estimating the firm-level total factor productivity. We also drop all companies with permanent negative total assets, an asset–liability ratio greater than one, and a negative owner’s equity. Finally, we exclude observations in the upper and lower 1% of each variable’s distribution. For the empirical analysis, the sample size of 2013 non-financial firms effortlessly matched the annual urban housing prices of 237 cities.

Explanatory variable: The explanatory variable “financialization decision” (Fi) takes on the value 1 if a firm exceeds the average financialization level of its industry; otherwise, Fi = 0. Note that we assume that financial assets held by firms above the average level of the industry in a given year are considered to be financialization, while financial assets held below the industry level in a given year are considered to be in a reasonable range. Generally, corporate financialization is defined as increased financial activity in the proportion of nonfinancial corporations’ financial assets [14]. Accordingly, financialization is characterized by the expansion of financial assets relative to the entity activity of nonfinancial firms. This paper uses the ratio of financial assets to total assets as a measure of financialization. The following main categories of financial assets are identified: (1) Trading financial assets, (2) available-for-sale financial assets, (3) held-to-maturity investments, (4) investment properties, (5) derivative financial instruments, and (6) long-term equity investments.

Interpreted variable: Total Factor Productivity (TFP) is used as the outcome variable. We estimate firm-level TFP using the semi-parametric approach developed by [32].

Control variables: The control variables include the determinants of firms’ TFP to be consistent with the literature [2,3], and we control the enterprise characteristic variables and macro-level factors that may affect the TFP, including firm size (Size), profitability (Roa), growth (Growth), and financial leverage (Lev); macro-level factors include financial deepening (M2/GDP). Table 1 reports the variable definition and descriptions.

Instrument variable: This paper selects the average city-level housing prices as instruments. The reason for the instrument set is as follows: The housing market in China is extremely important since changes in housing prices can have considerable effects on firms’ financial constraints, which are often associated with investment capacity [33]. Moreover, the real estate boom will promote a large amount of capital in the form of derivative financial assets into the real estate industry, which promotes corporate financialization. We expect that the housing boom grants external finance that may affect financial flexibility, and consequently, firms may exhibit a willingness to choose financialization. This is based on the notion that the value of financial flexibility to a firm is associated with the importance of higher investment risk-taking behavior [34]. Hence, there is reason to believe that the external effects of the housing boom, which shock the firm, may affect corporate financialization.

We note that the housing boom may affect firm productivity through the channel of financial constraints [35]. To rule out the potential correlation between housing prices and the TFP, which may jeopardize the validity of our instrument, we also control this highly relevant factor (SA). We follow the method proposed by Hadlock and Pierce [36], who used firm size and age (denoted as Size and Age, respectively) to predict the effectiveness of financing constraints, and then proposed the excluding endogenous financing constraint variable indexes of SA. The calculation formula is as follows: SA = −0.737 × Size + 0.043 × Size2 − 0.04 × Age.

In the following table, the data descriptions are reported.

Table 1. Data description.

Variable	Definition
TFP	LP Method
Fi	Fi = 1, if corporate financialization greater than median of industry financialization, otherwise Fi = 0.
Size	Logarithm of total assets
Cashflow	Logarithm of net cash flows from operating activities
Roe	Net profit/total assets
Age	Current year—year of establishment of each company
Growth	Annual growth rate of operating income
Tng	Fixed assets/total assets
Lev	Total liabilities/total assets
Top	Number of shares held by the largest shareholder/total share capital
SA	SA = −0.737 × Size + 0.04 × Size2 − 0.04 × Age
Shibor	The shanghai interbank offered rate
M2/GDP	M2/GDP
R&D	R&D investment/total assets

Table 1. Data description.

Variable	Definition
TFP	LP Method
Fi	Fi = 1, if corporate financialization greater than median of industry financialization, otherwise Fi = 0.
Size	Logarithm of total assets
Cashflow	Logarithm of net cash flows from operating activities
Roe	Net profit/total assets
Age	Current year—year of establishment of each company
Growth	Annual growth rate of operating income
Tng	Fixed assets/total assets
Lev	Total liabilities/total assets
Top	Number of shares held by the largest shareholder/total share capital
SA	SA = −0.737 × Size + 0.04 × Size2 − 0.04 × Age
Shibor	The shanghai interbank offered rate
M2/GDP	M2/GDP
R&D	R&D investment/total assets

The descriptive statistics are presented in

Table 2. Descriptive statistics.

	Fi = 1		Fi = 0
Variable	Obs	Mean	Obs	Mean
TFP	7426	2.510	7500	2.562
Houseprice	7735	9.076	7966	8.932
Cashflow	8353	0.041	8606	0.038
Size	8353	22.14	8606	21.91
Lev	8353	0.443	8605	0.480
Roa	8353	0.029	8606	0.034
Age	8353	11.47	8608	9.134
Tng	8353	0.524	8605	0.472
Top	8353	33.78	8608	34.75
SA	8353	2.892	8606	2.758
Growth	8266	0.316	8349	0.348
Shibor	8353	0.040	8608	0.040
M2/GDP	8353	0.055	8608	0.055
R&D	5043	0.039	5661	0.040

. We report the mean value for the samples by binary choice. We classify a firm as financialization (un-financialization) if its degree of financialization is in the top (bottom) dichotomy of the ratio of its industry. The mean TFP in the group for Fi = 0 is slightly higher than that of Fi = 1. Firms with higher Cashflow, lower Roa, as well as lower Growth are more likely to choose financialization. Across other control variables, such as Size, Tng, and Top, there are no significant differences between financialization and un-financialization firms.

4. Empirical Results and Discussion

4.1. Instrumental Variable Estimates

The main objective of this paper is to examine the heterogeneous effect of financialization on TFP using the MTE approach. However, it is informative to first consider the standard estimation methods. We start with ordinary least squares (OLS) estimations as a benchmark, and

Table 3. OLS and IV estimates.

Variable	(1)	(2)	(3)
	TFP	Fi	TFP
Houseprice		0.139 ***
Fi		(4.46)
	0.012		−0.049
	(0.011)		(−0.32)
Cashflow	0.025	−0.012	0.059
	(0.050)	(−0.20)	(1.46)
Size	0.041 ***	−0.004	0.039 ***
	(0.014)	(−0.49)	(6.17)
Lev	−0.129 **	0.181 ***	−0.128 **
	(0.051)	(3.46)	(−2.75)
Roa	0.225 ***	−0.061	0.210 ***
	(0.082)	(−0.87)	(4.23)
Age	−0.001	−0.005	−0.0003
	(0.002)	(−0.07)	(−0.01)
Tng	0.043	−0.037	0.038
	(0.037)	(−0.88)	(1.26)
Top	0.000	−0.003 ***	0.0007
	(0.001)	(−4.22)	(1.19)
SA	−2.748 ***	−0.010	−2.071 ***
	(0.241)	(0.007)	(0.105)
Growth	−0.002	0.001	−0.005
	(0.007)	(0.007)	(0.005)
Shibor	−2.772	−6.801	1.359
	(9.288)	(−0.03)	(0.01)
M2/GDP	−2.892	−0.983	−2.095 *
	(1.943)	(−0.70)	(−2.18)
Constant	2.073 ***	-	-
	(0.321)
Cragg-Donald-F		19.86
IndustryFE	Yes	Yes	Yes
YearFE	Yes	Yes	Yes
CityFE	Yes	Yes	Yes
Observations	13,660	12,732	12,732
R-squared	0.012	0.007	0.007

Note: Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

column (1) gives the OLS estimates. As opposed to OLS, standard IV analysis estimates a causal effect without assuming equal potential outcomes for treated and untreated firms, and the results are presented in columns (2) and (3).

In column (1) the OLS coefficient on Fi is 0.012 and is not statistically significant. Due to self-selection, the possible problem with the OLS regression is that the unobservable characteristics that may affect corporate financialization are correlated with TFP. Column (2) presents the results for the first-stage estimates; the results are exactly as expected, and the coefficients of the instrument point in the expected direction and are highly significant. The coefficient is statistically insignificant in column (3), and it appears that financialization choice has no significant impact on TFP when the full sample is considered. However, this overall insignificance hides firm-specific effects, and this seems to be a plausible effect. It is due to the fact that the regression results using instrumental variables show the local average treatment effect (LATE). When there are strong heterogeneous treatment effects in the sample (some subsamples show positive LATE, others do not), the superposition of policy effects across samples can cancel each other out. This is an important property that the local average treatment effect, as identified by conventional two-stage least-squares methods, would miss.

Since the setting of instrumental variable regression identifies LATE for different subsamples, it may mask the inherent heterogeneous effect on TFP due to firms’ different propensities for financialization. The MTE approach utilized in this paper may overcome the shortage of the traditional IV method and is able the reveal the true heterogeneous effects.

4.2. Marginal Treatment Effects

Figure 1 shows the distribution of the propensity scores indicating whether firms choose financialization in estimating the MTE. For both groups, the propensity score varies from 0 to approximately 1. Furthermore, the propensity score has common support at almost every interval. In order to determine local effects, the variation in the propensity score where the effects of the X variables are integrated out is used.

We can see that the distributions of propensity scores between the two groups are significantly different under the current model setting. This implies that the current setting can better distinguish the differences in financialization choices between the two groups of firms and ensure the validity of the subsequent MTE estimation.

Table 4. Estimated treatment parameters for main results.

Variable	(1)	(2)	(3)	(4)
TFP	β0	β1 − β0	k	effect
Cashflow	−0.492 **	1.521 ***
	(0.247)	(0.453)
Roa	0.906 ***	−1.138 *
	(0.325)	(0.582)
Age	0.034 *	−0.102 ***
	(0.018)	(0.034)
Tng	0.106	2.060 ***
	(0.126)	(0.305)
mills			2.590 *
			(1.512)
ATE				0.239
				(0.381)
ATT				−1.620
				(1.232)
ATU				2.136 *
				(1.210)
LATE				1.798 **
				(0.814)
mprte1				0.289
				(0.367)
mprte2				0.265
				(0.380)
mprte3				0.718
				(0.455)
_cons	--	−9.607
	--	(15.330)
CityFE	Yes	Yes
Observations	12,483	12,483

Note: Standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01.

shows parametric estimates of the MTE by levels of Fi. The selection of unobserved gains can be measured by testing whether the slope of MTE is zero. Column (1) β₀ represents Fi = 0 (firms that do not choose financialization), and in column (2), β₁ − β₀ represent the difference between the binary choices. The coefficients show that corporate observable characteristics (such as Cashflow, Roa, Age, and Tng) play a significant role in determining firm TFP.

The latter two columns of estimates present whether the treatment effect of TFP varies with the choice of financialization. We calculate the average treatment effect (ATE), average treatment effect on the treated (ATT), average treatment effect on the untreated (ATU), and local average treatment effect (LATE). The estimation results in column (4) indicate that there is significant heterogeneity in the MTE estimates with the choice of the firm’s financialization. Specifically, the ATU coefficient is positively significant at the 10% level, indicating that there is a positive effect on TFP if non-financial firms choose financialization. The ATT estimated coefficient is insignificantly negative. Due to the superposition of the effects of those subsamples, the average treatment effect (ATE) does not pass the statistical significance test. This is not to say that corporate financialization has no meaningful effect on TFP. In fact, the local average treatment effect (LATE) is significantly positive at the 10% level. The evidence shows that, influenced by the exclusionary instrumental variable (housing price), the option change from Fi = 0 to Fi = 1 will have a positive impact on TFP. There are several possible explanations for this finding: A reasonable level of financialization may effectively promote technology upgrades and improve TFP [37], while excessive financialization is likely to lead to the misallocation of productivity factors, which ultimately affects the production efficiency.

Figure 2 shows the estimated MTE, and it provides an overview of the unobserved heterogeneity. The TFP associated with increases in the propensity score that result from the gradual shift in the instrument are informative of the treatment effects of each of the shifted types, and thus the marginal TFP increase at a given point identifies the MTE for each type.

The slope of the MTE curve reveals the selection pattern in unobserved characteristics, which suggests a reverse selection on gains in unobserved characteristics. Specifically, low-resistance firms (who are more likely, due to unobserved reasons, to participate in the treatment) have a negative treatment effect, and high-resistance firms have a positive treatment effect. Thus, it would seem firms with high resistance to financialization can gain productivity by choosing financialization. In contrast, firms with a low resistance to financialization that choose financialization show a significant reduction in TFP.

4.3. Counterfactual Analysis

MTEs can be utilized to simulate the effects of policy changes, and we will discuss the average effects of alternative policy reforms. By comparing the reaction of house price changes to shocks to firms’ financialization decisions in the counterfactual and benchmark cases, we investigate whether housing price regulatory policy may impact firm productivity. In China, the government retains far more control over future housing prices and impacts all participants in the housing market. Thus, it is interesting to analyze the effects of different counterfactual prices, and we focus our attention on housing policy interventions that can be taken to help corporate firms withstand the impact of falling TFP.

Two facts about the housing prices in China are largely uncontroversial: On one hand, urban housing prices diverge greatly across regions based on the level of economic development. According to government data, the housing prices in tier-one cities have skyrocketed in the past decade. While the boom is extreme in top cities, prices in second-and third-tier cities have remained sluggish or have even fallen. On the other hand, the major difference is the average growth rates in top cities when compared with those in second and third-tier cities. According to data from the National Bureau of Statistics (NBS), house price growth in tier-one cities has far outpaced that of the rest of the country. In the latest housing boom that started in late 2015, the average growth rate of housing prices in tier-one cities experienced more than a 40% increase on a year-over-year basis. So, we set the counterfactual scenarios based on a 40% increase/decrease in housing prices.

Therefore, in this paper, the counterfactual is set to the following three scenarios: In scenario 1, housing prices in first-tier cities decrease by 40% during the sample period; in scenario 2, housing prices in second-and third-tier cities decrease by 40%; in scenario 3, housing prices in first-tier cities decrease by 40% and those in second-and third-tier cities increase by 40%. We conclude by comparing the effects in our baseline scenario to those of counterfactual scenarios. Scenario 3 implies regional differences in housing prices have decreased, meaning the trend is in a suitable and reasonable range.

Table 5. Counterfactual results.

Counterfactuals	PRTE
housing prices of tier one decrease by 40%	−0.157 (0.374)
housing prices of tiers two & three decrease by 40%	0.453 (0.377)
housing prices of tier one decrease by 40%; housing prices of tiers two & three increase by 40%	0.695 * (0.392)

Note: Standard errors in parentheses. * p < 0.10.

shows the counterfactual estimation results. In this study, we adjust for the strength of house price growth to demonstrate the importance of smooth fluctuations in the housing market providing a strong effect on TFP growth. As shown by PRTE coefficients, the evidence shows that a relatively balanced growth in housing prices in second-and third-tier cities can neutralize the inhibitory effect of financialization on firms’ productivity and increase the productivity of firms generally. However, there is no significant improvement in TFP when observed house prices fluctuate to a greater degree.

These findings have important policy implications. The results reveal that in order to alleviate adverse effects on firms’ productivity, housing prices with relatively balanced geographic distributions could be expected to counteract speculative investment behavior. Only stabilizing the housing market can benefit the allocation of capital across firms for aggregate productivity. Overall, our findings are supportive of TFP loss being mitigated by a stable fluctuation of housing growth, which helps policymakers to eliminate the side effects.

Figure 3 displays the predicted effect of the third scenario, which compares the MTEs under the baseline model with models under the counterfactual scenario. We can clearly observe that the MTE curve for PRTE is slightly shifted upward compare with the benchmark MTE curve. Note that a fairly stable housing market somewhat moderates the TFP loss, and the rise in TFP strengthens.

4.4. Robust Test

To check the robustness of the MTE results, we choose a higher quantile (75%) of financialization to define an alternative measurement of the binary financialization choice (Fi). For the sake of brevity, we only show the estimated MTE curve. As shown in Figure 4, the distribution of marginal treatment effects is remarkably similar to the baseline case.

5. Mechanism

Given the heterogeneous effects of corporate financialization on TFP just documented, we focus our attention on the negative impact of financialization choice on TFP, which may be particularly pronounced for high-resistance firms, that is, firms that are drawn from the left side of the “financialization choice” distribution. We next use the MTE model to explore the potential channel through which corporate financialization causes TFP loss.

Innovation performance is one of the key determinants to promote firm productivity. From the empirical point of view, financialization may affect firms’ productivity through technological improvement [38]. The natural assumption is it takes time for R&D to result in new technologies that become implementable and can be used in production, which leads to the idea that R&D can be seen as a long-term choice. Moreover, recent research indicates that there is a strong negative correlation between financial asset allocations and firms’ innovations [4]. Similar to previous studies, we also try to substantiate our results by looking at the potential mechanism of the average effects from the view of innovation.

Table 6. IV estimates of R&D.

Variable	(1) Fi	(2) R&D
Houseprice	0.146 ***
	(0.037)
Fi		−0.049 **
		(−0.237)
Cragg-Donald-F	15.36
IndustryFE	Yes	Yes
YearFE	Yes	Yes
CityFE	Yes	Yes
Observations	9352	9352
R-squared	0.29	0.29

Note: Standard errors in parentheses. ** p < 0.05, *** p < 0.01.

column (1) reports the first-stage results of the 2SLS estimations. The coefficients of the instrument point in the expected direction and are highly significant. As to be expected, they barely change across the outcome variables. Column (2) reports a significant negative correlation between corporate financialization and R&D.

In particular, in

Table 7. Treatment effects parameters of R&D.

ATE	−0.017 (0.036)
ATT	−0.215 * (0.120)
ATU	0.164 (0.165)

Note: Standard errors in parentheses. * p < 0.10.

we calculate the average treatment effect (ATE), average treatment effect on the treated (ATT), and average treatment effect on the untreated (ATU). This large range of heterogeneity in the treatment effect due to unobserved characteristics would not be visible if looking only at aggregate treatment effects such as ATE. As we can see, the estimates of ATE are statistically insignificant. In contrast, the ATT, which indicates that firms with a low resistance to financialization make the choice of financialization, significantly decreases the R&D.

The estimated effects applied to R&D for each U_i on the MTE curve are shown in Figure 5, which is consistent with the MTE pattern of TFP. Given the available evidence, we find it quite likely that the reduction in innovation inputs for low-resistance firms is at least part of the story for TFP loss by choice of financialization.

6. Conclusions

This study contributes to the growing literature that focuses on the effects of corporate financialization. Particularly, we use the Marginal Treatment Effect framework introduced by Heckman and Vytlacil [30] to estimate the effect of corporate financialization on TFP. This method accounts for corporate self-selection under essential heterogeneity. To identify the relationships between financialization choice, housing prices, and the effect on TFP, we use an instrument that exploits exogenous variation in the housing boom. In addition to estimating the marginal effect of corporate financialization on TFP, we also assess the TFP according to a given counterfactual. In our next stage, we will analyze the potential mechanism of the average effects from the view of innovation.

The findings of our research lead to the following empirical and policy implications. First, our empirical evidence suggests that there is a substantial heterogeneous effect of corporate financialization on TFP. We note that firms with high resistance to financialization can increase productivity by choosing financialization. In contrast, firms with low resistance to financialization that choose financialization show a significant reduction in TFP.

Second, regarding corporate innovation, we find a negative impact on financialization and R&D, which is the potential channel to depress aggregate TFP.

Lastly, based on the results of counterfactual analysis, our findings are supportive of TFP loss being mitigated by a stable path of housing growth, which helps policymakers to eliminate the side effects. Policymakers should also keep a close eye on the housing market, considering the price adjustment policies useful in preventing adverse TFP loss triggered by excessive corporate financialization choices.

Further studies could extend our work by proving the causal relationship between corporate financialization in corporate risk-taking and housing market stability.