Asymmetric Cost Behavior and Dividend Policy

Abstract

A prevalent phenomenon documented recently is that costs are sticky, i.e., they fall less for sales decreases than they rise for equivalent sales increases. We examine the effect of this asymmetric cost behavior on a firm’s dividend policy. Given managers’ reluctance to cut dividends, we predict that firms with higher resource adjustment costs and stickier costs pay lower dividends than their peers. We find evidence consistent with this prediction. Further, using a regression discontinuity design that exploits variation in labor adjustment costs generated by close-call union elections, we provide evidence suggesting that the negative relation between cost stickiness and dividend payouts is driven by resource adjustment costs. Our paper sheds new light on the determinants of dividend policy and demonstrates the role of cost behavior in corporate decisions.

Keywords: Asymmetric cost behavior; cost stickiness; dividend payout; adjustment cost

JEL Classification: G31; G35; J51; M41

1   

1. Introduction

What determines a firm’s dividend policy? Since Lintner (1956) and Miller and

Modigliani (1961), financial economists have proposed a number of economic and behavioral

factors that determine a firm’s dividend policy. DeAngelo, DeAngelo, and Skinner (2008)

develop an asymmetric information framework that combines Myers and Majluf’s (1984)

security valuation costs with Jensen’s (1986) agency costs, and conclude that “reported earnings

are the key driver of firms’ payout policy” (p. 161). Because costs are a fundamental determinant

of earnings, cost behavior can have a first-order impact on a firm’s dividend policy. In this paper,

we examine whether an important feature of cost behavior, cost stickiness, affects a firm’s

dividend payout.

Asymmetric cost behavior, i.e., costs fall less for sales decreases than they rise for

equivalent sales increases, is a prevalent phenomenon documented recently (see Banker and

Byzalov (2014) for a review). Intuitively, this cost behavior stems from differential managerial

responses to sales changes: in the presence of resource adjustment costs, managers retain slack

resources and do not cut costs proportionally when sales decrease, but they tend to add required

resources and increase costs proportionally when sales increase. This cost model is based on the

two primitives of cost behavior (adjustment costs and managerial decisions), and is broader than

the traditional bifurcation of costs into fixed and variable components. Many resources are

neither predetermined (i.e., fixed) nor mechanically determined (i.e., variable). A case in point is

labor, a key ingredient of a firm’s production function. Firms, when adjusting the amount of

labor to be used, have to incur firing costs for existing employees and/or searching and training

costs for new employees. Such labor adjustment costs are substantial, but neither small enough to

2   

make labor costs variable nor large enough to make them fixed.  Anderson, Banker, and

Janakiraman (2003) refer to these types of resources as “sticky” resources.

The observed degree of cost stickiness reflects the joint effect of three major

determinants of a corporate manager’s cost management decisions (Banker, Byzalov, Fang, and

Liang, 2018), the economic nature of which has different implications for the firm’s dividend

policy. The first determinant is the magnitude of resource adjustment costs (e.g., Anderson,

Banker, and Janakiraman, 2003; Banker, Byzalov, and Chen, 2013). In the presence of higher

adjustment costs, firms are less willing to cut or expand resources. However, because firms

cannot fully meet the increased demand unless they add the needed resources, this effect of

adjustment costs is likely to be stronger for resource reduction than for resource expansion. Thus,

firms with higher adjustment costs are likely to exhibit greater cost stickiness. When sales

decrease, firms with higher adjustment costs cut fewer resources and suffer a bigger decline in

earnings than their peers. Since managers are generally reluctant to cut dividends (Lintner, 1956;

DeAngelo, DeAngelo, and Skinner, 2008), they would choose a lower dividend payout to start

with in the presence of higher resource adjustment costs. Hence, under the adjustment cost

explanation, the observed degree of cost stickiness would be negatively related to the level of

current dividend payout.

The second determinant of cost management decisions is managerial expectations for

future demand (e.g., Anderson, Banker, and Janakiraman, 2003; Banker, Byzalov, Ciftci, and

Mashruwala, 2014). Optimistic expectations about future demand lead managers to retain unused

resources when sales decrease because these resources are deemed to be needed in the future.

This proclivity to retain resources results in a greater degree of cost stickiness. To the contrary,

pessimistic expectations about future demand lead managers to aggressively cut unused

3   

resources when sales decrease because these resources are deemed to be redundant in the future.

This proclivity to cut resources results in a lower degree of cost stickiness. Hence, the

managerial expectation explanation of cost behavior suggests that the managers of firms with

stickier costs are more optimistic and expect a stronger earnings-generating ability than other

managers. They are thus likely to pay more dividends in the current period.

The third determinant of cost management decisions is managers’ self-serving incentives

and the associated agency problems. Managers have incentives to retain resources within their

firms to enjoy private benefits (Jensen, 1986). This empire-building behavior can result in both a

greater degree of cost stickiness (Anderson, Banker, and Janakiraman, 2003; Chen, Lu, and

Sougiannis, 2012) and a lower dividend payout. However, other incentives of managers, such as

their desire to meet earnings targets, might result in excessive cost cutting and a lower degree of

cost stickiness (Kama and Weiss, 2013). Thus, the relation between cost stickiness and dividend

payouts under the agency explanation is ambiguous.

The observed degree of cost stickiness reflects the joint effect of adjustment costs,

managerial expectations, and agency conflicts. Thus, the relation between cost stickiness and

dividend payouts depends on the relative strength of these different underlying economic forces

and is therefore unclear ex-ante. In addition, if a firm’s reported earnings are a sufficient statistic

for its cost behavior, cost stickiness would have little incremental effect on the firm’s dividend

payout after we control for the effect of reported earnings.

To understand the relation  between cost stickiness and dividend payouts, we start by

estimating an ordinary least squares (OLS) regression of a firm’s dividend payout (in year t) on

its degree of cost stickiness (based on 16 quarterly observations of sales and costs from year t–3

to year t) after controlling for common determinants of dividend payouts, such as firm size, the

4   

level and volatility of profitability, financial leverage, growth opportunities, cash holdings, and

asset tangibility, as well as industry and year fixed effects. Following Anderson, Banker, and

Janakiraman (2003), we measure cost stickiness as the degree of asymmetry in cost responses to

decreases versus increases in sales. The results show that firms with stickier costs exhibit a lower

dividend payout, consistent with the adjustment cost and agency explanations of cost behavior.1

These results also suggest that a firm’s reported earnings are not a sufficient statistic for its cost

behavior with respect to the firm’s dividend payout.

We conduct two subsample analyses to further understand our baseline findings. First, if

the negative relation between cost stickiness and dividend payouts is driven by resource

adjustment costs, dividend payouts of firms with ample financial resources would be less likely

to be affected by the asymmetry in their cost behavior. Consistent with this prediction, we find

that the negative relation between cost stickiness and dividend payouts is only present in firms

with relatively low (i.e., below median) cash holdings. Second, if managers’ empire-building

behavior drives the documented relation between cost stickiness and dividend payouts, this

relation would be dampened in firms with strong governance that constrains this managerial

misbehavior. However, the negative relation between cost stickiness and dividend payouts is

statistically significant only in the subsample of firms with strong corporate governance,

inconsistent with the agency explanation.

In summary, these results based on OLS regressions show a negative relation between

cost stickiness and dividend payouts and suggest resource adjustment costs as the major driver of

                                                            1 Our inferences remain unchanged when we examine different measures of dividend payouts and cost stickiness. We focus on a firm’s asymmetric cost behavior instead of its cost-sales sensitivity (i.e., operating leverage) because it is unclear why a firm’s dividend payout should be affected by its cost-sales sensitivity for sales increase. We verify in our data that a firm’s dividend payout is not correlated with its cost-sales sensitivity for sales increase. Further, our results continue to hold after we directly control for the cost-sales sensitivity for sales increase in our regressions.

5   

this relation. To provide further evidence that links adjustment costs, cost stickiness, and

dividend payouts, we use a regression discontinuity (RD) design that exploits variation in labor

adjustment costs generated by close-call union elections.2 We collect data on union elections

from the National Labor Relations Board (NLRB). In these secret-ballot elections, a simple

majority (50% plus one vote) in favor of unionization is required to win. While unionization is

clearly nonrandom, the RD approach exploits locally exogenous variation in the union status

around the 50% cutoff and thus the labor adjustment costs generated by elections that pass or fail

by a small margin of votes. This approach compares firms that barely pass union elections to

firms that barely fail the elections. As shown in prior literature (e.g., DiNardo and Lee, 2004;

Lee and Mas, 2012; Bradley, Kim, and Tian, 2017), for these close-call union elections, passing

is very close to an independent, random event, and is unlikely to be correlated with unobservable

firm characteristics. Therefore, we use the RD design to test the prediction based on the

adjustment cost explanation of cost behavior.

We first show that the degree of cost stickiness in the four years subsequent to the union

election is higher for firms that pass a union election than for firms that fail an election. This

result supports our use of labor power as a proxy for labor adjustment costs. Next, we find that

unionization has a negative effect on a firm’s dividend payout. Further, to tighten the link

between cost stickiness and dividend payouts in our RD design, we conduct a two-step mediation

analysis, and show that unionization has a significant, negative effect on a firm’s dividend

payout that is explained by its cost behavior. In addition, we find that the negative effect of

unionization on dividend payouts is concentrated in firms with high labor intensity or high ex-

                                                            2 Prior studies use different variables to proxy for unobservable adjustment costs. For example, Anderson, Banker, and Janakiraman (2003) use firm-level asset- and employee-intensity, and Banker, Byzalov, and Chen (2013) use country-level employment protection. In our RD analysis, we use firm-level labor power to proxy for unobservable adjustment costs.

6   

ante degree of cost stickiness. Together, these results provide evidence that firms pay out lower

dividends in the presence of higher resource adjustment costs and stickier costs.3

Our paper mainly makes contributions to two strands of literature. First, our results add a

new angle to the literature on asymmetric cost behavior (see Banker and Byzalov (2014) and

Banker, Byzalov, Fang, and Liang (2018) for thorough reviews and references). Prior research in

this growing area highlights that cost behavior can affect earnings quality, earnings prediction,

detection of earnings manipulation, and analysts’ earnings forecasts (e.g., Banker and Chen,

2006; Weiss, 2010; Kama and Weiss, 2013; Banker, Basu, Byzalov, and Chen, 2016). Our paper

links cost behavior to a firm’s dividend policy, demonstrating that this accounting research topic

could have important implications for research in corporate finance.

Second, our paper is related to the large literature on dividend policy (see Lease, John,

Kalay, Loewenstein, and Sarig (2000), DeAngelo, DeAngelo, and Skinner (2008), and Farre-

Mensa, Michaely, and Schmaltz (2014) for excellent reviews). This literature identifies a variety

of potential factors influencing a firm’s dividend policy, including corporate taxes, signaling

motives, agency considerations, compensation practices, and management incentives. 4 Brav,

Graham, Harvey, and Michaely (2005) provide survey evidence that financial executives do not

consider signaling, clientele effects, catering, or agency conflicts as explanations for dividend

levels but believe that the perceived stability of future earnings affects dividend policy. We

provide empirical evidence suggesting that a firm’s reported earnings are not a sufficient statistic

                                                            3 We do not claim that adjustment costs are the only or the most important channel through which unionization affects corporate policies. However, our results highlight the economic significance of this channel, which has been largely ignored in the prior literature on labor unions (e.g., Ruback and Zimmerman, 1984; DiNardo and Lee, 2004; Lee and Mas, 2012; Bradley, Kim, and Tian, 2017; Cheng, 2017; Campello, Gao, Qiu, and Zhang, 2018). 4 See, for example, Watts (1973), Black (1976), Gonedes (1978), Bhattacharya (1979), John and Williams (1985), Miller and Rock (1985), Healy and Palepu (1988), Smith and Watts (1992), DeAngelo, DeAngelo, and Skinner (1992, 1996, 2000, 2004), Fama and French (2001, 2004), DeAngelo, DeAngelo, and Stulz (2006), Skinner (2008), Skinner and Soltes (2011), Hanlon and Hoopes (2014), and Floyd, Li, and Skinner (2015).

7   

for its cost behavior and that cost stickiness is an important determinant of a firm’s dividend

payout.

The rest of our paper proceeds as follows. Section 2 describes the data, variable

measurement, and summary statistics of the variables used in our OLS analysis. Section 3

presents the OLS results. Section 4 discusses the methodology, data, and diagnostic tests for our

RD analysis. Section 5 reports the results of our RD analysis. Section 6 concludes.

2. Data, variable measurement, and descriptive statistics for the OLS analysis

We start by estimating an OLS regression of a firm’s dividend payout on its degree of

cost stickiness after controlling for common determinants of dividend policy. Our main data

source is Compustat. Following Noreen and Soderstrom (1997) and Anderson, Banker, and

Janakiraman (2003), we estimate firm i’s cost stickiness in year t by running the empirical model

below with its 16 previous quarters of data (year t–3 to year t):

ΔLnSG&A = β0 + β1×ΔLnSales + β2×Decrease×ΔLnSales + µ, (1)

where ΔLnSG&A is the log-change in quarterly selling, general, and administrative expenses (i.e.,

costs), ΔLnSales is the log-change in quarterly sales, and Decrease is a dummy variable that

equals one if ΔLnSales is less than zero, and zero otherwise.5 β1 (or β1 + β2) measures the

percentage change in costs for a 1% increase (or decrease) in sales. β2 captures the degree of

asymmetry in cost behavior with respect to changes in sales. We define our cost stickiness

                                                            5 This empirical model is used to estimate how costs respond to changes in sales regardless of the drivers of sales changes (e.g., macroeconomic conditions, industry trends, or seasonality). Using 16 quarters of data to estimate firm responses is not uncommon in prior research. For example, De Franco, Kothari and Verdi (2011) use such a model to estimate how a firm’s reported quarterly earnings respond to economic news (as proxied by the stock return) during the quarter.

8   

measure (CostStickiness) as the negative of β2, so that a larger value of CostStickiness indicates a

higher level of cost stickiness.

Our main measure of dividend payouts is Div/Inc, which is the total annual cash

dividends for common stock divided by the income before extraordinary items. In our OLS

analysis, we control for a wide spectrum of firm characteristics that have been shown by prior

literature to affect a firm’s dividend payout. First, we control for firm size, measured by the

natural logarithm of book assets (Size). Further, since profitable firms tend to pay more dividends

to reduce the agency costs of free cash flows (Jensen, 1986), we control for a firm’s profitability

(ROA). Similarly, as debt is an alternative way of disciplining managers (other than paying out

dividends) to mitigate the agency problem of free cash flows (Jensen, 1986; Berger, Ofek, and

Yermack, 1997), we also control for financial leverage (Leverage) in our analysis. In addition,

we include market-to-book ratio (M/B) in our regression analysis to control for a firm’s

investment opportunities. Some studies (e.g., Jagannathan, Stephens, and Weisbach, 2000) have

shown that greater uncertainty about a firm’s cash flows could increase the demand for

precautionary cash holdings and reduce the level of payouts. Therefore, we use the standard

deviation of quarterly ROA (over year t–3 to year t) as a proxy for such cash flow uncertainty

(ROAVol). Finally, we control for a firm’s cash holdings (Cash) and its asset tangibility

(Tangibility) because both could affect dividend payouts (e.g., Jagannathan, Stephens, and

Weisbach, 2000; John, Knyazeva, and Knyazeva, 2011). The definitions of all variables are

provided in the Appendix.6 We drop firm-years with negative income but positive dividends

because it is difficult to interpret dividend payout ratios in such cases (Chay and Suh 2009).

                                                            6 To reduce the influence of extreme values, we winsorize all explanatory variables at the 1% level. The top one percentile of Div/Inc equals 297% due to some firms having extremely small profits, and thus we winsorize this

9   

Table 1 presents descriptive statistics of cost stickiness, dividend payouts, as well as

control variables mentioned above. Our sample is the population of U.S. public firms from 1978

to 2016.7 The average cost stickiness is 0.053, with a standard deviation of 1.469. The average

dividend payout ratio is 13.88% with a standard deviation of 24.28%. Regarding control

variables, an average firm has book assets of $2.13 billion, ROA of 8.9%, leverage ratio of

24.3%, market to book ratio of 1.77, ROA volatility of 2.5%, cash holdings of 14.7%, and asset

tangibility of 30.4%.

3. Results of the OLS analysis

3.1 Baseline OLS results

In our baseline OLS regressions, the dependent variable is the dividend payout ratio

(Div/Inc) and the key independent variable is the cost stickiness measure that indicates the

degree of asymmetry in cost responses to decreases versus increases in sales (CostStickiness).

We control for common determinants of dividend payouts discussed in Section 2. We also

include industry fixed effects to control for unobservable time-invariant industry characteristics,8

and include year fixed effects to absorb intertemporal variation that could affect the relation

between the degree of cost stickiness and dividend payouts. Standard errors are corrected for

heteroskedasticity and cross-sectional and time-series correlations using a two-way cluster at the

firm and year levels.

                                                                                                                                                                                                variable at the 5% level. Our inferences remain unchanged when we use alternative measures of dividend payouts (winsorized at the 1% level) that do not suffer from this potential small-denominator problem (see Section 3.3). 7 We choose this sample period to be consistent with our regression discontinuity analysis (to be discussed in detail below), which uses union election data from 1977 to 2012 and analyzes dividend payouts over a four-year period after the election (i.e., 1978 to 2016). 8 We use the industry classification based on Fama-French 48 industries: http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/Data_Library/det_48_ind_port.html.

10   

Table 2 presents the baseline OLS results. Column 1 uses the full sample, and column 2

restricts the sample to firms that pay dividends at least once over the sample period. As shown in

both columns, there is a significantly negative relation between cost stickiness and dividend

payouts, consistent with the adjustment cost and agency explanations of cost behavior. Because

we control for a firm’s level of profitability (ROA) and its earnings volatility (ROAVol), the

driver for this result is not the relation between costs and earnings in the same period. For firms

with the same level and volatility of profitability, those with greater cost stickiness pay lower

dividends in the current period. This relation between cost stickiness and dividend payouts is not

subsumed by traditional economic variables that are used to explain a firm’s payout policy.9

3.2 Subsample analyses

We conduct two subsample analyses to further understand our baseline findings. First, for

firms with ample financial resources, their dividend payouts are less likely to be affected by

resource adjustment costs. Thus, the adjustment cost explanation of cost behavior predicts that

the negative relation between cost stickiness and dividend payouts is only present in the

subsample of firms with fewer financial resources (e.g., lower cash holdings).

To test this prediction, we split our sample into two equally-sized subsamples based on

the median cash holdings and perform our baseline OLS regressions in these subsamples. Table

3 reports the results, with column 1 for firms with higher-than-median cash holdings and column

2 for firms with lower-than-median cash holdings. Consistent with the adjustment cost prediction,

cost stickiness exhibits a significantly negative association with dividend payouts only in the

                                                            9 In untabulated analysis (see the Online Appendix), we include in the model the cost-sales sensitivity for sales increases (i.e., β1 in Equation (1)), and do not find evidence that this variable affects a firm’s dividend payout. More importantly, both the magnitude and statistical significance of the coefficient on CostStickiness remain largely unchanged.

11   

subsample of firms with lower cash holdings, whereas the relation between cost stickiness and

dividend payouts is statistically insignificant and economically small in the other subsample.10

Second, managers may prefer to reduce corporate payouts due to their private benefits of

control, which motivate them to keep corporate resources within the firm rather than distributing

these resources back to shareholders. This empire-building behavior can result in both sticky

costs (Anderson, Banker, and Janakiraman, 2003; Chen, Lu, and Sougiannis, 2012) and low

dividend payouts. If this agency explanation of cost behavior drives our results, the negative

relation between cost stickiness and dividend payouts would be more pronounced in firms with

weaker corporate governance.

To test this prediction, we split our sample into two equally-sized subsamples based on

the median institutional ownership and perform our baseline OLS regressions in these

subsamples. Higher institutional ownership indicates stronger shareholder monitoring and better

corporate governance.11 Columns 3 and 4 of Table 3 report the results. Opposite to the empire-

building prediction, cost stickiness exhibits a significant, negative relation with dividend payouts

only in the subsample with higher institutional ownership (which indicates better corporate

governance), whereas its relation with dividend payouts is statistically insignificant and

economically small in the other subsample (with poorer corporate governance and worse agency

problems).12 These findings suggest that our baseline OLS results are more likely to be driven by 

managerial responses to adjustment costs than by their empire-building incentives.

                                                            10 The p-value of the difference in coefficients between the subsamples is 0.147. 11 We use institutional ownership as a proxy for corporate governance rather than the Governance Index (G-Index) developed by Gompers, Ishii, and Metrick (2003) or the Entrenchment Index (E-index) developed by Bebchuk, Cohen, and Ferrell (2009) because these measures are only available for a selective group of firm-years. Hence, using these alternative corporate governance measures will substantially limit the sample size and reduce the power of our tests. 12 The p-value of the difference in coefficients between the subsamples is 0.062.

12   

3.3 Robustness checks

We perform a battery of additional analyses to ascertain the robustness of our baseline

results. First, we use four alternative measures of dividend payouts as the dependent variable in

our baseline regression, including the dividend to sales ratio (Div/Sales), the dividend to assets

ratio (Div/Asset), the dividend to market value ratio (Div/MV), and the natural logarithm of one

plus the dollar value of cash dividends paid (LnDiv). These alternative measures allow us to

include all loss firms in our sample and are not subject to the small-denominator problem facing

the dividend payout ratio (Div/Inc). As shown in Table 4, cost stickiness is negatively correlated

with all these alternative measures of dividend payouts.

Second, we examine alternative measures of cost stickiness. While our main analysis

defines a firm’s cost stickiness based on selling, general, and administrative expenses (following

Anderson, Banker, and Janakiraman, 2003), we construct the cost stickiness measure using two

other common definitions of costs such as operating costs (Kama and Weiss, 2013) and total

costs (Rouxelin, Wongsunwai, and Yehuda, 2018). Table 5 shows that our baseline OLS results

continue to hold when we use these alternative measures of cost stickiness.

Third, prior research suggests that firms may substitute dividends with stock repurchases

in response to transitory earnings shocks (e.g., Guay and Harford, 2000; Skinner, 2008; Floyd,

Li, and Skinner, 2015). Hence, we repeat our baseline analysis using total payouts (i.e., the sum

of annual cash dividends and stock repurchases) as the dependent variable.13 In untabulated

analysis (contained in the Online Appendix), we find that firms with stickier costs have a

significantly lower total payout (including both dividends and share repurchases), suggesting that

asymmetric cost behavior affects the total amount, instead of just the form, of a firm’s payout.

                                                            13 We follow Grullon and Michaely (2002) to define repurchases as the total expenditures on the purchase of common and preferred stocks minus any reduction in the redemption value of the net preferred stocks outstanding.

13   

In summary, the above results based on OLS regressions provide evidence suggesting

that the negative relation between cost stickiness and dividend payouts is driven by resource

adjustment costs. To provide further evidence that links adjustment costs, cost stickiness, and

dividend payouts, we use a regression discontinuity (RD) design that exploits variation in labor

adjustment costs generated by close-call union elections.

4. Methodology, data, and diagnostic tests for the regression discontinuity (RD) analysis

4.1 Methodology of the RD analysis

A key determinant of the degree of cost stickiness is the magnitude of resource

adjustment costs (Banker and Byzalov, 2014). In the analysis below, we use labor power as a

proxy for a firm’s resource adjustment costs. Organized labor, such as unions, makes wages

stickier and layoffs more costly, thereby increasing the adjustment cost of a firm’s labor stock.

Furthermore, unions frequently intervene in a firm’s restructuring activities, such as blocking

plant closures to save workers’ jobs, which makes it harder for the firm to adjust its physical

capital.14 As a result, labor unions increase a firm’s resource adjustment costs and its degree of

cost stickiness. In response to stickier costs induced by greater labor power, firms would set a

lower dividend payout.

To identify the effect of a firm’s labor power on its cost stickiness (and in turn its

dividend policy), we adopt a RD design, in which unionization status to our sample firms is

                                                            14 Firms may respond to union activism by hiring temporary workers or closing stores. However, these actions often encounter strong resistance from labor unions. For example, Walmart reopened five U.S. stores whose closure was claimed by workers as a retaliation against them for organizing (Layne, 2015). Further, the Supreme Court of Canada ruled that Walmart violated Quebec’s labor code when it closed a unionized store (Marin, 2014). Similarly, Hugo Boss lost an outsourcing battle against the union in Cleveland (Putre, 2011). Therefore, we view it as an empirical question whether unionization leads to stickier costs, which we test for in our paper.

14   

based on a simple majority (50% plus one vote) passing rule.15 The RD design exploits a unique

feature of the union election data—we observe the percentage of votes in favor of unionization in

every union election.

The intuition behind our RD strategy is as follows. Elections that pass (leading to

unionization and hence enhancing labor power in the firm) or fail (not leading to new unions

inside the firm) with votes within a narrow bandwidth around the 50% threshold should follow

the pattern of a quasi-randomized experiment. Essentially, this empirical approach compares

subsequent cost stickiness and dividend payouts of firms that barely pass union elections to those

of firms that barely fail the elections. For a close-call union election, unionization is locally

exogenous in the sense that it is unlikely to be systematically correlated with unobservable firm

characteristics that could affect a firm’s cost stickiness or dividend policy. In other words, the

assignment of treatment (i.e. unionization status) to the set of firms near the 50% threshold is

likely to be random, which helps us identify the effect of labor power on cost stickiness and

dividend payouts.

Another advantage of the RD design is that we do not have to include observable

covariates (i.e., control variables) in our analysis (as in a standard multiple regression

framework), because firms with votes falling in a narrow band around the threshold are likely to

be similar in all dimensions of characteristics. Hence, firm covariates are unnecessary for

identification (see earlier survey papers on the RD approach, such as Imbens and Lemieux (2008)

and Lee and Lemieux (2010), for a more detailed discussion on this less stringent requirement).

                                                            15 As described in detail in DiNardo and Lee (2004), a prototypical union election scenario is as follows: (1) a group of employees decide to try to form a union; (2) the employees petition the NLRB to hold an election; (3) the NLRB makes a ruling on whether the people the union seeks to represent have a “community of interest,” a coherent group for the purposes of bargaining; and (4) the NLRB holds a secret-ballot election at the work site.

15   

4.2 Union election data

We collect union election data, including the closing date of an election, the number of

eligible participants/voters, and the outcome of the election, from the National Labor Relations

Board (NLRB). Our sample covers union elections held between 1977 and 2012.16 We drop

those elections either with a missing voting outcome or with fewer than 50 eligible participating

employees.17 We manually match our union election sample to Compustat by company name and

address so that we can extract relevant financial statement information and other firm

characteristics from Compustat.

Our sampling procedure results in 2,492 unique union elections, 28.2% of which are

passed in favor of unionization. The mean percentage of votes in an election that are in favor of

unionization is 43.1%, with a standard deviation of 21%.

4.3 Diagnostic tests

The success of a RD strategy hinges on the satisfaction of two key identifying

assumptions, namely, agents’ imprecise manipulation of the forcing variable and the absence of

discontinuity in predetermined firm characteristics around the cutoff point. We perform two sets

of diagnostic tests in this section to ensure that these identifying assumptions are not violated in

our setting.

                                                            16 Data over the 1977-1999 period are used by Holmes (2006) and available from Thomas Holmes’s website. Data over the 2000-2010 period are posted by the NLRB (http://www.data.gov/). Data after 2010 are manually collected from the NLRB (https://www.nlrb.gov/news-outreach/graphs-data/tally-of-ballots). 17 We focus on elections with at least 50 eligible participating employees because union elections with a smaller size likely have a negligible impact on corporate decisions. In addition, elections with fewer participants may also be subject to precise manipulation of votes, which violates the crucial identifying assumption of the RD design. This type of filter is commonly adopted by studies of labor union elections (e.g., Ruback and Zimmerman, 1984; Lee and Mas, 2012). Our inferences remain unchanged when we limit the sample to elections where at least 10% of a firm’s total workforce voted (see the Online Appendix).

16   

The assumption of imperfect control by any agent requires that agents (workers and firms)

in an election cannot precisely manipulate the forcing variable (i.e., the share of favorable votes)

near the known cutoff.18 The implication is that the distribution of the forcing variable should not

have any sizable jumps around the 50% threshold. To check the validity of this assumption, we

perform a formal statistical analysis, developed by McCrary (2008), to test for discontinuity in

the density of the vote shares. The z-statistic for the McCrary test of discontinuity is 1.295 (the

coefficient of estimate is 0.272 with a standard error of 0.210), which is statistically insignificant.

Thus, we are unable to reject the null hypothesis that the density function at the cutoff is

continuous, suggesting that no agents have precisely manipulated the votes around the known 50%

threshold to achieve their desired unionization status. Our finding of no precise manipulation

around the known cutoff is consistent with prior literature using the same election data from the

NLRB (e.g., DiNardo and Lee, 2004; Lee and Mas, 2012; Bradley, Kim, and Tian, 2017).

The other important assumption of the RD design is that there is no discontinuity in firm

characteristics other than the unionization status across the known cutoff point. In other words,

firms close to the left and the right of the cutoff point (i.e., those with vote shares slightly above

or below the 50% threshold) should be similar in terms of observable, predetermined

characteristics that might affect the outcome (cost stickiness and dividend payouts) and/or the

assignment variable (vote shares). If there are any significant jumps in the distribution of these

                                                            18 Note that this assumption does not require the absence of vote manipulation in the elections. In reality, both firms and unions run campaigns to push for success, which leads to some degrees of the manipulation of shares. However, as long as agents do not have precise control over the forcing variable (even though some manipulation exists), an exogenous discontinuity still allows for random assignment of treatment (Lee, 2008).

17   

firm characteristics near the 50% threshold, then the treatment effect we observe using the RD

design could be biased.19

We perform a formal diagnostic test for this assumption by running nonparametric local

linear regressions on various firm characteristics at the predetermined year (i.e., the year before

the reported closing date of the union election). The nonparametric local linear regression is the

most stringent RD model that does not consider all elections in the sample but only examines

union elections in the vicinity of the 50% threshold within a certain bandwidth. We use the

optimal bandwidth suggested by Imbens and Kalyanaraman (2012) that minimizes the mean

squared error in a sharp RD setting.20 Compared to the global polynomial method, the local

linear estimation model has better local fitness (Bakke and Whited, 2012), more attractive rate

optimality, and superior bias properties (Fan and Gijbeles, 1992; Hahn, Todd, and van der

Klaauw, 2001). We use the triangular kernel because the statistics literature has shown that a

triangular kernel is optimal for estimating local linear regressions at the boundary due to its

greater weight on observations closer to the cutoff point (Fan and Gijbeles, 1992).

In untabulated analysis (contained in the Online Appendix), we find that none of the local

linear RD estimates for the ex-ante firm characteristics (Size, ROA, Leverage, M/B, ROAVol,

Cash, Tangibility) are statistically significant, suggesting that there is no discontinuity in the

distribution of these variables around the known 50% threshold. More importantly, the

predetermined values of our key variables, CostStickiness (the ex-ante measure estimated from

year t–3 to year t) and Div/Inc measured at year t–1 do not exhibit discontinuity around the                                                             19 Note, however, that this assumption is much less restrictive than the textbook assumptions regarding endogeneity (such as the exclusion restriction) in that it does not require those predetermined characteristics to be exogenous: as long as they are determined prior to the assignment variable (the voting share) and continuously distributed around the cutoff point (i.e. with no jumps), then the RD procedure will yield valid and consistent estimates. See Lee and Lemieux (2010) for a more detailed discussion of this assumption and related tests. 20 A sharp RD design requires a discontinuity in the probability of assignment from zero to one around the cutoff. A fuzzy RD design only requires a different probability of assignment around the cutoff.

18   

narrow bandwidth of the cutoff point, suggesting that our main RD results are unlikely to be

driven by ex-ante differences in cost stickiness or payout policy between firms passing union

elections and those whose union elections fail.

In summary, the above two sets of diagnostic tests show that the key identifying

assumptions of the RD design, namely, agents’ imprecise manipulation of the forcing variable

and the absence of discontinuity in predetermined firm characteristics, are not violated,

supporting that the variation in unionization status is as good as that from a randomized

experiment (Lee, 2008).

5. Results of the RD analysis

5.1 Main RD results

In this section, we report the main RD results. We aim to consider the long-term effect of

increased labor adjustment costs on firms’ cost stickiness and their dividend payouts. Hence, we

examine how unionization affects the cost stickiness measure estimated using the 16 quarterly

observations after the closing year of the union election (year t+1 to year t+4) as well as the

average dividend payout during the same four-year window subsequent to the union election.

Before we explore the effect of labor power on cost stickiness and dividend payouts in a

rigorous regression framework, we first visually check the relation between passing a union

election and these variables in Figure 1. Panel A of the figure plots ex-post cost stickiness (i.e.,

cost stickiness estimated over the four-year window subsequent to union elections), and Panel B

plots the dividend payout ratio (Div/Inc). The horizontal-axis depicts the forcing variable, vote

share, which is the percentage of votes in favor of unionization. To the left of the 50% cutoff

point, firms fail to unionize after the labor union election; to the right of the cutoff, firms succeed

19   

in becoming unionized. The spectrum of vote share is divided into 50 equally-spaced bins (with a

bin width of 2%).21 The dots in the graphs represent the average plotted variable (cost stickiness

or dividend payout ratios) in each bin, and the solid curve is the result of a fitted cubic

polynomial (with a 90% confidence interval). From Figure 1, we observe discontinuity in both

the ex-post cost stickiness and the average post-election dividend payout across the cutoff during

the four years post the union election: there is a significant increase in cost stickiness and a

significant drop in dividend payouts when moving the vote share from the left to the right of the

50% threshold. These patterns suggest a positive effect of labor power on cost stickiness and a

negative effect of labor power on dividend payouts.

Next, we carry out a formal RD analysis for firms’ ex-post cost stickiness and dividend

payouts after union elections using a nonparametric local linear regression (local RD) model

with the optimal bandwidth following Imbens and Kalyanaraman (2012) and the triangular

kernel (Fan and Gijbeles, 1992). 22 Panel A of Table 6 reports the results for ex-post cost

stickiness. The coefficient estimate on Unionization is positive and significant at the 1% level.

Specifically, during the four-year window post the election, firms that win a union election have

a cost stickiness measure that is 0.556 higher than those that fail to pass a union election, which

is economically sizable relative to the standard deviation of this variable for the entire sample

(i.e., 1.469).

We repeat the above local RD analysis for the average dividend payout ratio during the

four-year window after the union election. Panel B of Table 6 reports the results. The coefficient

estimate on Unionization is negative and significant at the 5% level, indicating a negative effect

                                                            21 Alternative choice of bin widths does not change our results qualitatively. 22 To check the robustness of our findings, we also adopt a rectangular kernel (no weighting) in local linear estimation in untabulated analysis and obtain qualitatively similar results (see the Online Appendix).

20   

of labor power on dividend payouts. In terms of the economic magnitude, the coefficient

estimate (i.e., –5.801) suggests that during the four-year window post the elections, firms that

win a union election have a dividend payout ratio that is 5.8 percentage points lower than those

that fail an election.

5.2 Mediation analysis based on predicted dividend payouts

To tighten the link between cost stickiness and dividend payouts in our RD design, we

conduct a two-step mediation analysis in this section. First, we regress a firm’s average post-

election dividend payout (over year t+1 to year t+4, with year t being the closing year of the

union election) on its degree of cost stickiness over the same window. Then, we decompose the

average dividend payout into two components: a predicted component based on the firm’s cost

behavior (i.e., the fitted value of the dividend payout from the above regression), and a residual

component that is orthogonal to its cost behavior (i.e., the residual value of the dividend payout

from the above regression). To facilitate the comparison of their relative magnitudes, we

standardize these two components to have a mean of zero and a standard deviation of one.

Second, we repeat our nonparametric local linear regressions for these two components

of a firm’s average post-election dividend payout. The results reported in Table 7 show that

unionization has a significant, negative effect on a firm’s dividend payout that is explained by its

cost behavior (i.e., the predicted average dividend payout). This result tightens the link between

cost stickiness and dividend payouts in our RD design.23

                                                            23 Our results are unlikely to be driven by higher wages after unionization: DiNardo and Lee (2004) find that the impacts of unionization on wages are close to zero. Our results that unionization increases cost stickiness suggests that workers are less likely to get laid off (not necessarily paid more) after unionization. Our results are also unlikely to be driven by the possibility that firms commit to a lower dividend payout to demonstrate their need for union concessions (DeAngelo and DeAngelo, 1990, 1991): in their thorough literature review, DeAngelo, DeAngelo, and Skinner (2008) conclude that this managerial signaling motive has at best minor influences on payout policy.

21   

5.3 Subsample analyses

In this section, we perform two subsample analyses in the RD setting to examine whether

the effect of unionization on dividend payouts is concentrated in firms in which unionization can

significantly influence cost behavior.

First, we predict that the negative effect of unionization on dividend payouts is

concentrated in firms with higher labor intensity. This is because the costs of firms that rely more

on human capital (as opposed to physical capital) in their business operations are more closely

related to labor. As a result, these firms would face a greater cash flow risk after enhanced labor

power and thus would pay out lower dividends.

To test this prediction, we split our sample firms in the RD setting into two groups based

on their median ex-ante industry-adjusted labor intensity (measured in the year before union

elections). Labor intensity is defined as the number of employees over sales and further adjusted

by industry. We report the results in Table 8. Column 1 (2) in Panel A reports the result of

nonparametric local linear regressions for firms with higher-than-median (lower-than-median)

industry-adjusted labor intensity.

Consistent with our prediction, the negative effect of unionization on dividend payouts is

mostly concentrated in firms with higher labor intensity. On the contrary, the RD estimates for

firms with low labor intensity are much smaller in magnitude and statistically insignificant. This

finding suggests that the cost behavior of firms with low labor intensity is less affected by union

elections, making it less necessary for them to constrain their dividend payouts.

Second, we predict that the negative effect of unionization on dividend payouts is

concentrated in firms with a higher ex-ante degree of cost stickiness. Firms whose costs are

already sticky to begin with would face a greater cash flow risk after the enhanced union labor

22   

power makes the costs even stickier, and they would thus commit to paying out less. To test this

prediction, we first calculate, for each union election, the ex-ante cost stickiness level by using

the 16 most recent quarterly observations up to the election year. We then split the sample into

two halves based on the sample median of ex-ante cost stickiness, and carry out our RD analysis

for each subsample. We report the estimation results in Panel B of Table 8.

Consistent with our prediction, the coefficient estimate on Unionization is negative and

significant in column 1 (i.e. for firms with high ex-ante cost stickiness), suggesting that the

negative effect of labor power on dividend payouts is concentrated in firms with high ex-ante

degree of cost stickiness. On the contrary, the RD estimate for firms with low ex-ante cost

stickiness (in column 2 of the panel) is smaller in magnitude than that in column 1 and

statistically insignificant. This finding is possibly due to the fact that firms with an ex-ante low

level of cost stickiness have flexible operations to begin with and thus enhanced labor power has

no material effect on their cash flow risk, making it less necessary for them to keep a lower

dividend payout.

Together, the results based on the RD analysis provide evidence that firms pay out lower

dividends in the presence of higher resource adjustment costs and stickier costs.

5.4 Robustness checks

In this section, we discuss results for robustness checks in the RD setting. First, we check

whether our local linear regression estimates are sensitive to the choice of bandwidths, which

reflects a classical tradeoff between bias and precision. On the one hand, a wider bandwidth

makes use of more observations within the local neighborhood of the cutoff and thus yields more

precise estimates. On the other hand, a wider bandwidth could introduce more noise and bias into

the estimation because the use of more “non-local” observations away from the cutoff can make

23   

the linear approximation inaccurate. The converse is true for a narrower bandwidth. To address

the concern that our main local RD results in Table 6 are driven by the bandwidth we have

chosen, we plot the estimated local RD coefficients along with their 90% confidence intervals

(on the vertical axis) as a function of the chosen bandwidth (on the horizontal axis) in Figure 2.

A value of “100” on the horizontal axis represents the optimal bandwidth suggested by Imbens

and Kalyanaraman (2012). “200” means 200% of (i.e., two times of) the optimal bandwidth,

“300” means 300%, and so forth. Panel A plots ex-post cost stickiness, and Panel B plots the ex-

post average Div/Inc. The local RD estimates are almost always positive (negative) and stable for

cost stickiness (dividend payout ratios) over the whole spectrum of bandwidth choices. These

results show that our local linear RD estimates are unlikely to be driven by any specific choice of

bandwidths.

Second, if our RD estimation truly reflects a positive (negative) effect of labor power on

cost stickiness (dividend payouts), we should not observe a similar effect if we artificially

assume a threshold other than 50% that determines union election outcomes. Hence, we run

placebo tests to check whether we still observe discontinuity in cost stickiness or dividend

payouts at randomly selected thresholds that are different from the true 50% threshold. We run

this placebo test 5,000 times and plot a histogram of the distribution of the corresponding local

RD estimates in Figure 3. The vertical dashed line stands for the value of the local RD estimate

obtained using the true cutoff point of 50%. Both histograms in Figure 3 are approximately

centered around 0, suggesting that the positive (negative) effect of labor power on cost stickiness

(dividend payouts) is absent if we artificially pick a cutoff point other than 50%. This placebo

analysis enhances our confidence in the RD procedure and the resulting estimates, because it

rules out chances as an explanation for our main findings in the previous sections.

24   

Third, we examine alternative measures of cost stickiness and dividend payouts as in

Tables 4 and 5 using the local RD analysis. The untabulated analysis (contained in the Online

Appendix) shows that our main RD results continue to hold in these robustness tests.

6. Conclusion

We show that asymmetric cost behavior, a prevalent phenomenon documented in the

recent accounting literature, is an important determinant of a firm’s dividend policy. We find that

firms with stickier costs pay out lower dividends than firms with less sticky costs. We further

show that this negative relation is only present in the subsample of firms with fewer financial

resources. In addition, we do not find evidence that our results are driven by firms with weaker

corporate governance, suggesting that the negative relation between cost stickiness and dividend

payouts reflects a firm’s response to adjustment costs than its agency conflicts.

Further, we use a regression discontinuity design that exploits variation in labor

adjustment costs generated by close-call union elections. We show that unionization has a

significantly negative effect on a firm’s dividend payout that is explained by its cost behavior

and that the negative effect of unionization on dividend payouts is concentrated in firms with

high labor intensity or high ex-ante degree of cost stickiness. Our paper sheds new light on the

determinants of a firm’s dividend payout and the role of cost behavior in corporate finance

decisions.

25   

Appendix: Definition of Variables

Variable Definition

Measures of cost stickiness and dividend payouts

CostStickiness Cost stickiness in year t is estimated by running the following empirical model with its 16 previous quarters of data (year t–3 to year t): ΔLnSG&A = β0 + β1×ΔLnSales + β2×Decrease×ΔLnSales + µ, where ΔLnSG&A is the log-change in quarterly selling, general, and administrative expenses (i.e., costs), ΔLnSales is the log-change in quarterly sales, and Decrease is a dummy variable that equals one if ΔLnSales is less than zero, and zero otherwise. Coststickiness is the negative of β2.

CostStickiness (OperatingCost)

It is similarly defined as CostStickiness but uses operating cost (measured as the difference between sales (salesq) and operating income after depreciation (oiadpq)) as the cost measure instead of the selling, general and administrative expenses (xsgaq).

CostStickiness (TotalCost)

It is similarly defined as CostStickiness but uses total cost (measured as the sum of cost of goods sold (cogsq) and the selling, general and administrative expenses (xsgaq)) as the cost measure instead of the selling, general and administrative expenses (xsgaq).

Div/Inc (%) Common dividends (dvc) divided by income before extraordinary items (ibcom) multiplied by 100 (i.e., in percentage points). Observations with negative ibcom and positive dvc are dropped.

Div/Sales (%) Common dividends (dvc) divided by sales (sale) multiplied by 100 (i.e., in percentage points).

Div/Assets (%) Common dividends (dvc) divided by book value of total assets (at) multiplied by 100 (i.e., in percentage points).

Div/MV (%) Common dividends (dvc) divided by market value of equity (prcc_f×csho) multiplied by 100 (i.e., in percentage points).

LnDiv The natural logarithm of one plus common dividends (dvc).

Measures of firm characteristics Assets Book value of total assets (at). Size The natural logarithm of book value of total assets (at). ROA Return on assets, defined as operating income before depreciation (oibdp)

divided by book value of total assets (at). Leverage Leverage ratio, defined as book value of long-term debt (dltt) divided by

book value of total assets (at). M/B Market to book ratio, defined as market value of equity (prcc_f×csho)

plus book value of assets (at) minus book value of equity (ceq) minus deferred taxes (txdb) divided by book value of total assets (at).

26   

ROAVol The standard deviation of quarterly return on assets over the most recent 16 quarters (year t–3 to year t). Quarterly return on assets is defined as quarterly operating income before depreciation (oibdpq) divided by quarterly book value of total assets (atq).

Cash Cash holdings, defined as cash and short-term investments (che) divided by book value of total assets (at).

Tangibility Tangible assets, defined as net property, plant and equipment (ppent) divided by book value of total assets (at).

InstOwn Institutional ownership over the fiscal year, calculated as the arithmetic mean of the four quarterly institutional holdings reported by 13F filings.

 

 

27   

References

Anderson, M. A., R. D. Banker, and S. Janakiraman. 2003. Are selling, general, and administrative costs sticky? Journal of Accounting Research 41: 47–63.

Bakke, T. and T. Whited. 2012. Threshold events and identification: A study of cash shortfalls. Journal of Finance 67: 1083–1111.

Banker, R. D., S. Basu, D. Byzalov, and J. Chen. 2016. The confounding effect of cost stickiness on conservatism estimates. Journal of Accounting and Economics 61: 203–220.

Banker, R. D. and D. Byzalov. 2014. Asymmetric cost behavior. Journal of Management Accounting Research 26: 43–79.

Banker, R. D., D. Byzalov, and L. Chen. 2013. Employment protection legislation, adjustment costs and cross-country differences in cost behavior. Journal of Accounting and Economics 55: 111–127.

Banker, R. D., D. Byzalov, M. Ciftci, and R. Mashruwala. 2014. The moderating effect of prior sales changes on asymmetric cost behavior. Journal of Management Accounting Research 26: 221–242.

Banker, R. D., D. Byzalov, S. Fang, and Y. Liang. 2018. Cost management research. Journal of Management Accounting Research, forthcoming.

Banker, R. D. and L. Chen. 2006. Predicting earnings using a model based on cost variability and cost stickiness. The Accounting Review 81: 285–307.

Bebchuk, L., A. Cohen, and A. Ferrell. 2009. What matters in corporate governance? Review of Financial Studies 22: 783–827.

Berger, P. G., E. Ofek, and D. Yermack. 1997. Managerial entrenchment and capital structure decisions. Journal of Finance 52: 1411–1438.

Bhattacharya, S. 1979. Imperfect information, dividend policy, and “the bird in the hand” fallacy. Bell Journal of Economics 10: 259–272.

Black, F. 1976. The dividend puzzle. Journal of Portfolio Management 2: 5–8.

Bradley, D., I. Kim, and X. Tian. 2017. Do unions affect innovation? Management Science 63: 2251–2271.

Brav, A., J. R. Graham, C. R. Harvey, and R. Michaely. 2005. Payout policy in the 21st century. Journal of Financial Economics 77: 483–527.

Campello, M., J. Qiu, J. Gao, and Y. Zhang. 2018. Bankruptcy and the cost of organized labor: Evidence from union elections. Review of Financial Studies 3: 980–1013.

Chay, J. B. and J. Suh. 2009. Payout policy and cash-flow uncertainty. Journal of Financial Economics 93: 88–107.

Chen, C. X., H. Lu, and T. Sougiannis. 2012. The agency problem, corporate governance, and the asymmetrical behavior of selling, general, and administrative costs. Contemporary Accounting Research 29: 252–282.

28   

Cheng, L. 2017. Organized labor and debt contracting: Firm-level evidence from collective bargaining. The Accounting Review 92: 57–85.

DeAngelo, H. and L. DeAngelo. 1990. Dividend policy and financial distress: An empirical investigation of troubled NYSE firms. Journal of Finance 45: 1415–1431.

DeAngelo, H. and L. DeAngelo. 1991. Union negotiations and corporate policy: A study of labor concessions in the domestic steel industry during the 1980s. Journal of Financial Economics 30: 3–43.

DeAngelo, H., L. DeAngelo, and D. J. Skinner. 1992. Dividends and losses. Journal of Finance 47: 1837–1863.

DeAngelo, H., L. DeAngelo, and D. J. Skinner. 1996. Reversal of fortune: Dividend signaling and the disappearance of sustained earnings growth. Journal of Financial Economics 40: 341–371.

DeAngelo, H., L. DeAngelo, and D. J. Skinner. 2000. Special dividends and the evolution of dividend signaling. Journal of Financial Economics 57: 309–354.

DeAngelo, H., L. DeAngelo, and D. J. Skinner. 2004. Are dividends disappearing? Dividend concentration and the consolidation of earnings. Journal of Financial Economics 72: 425–456.

DeAngelo, H., L. DeAngelo, and D. J. Skinner. 2008. Corporate payout policy. Foundations and Trends in Finance 3: 95–287.

DeAngelo, H., L. DeAngelo, and R. Stulz. 2006. Dividend policy and the earned/contributed capital mix: A test of the life-cycle theory. Journal of Financial Economics 81: 227–254.

De Franco, G., S. Kothari, and R. S. Verdi. 2011. The benefits of financial statement comparability. Journal of Accounting Research 49: 895–931.

DiNardo, J. and D. S. Lee. 2004. Economic impacts of new unionization on private sector employers: 1984–2001. Quarterly Journal of Economics 119: 1383–1441.

Fama, E. F. and K. R. French. 2001. Disappearing dividends: Changing firm characteristics or lower propensity to pay? Journal of Financial Economics 60: 3–40.

Fama, E. F. and K. R. French. 2004. New lists: Fundamentals and survival rates. Journal of Financial Economics 73: 229–269.

Fan, J. and I. Gijbels. 1992. Variable bandwidth and local linear regression smoothers. Annals of Statistics 20: 2008–2036.

Farre-Mensa, J., R. Michaely, and M. Schmalz. 2014. Payout policy. Annual Review of Financial Economics, Volume 6, edited by Andrew W. Lo and Robert C. Merton. Palo Alto, CA: Annual Reviews: 75–134.

Floyd, E., N. Li, and D. J. Skinner. 2015. Payout policy through the financial crisis: The growth of repurchases and the resilience of dividends. Journal of Financial Economics 118: 299–316.

Gompers, P. A., J. L. Ishii, and A. Metrick. 2003. Corporate governance and equity prices.

29   

Quarterly Journal of Economics 118: 107–155.

Gonedes, N. J. 1978. Corporate signaling, external accounting, and capital market equilibrium: Evidence on dividends, income, and extraordinary items. Journal of Accounting Research 16: 26–79.

Grullon, G. and R. Michaely. 2002. Dividends, share repurchases, and the substitution hypothesis. Journal of Finance 57: 1649–1684.

Guay, W. and J. Hartford. 2000. The cash-flow permanence and information content of dividend increases versus repurchases. Journal of Financial Economics 57: 385–415.

Hahn, J., P. Todd, and W. van der Klaauw. 2001. Regression discontinuity. Econometrica 69: 201–209.

Hanlon, M. and J. L. Hoopes. 2014. What do firms do when dividend tax rates change? An examination of alternative payout responses. Journal of Financial Economics 114:105–124.

Healy, P. M. and K. G. Palepu. 1988. Earnings information conveyed by dividend initiations and omissions. Journal of Financial Economics 21: 149–176.

Holmes, T. 2006. Geographic spillover of unionism. Staff Report 368, Federal Reserve Bank of Minneapolis.

Imbens, G. W. and K. Kalyanaraman. 2012. Optimal bandwidth choice for the regression discontinuity estimator. Review of Economic Studies 79: 933–959.

Imbens, G. W. and T. Lemieux. 2008. Regression discontinuity designs: A guide to practice. Journal of Econometrics 142: 615–635.

Jagannathan, M., C. P. Stephens, and M. S. Weisbach. 2000. Financial flexibility and the choice between dividends and stock repurchases. Journal of Financial Economics 57: 355–384.

Jensen, M. C. 1986. Agency costs of free cash flow, corporate finance, and takeovers. American Economic Review 76: 323–329.

John, K., A. Knyazeva, and D. Knyazeva. 2011. Does geography matter? Firm location and corporate payout policy. Journal of Financial Economics 101: 533–551.

John, K. and J. Williams. 1985. Dividends, dilution, and taxes: A signaling equilibrium. Journal of Finance 40: 1053–1070.

Kama, I. and D. Weiss. 2013. Do earnings targets and managerial incentives affect sticky costs? Journal of Accounting Research 51: 201–224.

Layne, N. 2015. Wal-Mart to reopen five U.S. stores at center of union complaint. Reuters, September 2.

Lease, R. C., K. John, A. Kalay, U. Loewenstein, and O. H. Sarig. 2000. Dividend Policy: Its Impact on Firm Value. Boston, MA: Harvard Business School Press.

Lee, D. S. 2008. Randomized experiments from non-random selection in U.S. house elections. Journal of Econometrics 142: 675–697.

Lee, D. S. and T. Lemieux. 2010. Regression discontinuity designs in economics. Journal of

30   

Economic Literature 48: 281–355.

Lee, D. S. and A. Mas. 2012. Long-run impacts of unions on firms: New evidence from financial markets, 1961–1999. Quarterly Journal of Economics 127: 333–378.

Lintner, J. 1956. Distribution of incomes of corporations among dividends, retained earnings, and taxes. American Economic Review 46: 97–113.

Marin, S. 2014. Supreme Court of Canada: Wal-Mart must pay for closing unionized store. The Canadian Press, August 27.

McCrary, J. 2008. Manipulation of the running variable in the regression discontinuity design: A density test. Journal of Econometrics 142: 698–714.

Miller, M. and F. Modigliani. 1961. Dividend policy, growth, and the valuation of shares. Journal of Business 34: 411–433.

Miller, M. and K. Rock. 1985. Dividend policy under asymmetric information. Journal of Finance 40: 1031–1051.

Myers, S. C. and N. S. Majluf. 1984. Corporate financing and investment decisions when firms have information that investors do not have. Journal of Financial Economics 13: 187–221.

Noreen, E. and N. Soderstrom. 1997. The accuracy of proportional cost models: Evidence from hospital service departments. Review of Accounting Studies 2: 89–114.

Putre, L. 2011. How Hugo Boss lost a Cleveland union battle. Pacific Standard Magazine, March 9.

Rouxelin, F., W. Wongsunwai and N. Yehuda. 2018. Aggregate cost stickiness in GAAP financial statements and future unemployment rate. The Accounting Review 93: 299–325.

Ruback, R. and M. Zimmerman. 1984. Unionization and profitability: Evidence from the capital market. Journal of Political Economy 92: 1134–1157.

Skinner, D. J. 2008. The evolving relation between earnings, dividends, and stock repurchases. Journal of Financial Economics 87: 582–609.

Skinner, D. J. and E. Soltes. 2011. What do dividends tell us about earnings quality? Review of Accounting Studies 16: 1–28.

Smith, Jr., C. W. and R. L. Watts. 1992. The investment opportunity set and corporate financing, dividend, and compensation policies. Journal of Financial Economics 32: 263–292.

Watts, R. L. 1973. The information content of dividends. Journal of Business 46: 191–211.

Weiss, D. 2010. Cost behavior and analysts’ earnings forecasts. The Accounting Review 85: 1441–1474.

31   

Figure 1: Regression discontinuity plots for ex-post cost stickiness and dividend payouts This figure presents regression discontinuity plots using a fitted cubic polynomial estimated with a 90% confidence interval around the fitted value. The horizontal-axis is union vote share (i.e., the percentage of votes in an election in favor of unionization). The dots depict the ex-post CostStickiness or average Div/Inc, defined in the Appendix, in each of the 50 equally-spaced bins (with a bin width of 2%).

Panel A: Ex-post CostStickiness

Panel B: Average Div/Inc

32   

Figure 2: Alternative RD bandwidths This figure plots the estimated local RD coefficients along with their 90% confidence intervals (on the vertical axis) against alternative values of bandwidths (on the horizontal axis). A value of “100” on the horizontal axis represents the optimal bandwidth suggested by Imbens and Kalyanaraman (2012). “200” means 200% of (i.e., two times of) the optimal bandwidth, “300” means 300%, and so forth. The dependent variables in the two panels are ex-post CostStickiness and average Div/Inc, respectively.

Panel A: Ex-post CostStickiness

Panel B: Average Div/Inc

33   

Figure 3: Placebo tests This figure plots a histogram of the distribution of the local RD estimates from 5,000 placebo tests. The horizontal-axis represents the RD estimates from the placebo tests that artificially select an alternative threshold other than 50%. The dashed vertical line represents the RD estimate at the true 50% threshold. The vertical-axis in the two panels are the densities of the estimated coefficients for ex-post CostStickiness and average Div/Inc, respectively.

Panel A: Ex-post CostStickiness

Panel B: Average Div/Inc

34   

Table 1: Descriptive statistics of the variables used in the OLS analysis This table reports the descriptive statistics of the variables for the population of U.S. public firms from 1978 to 2016. All variables are defined in the Appendix. Variable Mean P25 Median P75 Std Dev N CostStickiness 0.053 -0.484 0.027 0.605 1.469 72,238 Div/Inc (in %) 13.876 0.000 0.000 20.893 24.283 72,238 Assets(in $billion) 2.125 0.046 0.231 1.110 6.285 72,238 ROA 0.089 0.059 0.119 0.175 0.183 72,238 Leverage 0.243 0.046 0.201 0.356 0.240 72,238 M/B 1.771 1.021 1.355 1.981 1.330 72,238 ROAVol 0.025 0.009 0.015 0.028 0.030 72,238 Cash 0.147 0.024 0.081 0.211 0.168 72,238 Tangibility 0.304 0.109 0.241 0.447 0.240 72,238

35   

Table 2: OLS regressions of dividend payouts on cost stickiness This table presents the results of OLS regressions of dividend payouts on cost stickiness and control variables. Cost stickiness in year t is estimated using the 16 most recent quarterly observations (the quarterly data from year t–3 to year t). Column 1 reports results using the full sample and column 2 reports results using firms that pay dividends at least once during the sample period. All variables are defined in the Appendix. The t-statistics, presented in parentheses below the coefficients, are corrected for heteroskedasticity and cross-sectional and time-series correlations using a two-way cluster at the firm and year levels. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Dep. Var. Div/Inc

(1) (2)

CostStickiness -0.182** -0.293** (-2.365) (-2.042) Size 2.893*** 1.430*** (13.636) (3.822)ROA 4.671** -0.795 (2.547) (-0.106) Leverage -9.554*** -11.059*** (-8.512) (-3.216) M/B 1.101*** 0.126 (5.062) (0.243)ROAVol -14.173* -154.100*** (-2.022) (-4.567) Cash -3.540** 0.379 (-2.685) (0.117) Tangibility 8.078*** 20.744*** (5.479) (7.272) Year FE Yes Yes Industry FE Yes Yes Observations 72,238 39,408 R-squared 0.194 0.088

36   

Table 3: Subsample analysis based on cash holdings or institutional ownership This table presents the OLS regression analysis of dividend payouts on cost stickiness for subsamples split by cash holdings or institutional ownership. Columns 1 and 3 report the results for firms with higher-than-median cash holdings and institutional ownership, respectively. Columns 2 and 4 report the results for firms with lower-than-median cash holdings and institutional ownership, respectively. Cost stickiness in year t is estimated using the 16 most recent quarterly observations (the quarterly data from year t–3 to year t). All variables are defined in the Appendix. The t-statistics, presented in parentheses below the coefficients, are corrected for heteroskedasticity and cross-sectional and time-series correlations using a two-way cluster at the firm and year levels. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Dep. Var. Div/Inc

Subsample Var. Cash InstOwn

High Low High Low (1) (2) (3) (4)

CostStickiness -0.130 -0.245** -0.238** -0.079 (-1.286) (-2.127) (-2.082) (-0.791) Controls Yes Yes Yes Yes Year FE Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Observations 36,119 36,119 36,119 36,119 R-squared 0.183 0.208 0.213 0.217

37   

Table 4: OLS regressions of alternative dividend payout measures on cost stickiness This table presents the results of OLS regressions of alternative dividend payout measures on cost stickiness and control variables. Cost stickiness in year t is estimated using the 16 most recent quarterly observations (the quarterly data from year t–3 to year t). All variables are defined in the Appendix. The t-statistics, presented in parentheses below the coefficients, are corrected for heteroskedasticity and cross-sectional and time-series correlations using a two-way cluster at the firm and year levels. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Dep. Var. Div/Sales Div/Assets Div/MV LnDiv

(1) (2) (3) (4)

CostStickiness -0.023** -0.018*** -0.019*** -0.024*** (-2.113) (-2.787) (-2.733) (-4.553) Controls Yes Yes Yes Yes Year FE Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Observations 75,447 75,449 75,439 75,449 R-squared 0.141 0.174 0.172 0.464

38   

Table 5: OLS regressions of dividend payouts on alternative cost stickiness measures This table presents the results of OLS regressions of dividend payouts on alternative cost stickiness measures and control variables. CostStickiness (OperatingCost) and CostStickiness (TotalCost) are defined similarly to CostStickiness but operating costs (sales minus operating income after depreciation) and total costs (costs of goods sold and SG&A) are used as the cost measure, respectively. Cost stickiness in year t is estimated using the 16 most recent quarterly observations (the quarterly data from year t–3 to year t). All variables are defined in the Appendix. The t-statistics, presented in parentheses below the coefficients, are corrected for heteroskedasticity and cross-sectional and time-series correlations using a two-way cluster at the firm and year levels. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Dep. Var. Div/Inc

(1) (2)

CostStickiness -0.717*** (OperatingCost) (-3.962) CostStickiness -0.552*** (TotalCost) (-3.632) Controls Yes Yes Year FE Yes Yes Industry FE Yes Yes Observations 71,967 72,214 R-squared 0.195 0.194

39   

Table 6: Regression discontinuity analysis on ex-post cost stickiness and dividend payouts This table presents the results of nonparametric local linear regression using the optimal bandwidth following Imbens and Kalyanaraman (2012) and the triangular kernel. Panels A and B report the results for firms’ ex-post cost stickiness and dividend payouts after union elections, respectively. Cost stickiness is calculated using the 16 quarterly observations after the closing year of the union election (year t+1 to year t+4). Dividend payouts are measured as the average over the same four-year window subsequent to the union election. All variables are defined in the Appendix. The z-statistics are presented in parentheses below the coefficients. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Panel A: Cost stickiness after the election Dep. Var. CostStickiness Unionization 0.556*** (2.802) Observations 2,492 Bandwidth +/–0.152 Panel B: Dividend payouts after the election Dep. Var. Average Div/Inc Unionization -5.801** (-1.989) Observations 2,492 Bandwidth +/–0.155

40   

Table 7: Regression discontinuity analysis on predicted dividend payouts This table presents the results of nonparametric local linear regression discontinuity analysis for firms’ predicted and residual average post-election dividend payouts over year t+1 to year t+4 (with year t being the closing year of the union election) using the optimal bandwidth following Imbens and Kalyanaraman (2012) and the triangular kernel. The predicted and residual average dividend payouts are calculated from a simple OLS model that regresses average dividend payouts on the cost stickiness measure computed over the same window (i.e., from year t+1 to year t+4) controlling for year and industry fixed effects. Both average dividend payout measures are standardized to have a mean of zero and a standard deviation of one. All other variables are defined in the Appendix. The z-statistics are presented in parentheses below the coefficients. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Dep. Var. Predicted Average Div/Inc Residual Average Div/Inc Unionization -0.246** -0.090 (-2.300) (-1.471) Observations 2,492 2,492 Bandwidth +/–0.183 +/–0.144

41   

Table 8: Subsample RD analysis based on ex-ante industry-adjusted labor intensity or cost stickiness This table presents the results of regression discontinuity analysis on average post-election dividend payouts for subsamples partitioned by ex-ante industry-adjusted labor intensity or cost stickiness. The dependent variable is the average dividend payouts over year t+1 to year t+4 (with year t being the closing year of the union election). Ex-ante labor intensity is defined as the industry-adjusted number of employees over sales in the year before the union election. Ex-ante cost stickiness is calculated using the 16 most recent quarterly observations up to the election year. We split the sample into two subsamples based on the median industry-adjusted labor intensity or cost stickiness. Column 1 (2) in Panels A and B reports the result of nonparametric local linear regression for firms with higher-than-median (lower-than-median) industry-adjusted labor intensity or cost stickiness using the optimal bandwidth following Imbens and Kalyanaraman (2012) and the triangular kernel. All variables are defined in the Appendix. The z-statistics are presented in parentheses below the coefficients. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Panel A: Dividend payouts after the election based on ex-ante industry-adjusted labor intensity Dep. Var. Average Div/Inc (1) (2) High Labor Intensity Low Labor Intensity Unionization -13.466*** 0.999 (-2.651) (0.178) Observations 1,246 1,246 Bandwidth +/–0.152 +/–0.157 Panel B: Dividend payouts after the election based on ex-ante cost stickiness Dep. Var. Average Div/Inc (1) (2) High Ex-ante Cost Stickiness Low Ex-ante Cost Stickiness Unionization -9.470** -3.590 (-2.045) (-0.822) Observations 1,246 1,246 Bandwidth +/–0.137 +/–0.197

42   

ONLINE APPENDIX

(NOT INTENDED FOR PUBLICATION)

43   

Table A1: OLS regressions of total payout measures on cost stickiness This table presents the results of OLS regressions of alternative total payout measures on cost stickiness and control variables. Total payouts (Total) equal the sum of annual cash dividends and stock repurchases. We follow Grullon and Michaely (2002) to define repurchases as the total expenditures on the purchase of common and preferred stocks minus any reduction in the redemption value of the net preferred stocks outstanding. The total payout measures include the total payout to income ratio (Total/Inc), the total payout to sales ratio (Total/Sales), the total payout to assets ratio (Total/Asset), the total payout to market value ratio (Total/MV), and the natural logarithm of one plus the dollar value of total payout (LnTotal). The first three ratios are expressed in percentage points. Cost stickiness in year t is estimated using the 16 most recent quarterly observations (the quarterly data from year t–3 to year t). All other variables are defined in the Appendix. The t-statistics, presented in parentheses below the coefficients, are corrected for heteroskedasticity and cross-sectional and time-series correlations using a two-way cluster at the firm and year levels. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Dep. Var. Total/Inc Total/Sales Total/Assets Total/MV LnTotal (1) (2) (3) (4) (5) CostStickiness -0.517*** -0.048** -0.036*** -0.049*** -0.028*** (-3.632) (-2.325) (-2.819) (-3.692) (-5.038) Controls Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Observations 72,238 75,447 75,449 75,439 75,449

44   

Table A2: OLS regressions after controlling for cost-sales sensitivity for sales increase This table presents the results of OLS regressions after controlling for cost-sales sensitivity for sales increase (i.e., β1 in Equation (1)). The dependent variable is dividend payout ratio in column 1 and total payout ratio in column 2, where the total payout ratio is defined as the sum of annual cash dividends and stock repurchases divided by income before extraordinary items. Both Div/Inc and Total/Inc are expressed in percentage points. CostStickiness and CostSalesSensitivity in year t are estimated using the 16 most recent quarterly observations (the quarterly data from year t–3 to year t). All other variables are defined in the Appendix. The t-statistics, presented in parentheses below the coefficients, are corrected for heteroskedasticity and cross-sectional and time-series correlations using a two-way cluster at the firm and year levels. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Dep. Var. Div/Inc Total/Inc (1) (2) CostStickiness -0.270** -0.418** (-2.345) (-2.112) CostSalesSensitivity 0.273 -0.308 (0.908) (-0.547) Controls Yes Yes Year FE Yes Yes Industry FE Yes Yes Observations 72,238 72,238 R-squared 0.194 0.163

45   

Table A3: Regression discontinuity analysis on predetermined firm characteristics This table presents the results of nonparametric local linear regression for firm characteristics measured at the predetermined year (the year before the reported closing date of the union election), using the optimal bandwidth following Imbens and Kalyanaraman (2012) and the triangular kernel. Ex-ante cost stickiness is calculated using the 16 most recent quarterly observations up to the election year (year t–3 to year t). All variables are defined in the Appendix. The z-statistics are presented in parentheses below the coefficients. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Variable Coefficient z-statistic Ex-Ante CostStickiness -0.147 -0.323 Div/Inc -0.913 -0.273 Size -0.148 -0.553 ROA 0.002 0.240 Leverage 0.018 1.036 M/B -0.032 -0.421 ROAVol 0.002 1.232 Cash 0.069 0.712 Tangibility 0.018 0.514  

 

 

 

 

 

 

 

 

46   

Table A4: Regression discontinuity analysis on ex-post cost stickiness and dividend payouts using the rectangular kernel This table presents the results of nonparametric local linear regression using the optimal bandwidth following Imbens and Kalyanaraman (2012) and the rectangular kernel. Panels A and B report the results for firms’ ex-post cost stickiness and dividend payouts after union elections, respectively. Cost stickiness is calculated using the 16 quarterly observations after the closing year of the union election (year t+1 to year t+4). Dividend payouts are measured as the average over the same four-year window subsequent to the union election. All variables are defined in the Appendix. The z-statistics are presented in parentheses below the coefficients. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Panel A: Cost stickiness after the election Dep. Var. CostStickiness Unionization 0.546*** (-2.711) Observations 2,492 Bandwidth +/–0.120 Panel B: Dividend payouts after the election Dep. Var. Average Div/Inc Unionization -7.447** (-2.494) Observations 2,492 Bandwidth +/–0.122

47   

Table A5: Regression discontinuity analysis on ex-post cost stickiness and dividend payouts using elections with at least 10% workforce participation This table presents the results of nonparametric local linear regression using elections in which more than 10% of a firm’s workforce participated. We use the optimal bandwidth following Imbens and Kalyanaraman (2012) and the triangular kernel. Panels A and B report the results for firms’ ex-post cost stickiness and dividend payouts after union elections, respectively. Cost stickiness is calculated using the 16 quarterly observations after the closing year of the union election (year t+1 to year t+4). Dividend payouts are measured as the average over the same four-year window subsequent to the union election. All variables are defined in the Appendix. The z-statistics are presented in parentheses below the coefficients. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Panel A: Cost stickiness after the election Dep. Var. CostStickiness Unionization 0.842** (1.966) Observations 363 Bandwidth +/–0.172 Panel B: Dividend payouts after the election Dep. Var. Average Div/Inc Unionization -10.296** (2.802) Observations 363 Bandwidth +/–0.371

48   

Table A6: Regression discontinuity analysis on alternative measures of ex-post cost stickiness and dividend payouts This table presents the results of regression discontinuity analysis for firms’ alternative measures of ex-post cost stickiness and dividend payouts after union elections using the optimal bandwidth following Imbens and Kalyanaraman (2012) and the triangular kernel. Cost stickiness is calculated using the 16 quarterly observations after the closing year of the union election (year t+1 to year t+4). Dividend payouts are measured as the average over the same four-year window subsequent to the union election. All variables are defined in the Appendix. The z-statistics are presented in parentheses below the coefficients. ***, **, and * denote significance at the 1%, 5%, and 10% levels for two-tailed tests, respectively. Variable Coefficient z-statistic CostStickiness(OperatingCost) 0.200*** 2.887 CostStickiness(TotalCost) 0.111** 2.028 Div/Sales -0.398*** -2.000 Div/Assets -0.297** -2.004 Div/MV -0.515** -2.564 LnDiv -0.534** -2.518