Abstract
This paper presents the first causal evidence of the role that charity managers play in the financial performance of a charity. Using a novel data set of Canadian charities, I find that a one standard deviation increase in manager ability leads to a 0.516 standard deviation increase in total revenue, which amounts to more than $650,000. Married couples are found to have a larger effect than individuals, and female managers are shown to have a larger effect than male managers. I then present extensions that show that the baseline model overestimates the effect of the managers.
I. Introduction
Charities play an important role in society by providing many goods and services that contribute to the building of our communities, and by servicing the needs of those facing financial, personal, and other hardships. Donors, be they private individuals or government agencies, must choose the charities to which they donate, and their donation decisions may be influenced by the perceived impact that their gifts and grants has on servicing those in need.1 There is, however, limited quantitative work on charity performance. Past empirical work has focused on measuring objective functions of charities. Steinberg (1986) shows empirically that there is great heterogeneity in the measured objective functions of charities, and one possible explanation for this heterogeneity is differences in the charity managers (Rose-Ackerman 1987; Besley and Ghatak 2005). How much of the differences in a charity’s financial performance can be attributed to the manager? This work provides evidence that the manager can make a significant difference in a charity’s financial performance.
In any organization, leadership matters. Whether the setting involves CEOs setting firm policies that influence profits and acquisitions or principals setting school policies that influence student behavior and teacher hiring practices, empirical evidence of the effect of leadership is abundant.2 Leadership may also matter for nonprofit firms, yet the empirical evidence on the impact of leaders in the nonprofit sector is lacking. Besley and Ghatak (2005) suggest that the managers at for-profit firms are likely different than those at nonprofit firms. This study fills this gap by employing a fixed-effects model, often used in studies of leadership in other settings, to estimate the effect of charity managers.
Only recently researchers have begun to show that charities also play a role in the market for donations. For example, Andreoni and Payne (2003) show that much of the crowding out observed from government grants is actually due to reduced fundraising efforts by the charity. However, little work has been done to understand how charities operate and what factors from within the charity affect their finances. This work sheds light on one factor that affects how a charity operates, the charity’s manager.
Charities present a unique case in which to study leadership because, unlike competitive firms who maximize profits, charities are much more heterogeneous in their objective functions (Steinberg 1986). This presents a problem not only for charities, but for other public sector settings where objective functions are also heterogeneous, and data are not readily available that would help us to measure performance. Weisbrod (1991) points out that, “A great deal of literature in public administration highlights the ‘vagueness and intangibility’ of public-sector outputs; it underscores ‘the unique difficulties in specifying and quantifying performance measures in the public sector.’” Charities are a perfect setting to study public-sector manager performance because of the availability of data that allows me to measure the performance of managers. Tinkelman (2004) uses data from charities in the United States to estimate charities’ fundraising elasticities (the change in donations received as fundraising expenditures change). He then argues that these fundraising elasticities can show whether charities are net revenue maximizers (elasticities close to one) or total revenue maximizers (elasticities close to zero), which in turn reveals the charity manager’s objective function. He finds that fundraising elasticities for charities differ by charity type, but that they typically fall somewhere between zero and one, with many falling between 0.07 and 0.27.
Previous studies (for example, Tinkelman 2004; Sieg and Zhang 2012) have attempted to quantify the effect of charity managers but have faced the problem of using data sets that contain information about charity performance but no information about the managers, which limits their ability to measure the effect of the manager conclusively. This work overcomes this problem by using a data set that contains both the identities of charity managers and the charity’s financial information. This allows me to track charity managers across different charities and to present the first causal evidence of the effect of charity managers.
Since the objective functions of charities are so heterogeneous, the question then arises as to how we should measure a successful manager. Ideally we would use some outcome measure that the charity is trying to achieve (reducing poverty, reducing homelessness, alleviating unemployment, etc.), but the effect of a particular charity on these outcomes is extremely difficult to isolate. Instead, I use measures of financial performance for the charity, which is a common proxy to measure a charity’s success. In estimating the objective function of charities, Steinberg (1986) found that most charities were either total budget maximizers or service maximizers. This suggests that we can use the total size of the budget (total revenue) and the amount spent on providing programs and services (program expenditures) as two measures of the charity’s objective function. From this we can estimate the manager’s ability by estimating the extent to which the manager influences these measures.
The empirical analysis uses a fixed effects widely used in the labor literature and popularized by Abowd, Kramarz, and Margolis (1999) (referred to from here forward as AKM). This method has been used to measure worker productivity in many settings, including for-profit firms and schools. The AKM method requires movement of workers across firms, which is where the data used for this paper is especially useful. The analysis uses a data set that contains financial information over a span of 23 years for a panel of 460 religious-based social service charities that are all connected to a national organization. The managers are moved across different charities over time, allowing us to observe more than 1,100 different managers that move between different charities. The movement of managers is a policy followed strictly by the organization and co-ordinated by the national head office. While this policy for the movement of manager means that the change of managers is nonrandom, most other studies that use the AKM method with more typical employer–employee matched data face a similar but slightly different problem in that decentralized workers choose jobs for nonrandom reasons, so nonrandom assignment is common in the literature.
I find that a one standard deviation increase in manager estimated ability leads to a 0.516 standard deviation increase in total revenue, which amounts to more than $650,000, while a one standard deviation in manager estimated ability leads to a 0.635 standard deviation increase in program expenditures, which amounts to more than $600,000. Replacing an average manager with a good manager raises total revenue by 0.139 standard deviations, which translates to a nearly $200,000 increase in yearly revenue, while program expenditures rise by 0.155 standard deviations, which amounts to more than $150,000. I then show which manager characteristics are associated with high-ability managers. I find evidence that both married couples and individuals are effective managers, depending on the measure used as the outcome. I find that female managers are of higher ability than male managers for both measures.
I then estimate how much of the variance is explained by the charity effect, the manager effect, and the covariance between them and find that the charity effect accounts for a significant amount of the total variance in the outcome measures, and that the covariance between the manager and charity effects is large and negative. I further explore the extent to which these results are due to match effects between managers and charities. The results suggest that these match effects are negative for both total revenue and program expenditures, suggesting that managers’ effectiveness varies across charities. Finally, I estimate a model that allows for the effect of the manager to persist beyond their tenure at a given charity. These estimation results suggest that the baseline specification tends to overestimate the effect of the managers by nearly 20 percent and highlights the need to take into account the longer-term effect of managers on charity outcomes.
This work makes several key contributions to the literature on charities. First, I build an extensive data set that matches information about the finances of a charity, characteristics of their managers, and information about the neighborhoods in which they operate. This data set allows me to use a framework developed in other studies of leadership to estimate the causal impact of the manager on a charity’s finances. Second, I propose an addition to the standard method used in the labor literature that allows me to measure the lasting impact of a manager on a charity’s finances. These provide insights into the internal factors that affect a charity’s performance.
The paper proceeds as follows: Section II describes the data used for the empirical analysis, and Section III presents the empirical framework and details the identification strategy. Section IV presents the baseline results. Section V discusses extensions to the baseline model, and Section VI concludes.
II. Data
A. Sample
For reasons discussed in more detail in Section III, the empirical strategy requires managers to be observed at more than one charity. In order to address this need, I select a sample of charities in Canada that all belong to a single national organization. This allows us to follow managers across different charities with a great deal of certainty that they are the same manager, which is of vital importance for proper identification of the manager’s ability.
The organization used for the analysis is a national religious-based social service organization and is one of the largest providers of social services in Canada. The organization has a national head office, as well as seven regional head offices and more than 600 charities across Canada. These centers provide a relatively homogeneous set of social services, including church activities, food banks, counseling, homeless shelters, and/or other social services. On average, charities report that 50 percent of their emphasis (in terms of time and resources) is placed on religious services, with the remaining 50 percent composed of social services, suggesting that these locations are not simply churches providing services to their own members.3
There is a national director of the organization and regional directors at each regional head office. As this organization is also a church, the majority of directors in the sample are clergy, and most of these clergy are married (that is, both spouses are clergy). Some of the individual managers are clergy as well, and some are nonclergy employees.4 The clergy are employed by the national head office, which enables the head office to move them among charities. This last feature of the organization’s structure is the key that allows me to use this data for the purposes of this analysis and allows for identification of the effect of both clergy and nonclergy managers. I will only use the local charities in the data set, as regional and head offices perform mostly administrative tasks to support the local centers.
One of the main obstacles to studying charity leadership in the past has been the inability to find data that contains information about both the charity’s manager and the charity’s finances. Using data from three different sources, I am able to construct a data set that not only matches charity managers and charities, but also contains detailed information about the neighborhoods in which the charities are operating.
The first source of data comes from charity information returns (T3010) filed by Canadian charities with the Canada Revenue Agency (CRA) each year. In order to maintain its charitable status, each charity must file an information return every fiscal year.5 The T3010 contains detailed breakdowns of revenues, expenditures, assets, and liabilities, as well as location information for the charity.6 Importantly, the information return also requires charities to provide the names of the directors and managers of the charity. From these names I am also able to identify the gender of the manager.7 I am also able to identify if the manager is an individual or a married couple that comanages the charity.
The second data source is the Canadian census, which is collected every five years. I use measures from the 1991, 1996, 2001, 2006, and 2011 censuses and linearly interpolate these measures for the years between the census years. The census measures are constructed at the neighborhood level, which is defined by the first three characters of the postal code, known as the forward sortation area (FSA). Each FSA has around 8,000 households on average. The measures used from the census are designed to help control for the socioeconomic characteristics of the neighborhood in which the charity operates.
The final source of data is an internal publication from the organization and contains detailed information about the locations of each charity, as well as the manager assigned to that charity for the year. These data are mainly used for verification purposes to ensure that the physical location of the charity is accurate and that the assignment of managers to charities is correct.
The financial data from the T3010 spans the years 1992–2014, but the form only includes a section for the information on managers from 1997 onwards. Using supplemental data obtained from the organization, I am able to match managers to charities for the period 1992–1997, meaning that the final data set of managers and financial information spans 1992–2014. The original sample contains 608 different charities. Of these, 70 are removed because they are not social service providers (such as administrative arms), 52 are removed because they do not have at least three years of data in which the manager can be identified, and 26 are removed because they never have a manager that has been at another charity. This leaves a final sample of 460 charities, with a total of 8,909 observations and 1,103 managers.
It is worth noting here what is meant when I refer to one manager in the analysis. As mentioned previously, some of the managers are married couples. Since most of these married couples are only ever observed together at the same charity, it is not possible to separate their individual effects. For this reason, I treat married couples as one manager, so each manager will either be an individual or the couple as a whole. It is also important to note that those that I may identify as an individual manager may in fact be married, but their spouse is simply not also a manager. If the couple is at any point assigned to separate charities they are each treated as new managers for these purposes.
B. Summary Statistics
Table 1 shows the summary statistics for charities. Panel A shows the summary statistics for the full sample used for analysis, where all financial figures are in 2014 dollars. The mean charity has $762,000 in total revenue over the sample period, with a minimum of $34,800 and a maximum of more than $10 million, and a variance of more than $1.3 million. On average, total revenue consists of private donations (21.1 percent), revenue from government (42.5 percent), revenue from other charities (10.7 percent), and the remaining 25.7 percent from various other sources, such as sales and investment income. The average charity has program expenditures of $555,000, meaning that the average charity spends approximately 73 percent of their total revenue on providing the charitable good.8 We can see that the average charity has a little more than five different managers, with an average tenure of slightly fewer than four years.
Summary Statistics for Charities
Panel B reports the summary statistics when we remove the top 10 percent of charities in terms of average total revenue. The mean total revenue has fallen by nearly 50 percent, with the mean now $394,500 for total revenue and $271,400 for program expenditures. The standard deviations have also fallen substantially to $369,700 for total revenues and $275,000 for program expenditures. For the manager movement statistics, they are exactly the same as for the full sample. So while there are obvious monetary differences among the charities in the bottom 90 percent, the manager movement patterns appear to be nearly identical.
Panel C of Table 1 reports summary statistics for the neighborhoods in which the charities operate that are used as control measures in the analysis. This shows that the neighborhoods in which the charities operate are diverse in measures of population size, income, education, and other characteristics. I include information about the unemployment rate, the percent of the population who are immigrants, and the percent who are low income to try to control for the size of population that is potentially served by the charities.
Turning to statistics about the managers and changes in leadership, we can see in Table 2 that the average manager appears for almost eight years in the data set, is observed at 2.2 different charities, and stays at a charity for four years. On average across all charities, 22.7 percent of managers change charities in a given year. At the extremes, the longest tenured manager at a single charity is at that charity for 21 of the 23 years in the sample, and the manager who has managed the most charities has run eight different charities. The statistics are then broken down by the observable characteristics that I have for the managers. First, we know whether the manager is a married couple or an individual. Of the 1,103 managers in the sample, 782 are married couples and 321 are individuals. On average, the married couples are in my sample for longer, spend more time at each location, and are at more locations than individuals. This comes from the fact that married couples are observed twice as long on average as individual managers (9 years vs. 4.5 years). There are also stark differences in the size of the charities that individuals manage, with the total revenue and program expenditures being twice as large for charities managed by individuals versus those that are managed by married couples. In Columns 4 and 5, the individuals are then broken down further by gender, which displays further differences. The most striking difference is that, on average, males manage charities that are nearly twice as large as the charities managed by females in terms of both total revenue and program expenditures.
Summary Statistics for Managers
These summary statistics show that there is great variation across charities in the financial measures, as well as many changes in leadership for the charities over the sample. The next section outlines the empirical strategy and details how changes in charity managers can be used to identify the effect of the managers on these financial measures.
III. Empirical Framework
To test whether charity managers have an effect on the financial outcomes of the charity I use the AKM method, which estimates the following:
(1)
Yjt
is the financial measure of interest for charity j in period t, αt
are year fixed effects, γj
are charity fixed effects, Xs are characteristics of the neighborhood in which the charities operate, λ are manager fixed effects, and rjt
is the error term. In this case, the manager effect is properly identified if it is uncorrelated with the error term. However, it is possible that the error term contains a charity–manager match effect μMatch
such that
In this case, the manager effect will be biased if it is correlated with either component of the error term. In Section V.A I estimate the full model in Equation 1 with rjt replaced by μMatch + εjt and estimate all three components (charity, manager, and match effects), using the method found in Woodcock (2015) and Jackson (2013). The other type of bias is tested for in Section III.B, where I test to see whether changes in the outcome variables are associated with changes in managers, which is to test whether the manager fixed effect is correlated with the error terms.
For now we focus on estimating the baseline model found in Equation 1 using the AKM method. In order to separately estimate λm and γj we require that there are managers who move between different charities. Specifically, Equation 1 is estimated over a connected group of charities and managers. A clever analogy from Mansfield (2015), adapted to this paper’s context, explains the connected groups in the following way: imagine the group as a connected graph where charities are vertices and managers are edges. Any two charities in the group are connected by a manager that has been at both charities; any two managers in the group are connected by a charity where they have both been managers.
Within these groups, the charity and manager fixed effects are overparameterized. In most cases the way this is addressed is by leaving out an arbitrary manager fixed effect. This does not work in this context since the manager fixed effects are the parameter of interest, and all other fixed effects are interpreted as deviations from whichever fixed effect is left out. Leaving out an arbitrary fixed effect will produce greatly different results depending on which fixed effect is omitted. Following McCaffrey et al. (2012), I restrict the mean of the manager fixed effects to be zero within each group, so the fixed-effect estimates are interpreted as deviations from the mean manager of each group. Of the managers in the sample, 97 percent are connected in a single group. Following the earlier referenced Mansfield (2015) example, this means that there is a single group of managers (edges) and charities (vertices) that connects 97 percent of managers in the sample. Following Card, Heining, and Kline (2013), I only use this large connected group for the estimation, meaning that the manager fixed effects are not estimated for the other 3 percent of the sample. All estimates are reflected as deviations from the mean manager of this single group.
A. Endogenous Manager Changes
A possible threat to the identification of the manager fixed effects is the presence of endogenous changes in managers. The concern about endogenous mobility comes from the way in which the manager fixed effects are identified. If a manager is observed at more than one charity, then their ability can be compared to the ability of the other managers at those charities. The main cause for concern with endogenous manager movement is for a manager to move to or from a charity that is about to either boom or bust. This would alter the manager’s ability to influence the financial outcomes and thus would bias the estimates of the manager fixed effects. I test for the possibility of this in Section III.B.
It is important to note that certain types of movements of managers that may seem endogenous do not bias the estimates of the effect of the managers in this setting. An example of this is if managers systematically move from worse charities to better charities or vice versa. This is because the true quality of the charity is already accounted for in the charity fixed effect. Therefore, the manager fixed effect will still be measuring the true quality even if the managers are moving from worse to better charities.
Managers who are of higher or lower ability moving more often is also not cause for concern for similar reasons. Note, however, that this is only the case for time-invariant characteristics of the charity that are captured by the charity fixed effect. For example, if a particular charity is able to grow revenues more quickly, this will not be picked up by the charity fixed effect and may bias the results. This is what I test for in Section III.B.
To better understand how the changes in managers may be of concern it is helpful to know more about the institutional details of the organization. Each regional director of the organization is responsible for determining if a charity in their region is to undergo a change in leadership. The regional director makes this decision with input from the current manager of the charity, and if the charity is a church, input from the church’s membership. While some changes happen throughout the fiscal year, the vast majority of changes occur the last week of June or the first week of July, depending on the year.9
Once the regional director decides that a particular charity will change leadership, the list of managers that are to be moved from those charities is sent to the national head office. There, the human resources department and the senior leadership team at the national head office decide where the managers will move. This matching of vacancies with available managers also involves flows into and out of the pool of potential managers. Normal transitions in and out of the pool of potential managers include retirements, new hires, managers who take positions at charities outside of Canada with a related organization, and managers who take administrative positions within the national or regional head offices.10
It is likely the case that the movement of managers between charities is not truly random. Informal conversations with those involved in the process of assigning available managers to available positions suggest there is evidence that the organization makes an attempt to assign managers to charities that they feel would be a good match. Whether a good match for the organization can be measured as a good match in terms of financial outcomes is not inherently clear. This attempt to match managers to charities further motivates the analysis in Section V.A, which tests for the presence of match effects; this is, however, a separate issue from the endogenous movement of managers tested here.
Knowing what we know about the process of deciding which managers are to be moved, it is not clear if we should be concerned about endogenous changes of managers biasing the estimate of the manager fixed effects. Nevertheless, I will still test for endogenous changes in the next section.
B. Testing for Endogenous Manager Changes
If changes in managers are not exogenous, the estimates of the effect of the managers may be biased. As discussed above, one potential threat to estimating the effect of managers properly is that managers move just before or after a charity is about to boom or bust. If we can predict the changes in managers using the leads and lags of changes in the financial outcomes used for analysis, then this might suggest that the changes in managers are related to booms and busts for the charities, which might bias the results. Table 3 shows the estimates for whether a charity in year t+1 has a different manager than in year t, estimated using OLS. As covariates I use changes in both total revenue and program expenditures, and their lags, as well as all of the control variables. The change is calculated as the percent change from year t − 1 to year t, and the first lag is the percent change from year t − 2 to year t − 1. I separate out positive and negative changes to allow for the differential effect that these may have on the probability of a manager moving.
Predicting Changes in Leadership
The first column of Table 3 shows the results using only the current changes and no lag. We can see that a 1 percent increase in total revenue has no meaningful effect on the probability of a manager moving, while a 1 percent decrease in total revenue increases the probability of a manager moving by 0.18 percentage points. For program expenditure, a 1 percent increase in program expenditure is associated with a 0.03 percentage point increase in the probability of a manage moving, while a 1 percent decrease in program expenditure is associated with a 3.6 percentage point decrease in the probability of a manager moving.
In Column 2, a one-period lag is added for each measure to try to account for possible patterns in the years before a change in leadership takes place. The coefficients on the current year changes are all very similar to the first column and have similar significance levels as before. For the lags, the coefficient for an increase in total revenue is both small and statistically insignificant, while a 1 percent decrease in total revenue in the previous year is associated with a 0.37 percentage point decrease in the probability a manager moves. For program expenditures, the coefficient on the first lag for a positive change is very small (0.01 percentage points), and for a negative change it is statistically insignificant.
For both specifications, the R2 is less than 0.01, meaning that less than 1 percent of the variation in these changes in managers can be accounted for by changes in the outcome measures in the years surrounding the movement of managers. This provides evidence that the changes in managers are not strongly affected by booms or busts in total revenue or program expenditures and that the estimates of the effect of the manager are likely not biased from the possibility of endogenous movements of managers across charities.
IV. Baseline Results
The baseline results are presented in Table 4 and Figure 1. The coefficients of interest are associated with the manager fixed effects. Given the number of observations, it would be impractical to report the estimates for every fixed effect, so instead I only report the standard deviation, adjusted standard deviation, different percentiles of the distribution, and the difference between the 75th percentile and the median. I also report the p-value from the F-test of the joint significance of all of the manager fixed effects. Recall the fixed effects are recentered so that the mean is zero.
Kernel Density Distributions for Manager Fixed Effects
Baseline Estimates of Manager Fixed Effects
To account for sampling error in the estimates of the manager fixed effects I adjust the standard deviation following the method in Aaronson, Barrow, and Sander (2007) to better reflect the true variance in the manager effects. Assume that the estimated manager fixed effect, 
, is composed of the true manager fixed effect λM
and normally distributed independent disturbance with mean zero, ε, such that 
. Since the two components are independent, it follows that the variance of the manager effect is 
. To obtain the true variance of the manager effect, we need to subtract the variance of the sampling error term from the measured manager effect variance. Since 
is not observable, I estimate it by finding the mean of the square of the standard errors of 
, which is the estimated variance. The reason I adjust for the sampling error is to try to get rid of some of the randomness associated with the revenues and expenditures.
Columns 1 and 2 of Table 4 report the estimated manager fixed effects for the full sample. Total revenue and program expenditures have been normalized to mean zero and variance one, so these fixed-effects estimates are interpreted in standard deviations. Focusing on the adjusted standard deviation, we can see that for total revenue, a one standard deviation increase in manager ability leads to a 0.516 standard deviation increase in total revenue, while replacing an average manager with a good manager (the difference between 75th percentiles and the median) increases total revenue by 0.139 standard deviations. For program expenditures, a one standard deviation increase in manager ability leads to a 0.635 standard deviation increase in program expenditures, and that replacing an average manager with a good manager raises program expenditures by 0.155 standard deviations. The p-value for the F-tests of the joint significance of all manager fixed effects strongly rejects the hypothesis of no manager fixed effects for both total revenue and program expenditures.
These results show that a manager can have a significant impact on all of these different measures of revenue and expenditure, but how large are these estimates in terms of dollars? Moving from an average manager to a good manager results in an increase of nearly $200,000 in total revenue, which represents more than 25 percent of the mean total revenue for charities. For program expenditures, moving from an average manager to a good manager raises program expenditures by approximately $150,000.
It’s possible that most of these results are being driven by the extremely large charities. From Table 1 we see that the largest charity averages more than $10 million in total revenue, but that the average is much lower at around $762,000. To account for this, I remove the top 10 percent of charities in terms of total revenue and perform the analysis with the bottom 90 percent of charities. Columns 3 and 4 of Table 4 present these results. The adjusted standard deviation for total revenue has not changed from the full sample estimates, while it has fallen from 0.635 standard deviations to 0.535 standard deviations for program expenditures. The difference between an average manager and a good manager has more than doubled for both total revenue and program expenditures. While the estimates themselves have either not changed or risen, the standard deviations of both total revenue and program expenditures for this sample is much smaller than it was for the full sample ($1.3 million versus $370,000), so the corresponding dollar values are smaller.11 Moving from an average manager to a good manager raises total revenue by $100,000 and program expenditures by $93,000 in this sample.
Next I wish to see whether there is any correlation between the managers’ effects when using total revenue and program expenditures. If the correlation is negative, then this would suggest there are different objectives that the managers are trying to maximize (as suggested by Steinberg 1986). If the correlation is high and positive, then that would suggest that estimates using either measure reveals the manager’s ability. Figure 2 presents plots of the manager fixed effects for both total revenue and program expenditures, along with the correlation. Panel A shows the plot for all managers, with a correlation of 0.792. Panels B and C show the plots separately for married couples and individuals, with a smaller correlation for individuals (0.748) than for married couples (0.846). Finally, Panels D and E show that the correlation is stronger for females (0.866) than for males (0.823). Overall, the correlation is quite strong for the whole sample. This suggests that there are not different styles in managers, but that both measures reveal the manager’s ability.
Correlation of Fixed Effects
A. Variance Decomposition
Following Card, Heining, and Kline (2013), in this section I show the amount of variation in Equation 1 that can be explained by each of the components. This variance decomposition shows the magnitude of the effect of the manager, the charity, and the covariance between them. This can be written as the following:
(2)
Table 5 shows the estimates for the sample analogue for each component in Equation 2. For both total revenue and program expenditures, the share of the total variance explained by the charity effect is very large: 94.2 percent for total revenue and 97.5 percent for program expenditures. The share of the total variance explained by the manager effect is much smaller for both total revenue and program expenditures. Of note is that the covariance between the manager and charity effects is large and negative (−12.5 percent for total revenue and −20.4 percent for total expenditures), suggesting that “good” managers are placed at “bad” charities. This will be explored more in Section V.A.
Variance Decomposition
B. Patterns in Leadership
The last section showed that the managers who have a high estimated ability determined using total revenue tend to have a high estimated ability determined using program expenditures as well. This section explores whether we see differences in estimated abilities using characteristics of the managers. There are two characteristics of the manager available in the data that we can use in order to help understand what attributes are associated with a good manager. The first is whether the manager is a married couple or an individual. Second, for the individual managers, the gender is identified from their first name. Using these two characteristics of the manager I plot the densities of the distribution of fixed-effect estimates for each outcome. Figure 3 plots the distributions by whether the manager is a married couple or an individual, and Figure 4 plots the distributions by gender. In addition, I perform the the Kolmogorov–Smirnov test for equality of distributions for each outcome measure, and Table 6 reports the p-values from these tests.
Kernel Density Distributions for Manager Fixed Effects for Individuals vs. MarriedCouples
Kernel Density Distributions for Manager Fixed Effects by Gender
Kolmogorov–Smirnov Test p-Values
From Figure 3 we first observe that most of the fixed effects are concentrated around the mean of zero. We can also note that from the plots of the distributions it is not obvious that the distribution of fixed effects for married couples and individuals are different. Instead we turn to Panel A of Table 6. The null hypothesis for the Kolmogorov–Smirnov test is that each sample is drawn from the same distribution (in other words that the distributions are equal), and the p-values are reported in the table. Looking first at the test for the equality of distributions, we reject the hypothesis that the distributions are equal for both total revenue and program expenditures. The second test tests whether the distribution for one group contains smaller values than for the other group. In this case, I report both the test for whether managers who are a married couple have smaller values than managers who are individuals and vice versa. For total revenue we reject the hypothesis that married couples contain smaller values and fail to reject the hypothesis that individuals contain smaller values. This suggests that married couples have a higher estimated ability when using total revenue. For program expenditures, while the test of equality of distributions is rejected, the test for whether married couples have smaller values and the test for whether individuals have smaller values both are rejected. This means that there are differences in the estimated abilities of managers, but that neither group contains larger or smaller values than the other.
Figure 4 shows that the distribution of manager fixed effects is more noticeably different for gender than it was for married couples vs. individuals. It is not clear, however, which managers have larger values for the manager fixed effects. For each outcome the fixed effects for female managers have a larger number centered at the mean of zero, but from the graphs it is not apparent that there are patterns in the tails of the distribution. In order to analyze this we turn again to the Kolmogorov–Smirnov test. Panel B of Table 6 shows that for both total revenue and program expenditures the hypothesis that the distributions are equal is rejected, the hypothesis that females contain smaller values is strongly rejected, and that we fail to reject the hypothesis that males have smaller values. This shows that females are better managers than males no matter which outcome we use to measure the manager’s ability.
These figures and tests suggest that there is no clear pattern for whether managers who are married couples are of higher ability than those who are individuals. However, for individual managers it is clear that the managers who are female are of higher ability than those who are male. It is important to note, however, that these should only be considered correlations, and no causal implications should be taken from these findings—there are likely unobservables affecting the correlations found here, and it is difficult to interpret the causal effects of gender in any context.
V. Extensions
A. Match Effects
Part of what is being measured in the baseline specification may be match effects between the charity and the manager. These match effects may arise if managers are particularly effective at certain charities (positive match effects), if managers are less effective at different charities (negative match effects), or if there is no change in the manager’s effectiveness across different charities (no match effects). Jackson (2013) and Woodcock (2015) present a way to measure the match effects, but this method requires a great deal of data. In other literature that uses the AKM method with employee–employer matched data, there are often millions of observations that allow for accurate estimation of all three effects (person, firm, and match). In the case here, we only have thousands of observations, not millions. Nevertheless, I present the results from this estimation, with the results to be interpreted cautiously.
The method uses a hybrid random effects model that uses a fixed-effect estimator to first estimate the following model:
(3)
The residuals from this estimation are then used in a maximum likelihood model, where the manager, charity, and match effects are treated as random effects, and their variances are estimated. Technically, they could be estimated using a fixed-effects model, but as noted by Woodcock (2015), the fixed-effects model is overly restrictive when identifying the match effects, and the hybrid random effects model is identified with fewer restrictions on the orthogonality of the manager, charity, and match effects.
The results from this estimation are found in Table 7. Panel A shows the baseline estimates of the manager and charity effects without match effects. Panel B shows the results from this estimation that includes match effects. The charity effect has decreased slightly for both total revenue and program expenditures. What is most striking, however, is the size of the manager effect in the model with match effects, which has decreased by more than 80 percent for both total revenue and program expenditures. This suggests that much of the effect that was attributed to the manager is actually a result of the match between the manager and the charity. In fact, the size of the match effects are nearly three times as large as the manager effect in this model. Again, these results should be treated with caution because the method used for them usually requires a great deal of data and the estimates are consistent asymptotically. At the very least, combined with the variance decomposition shown in Section IV.A, they show that the match between the charity and manager is an important part of the effects measured.
Match Effects
B. Persistence of the Effect of Managers
Managers may influence outcomes for the charity beyond their tenure at the charity. For revenues, managers can establish contacts with individuals who commit to give a regular donation or can apply for multiyear government grants, both of which may last beyond the time when the manager leaves the charity. Likewise, for expenditures, the manager can enact programs that continue after they leave the charity. To account for this persistence I estimate the following model using fixed effects:
(4)where α, γ, and X are all the same as in the baseline model (Equation 1), t′ is the time when the manager leaves the charity, s is the time since the manager has left the charity (either three or four in this case), and β is a linear decay. This allows for the effect of the current manager λm,t to be separated out from the effect of the past manager λm,(t-t′).
Table 8 reports the baseline results as well as the results from using a linear decay of three and four years. For total revenue, there is a slight decline in the adjusted standard deviation in both the three-year and four-year decay from the baseline results, from 0.516 in the baseline to 0.442 (three years) and 0.431 (four years). The same pattern is seen for program expenditures: the baseline results overestimate the impact of the manager, with the adjusted standard deviation falls from a baseline of 0.635 to 0.581 (three years) and 0.571 (four years). This suggests that the baseline specification overestimates the effect of the manager by attributing some of what the past manager did to the current manager. The results from this specification suggest that accounting for the persistence of the effect of the managers is important for estimating their ability. Since the managers are choosing policies that may affect the charity in the long term, it is not surprising that we see their effect persist beyond their tenure at a charity.
Persistence of the Effect of Managers
VI. Conclusion
This paper provides the first causal evidence of the effect that charity managers have on the finances of their charity. Using data that match managers with financial information from charities, I find that one standard deviation increase in manager ability leads to a 0.516 standard deviation increase in total revenue and 0.635 standard deviations for program expenditures.
Relating these fixed effects to characteristics of the managers shows that there are significant differences in the distribution of the fixed effects for managers who are married couples versus individuals, and also for male versus female managers. Married couples are shown to be higher-ability managers when using total revenue. When comparing the male and female managers, the female managers are shown to be of higher ability for both total revenue and program expenditures. There may be a number of reasons for this finding, including self-selection into this profession.
Extending the baseline model to allow for match effects between managers and charities, I find that once accounting for the charity–manager match effects, the effect of the manager drops by more than 80 percent. A second extension allows for the effect of the manager to persist beyond the tenure of a manager at a charity. In particular, I find that the effect of the managers are smaller for total revenue and program expenditures than in the baseline model. These results show that the baseline model can overestimate the effect of the manager by nearly 20 percent when not taking the persistence into account. This persistence is likely to be seen in other settings where leaders set policies, such as principals and CEOs, and should be taken into account in future studies in those contexts.
These two extensions suggest that while the baseline model has proved useful for studying leadership in other settings, in this setting the match effects and persistence of manager effects need to be taken into account to measure the ability of the manager properly. The match effects are particularly striking in how much they decrease the manager effect once they are included. This suggests that in the setting used in this study, the interaction between the manager and the charity effect is particularly important. Future work should examine whether these effects are important for other types of charities (for example, education or environmental charities) and in other settings (for example, nonreligious organizations).
There may also be concerns about the generalizability of the results, given the particular setting in which the effect of the managers is studied. One concern might be that the effects we are seeing are driven by the fact that donors to the charities may be in large part congregation members for the church-based locations, and they are also consuming the services. This may be true in certain cases, but on average, only 20 percent of the revenue comes from private donations, and results using total revenue less private donations produce similar estimates.12
We may also be concerned that because these charities are religious based, the managers and/or charities may be different from the average manager or charity in a number of ways. However, these differences likely make the estimates of the effect of charity managers a lower bound for a number of reasons. First, we likely see more bad managers continue in the sample than in the typical charity because of the nature of the organization. There is also a barrier to entry for many of the manager positions at charities, as some require the manager to be an ordained member of the clergy. Second, the manager’s pay is based solely on their years of experience and is not tied to financial performance in any way. Since there are no financial incentives for the manager, then the estimates are likely smaller than they would be for managers who have pay tied to their performance.
For future research on the effect of charity managers in a nonreligious setting, the largest obstacle to overcome is finding a source of matched employee–employer data where the same manager can be tracked between charities. One potential source of this may be scraped LinkedIn data that are matched to publicly available charity tax filings. As Edelman (2012) highlights, use of online sources such as publicly available LinkedIn data may become increasingly more easily accessible.
Footnotes
The author thanks Abigail Payne, Katherine Cuff, Michael Veall, Justin Smith, Ross Hickey, Iryna Khovrenkov, Nicolas Lawson, Thomas Downes, as well as participants of McMaster University’s seminar series, USC Price School of Public Policy’s seminar series, Melbourne Institute’s seminar series, the Canadian Economics Association 2016 Annual Meetings, Canadian Public Economics Group 2016 Annual Meetings, and the National Tax Association 2016 Annual Conference for their helpful comments. All remaining errors are those of the author. This research has been funded with a grant from the Social Science and Humanities Research Council of Canada. Data are available from https://seal.mcmaster.ca/download-data.
↵1. For a summary of previous work on the many motivations for charitable giving see Andreoni (2006) and Andreoni and Payne (2013).
↵2. See Bertrand and Schoar (2003); Bloom and Van Reenen (2007); Malmendier and Tate (2009); and Goldfarb and Xiao (2011) for evidence on CEOs and Eberts and Stone (1988); Clark, Martorell, and Rockhoff (2009); Branch, Hanushek, and Rivkin (2012); Coelli and Green (2012); and Dhuey and Smith (2014) for evidence on school principals. There is also an extensive field of research that studies teachers.
↵3. These data come from the T3010 form described below. The measures are self-reported areas of emphasis for the charity.
↵4. Unfortunately, I am not able to identify those who are clergy and those who are nonclergy in my data.
↵5. The key benefit to charitable status is the ability to issue tax receipts for donations.
↵6. While the true location for each charity is confidential and not provided in the data, a mailing address is provided. For the sample used in the analysis the mailing address has been verified using publicly available address information and corrected to the physical location when necessary.
↵7. While names do not always identify gender, I verify the genders using supplemental information from the organization.
↵8. Program expenditure is one component of total expenditures, with the others being management and administration, fundraising, political activities, and gifts to other charities.
↵9. The fiscal period ends in March for each charity, so the leadership change happens three months into the fiscal year. The financial data contain only a year by year breakdown, so it would not be possible to identify which part of the revenues or expenditures can be attributed to which manager. Since the new manager is at the charity for nine months of the fiscal period, and especially during the busiest time for charities around the end of the calendar year, I make the decision to treat the charity as being under the leadership of the new manager for the entire fiscal period.
↵10. The pay for the managers is solely determined by the number of years of experience at the organization; there is no performance pay available to managers, so managers would have no incentive to try to move to good charities for pay reasons.
↵11. Also note that the bottom 90 percent has less than 90 percent of the charities of the full sample (460 in full sample and 409 in the 90 percent). This comes from the construction of the connected group of charities and managers.
↵12. Results are not displayed here but are available from the author upon request.
- Received May 2017.
- Accepted June 2019.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.