Corporate Governance Indices and Construct Validity

Bernard Black is Nicholas D. Chabraja Professor at Northwestern University Pritzker Law School and Kellogg School of Management; and B. Burcin Yurtoglu is Professor and Chair of Corporate Finance at WHU—Otto Beisheim School of Management. This post is based on a recent paper by Professor Black; Professor Yurtoglu; Antonio Gledson de Carvalho, Assistant Professor at Fundação Getúlio Vargas School of Business at Sao Paulo; Vikramaditya Khanna, the William W. Cook Professor of Law at the University of Michigan Law School; and Woochan Kim is Associate Professor of Finance at Korea University Business School (KUBS). Related research from the Program on Corporate Governance includes Related research from the Program on Corporate Governance includes What Matters in Corporate Governance? by Lucian Bebchuk, Alma Cohen and Allen Ferrell (discussed on the Forum here).

A common strategy in corporate governance research is to build a corporate governance index and then see whether the index predicts firm value or performance. These indices are imperfect, but their use is widespread because researchers lack good alternatives. A major concern with governance indices is what they actually measure. The concept of governance is abstract and latent rather than concrete and observable, and we are not sure how to proxy for this vague concept using observable measures. This raises concerns about the degree to which the proxy (a governance index) measures what it claims to be measuring. The fit between the observable proxy or “construct” (the governance index) and the underlying concept (governance) is known as construct validity. This core issue is rarely addressed in corporate governance research. Larcker, Richardson and Tuna (2007) and Dey (2008) are exceptions.

In our paper we discuss what can usefully be said about which of the many possible governance indices are sensible constructs. We conduct an exploratory analysis of how to tackle this question, using tools drawn from the causal inference, education and psychology literatures.

The often-used Gompers, Ishii and Metrick (2003) “G” index illustrates the central role that governance indices play in corporate governance research and how central unaddressed issues of construct validity are to index construction. They create a governance index with 24 equally weighted elements that measure takeover defenses and provide evidence that this construct predicts firm value and performance. Bebchuk, Cohen, and Ferrell (2009) criticize this index and argue that only six elements, which they use to build their own “E” index, predict firm value and performance; the remainder are noise. Straska and Waller (2014) beg to differ, and report evidence that the 18 measures that Bebchuk, Cohen and Ferrell want to drop from the G index, treated as an “O” (for other) index, predict takeover likelihood. Karpoff, Schonlau and Wehrly (2016) build yet a different subset of the G-index elements, which they call the “D” index, that also predicts takeover likelihood. The confusion would be compounded if one considered takeover defense elements not in the original G index, or sought to build a broader governance index not limited to takeover defenses.

As the basis for our own analysis, we begin with our own prior work (Black et al. 2014 and 2016), in which we build governance indices in four major emerging markets (Brazil, India, Korea, and Turkey). In those studies, we argue that using a “common index” that relies on the same set of governance “elements” in each country—as massively multicountry studies typically do—is likely to yield poor constructs. As an example, consider board independence, often seen as a central component of corporate governance. Typical levels of board independence vary greatly across countries. Many Brazilian and Turkish firms have no independent directors at all. Korean firms are required to have a minimum of 25% independent directors, and Indian firms must have either a majority of independent directors or else at least one-third independent directors plus a non-executive board chair. Thus, a board structure element that asks whether a firm has one independent director is useful in Brazil and Turkey, but meaningless in India and Korea. Conversely, an element that asks whether a firm has a majority of independent directors is useful in India and Korea, but of limited value in Brazil and Turkey, where very few firms have a majority of independent directors. To use the fraction of independent directors as a governance element would also be misleading. The effect in Brazil and Turkey of firms moving from zero to one independent director may be very different from the effect in India or Korea firms moving from three to four independent directors (in these countries, a minimum percentage is required by law; Turkey added a minimum independence requirement during our sample period); or the effect from moving from a minority to a majority of independent directors.

Instead of assuming that the same elements have the same meaning in different countries, we accept that the meaning of the same element will often differ across countries. We attempt to build different constructs in each country, that are likely to proxy for similar underlying governance aspects. More specifically, we first identify a limited number of general aspects of governance, using a combination of our own judgment, the available empirical evidence and such corporate governance theory as exists: board structure, disclosure, shareholders rights, related party transactions and ownership structure. Next, for each country, we identify elements (observable variables) that are “meaningfully” related to each of the general aspects. The elements used in each country reflect a combination of local norms, local institutions and local data availability. We use these elements to build proxies for the general aspects of governance. We call these proxies “subindices.” We then build each overall country governance index (CGI) as an equally weighted average of the subindices.

How well does a particular construct (a board structure subindex in a particular country, say) represent the corresponding general aspect of governance (board structure)? We cannot assess the validity of board structure subindex, seen as a construct, simply by asking whether this subindex empirically predicts an outcome of interest (we focus on Tobin’s q). If board structure subindex predicts the outcome, it could still be a poor construct, which is measuring something else—perhaps about “governance,” perhaps not—or is simply correlated with an omitted variable which is the “true” predictor of the outcome. Board structure index could also be a useful construct, yet fail to predict the outcome because the underlying theory that posits a relationship between the general aspect (board structure) and the outcome is wrong. Therefore, predictive power is neither necessary nor sufficient, as a test for construct validity.

We pursue two approaches for assessing construct validity. First, we measure Cronbach’s α scores, both for subindices (comprised of elements) and overall indices (comprised of subindices). Cronbach’s α measures the inter-item correlation among the elements of an index. If the elements of a subindex collectively contribute to measuring the same general aspect of governance, one would expect those elements to be positively correlated and to yield a reasonably high Cronbach’s α. At the same time, overly high inter-element correlations suggest that two elements are not sufficiently distinct and are capturing the same concept. Furthermore, if subindices in fact capture distinct aspects of governance, Cronbach’s α across subindices cannot be extremely high.

Our second approach uses principal component analysis (PCA) as an alternative procedure to compute subindices. PCA consists of finding clusters (principal components) of related elements. Each component consists of a group of elements that correlate more among themselves than with other elements not belonging to that component. In this fashion, elements are aggregated according to their statistical properties rather than by prior leads from theory or previous empirical evidence. If our subindices (based on prior knowledge) are good constructs, one would expect that components will be loaded mostly or entirely of elements from a single subindex. We also perform regression analysis with firm fixed effects and extensive covariates to test the predictive power of subindices vis-à-vis components.

We find in all four countries that overall indices that are calculated as the average of subindices present reasonable construct validity. The mean correlations across subindices are moderate, suggesting that the subindices in fact capture different aspects of governance. Conversely, these correlations suggest that inference from a narrow index, a single subindex, or, even worse, a single element, likely suffers from omitted variable bias, because of the omission of important aspects of governance.

At the subindex level construct validity in often reasonable, but we find exceptions, where one has less confidence that a subindex is measuring a coherent underlying governance aspect. One can also use the construct validity analysis as a guide to how to build indices and subindices. We rely on that analysis to guide an effort to divide Board Structure Subindex into sub-subindices for Board Independence and Board Committees.

We find that regressions of Tobin’s q on principal components, while informative, are not a substitute for regressions on carefully built subindices. Instead the subindices often have greater statistical power in predicting Tobin’s q.

The full paper is available for download here.

Both comments and trackbacks are currently closed.