Empirical Methods for Corporate Finance. Panel Data, Fixed Effects, and Standard Errors

Transcription

1 Empirical Methods for Corporate Finance Panel Data, Fixed Effects, and Standard Errors

2 The use of panel datasets Source: Bowen, Fresard, and Taillard (2014) 4/20/2015 2

3 The use of panel datasets Source: Bowen, Fresard, and Taillard (2014) 4/20/2015 3

4 Basics (As said earlier) There are three main causes of trouble (violation of E(u x) = 0 in OLS) in empirical corporate finance Endogenous explanators Mismeasured explanators Omitted factors (Refresher) Some variable(s) that the econometrician does not observe may be correlated with the dependent and explanatory variable(s) Mainly due to the considerable (unobserved) heterogeneity present in many corporate finance settings Firms can differ in so many ways! (think about it for one second) Panel data can sometimes offer a partial (neither complete nor costless) solution to this problem 4/20/2015 4

5 What is a panel dataset? In panel data, individuals (persons, firms, cities, countries, ) are observed at serveral point in time (days, year, before and after a treatment, ) Notatation: We observe N firms over T period If each firm (i) is observed in all periods (t), the panel is balanced Two main interests in the econometrics of panel data (1) Exploit panel data to control for unobserved time invariant heterogeneity in cross sectional models («fixedeffects») (2) Disentangling components of variance and estimating the transition probability among states (study the dynamics of cross sectional populations) («random effects») In most corporate finance applications, we rely on (1) and estimate fixed effects models 4/20/2015 5

6 General structure of a panel The T observations for firm i can be written as And NT observations for all firms and time periods as 4/20/2015 6

7 Econometric specification: the logic Suppose a cross sectional model of the form («the truth») yi 1 xi 1 ci ui 1 with E( ui 1 xi 1, ci ) 0 If c i is observed then β can be identified from a multiple regression of y on x and c If c i is not observed, causal identification of β requires Lack of correlation between x i1 and c i («random effect») Cov( xi1, yi1) Var( x ) i1 An instrument (z i ) uncorrelated with u i1 and c i Cov( zi Cov( z 1 i, y, x i1 i1 ) ) 4/20/2015 7

8 Econometric specification: the logic Suppose that neither of these two options is available but we observe y i2 and x i2 for the same firm in a second period (T=2) y i2 xi2 ci ui2 with E( uit xi1, xi, 2, ci ) 0 Then β is identified in the regression in first difference even if c i is not observed y y ( xi2 xi1) ( ui,2 u 1) i2 i1 i Taking the first difference eliminate the time invariant unobserved heterogeneity! And Cov( x i2 Var( x, y i2 ) i2 ) The variation that identifies β is within firm variation 4/20/2015 8

9 Illustration This is the line we would fit if we do not account for individual effects (Pooled OLS on a large cross section) This is the «true» unbiased slope (that account for the individual effects) Estimation with pooled OLS is biased and inconsistent! 4/20/2015 9

10 Why might fixed effects arise? Any time invariant individual characteristic that cannot be observed in the data at hand could contribute to the presence of fixed effects In regressions aimed at understanding firm behavior, specific sources of fixed effects depend on the application In capital structure regression (e.g. leverage) a fixed effect might be related to Unobserved technological differences across firms Unobserved ability of the CEO or people deciding on financial policies In general, a fixed effect can capture any low frequency unobserved variable 4/20/

11 Fixed effect models We focus here on static fixed effect models (no lagged dependent variables). The basic model is: ' yit xit ci uit witht 1,...,T Assumption (A1): E(ui xi,ci ) 0 (t 1,...,T) The error at any period is uncorrelated with past, present and future values of x (strict exogeneity assumption) E.g. current values of x are not influenced by past errors Assumption (A2): Var(u x i i,c i ) σ 2 I T Errors are conditionally homoskedastic and not serially correlated Treatment with heteroskedastic is relatively easy (e.g. available in Stata) 4/20/

12 Estimation with fixed effects (1) Under A1 (mean independence), we can estimate the model in first difference by OLS (convenient) OLS estimator will be unbiased and consistent (for large N) This is rarely the case (Cov( u it, u it-1 )isnot zero!) If Cov( u it, u it-1 )isnot zero, the optimal estimator is given by generalized leastsquares (GLS) Take deviation from the mean! Estimate the following (within or FE) specification ~ y it ~ x u~ ' 1 it it with ~ yit yit T T t 1 y it y it y The variation that identifies β is within firm variation Note that the firm specific effect (c i ) dissapears (constant) Time invariant hetegeneity is solved 4/20/

13 Estimation with fixed effects (2) The FE estimator is asymptotically normally distributed so that the usual OLS inference can be applied The usual tests (t, Wald, ) can be used in large and small samples In practice, the error is often likely serially correlated (T >2). This needs to be corrected (using cluster robust standard errors) Note that time invariant regressors (e.g. the constant) cancel so their effect cannot be estimated with a within estimator The FE estimator is numerically identical to pool OLS including a set of N-1 dummy variables which identifies the firms and hence N-1 parameters («Least Squares Dummy Variables Estimator» LSDV) Get estimates for c i The LSDV estimator is generally not consistent as the number of parameter goes to infinity F test for the joint significance of fixed effects (this is useful) 4/20/

14 Alternative: random effects In the FE specification, the firm specific effect (c i ) is allowed to be correlated with the explanatory variables (Note that this is the root of the problem!) If the firm specific effect (c i ) is uncorrelated with the explanatory variables (past, present and future), we have a random effect (RE) specification In RE specification, the emphasis is on the error term The error term has two components (c i +u i ) that are unrelated to the regressors No problem to identify β (identified in the cross section) but the panel structrue is used to identify the variance of c i and u i Used to separate out permanent from transitory components of variation Not many application of RE models in corporate finance 4/20/

15 Fixed or random effects? Intuitively, we should opt for a RE specification if one can be sure that the firm specific effect really is unrelated to the explanatory variables This can be tested using a (Durbin Wu ) Hausmann test Comparision of the FE and RE estimators: H ˆ ˆ ˆ( ˆ ˆ( 2 ( IV OLS )'[ V IV ) V OLS )]( IV OLS ) ~ J ˆ ˆ ˆ Where J is the number of time varying regressors The null hypothesis is that the firm specific effect (c i ) is uncorrelated with the regressors and the errors are equicorrelated 4/20/

16 Implementation in Stata (1) 4/20/

17 Implementation in Stata (2) 4/20/

18 Implementation in Stata (3) 4/20/

19 Practical issues (1) Most of the empirical corporate finance use fixed effects. Is that fine? The answer is not obvious. The use of FE deserves careful thoughts Always try estimations with fixed and random effects and check for statistical significance of the FE Check if the inclusion of fixed effects changes the magnitude of the coefficients in an economically meaningful way The inclusion of fixed effects reduces efficiency Even if the Hausman test rejects the null of random effects, if the economic significance is little changed, using OLS can still be valid E.g. Lemmon, Roberts and Zender (JoF 2008) 4/20/

20 Practical issues (2) Warning: Including fixed effects can exacerbate measurement problems If the dependent variable is a first differenced variable (e.g. investment or change in cash holdings) and if the fixed effect is related to level of the dependent variable, then the fixed effect has already been differenced out of the regression Again, using fixed effect reduces efficiency E.g. fixed effects rarely tend to make important qualitative differences in investment equations If the research question is aimed at understanding cross sectional variation in a variable, then fixed effects defeat this purpose 4/20/

21 Coles and Li (2011) 4/20/

22 4/20/

23 4/20/

24 4/20/

25 4/20/

26 4/20/

27 4/20/

28 Graham and Leary (2012) (By traditional determinants) 4/20/

29 Graham and Leary (2012) 4/20/

30 Graham and Leary (2012) 4/20/

31 Application: Managing with style Bertrand and Schoar, 2003, Managing with style: The effect of managers on firm policies (Quarterly Journal of Economics) Question: How much do individual managers matter for firm behavior and economic performance? Motivation: Previous studies focused on firm, industry, and market level characteristics to explain corporate behaviors View in the business press is that (certain) managers are key factors in corporate practices (e.g. Steve Jobs) Can add to our understanding of corporate finance Empirical challenge: Identify/measure the (marginal & causal) impact of managers on firms Empirical strategy: Use managers fixed effects to uncover their impact on various policies (no causal statement here) Results: Managers traits appear to be related to firms choices and performance 4/20/

32 Location in the field Governance Real decisions Financing Valuation Institutional framework: laws, regulations, taxes, markets, macro economy No causal effects here 4/20/

33 Why should manager matter? (in theory) No effect (Null hypothesis) In a purely neoclassical view of the firm, managers are homogenous and selfless inputs into the production process (remember your microeconomic classes ) Managers can differ (preferences, risk aversion, skills, ) but do not affect corporate policies Non zero effect (managers matter) Agency models (principal agent) allow for managers having discretion inside the firm (depend on governance practices only) Some models allow managers to differ Agency perspective (1): Managers may matter if the corporate control is poor or limited (improved) Neoclassical perspective (2): Firms (e.g. boards) select specific managers (endogenous matching) Both approaches predict that managers matter! 4/20/

34 Sample construction Construction of a manager firm matched panel dataset that tracks different managers across different firms and time (why is this key?) Data sources Forbes 800 files ( ) and Execucomp ( ) for the information about CEO (and other top executives) Focus on CEO but also CFO, COO and subdivision CEOs Restrict on the subset of firms for which at least one executive can be observed in at least one other firm (for minimum three years) The resulting sample contains about 600 firms and 500 managers 4/20/

35 Summary statistics (representative?) Firms in the sample are somewhat special (e.g. larger) due to data screening Executives from larger firms are more likely to move between COMPUSTAT firms (be in the sample) The effects documented in the paper may not generalize to smaller (private) firms 4/20/

36 Job transition A large majority of job moves are from «other» to «other» «other» corresponds to operationally important positions 117 CEO moved to another CEO position in another firm! (fired?) 4/20/

37 Panel specification The (FE) specification is as follows (on the CEO firm matched sample): The objective is the get an estimate of the different λs for different y (policies) They use the LSDV estimator (clear why?) They are not after causality (non random allocation of managers across firms) 4/20/

38 Managers fixed effects (investment) CEO fixed effect increase the Adjusted R 2 by 3% in the investment equation This represents the average marginal contribution of CEO fixed effects CEO fixed effect are jointly significant (F Tests are large ) 4/20/

39 Managers fixed effect (financing) CEO fixed effect has a rather small effect in leverage equation CEO fixed effect are jointly not significant (F Tests are super small) 4/20/

40 Managers fixed effects (performance) CEO fixed effects are significant in performance equations Controlling for many different things, some CEOs are related to better performance 4/20/

41 Magnitude of the managers fixed effects? Retrieve the estimated managers fixed effect from the LSDV estimation (e.g. in Stata) Some CEO are negatively related to performance! Important heterogeneity across managers 4/20/

42 Management styles? Can we detect systematic «styles» among managers fixed effect? (Use simple OLS regression) We do see consistent patterns (e.g internal vs external investment or debt vs cash) 4/20/

43 Fixed effects and compensation Managers fixed effect appear to be related to their compensation! Firms appear to pay a premium for managers who are associated with higher rates of return on assets! 4/20/

44 Managers traits? 4/20/2015 Empircal Corporate Finance 44

45 Conclusion/comments CEO appear to matter CEO Systematically related to corporate policies This has generated a lot of research on WHY do CEO matter (some explanations) Overconfident CEO Connected CEO (political, school, investment bankers, private equity) Financial expertise Original use of fixed effects to answer an interesting question (not a causal question though) This type of approach has been used by others in the literature 4/20/

46 Think about Standard Errors! Petersen (2009)

47 Influencial Research?

48 Potential Problem? In panel data, residuals may be correlated across firms and across time (not iid) Why? Consequences? OLS standard errors (hence t stats) can be biased May lead to wrong conclusion Let s see what happens and what can de done 4/20/

49 Pooled OLS Simple panel model (firm i, time t) Assume Cov(X, ε)=0 IID 4/20/

50 Unobserved Firm Fixed Effect Residual becomes: and Correlation within firms but not between firms Continue to assume independence between X and ε

51 True Standard Errors (OLS)? OLS understates true standard errors if ρs are not zero! + Increases in T

52 True Residual Structure

53 How bad it this? Simulation of a panel dataset with a firm fixed effect Repeat this several time to observe true SE SE of y=1 and SE of ε=2 R 2 =20% Change the fraction of the variance of y that is due to the firm fixed effect (0% 75%) Change the fraction of the variance of ε that is due to the firm fixed effect (0% 75%)

54 True SE OLS SE OLS SE cluster

55 Time effect? Petersen (2009) execute a similar analysis for Time fixed effect (cross firm residual correlation) Firm and Time fixed effect Temporary Firm fixed effect (decay) Bottom line: Everything potentially create BIASES

56 Application: Capital Structure

57 Take Away In the presence of firm fixed effect OLS SE are biased SE clustered by firm are unbiased In the presence of a time fixed effect OLS SE are biased Fama McBeth SE are unbiased In the presence of both Include time dummies and cluster by firm Double cluster by time and firm Learn about the structure of the data

58 Stata and SAS code and help /htm/papers/standarderror.html