Financial Markets and Human Capital Unit 02: Empirical Methods Ernst Maug University of Mannheim http://cf.bwl.uni-mannheim.de ernst.maug@uni-mannheim.de Secretary: Angelika Wolf-Tobaben cf.secretary@uni-mannheim.de Tel: +49 (621) 181-1951 Difference-in-differences analyses →What is a diff-in-diff analysis? - A statistical method to analyze the causal effect of a treatment on an outcome variable. →Idea: Emulate a randomized control trial (RCT) as in medicine, psychology, etc. - For example, test a new drug against a certain type of cancer - Give the drug to 100 patients (“treated” group) - Give a placebo to another 100 patients (“control” group) e.x.: standard medicine for cancer - Nobody knows in which group they are, nobody can select - Both groups are as similar as possible in terms of gender, age, history of diseases, etc. (hence, “random”) - Assumption: if the “treated” group were to receive the placebo, they would develop exactly like the control group © 2022 Ernst Maug Financial Markets and Human Capital: Internal Labor Markets 2 Quasi-random experiments →How would we do this in business, finance, economics - Have planning authority assign firms to “treatment” and “control” group - Make sure treated and control group are the same, no self-selection! - Prescribe treatment (e.g., merger, divestiture, investment,…) - It simply can’t be done… →Solution: Exploit quasi-experiments, events with differential impact on otherwise comparable samples: βͺ Labor protection legislation in some states but not in the others. βͺ Cash windfalls to some firms but not to the others. βͺ Failure of main banks for some firms but not for the others →Measure difference between “treated” group (subject to event) and “control” group (not subject to event) © 2022 Ernst Maug Financial Markets and Human Capital: Internal Labor Markets 3 Difference-in-differences: recall study on AWS shock from unit 1 →Compare the treated-control difference - before the event to - after the event - hence, difference (after minus before) of differences (treated minus control) →Study on impact of AWS shock on capital requirements of start-ups - Treated group: firms with high computing requirements (proxied by membership in 8 industries) - Control group: firms with lower computing requirements (proxied by membership in 18 industries) - Before = before 2006 (year Amazon Web Services became available to start-ups) - After = after 2006 →Compute: πππππ‘ππ πΆπππ‘πππ - Difference in start-up capital before 2006: βπΆππππ‘πππ΅πππππ = πΆππππ‘πππ΅πππππ − πΆππππ‘πππ΅πππππ πππππ‘ππ πΆπππ‘πππ - Difference in start-up capital after 2006: βπΆππππ‘πππ΄ππ‘ππ = πΆππππ‘πππ΄ππ‘ππ − πΆππππ‘πππ΄ππ‘ππ - Difference in differences: © 2022 Ernst Maug DD = βπΆππππ‘πππ΄ππ‘ππ − βπΆππππ‘πππ΅πππππ Financial Markets and Human Capital: Internal Labor Markets 4 Diff-in-diff in a graph →Hypothesis: if the AWS shock had not occurred, start-up capital requirements in the treated group would have developed exactly as those in the control group (remember RCT!) →This does not mean they would have been the same. →Only that their developments would have continued on the same paths as before →Assumption: deviations from pre-AWS path of treated group can be attributed to AWS shock Outcome variable: Start-up capital AWS becomes available Control group Treatment effect Treatment group Timeline before 2006 © 2022 Ernst Maug Financial Markets and Human Capital: Internal Labor Markets after 2006 5 Regression analysis →This can be analyzed using standard OLS regressions →Let πΆ - πΆπππ΅ππ : capital requirements of control (C) group before (t<2006) AWS shock π - πΆπππ΅ππ : capital requirements of treated (T) group before (t<2006) AWS shock πΆ - πΆπππ΄ππ‘ : capital requirements of control (C) group after (t≥2006) AWS shock π - πΆπππ΄ππ‘ : capital requirements of treated (T) group after (t ≥ 2006) AWS shock →Then - πππππ‘ππππ‘ πΈπππππ‘ = βπΆπππ − βπΆπππΆ π π πΆ πΆ = πΆπππ΄ππ‘ − πΆπππ΅ππ − πΆπππ΄ππ‘ − πΆπππ΅ππ © 2022 Ernst Maug Financial Markets and Human Capital: Internal Labor Markets 6 Diff-in-diff regressions E-C= beta difference = conterfactual change B-A = actual change in control D-E = delta treatment effect ( or change of treated- change of control) D-C = beta+ delta = change in the treated in step 4 after =1 and treatment = 1 but after*treatment = 0 > pretend that the shock had no impact bay setting condition 2 > hypothetical →Run regression: πΆπππ‘π = πΌ + π½ × π΄ππ‘πππ‘π +πΎ × πππππ‘ππππ‘π‘π +πΏ × π΄ππ‘πππ‘π × πππππ‘ππππ‘π‘π +π’π‘π where: dummy variables (0,1) π΄ππ‘πππ‘π = 1 if t ≥ 2006, zero otherwise πππππ‘ππππ‘π‘π = 1 if π = high internet exposure firm, zero otherwise AWS becomes available B →Then the coefficient estimates can be interpreted as follows: πΌ: average of control group before AWS shock π½: average change in control group (placebo effect) πΎ: average difference treated minus controls before AWS shock πΏ: treatment effect → Advantage: flexibility - can add further control variables and fixed effects to regression E π½ Control group A πΎ C D Treatment group before 2006 © 2022 Ernst Maug Financial Markets and Human Capital: Internal Labor Markets πΏ πΌ after 2006 7 Diff-in-diff regressions: What the coefficients mean (supplemental material) πΆπππ‘π = πΌ + π½ × π΄ππ‘πππ‘π +πΎ × πππππ‘ππππ‘π‘π +πΏ × π΄ππ‘πππ‘π × πππππ‘ππππ‘π‘π +π’π‘π →Subsamples - Subsample A: Control group before the shock (After = 0, Treatment = 0) βͺ πΆπππ‘π = πΌ + 0 + 0 + 0 + π’π‘π - Subsample B: Control group after the shock (After = 1, Treatment = 0) βͺ πΆπππ‘π = πΌ + π½ + 0 + 0 + π’π‘π - Subtract A from B: π − π = πΌ + π½ − πΌ = π½ = Impact of After on controls. - Subsample C: Treated group before the shock (After = 0, Treatment = 1) βͺ πΆπππ‘π = πΌ + 0 + γ + 0 + π’π‘π - Subtract A from C: C − A = πΌ + πΎ − πΌ = πΎ = Difference treated minus controls before - Subsample D: Control group after the shock (After = 1, Treatment = 1) βͺ πΆπππ‘π = πΌ + π½ + πΎ + πΏ + π’π‘π - Subtract C from D: π − π = πΌ + π½ + πΎ + πΏ − πΌ + γ = π½ + πΏ = Impact of After on the treated. - Treatment effect = (D – C) – (B – A): π½ + πΏ − π½ = πΏ =Impact of After on treated minus Impact of After on controls © 2022 Ernst Maug Financial Markets and Human Capital: Internal Labor Markets 8 Regression results All VCs Treatment × After πΏ Treatment πΎ Syndicate size Startup based in CA Startup based in MA Startup based in NY After π½ (1) -0.294*** (0.0663) -0.0293 (0.0504) 0.323*** (0.0200) 0.438*** (0.0773) 0.465*** (0.0797) 0.276*** (0.0743) 0.0741* (0.0435) (2) -0.269*** (0.0673) -0.0338 (0.0504) 0.325*** (0.0197) 0.424*** (0.0766) 0.453*** (0.0792) 0.290*** (0.0740) 0.179*** (0.0665) AWS becomes available β = 0.179 Control group γ = -0.338 Treatment group before 2006 δ = -0.269 α after 2006 / ** / *** = p values smaller than 10% / 5% / 1% (2) includes more control variables than (1) numbers refer to regression (2) two/three stars, depends on the number of beta hat* #standard deviation away from beta Standard deviations are in parentheses, as defined in the event study part. is the 0 inside or outside of the confidence interval From Ewens, Nanda, and Rhodes-Kopf (2017), Table III © 2022 Ernst Maug if we are in > we cannot reject H0 (more or less equal to 0) if we are out > we can reject H0 ("significantly different" so not equal to 0) the greater the number of set > teh smaller p > the more precise (smaller confidence interval ) Financial Markets and Human Capital: Internal Labor Markets 9 Postscript and summary importance of good estimates and good research design during the assignement →Event studies are a special type of diff-in-diff analysis - Outcome variable = stock return - Exogenous shock = event - Treated units = events - Control units = market index →Event studies and diff-in-diff analyses allow us to evaluate observational data as quasi-random experiments - Event / shock needs to be exogenous (no self selection) - Treated and controls have to be similar except for the event / shock →Powerful tools, but need to be handled with care! © 2022 Ernst Maug Financial Markets and Human Capital: Internal Labor Markets 10