Clustered Standard Errors 1. The Attraction of “Differences in Differences” 2. Grouped Errors Across Individuals 3. Serially Correlated Errors1. The Attraction of Differences in Differences Estimates  Typically evaluate programs which differ across groups, such as U.S. States e.g., effect of changes in state minimum wage laws or state welfare programs on earnings or unemployment  Treat selection (heterogeneity) bias by removing state effects (one “diff”)  Treat common economic fluctuations by removing year effects (the other “diff”)  Hence the appealing nickname “diffs in diffs”2. The Grouped Error Problem:  Binary covariates define groups within which errors are potentially correlated (e.g., cities, states, years, states after treatment, self-employed, etc..) - remember that errors contain unobserved variables  Yist = Ast + Bt + cXist + βIst + εist ,  s are groups (perhaps states)  t is time  I is an indicator for treatment, which occurs as the group x time level  ε is an error term, which is not necessary iid.2. Grouped Errors Across Individuals E.g., Minimum wages on NJ/Penn border  Card and Krueger (1994) looked at the effects of minimum wages on employment in fast-food restaurants near the NJ – Penn border.  Data collected before and after NJ raised its‟ minimum wage by 80 cents (in 1992).  i - restaurant, s – state, t – time  S=2, T=2, N is large.  They found small positive effects within a small confidence interval of zero.2. Grouped Errors Across Individuals E.g., Mariel Boatlift  Card (1990) looked at the effects of a surprise supply shock of immigrants to Miami due to a temporary lifting of emigration restrictions by Cuba in 1980.  He estimates the effect of the boatlift on unemployment and wages of low skill workers in Miami using four other cities as comparisons (Atlanta, Houston, LA and Tampa-St. Petersburg) with CPS data.  i - individual, s – city, t – time  S=5, T~=2, N is large.  He finds no statistically significant effect on employment or wages of the labor supply shock.2. Grouped Errors Across Individuals  How big does the number of groups (S, or S*T) have to be?  Yist = ast + dt + cZist + βIst + εist ,  Donald and Lang (2004): In the (plausible) case where we have some within-group correlation, and under generous assumptions the t-statistics converge to a normal distribution at rate S*T no matter what N is.  Intuition: Imagine that within s,t groups the errors are perfectly correlated. Then you might as well aggregate and run the regression with S*T observations.  Intuition: 2 step estimator  If group and time effects are included, with normally distributed group-time specific errors under generous assumptions, the t-statistics have a t distribution with S*T-S-T degrees of freedom, no matter what N is. (Table 3)  Donald-Lang suggested estimator has this flavor. (Table 3)  Alternative: collapse into s,t groups  3 issues: consistent s.e., efficient s.e. and distribution of t-stat in small samplesDistribution of t-ratio, 4 d.o.f, β = 0 When N=250 the simulated distribution is almost identical3. Correlations over time in panels  Yist = Ast + Bt + cXist + βIst + εist ,  S are groups (perhaps states)  t is time  I is an indicator for treatment, which occurs as the group x time level  Correlations within group, period (i.e., s,t) cells only is very restrictive.  In general we want to allow correlations over time as well (within s but not within t)Lots of DD papers T is large The variables tend to be serially corr. So are std. errors consistent?Placebo Binary “Laws”  Randomly choose a year between 79-99 & randomly assign a law to 25 states til end of 99  Rej. rate is % for which t>1.96Placebo Binary “Laws”  Type I error is worst when T is largeSolutions: AR(1) correction • N=50, T=21 • AR(1) biased for small T • Process looks more like AR(2)Solutions: Ignore TS Information • correct size but loss of power • Residual aggregation is a Frisch-Waugh exercise: first - regress on other variables, then - aggregate residuals before and after treatmentSolutions: “Cluster” within states (over time) • simple, easy to implement • Works well for N=10 • But this is only one data set and one variable (CPS, log weekly earnings) -Current Standard Practice  Be conservative: cluster by group or time (not the interaction) and report the larger std. error - note: this may get size and power wrong  Better.. you can cluster on both! Cameron, Gelbach, and Miller (2006, NBER Technical WP) method not coded in Stata yet, but you can get an .ado from Doug Miller‟s Stata page http://www.econ.ucdavis.edu/faculty/dlmiller/statafiles/  Do you have enough groups for a normal approximation? .. Check with a “Wild Bootstrap” Cameron, Gelbach, Miller (ReStat 2008); .do file on Miller„s page.  May be argument for using Newey-West std. errors.  Ask Gordon Dahl, who is working on a better methodExam ?  Wed Dec 7 in Granger room,