class: center, middle, inverse, title-slide .title[ # Misclassification in Difference-in-Differences Models ] .author[ ### Augustine Denteh and Desire Kedagni
] .date[ ### Comments from: Brantly Callaway
Southern Economics Association Conference, 2022 ] --- # Introduction `$$\newcommand{\E}{\mathbb{E}} \newcommand{\E}{\mathbb{E}} \newcommand{\var}{\mathrm{var}} \newcommand{\cov}{\mathrm{cov}} \newcommand{\Var}{\mathrm{var}} \newcommand{\Cov}{\mathrm{cov}} \newcommand{\Corr}{\mathrm{corr}} \newcommand{\corr}{\mathrm{corr}} \newcommand{\L}{\mathrm{L}} \renewcommand{\P}{\mathrm{P}} \newcommand{\independent}{{\perp\!\!\!\perp}}$$` <style type="text/css"> border-top: 80px solid #BA0C2F; .inverse { background-color: #BA0C2F; } .alert { font-weight:bold; color: #BA0C2F; } .alert-blue { font-weight: bold; color: blue; } .remark-slide-content { font-size: 23px; padding: 1em 4em 1em 4em; } .highlight-red { background-color:red; padding:0.1em 0.2em; } .assumption-box { background-color: rgba(222,222,222,.5); font-size: x-large; padding: 10px; border: 10px solid lightgray; margin: 10px; } .assumption-title { font-size: x-large; font-weight: bold; display: block; margin: 10px; text-decoration: underline; color: #BA0C2F; } </style> <br><br><br> <center><span class="alert">I totally agree with authors that misclassification is interesting/understudied in the context of DID</span></center> --- # Main comment about Parallel Trends My only substantive comment is that I was unsure why the parallel trends assumption (Assumption 1) used the observed treatment status rather than "true" treatment status, i.e., why: `$$\E[\Delta Y_t(0) | D=1] = \E[\Delta Y_t(0)|D=0]$$` `$$\textrm{rather than}$$` `$$\E[\Delta Y_t(0) | D^*=1] = \E[\Delta Y_t(0)|D^*=0]$$` -- To me: this seems like a surprising choice to me and asymmetric from the definition of `\(ATT := \E[Y_t(1) - Y_t(0) | D^*=1]\)`. -- I'm not sure how much this would change the conclusions in the paper (maybe not at all), but I encourage authors to think some about this case --- # Minor Comments 1 I think there is some <span class="alert">positive news</span> and perhaps it is worth emphasizing this more in the paper * Empirical researchers often have the intuition that misclassification `\(\implies\)` attenuation bias * The paper provides <span class="alert">precise conditions</span> under which misclassification leads to attenuation bias. To be fair, as emphasized in the paper, these are strong conditions, but <span class="alert">this is helpful for empirical researchers to know.</span> * Also, if you think of four groups divided by actual treatment status and misclassification status, it's <span class="alert">not sufficient that all four groups experience the same trends in untreated potential outcomes</span> (Assumption 2.2), you also need that misclassification is independent of treated potential outcome conditional on true treatment status --- # Minor Comments 2 The paper emphasizes "mechanical misclassifications" - The example of using a proxy to define treatment (e.g., a location) seems like a good one to me and seems common in applications - I have a harder time following the examples about differences between the timing of a policy being voted on and when it is actually implemented. - To me: this either has the flavor of anticipation - Or many applications are roughly: a researcher observes yearly data, but a policy was implemented in the middle of the year. Should that year be counted as treated or untreated? Is that misclassification or something else? --- # Minor Comments 3 * Important to be clear about what is different in this case relative to papers about misclassification with cross-sectional data and under unconfoundedness - Good idea to make clear that the contribution of the paper (which it is!) is more than just replacing `\(Y\)` (in levels) with `\(\Delta Y\)` (in differences) - Perhaps can emphasize more some features of DID: e.g., no one treated in the first period, targeting ATT, etc.