class: right, middle, inverse, title-slide .title[ # Lecture 4 - Causal Exaggeration ] .subtitle[ ##
Topics in Econometrics ] .author[ ### Vincent Bagilet ] .date[ ### 2024-10-02 ] --- class: right, middle, inverse # Exercise ## Treatment effect heterogeneity --- class: titled, middle # Simulating treatment effect heterogeneity - How did you model heterogeneity? - Did your model represent the true DGP? What happens if it does not? - Questions about calibration? - Bonus Question: how to proceed to keep track of all analyses (with varying DGP)? --- class: titled, middle # For next week - Calibration exercise - Any comments on the exercise? - Is there anything you would like me to discuss before the end of next class? --- class: right, middle, inverse # Summary from last week --- class: titled, middle # Summary from last week - Simulations: varying parameters and complexifying - Fire alarm break - Low statistical power leads to exaggerated significant estimates --- class: right, middle, inverse # Causal Exaggeration --- # The Confounding-Exaggeration Trade-Off .pull-left[ - Challenge of empirical economics: identifying causal effects - Only use part of the variation: the exogenous part - Reduces precision and statistical power - `\(\mathbb{E}[\hat{\beta}] = \beta\)` but `\(\mathbb{E}[\hat{\beta} | \text{ Significant}] \neq \beta\)` ] -- .pull-right[ <img src="data:image/png;base64,#images/intuition_trade_off.png" width="600" style="display: block; margin: auto;" /> ] --- # Unconfounded but Inflated - Problem: `\(\exists\)` **publication bias** -- - Can explain part of the replication crisis in economics -- - In economics, nearly 80% of estimates are exaggerated by a factor of 2 (!!) - Exaggeration could be on par with OVB -- - Trade-off confounders/exaggeration: **avoiding a bias creates another type of bias** -- - Shared mechanism (causal strategies limit variation) but specific channels -- - Thinking of causal stragtegies as control approaches ties all channels together --- class: titled, middle # A Potential Example of Exaggeration <img src="data:image/png;base64,#images/graph_he_th_annotated.png" width="900" style="display: block; margin: auto;" /> --- # Intuition for controls and exaggeration - Let `\(y_{i} = \alpha + \beta x_{i} + \delta w_{i} + u_{i}\)`, thus `$$\sigma_{ovb}^2 = \dfrac{\sigma^{2}_{u_{ovb}}}{n \ \sigma_{x}^{2}} = \dfrac{\sigma^{2}_{y^{\perp x}}}{n \ \sigma_{x}^{2}} \qquad \text{and} \qquad \sigma_{ctrl}^2 = \dfrac{\sigma^{2}_{u_{ctrl}}}{n \ \sigma_{x^{\perp w}}^{2}} = \dfrac{\sigma^{2}_{y^{\perp x, w}}}{n \ \sigma_{x^{\perp w}}^{2}}$$` -- - Controlling `\(\nearrow\)` variance (and exaggeration) if the fraction of the variance unexplained by `\(w\)` is smaller for `\(x\)` than for `\(y^{\perp x}\)` - `\(\Leftrightarrow\)` if it absorbs more of the variation in `\(x\)` than in the residual part of `\(y\)` ( `\(y^{\perp x}\)`) --- # RDD - **Identification approach**: only uses observations inside the bandwidth - Economics of education: impact of additional lessons on test scores <img src="data:image/png;base64,#images/DAG_RDD.png" width="500" style="display: block; margin: auto;" /> - `\(Final_{i} = \beta_{0} + \beta_{1} T_i + \eta Qual_{i} + \delta W_i^{3} + u_{i}\)` - `\(Qual_i = \mu + \gamma W_i^{3} + \epsilon_i\)` --- <img src="data:image/png;base64,#images/main_graph_RDD-1.png" width="800" style="display: block; margin: auto;" /> --- # IV - **Identification approach**: only uses the part of the variation in `\(x\)` explained by the instrument - Political economy: impact of turnout on vote share <img src="data:image/png;base64,#images/DAG_IV.png" width="700" style="display: block; margin: auto;" /> - `\(Turnout_{i} = \pi_{0} + \pi_{1} Rain_{i} + \gamma w_{i} + e_{i}\)` - `\(Share_{i} = \beta_{0} + \beta_{1} Turnout_{i} + \delta w_{i} + u_{i}\)` --- <img src="data:image/png;base64,#images/main_graph_IV-1.png" width="800" style="display: block; margin: auto;" /> --- # Exogenous shocks - **Identification approach**: only uses exogenous variation from shocks - Impact of air pollution (plant closure) on birthweight <img src="data:image/png;base64,#images/DAG_DiD.png" width="700" style="display: block; margin: auto;" /> - `\(bw_{z,t} = \beta_{0} + \beta_{1} T_{z, t} + \zeta_z + \tau_t + u_{z,t}\)` --- <img src="data:image/png;base64,#images/main_graph_shocks-1.png" width="800" style="display: block; margin: auto;" /> --- # Fixed Effects - **Identification approach**: only uses the remainin variation after partialling out fixed effects - Impact of high temperature on worker productivity <img src="data:image/png;base64,#images/DAG_fe-1.png" width="600" style="display: block; margin: auto;" /> --- <img src="data:image/png;base64,#images/burn.png" width="800" style="display: block; margin: auto;" /> --- <img src="data:image/png;base64,#images/main_graph_fe-1.png" width="800" style="display: block; margin: auto;" /> --- class: titled, middle # Gauging Risks of OVB and Exaggeration - In essence impossible to measure - For OVB: **sensitivity analyses** (e.g. Cinelli and Hazlett, 2020) - Power calculations: not only useful to know if we will be able to detect an effect - **Prospective simulations**: simulate the DGP - Helps think about effect sizes, relation between the variables, etc - **Retrospective calculations**: retrodesign(beta, se) --- class: titled, middle # Driver of the Trade-off .pull-left[ - Exaggeration `\(\Rightarrow\)` variance matters even **after** a significant estimate has been obtained - Barely significant result: sign of low power and potential exaggeration - Evaluate the variation used for identification - Visualize where the variation comes from ] .pull-right[ <img src="data:image/png;base64,#images/starwars.png" width="90%" style="display: block; margin: auto;" /> ] --- class: titled, middle # Structural Approaches - Without publication bias this issue disappears - Abandoning the 5% significance threshold - Interpretation of CI’s width to embrace uncertainty - Replication of studies with similar designs --- class: titled, middle # Summary - Lack of power and significance filter in economics - Recipe for exaggeration - What are the drivers? - This paper: **causal methods can exacerbate exaggeration** - Addressing the issue: - Power calculations - Highlighting where the variation comes from --- class: right, middle, inverse # Calibrating simulations --- class: titled, middle # Why calibrating? -- - Make simulations more realistic - But simulations will never be truly realistic - But can still allow to run some sort of **robustness check** - Again, --- layout:true # Fake data simulations --- ### Distributions of the variables - Emulate the distribution of variables in existing data sets <img src="data:image/png;base64,#images/distrib_logyield-1.png" width="600" style="display: block; margin: auto;" /> --- class: titled, middle ### Relationships between variables - Read the literature - Get a sense of **typical effect sizes and of relationships** between variables - Make assumptions on those relationships. Acknowledge them. - Complexify later if needed. You choose when you stop - Varying parameters values might change the distribution of "variables" (*eg* the error term) --- layout: false class: titled, middle # Real data simulations ### General approach - Start from an existing data set - Not yours. At least not the subset you are interested in - Try to pick a subset where there is not already a treatment effect - Define a treatment allocation mechanism - Add an artificial treatment effect to the outcome variable in our initial data set ( `\(y_i = Y_i(0)\)` ) `$$Y_i(1) = Y_i(0) + \beta_i T_i$$` --- # Real data simulations ### Complexifying - Only one artificial aspect: the treatment - We can play on only 2 components: -- .pull-left[ - **Who is treated?** *Treatment allocation* - Everyone - Only a subset of the population ] .pull-right[ - **How?** *Treatment effect* - Homogenous - Heterogenous but random - Some specific correlation structure ] --- class: right, middle, inverse # Summary --- class: titled, middle # Punchlines - Credibility revolution + convincing identification methods *BUT* replication failures - When power is low, significant estimates **are always** far from the true effect - The bias/variance trade-off can in fact be a bias/bias trade-off - **Variance matters**, even once a significant estimate has been obtained - Calibrate your simulation based on existing information ??? - Even if the distribution of our causal estimator is centered around the true effect, the distribution we actually draw estimates from may not be. --- class: right, middle, inverse # Thank you!