class: center, middle, inverse, title-slide # Applications of Present Bias ## EC895; Fall 2022 ### Prof. Ben Bushong ### Last updated September 26, 2022 --- layout: true <div class="msu-header"></div> <div style = "position:fixed; visibility: hidden"> `$$\require{color}\definecolor{yellow}{rgb}{1, 0.8, 0.16078431372549}$$` `$$\require{color}\definecolor{orange}{rgb}{0.96078431372549, 0.525490196078431, 0.203921568627451}$$` `$$\require{color}\definecolor{MSUgreen}{rgb}{0.0784313725490196, 0.52156862745098, 0.231372549019608}$$` </div> <script type="text/x-mathjax-config"> MathJax.Hub.Config({ TeX: { Macros: { yellow: ["{\\color{yellow}{#1}}", 1], orange: ["{\\color{orange}{#1}}", 1], MSUgreen: ["{\\color{MSUgreen}{#1}}", 1] }, loader: {load: ['[tex]/color']}, tex: {packages: {'[+]': ['color']}} } }); </script> <style> .yellow {color: #FFCC29;} .orange {color: #F58634;} .MSUGreen {color: #14853B;} </style> --- class: inverseMSU name: Overview # Some Comments ### On Feedback Recommended reading: *Thanks for the Feedback* by Stone and Heen. -- - Why is this so important? It's a **huge** part of the job. -- - Three types of feedback: "Coaching, appreciation, evaluation". - Learn to spot each and, if needed, steer conversations toward the one you have in mind. -- - Cultivate a "growth" identity. -- ### On Thinking in Experiments (Insert rant here) --- class: MSU # My Favorite "Evidence" <video width="448" height="256" controls> <source src="graphics/Marshmallow.mov" type="video/mov"> </video> --- class: MSU # Monetary Choices *Recipe from (Thaler, 1981a)* - Assume linear utility. - Back out discount factor using exponential model: `$$Y=\delta^{t}X\Rightarrow-\log(\delta)=\frac{\log(X/Y)}{t}$$` -- What `\(X\)` makes you indifferent between $15 today and $X in a month? - For `\(Y=15\)` and `\(X=20\)`, we get `$$\log(\delta)=\frac{\log(20/15)}{1/12}\approx\text{345% per year}$$` -- - What `\(X\)` makes you indifferent between $15 today and $X in 10 years? `$$\log(\delta)=\frac{\log(100/15)}{10}\approx\text{19% per year}$$` --- class: MSU # Note on Monetary Experiments There are a zillion such experiments; see (Frederick, Loewenstein, and O'Donoghue, 2002). -- But they are mostly bad. (Good intuitive evidence, bad actual evidence). --- class: MSU # Some Issues ### Money earlier vs. money later: many confounds See (Cohen, Ericson, Laibson, and White, 2016) for an excellent discussion of measuring time preferences. -- ### Issue #1: Hypothetical rewards - Subjects may not reveal true preferences. - Doesn't appear to matter much in practice (Johnson and Bickel, 2002). -- ### Issue #2: Unreliability of future rewards - Choose earlier reward since later rewards seem uncertain? - show effect of uncertainty. -- ### Issue #3: Transaction costs - May have to come back, fill out form, etc. to receive later reward. - (Andreoni and Sprenger, 2012a) attempt to deal with this issue. --- class: MSU # More Issues ### Issue #4: Investment vs. consumption - Subjects may perceive choice as an *investment* decision. - Put another way: money receipt at date `\(t\)` is not necessarily a utils event at date `\(t\)`. -- - Remember, discounting applies to **utility**, not money! - If agent is not liquidity-constrained, she should: 1. Maximize NPV of payments 2. Then pick indifference point of consumption using time preferences -- - NPV based on market interest rates, **not** time preferences - Some work in development economics arguing that money discounting picks up time-variation in liquidity constraints, rather than time preferences (Dean and Sautmann, 2014), (Cassidy, 2018). --- class: MSU # Money earlier vs. money later (cont) ### Issue #5: Curvature of utility function - Utility is not (necessarily) linear in money. - (Andreoni and Sprenger, 2012a) attempt to deal with this issue. -- ### Issue #6: Framing effects and demand characteristics - Menu of choices or set of questions may influence subjects' choices. - Procedures may bias subjects' responses by implicitly guiding their choices. --- class: MSU name: section2 # Evidence for Present Bias - (Non-evidence) Monetary choices: (Thaler, 1981a) - **Preference reversals:** (Read and van Leeuwen, 1998); (Read, Loewenstein, and Kalyanaraman, 1999a) - Convex time budgets using real effort choices: (Augenblick, Niederle, and Sprenger, 2015) - Demand for commitment: (Kaur, Kremer, and Mullainathan, 2015); --- class: MSU # Preference Reversals ### Recall: preference reversals are a direct prediction of quasi-hyperbolic models -- *What follows is from (Read, Loewenstein, and Kalyanaraman, 1999a)* - Choose among 24 movie videos - Some are "low brow" : *Four Weddings and a Funeral* - Some are "high brow" : *Schindler's List* -- (Yeah, I know. Not my labeling nor film choices.) -- - Picking for tonight: 66% choose low brow. - Picking for 7 days from now: 37% choose low brow. - Picking for 14 days from now: 29% choose low brow. --- class: MSU # More Evidence *from (Read and van Leeuwen, 1998)* <img src="graphics/ReadVL1.png" width="550px" style="display: block; margin: auto;" /> --- class: MSU # Patient Choices for Future <img src="graphics/ReadVL2.png" width="550px" style="display: block; margin: auto;" /> --- class: MSU # Impatient Choices for Today <img src="graphics/ReadVL3.png" width="550px" style="display: block; margin: auto;" /> --- class: MSU # Time Inconsistency <img src="graphics/ReadVL4.png" width="550px" style="display: block; margin: auto;" /> --- class: MSU # Aside: Evidence of Projection <img src="graphics/ReadVL5.png" width="550px" style="display: block; margin: auto;" /> --- class: MSU # Summary ### Recapping the Paper - Evidence of time inconsistency - But hard to estimate parameters ( `\(\beta,\delta\)` ) using this type of data -- - Potential confound: Can uncertainty about future preferences explain this result? If so, how? (see Strack and Taubinsky (2021) for more.) --- class: MSU name: section3 # Evidence for Present Bias - (Non-evidence) Monetary choices: (Thaler, 1981a) - Preference reversals: (Read and van Leeuwen, 1998); (Read, Loewenstein, and Kalyanaraman, 1999a) - **Convex time budgets using real effort choices:** (Augenblick, Niederle, and Sprenger, 2015) - Demand for commitment: (Kaur, Kremer, and Mullainathan, 2015); --- class: MSU # About the Paper - Measure time preferences using real-effort tasks **and** monetary amounts - Evidence of present bias for effort choices - Weak present bias for monetary choices... as in (Andreoni and Sprenger, 2012a). -- - Provide commitment technology and elicit demand for commitment - Significant demand for commitment for effort choices (at 0 price) - Link time preference estimates to demand for commitment - Estimated present bias predicts subsequent demand for commitment. --- class: MSU # About Multiple Price Lists (MPLs) - Following (Thaler, 1981a), MPLs were used for many years to estimate time preferences. - Choices between smaller, sooner rewards and larger, later reward. <img src="graphics/Andreoni1.png" width="550px" style="display: block; margin: auto;" /> -- - Impose linear utility over money, and ask subjects to solve: `$$\max_{c_{t},c_{t+k}}U(c_{t},c_{t+k})$$` `$$\text{s.t.}\quad((1+r)c_{t},c_{t+k})\in\{(m,0),(0,m)\}$$` --- class: MSU # Double Multiple Price Lists *from (Andersen, Harrison, Lau, and Rutström, 2008)* ### Restriction to corner solutions problematic with non-linear utility -- -Price list switching point reveals `\(u^{\prime}(c_{t})\approx\delta^{k}u^{\prime}(c_{t+k})\)`. -With linear utility function, no problem: `\(\delta_{L}\approx\left(\frac{c_{t}}{c_{t+k}}\right)^{1/k}\)` -With curvature, should instead estimate: `\(\delta_{C}\approx\left(\frac{u^{\prime}(c_{t})}{u^{\prime}(c_{t+k})}\right)^{1/k}\)` -Upward bias ( `\(\delta_{C}-\delta_{L})\)` ); overestimate discount rates. -MPLs generally yield very (too) high discount rates. --- class: MSU # Double Multiple Price Lists - Approach to take curvature of utility into account - Separately elicit risk preferences - Using both time and risk price lists, jointly estimate curvature and discounting parameters --- class: MSU # Convex Time Budget: *from (Andreoni and Sprenger, 2012a)* - Different approach: "convexify" experimental budgets - Allow for interior solution directly; ask subjects to solve: `$$\max_{c_{t},c_{t+k}}U(c_{t},c_{t+k})$$` `$$\text{s.t.}\quad(1+r)c_{t}+c_{t+k}=m$$` -- - Simple dynamic budget constraint; can estimate curvature of utility <img src="graphics/Andreoni2.png" width="450px" style="display: block; margin: auto;" /> --- class: MSU # Back to Previous Paper ### Augenblick, Niederle, and Sprenger (2015) Setup: - 102 UC Berkeley Xlab subjects recruited for 7 weeks. - Required to come to the lab in weeks 1, 4, and 7. - Weekly tasks (same day of week) online during other weeks - Pay largely at the end to reduce attrition: $100 completion payment, $10 otherwise -- - Subjects allocate, reallocate, and complete units of tedious tasks. - **Main question:** How to divide work between Day 1 and Day 2 - Interest rate of transferring work from Day 1 to Day 2 varied --- class: MSU # Experiment Details (cont) ### The task: boring Greek letter transcription task - Blurry Greek letters appear in transcription box (top). - Subject has to click on corresponding letters (bottom). <img src="graphics/ANS_setup1.png" width="550px" style="display: block; margin: auto;" /> -- - Why this task? Boring, relatively little learning over time, little intrinsic utility, (perhaps) not much variation in ability across individuals. --- class: MSU # Design ### Overview of experimental design <img src="graphics/ANS_setup3.png" width="550px" style="display: block; margin: auto;" /> -- - Two blocks of three weeks each - Block 1: weeks 1, 2, and 3 - Block 2: weeks 4, 5, and 6 - Block 2 same as Block 1, but with commitment technology - Payment in week 7 --- class: MSU # Identification ### What identifies present bias? -- - Week 1: choice between effort in week 2 and week 3 - (It is, after all, a choice between two future disutilities and the associated dates) -- - Week 2: choice between effort in week 2 and week 3 - Choice between a **present** and a future work date -- - Model prediction: present-biased individuals choose less effort in week 2 when choosing in week 2 (compared to choices in week 1). --- class: inverseMSU # Potential Confounds - (Systematic) misprediction about time-varying effort costs? - Learning (or mislearning; see Gagnon-Bartsch and Bushong 2022)? - Overoptimism about future opportunity cost of time? -- - ...but would the above mechanisms predict demand for commitment? --- class: MSU # More Details Subjects are asked to allocate tasks at various task rates. <img src="graphics/ANS_setup2.png" width="550px" style="display: block; margin: auto;" /> -- - Task rates `\(R\in\{0.5,0.75,1,1.25,1.5\}\)` correspond to different (gross) interest rates - Exactly one of all week 1 and week 2 decisions is implemented. - Randomization favors week 2 (90% of allocations). --- class: MSU # Results (Not shown, look at the paper: limited / weak evidence of present bias for money.) **Below:** strong evidence for present bias over tasks. <img src="graphics/ANS_results2.png" width="550px" style="display: block; margin: auto;" /> --- class: MSU # Results <img src="graphics/ANS_results4.png" width="700px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Surprising finding: almost no correlation between `\(\beta_{e}\)` and `\(\beta_{m}\)` <img src="graphics/ANS_results5.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Demand for commitment? - Subjects make 15 decisions between two allocation choices: - (Option 1) Allocation-that-counts comes from initial allocations with probability 0.1, plus earn $X. -- - (Option 2) Allocation-that-counts comes from initial allocations with probability 0.9, plus earn $Y. -- - Choices are made after initial allocations (week 1). - If `\(X \geq Y\)`, then choice of Option 2 implies demand for commitment. -- - Choosing Option 2 increases the chance that the choice set is fixed (to the initial choices) by 80 percentage points. - Price of commitment: `\(P_{C} = X - Y\)`. --- class: MSU # Understanding the Data ### Learning from Choices - From above: the price of commitment is `\(P_{C} = X - Y\)`. - What do we learn if an individual chooses Option 2 for `\(P_{C}<0\)`? -- - Nothing. May choose Option 2 just to get the payment `\(Y\)` rather than `\(X\)`. - What if she chooses Option 2 if `\(P_{C}=0\)`? -- - She restricts her choice set without receiving any compensation for it. - The person exhibits demand for commitment. However, she could just be (close to) indifferent between the initial and later choices. -- - What if they choose Option 2 if `\(P_{C}>0\)`? - The individual selects to restrict her choice set **and** to receive a lower compensation. - Implies positive willingness to pay for commitment --- class: MSU # Back to Results! ### Demand for commitment - 59% of subjects choose commitment when it is free (i.e. they choose Option 2 at `\(P_{C}=0\)`). - Subjects who commit exhibit significantly more present bias in previous choices. -- - Almost nobody chooses commitment at `\(P_{C}>0\)`. <img src="graphics/ANS_commitment2.png" width="450px" style="display: block; margin: auto;" /> --- class: inverseMSU # Assessment ### What makes this a great (and not so great) paper? - Big step forward in literature. Important questions and results - Addresses long-standing debate in the literature - Thoughtful design - Clear connection between theory and experiment -- - Replication exercise (Section 4.5) addresses several shortcomings - Inconsequential outcomes: no surprise that nobody is willing to pay for commitment. - Unknown whether we can predict real-world behavior using these parameters. - High complexity: potential for confusion, difficulty of using in low-literacy populations --- class: MSU # Let's Go Deeper! ### Naivete and sophistication - Do individuals understand their future present bias? - `\(\hat{\beta}\)`: prediction for future `\(\beta\)` - Enter Augenblick and Rabin (2018) --- class: MSU # Design - Experiment ran for seven days over six-week period (in lab first day and online remaining days) - On each experimental day: - Participants did some minimum work (to incur fixed costs of logging in) -- - Made committed choices for amount of work to do (on the same day and two future periods) at five wages each -- - One of the wages and work choices for that day was randomly selected to be implemented - Required work for today may be a choice you made previously (e.g. what you chose for today last week) or the one you made just now - Payments for all work + completion bonus one week after last work day --- class: MSU # Identification ### Present bias: `\(\beta\)` - Compare choices for present and future periods - Can compare choice for Day `\(t\)` made on Day `\(t-1\)` vs made on Day `\(t\)` itself -- ### Sophistication: `\(\beta_h\)` (aka `\(\hat{\beta}\)` in previous work) - Compare predictions of future choices to actual future choices -- ### Marginal disutility of effort `\(\gamma\)` (convex effort-cost curve) - Compare choices at different piece rates -- ### Projection bias parameter `\(\alpha\)` (more later) - Compare choices before vs. after mandatory minimum work --- class: MSU # Interface <img src="graphics/AugenblickRabin_Choices2.png" width="600px" style="display: block; margin: auto;" /> --- class: inverseMSU # Some Issues (?) ### Size of incentives for predictions of future choices - Small prediction-accuracy bonus? `\(\Rightarrow\)` Low incentive for accuracy -- - Participants may just not care about accuracy of predictions. -- - Large prediction-accuracy bonus? - May match immediate-work decisions to earlier predictions to receive the bonus. - Sophisticated present-biased participants may use predictions as commitment device. --- class: inverseMSU # Identification ### A Simple Approach -- - Test whether predictions are sensitive to size of bonus - Show that participants actually don't use this as a commitment device -- - Participants reminded of previous predictions by visual cue -- Note: this might have caused anchoring - But otherwise would have been hard to know whether individuals remembered predictions -- - No predictions for some days -- *Help:* Any better way to do this? --- class: MSU # Results <img src="graphics/AugenblickRabin_Results1.png" width="650px" style="display: block; margin: auto;" /> -- - Individuals choose 5.7 fewer tasks for present than for future. - Predictions about future work are largely in line with *current* decisions for future work (but not with future decisions). --- class: MSU # Results ### Stylized fact: people choose equal effort levels for any time in the future (not today) <img src="graphics/AugenblickRabin_Results2.png" width="650px" style="display: block; margin: auto;" /> -- - Evidence consistent with quasi-hyperbolic present bias rather than more continuous hyperbolic discounting - **Not necessarily** incompatible with evidence from Kaur, Kremer, and Mullainathan (2015) --- class: MSU # Results ### No connection between bonus level and predictions <img src="graphics/AugenblickRabin_Results3.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Aggregate estimates of `\(\beta\)`, `\(\beta_h\)`, and `\(\delta\)` <img src="graphics/AugenblickRabin_Results4.png" width="650px" style="display: block; margin: auto;" /> -- ### No evidence of sophistication! Cannot reject that `\(\beta_h = 1\)`. --- class: MSU # Results ### Individual estimates of `\(\beta\)` and `\(\beta_h\)`} <img src="graphics/AugenblickRabin_Results5.png" width="550px" style="display: block; margin: auto;" /><img src="graphics/AugenblickRabin_Results6.png" width="550px" style="display: block; margin: auto;" /> -- - Individual estimates of `\(\beta\)` and `\(\beta_h\)` are weakly correlated. - Suggests that maybe there is *some* degree of sophistication (if minor) --- class: MSU # Assessment and Open Questions ### Relationship to Augenblick, Niederle, and Sprenger (2015) - Why do ANS find demand for commitment (at zero price) if `\(\beta_h \approx 1\)`? - Why is there demand for commitment in some contexts if `\(\beta_h \approx 1\)`? -- ### Learning about present bias ( `\(\beta_h\)` ) - Is there learning about `\(\beta_h\)` over time? If not, why not? - Could investigate learning with similar design - Are individuals also naive regarding other, more important tasks? - Little work on learning self-control: Ali (2011) Le Yaouanq and Schwardmann (2019) --- class: MSU # Assessment (cont) - Not much other work in this area (yet) - Exception: Acland and Levy (2015) on gym attendance -- ... but this paper has major issues - Bai et al (2021) also attempt to estimate distribution of `\(\hat{\beta}\)` using continuous measure of "commitment amount" chosen by patients to incentivize future doctor visits. --- class: MSU # On Commitment ### Fundamental Question #1: Why so little commitment seen in the real world? - Naivete - Trade-off between flexibility and commitment in the presence of uncertainty (Laibson, 2015) -- Is offering commitment actually welfare-enhancing? . - (Bai, Handel, Miguel, and Rao, 2021), (John, 2015): Partial naifs take up commitment, but choose too-small penalties - Of course, this results from the fact that they under-estimate their future self-control problems - May be better off if not offered commitment at all --- class: MSU # On Commitment ### Fundamental Question #2: How can we improve the effectiveness (and takeup) of commitment contracts? - Provide experience with commitment devices? - Personalize or customize based on observed behavior or parameter estimates? -- **Actual class question:** Any ideas? --- class: MSU # (Bai, Handel, Miguel, et al., 2021) Starting plance: Is there demand for commitment contracts for hypertension management in rural India? - How do different types of theoretically-motivated commitment contracts perform? -- Authors measure take-up and follow-through to address this. -- - Do personalized vs. fixed contracts have different take-up and success rates? -- - What fraction of individuals do commitment contracts help, and what fraction do they hurt? -- ### Contribution Combining RCT + structural model, which allows: -- Behavioral parameter estimates, -- welfare analysis, -- and analysis of counterfactual contracts. --- class: MSU # The Setting Trying to encourage doctor visits by individuals at risk of hypertension. - Rising incidence in developing economies -- - India `\(\approx 25 \%\)`, Africa `\(\approx 30 \%\)` (country-level variation huge) -- Controllable using effective medicines - Yet only 10-20% actively manage in India (Mohan 2013) - WHO recommends bi-monthly visits when starting treatment in such conditions. --- class: MSU # Experiment ### Experimental Design N=1728, approx n=280 in each treatment arm: - **Control group** (pay as you go): Rs. 30 per visit -- - **Price subsidy** (50% off consultation fees): Rs. 15 per visit -- - **Commitment contracts**: Pay up-front for 3 future visits + possible additional "commitment" amount -- - **Personalized**: Choose commitment amount - **Fixed**: Commitment = Rs. 45 (take it or leave it) -- - **Bundled**: Commitment contracts + price subsidies -- - Randomized reminders --- class: MSU # Experimental Interventions <img src="graphics/experimental_interventions.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Experimental Design ### Theoretical motivation for design Personalized (self-designed) commitment contracts: - Consumer picks the commitment amount - Commitment amount paid back if follows through on visit - Higher up front payment increases ex post incentive to attend -- - Great for sophisticated individuals but could hurt over-confident folks. [Insert aside] -- - Why great for sophisticates: will use to reach first best - Partial naifs will choose commitment amounts that are too small (i.e. not enough to induce the desired behavior) - Full naifs won't demand commitment --- class: MSU # Experimental Design (cont) Fixed strong commitment contracts: Take it or leave it - Commitment amount set by us (the researchers), relatively large - Receive substantial amount back iff visit doctor on schedule - *My "ex ante" intuition:* more follow through, but lower demand? -- Combining strong contract with price subsidy - Entice naifs/partial naifs who expect to visit into contract - Ex-post, follow through only because of commitment --- class: MSU # Results ### Contract Take-up and Clinic Visits <img src="graphics/EHP-multi-panel_part1_2016-07.png" width="750px" style="display: block; margin: auto;" /> -- - Different rates of take-up; similar rates of service utilization --- class: MSU # Summary of Findings 1. Low to moderate take-up of commitment contracts - Largely in line with the literature -- 2. Bundling with price subsidies increases take-up of contracts -- 3. Poor follow-through on contracts - 62-77% fail to visit even once 4. Lower follow-through in personalized contracts -- ### WTF. Why do people sign up and then not visit the doctor? --- class: MSU # Potential Explanations ### Uncertainty, learning and beliefs about benefits - Shocks to costs of going to the doctor. -- (My comment: then let's take this seriously and calibrate. I doubt this is right.) -- - Social pressure inducing sign-ups - Severe self control problems but with partial naivete --- class: inverseMSU # Discussion ### Some of my conjectures of more promising approaches: - Provide direct incentives rather than commitment -- - Reduce "hassle" costs. -- - Highlight dimensions of provider quality that consumers value -- - Help calibrate beliefs about future utility. -- - Help patients understand the psychological costs of health (non)behaviors --- class: MSU # Self Control at Work ### (Kaur, Kremer, and Mullainathan, 2015) *discussion courtesy of Stefano* Setting: workers in India who are paid a piece rate `\(w\)` in a weekly paycheck -- - Since effort at work is immediate and benefits delayed, effort at work is an investment good - We'll examine a `\((\beta, \hat \beta, \delta)\)` model where, for ease, `\(\delta = 1\)`. -- - Consider effort at work `\(e\)`. - Exerting this effort yields overall costs `\(c(e)\)` , with `\(c'(\cdot) > 0\)`. -- - Assume (for now) `\(c'' >0\)` (convex costs). - To follow the math a bit, we'll introduce the special case `\(c(e) = \gamma e^2/2\)` -- - (Why divide by 2? To make the derivative easier!) --- class: MSU # SCaW Model We'll explore a setting with uncertainty over the agent's mapping between effort and productivity. -- Assume two "states": 1. High output `\(y_H\)` with probability `\(e\)`. In this case, agent earns `\(w_H\)` 2. Low output `\(y_L\)` with probability `\(1-e\)`. In this case, agent earns `\(w_L\)` -- Notice that this is only local approximation and binds us to `\(e \in [0, 1]\)`. --- class: MSU # SCaW Model Agents are paid at `\(t = 2\)` - If the agent is working in `\(t = 1\)`, she maximizes $$\max_e \left [ \beta (e w_H +(1−e) w_L]−c(e) \right ] $$ -- with fist-order condition: `$$\beta (w_H − w_L]−c'(e^*)=0$$` -- `\(\Rightarrow\)` Effort `\(e^*\)` increases in wage differential and in `\(\beta\)`. --- class: MSU # SCaW Model If `\(c(e) = \gamma e^2/2\)`: `$$e^* = \frac{\beta (w_H − w_L)}{\gamma}$$` -- **However:** If the agent is working in `\(t = 2\)` (same period as paydate), optimal effort changes. -- Now `\(e_2^*\)` solves $$\max_{e_2} [ e_2 w_H +(1−e_2)w_L]−c(e_2)] $$ -- and for the special case: `$$e_2^* = \frac{(w_H − w_L)}{\gamma}$$` -- **Prediction 1.** Effort is higher near payday for `\(\beta<1\)` -- (and independent of `\(\hat \beta\)`) -- From `\(t = 0\)` perspective, (perceived) utility `\(V_0\)` from working at `\(t = 1\)` is `$$V_0 =e^∗w_H +(1−e^∗)w_L −c(e^∗)$$` --- class: MSU # SCaW Model ### Let's ask: how does changing `\(w_L\)` alter incentives? -- Effect of altering `\(w_L\)` on `\(V_0\)` (expected welfare in `\(t=0\)` is: `$$\frac{d V_0}{d w_L} = (1-e^*) + \frac{d e^*}{d w_L}[(w_H − w_L)−c'(e^*)]$$` -- `$$= (1-e^*) + \frac{d e^*}{d w_L}\left [ \left (1 - \hat \beta \right )(w_H − w_L) \right ]$$` -- - *Why is this the case?* -- - But now note: first term is the direct effect on pay. -- Lowering `\(w_L\)` lowers pay and thus welfare. -- - The second term is the effect on incentives. This is zero for `\(\beta = \hat \beta = 1\)`, by the envelope theorem. -- ...but envelope theorem does not apply for `\(\hat \beta = 1\)`. Indeed, second term is negative! -- (Notice that it is `\(\hat \beta\)` which matters, since this is the value function from the `\(t = 0\)` perspective) --- class: MSU # SCaW Model ### Let's consider contrast between sophisticates and naifs When `\(\beta = \hat \beta\)`: `$$\frac{d V_0}{d w_L} = 1 - \frac{\beta (w_H − w_L)}{\gamma} - \frac{\beta (1-\beta) (w_H − w_L)}{\gamma}$$` -- - Observe that the second term becomes large as `\(\beta\)` goes below 1 and is highest at `\(\beta = \frac{1}{2}\)`. - If this second term is large enough, the implication is that the agent would demand a commitment device insofar as she prefers a **lower** `\(w_L\)`. -- (Take a hot minute to digest that result; it's surprising.) -- - **Prediction 2.** When `\(\beta < 1\)`, agents in period `\(t=0\)` may prefer a lower `\(w_L\)` as a form of a commitment device. -- - **Prediction 3.** If the population includes both types ( `\(\beta < 1\)` and `\(\beta = 1\)` ), demand for commitment should be associated with strong payday effects. --- class: MSU # SCaW Setting Full-time data entry workers in India - Job is primary source of earnings. - Data-entry office set up as part of study - Duration of study: 13 months -- Measuring output: - Number of accurate fields entered in a day - Accuracy measured using dual entry of data -- Incentives: - Workers paid a piece rate `\(w\)` in a weekly paycheck - No restriction on hours (can arrive, leave, take breaks at any time) - No penalties for absences --- class: MSU # Work Task <img src="graphics/Kaur_setup.png" width="550px" style="display: block; margin: auto;" /> -- - View scanned images on screen - Enter information into field in data entry software --- class: MSU # Design of Contracts ### Demand for commitment and dominated contracts <img src="graphics/Kaur_dominated_contracts.png" width="550px" style="display: block; margin: auto;" /> - Control contract: linear piece rate `\(w\)` -- - Dominated contract: `\(w\)` if production exceeds `\(T\)` (target); `\(w/2\)` if production is lower than `\(T\)`. - Worker chooses `\(T\)` in advance (can choose `\(T=0\)`). --- class: MSU # Design Overview <img src="graphics/Kaur_design_overview.png" width="550px" style="display: block; margin: auto;" /> --- class: MSU # Design (cont) ### Payday Effects Randomization of payday (Tues, Thurs, Sat) - Allows to control for day-of-the-week effects -- - *What does the exponential discounting model predict?* -- - Effort should be (close to) unrelated to paydays --- class: MSU # Results <img src="graphics/Kaur_payday_effects.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Demand for commitment (dominated contracts) - Workers selected dominated contracts 36% of the time. - Lower bound for the extent of time inconsistency. -- *Why?* -- - People may be naive. - People may think commitment device is ineffective. - People may prefer flexibility/be risk-averse. -- ### Impact of offering dominated contract - ToT effect of 6% for those who accept the dominated contract. - Increase corresponds to 18% increase in piece-rate wage. --- class: MSU # Results <img src="graphics/Kaur_demand_commitment.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results <img src="graphics/Kaur_dominated_impact.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results ### High payday workers more likely to select positive targets <img src="graphics/Kaur_payday_dominated.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results <img src="graphics/Kaur_impact_offer.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Main Contributions and Recap Significant demand for commitment in real-world setting in India - 36% of data-entry workers choose dominated contracts. -- - *Merely offering* commitment device increases productivity and earnings - Increases in output are calibrationally equivalent to an 18% raise in wage rate -- Provide strong evidence of payday effects - Effort increases as the (randomly assigned) payday gets closer. - Payday effects are predictive of subsequent demand for commitment. -- - More broadly: workers demand higher-powered incentives. -- `\(\Rightarrow\)` Same contract features that mitigate moral hazard can also mitigate self-control problems. --- class: MSU # Ashraf, Karlan, and Yin (2006) ### Setting: Personal Finance - Clients of Philippine bank - Randomly selected subset offered commitment savings product -- 1. **SEED Treatment:** Encouraged to save plus commitment offer: client can restrict access to deposits upon opening the account. -- 2. **Marketing Treatment:** Encouraged to save, no commitment offer -- 3. **Control Treatment:** Standard savings account (same interest rate) --- class: MSU # Clients' Savings Goals <img src="graphics/AKY_goals.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results - 24% take-up rate of SEED account (either date-based or amount-based goal) - Offering commitment savings product significantly increased savings. -- - Survey responses to "money now vs. money later" question predicts commitment take-up. --- class: MSU # Results ### Offering commitment savings product increased savings (ITT analysis) <img src="graphics/AKY_results.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Schilbach (2019) ### Alcohol consumption and rickshaw drivers in India (more broadly: day-drinking among low-income workers) <img src="graphics/rickshaw_peddler.jpg" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Setting ### Day-drinking among cycle-rickshaw drivers in Chennai <img src="graphics/Schilbach_breathalyzer_scores.png" width="650px" style="display: block; margin: auto;" /> - Breathalyzer scores from tests during **regular work hours** --- class: MSU # Design ## Incentives for sobriety: - Three-week field experiment with low-income workers in India - Visit study daily between 6 pm and 10 pm - Short survey and breathalyzer test - Financial incentives for sobriety for random subset of individuals --- class: MSU # Design ### How to measure demand for commitment? - Majority of people say they would like to drink less. - Majority also say their lives would be better if all liquor stores closed. -- But we want evidence of demand for commitment. - Antabuse - Dominated contracts --- class: MSU # Design ### Demand for Incentives - **Option A:** incentives for sobriety - Same payment structure as below - Rs. 60 if BAC `\(>\)` 0, Rs. 120 if BAC `\(=\)` 0 -- - **Option B:** incentives regardless of BAC <img src="graphics/Schilbach_incentives.png" width="350px" style="display: block; margin: auto;" /> --- class: MSU # Design: Timeline <img src="graphics/Schilbach_timeline.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Incentives reduce (only) daytime drinking <img src="graphics/Schilbach_drinking_effect.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Sobriety incentives and savings commitment are substitutes <img src="graphics/Schilbach_sub.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Persistently non-zero demand for commitment <img src="graphics/Schilbach_choices_weeks123.png" width="650px" style="display: block; margin: auto;" /> --- class: MSU # Recapping the Study - Intervention successful in reducing day drinking - Day drinking falls by about one third in treatment group. - But individuals mostly substitute to drinking later at nights. -- - Persistently high demand for commitment to increase sobriety - About 50% choose options that are dominated in study payments - About one third are willing to forego 10% of their daily earnings to receive incentives for sobriety -- - Results consistent with naive quasi-hyperbolic discounting --- class: MSU # Overall Results <img src="graphics/commitment_recap.png" width="750px" style="display: block; margin: auto;" /> --- class: MSU # Patience and Wealth ### From (Epper, Fehr, Fehr-Duda, Kreiner, Lassen, Leth-Petersen, and Rasmussen, 2018) - Sample of 3600 middle-aged Danish adults - Money earlier or later is the elicited measure of time preferences (0 vs. 8 weeks; 8 vs 16 weeks) - Link to Danish administrative data (high-quality data on wealth and income, employment etc.) -- - *Findings:* 1/3 most patient are 6 percentiles higher in wealth distribution than 1/3 least patient - Similar to comparing 1/3 least educated with 1/3 most educated -- - **Questions:** Reverse causality? Liquidity constraints? --- class: MSU # What Determines `\(\beta\)` ? ### If `\(\beta\)` is not fixed within person, what determines fluctuations? -- - Visceral factors, emotions (pain, hunger, anger, etc.) - Cues (Laibson (2001); Bernheim and Rangel (2004) - Willpower, ego depletion (Muraven and Baumeister, 2000) - Sleep (Bessone et al. 2021 QJE) -- - Stochastic discount factors? - Individual differences in mean *and* variance of `\(\beta\)` (and `\(\hat{\beta}\)`)? -- - Good-specific discount factors (Banerjee and Mullainathan, 2010) - Some goods are more tempting than others. - Set of temptation good may systematically vary across people. -- - Teaching willpower? Baumeister's idea of willpower as a muscle --- class: MSU # (Alan and Ertac, 2018) ### An RCT on increasing patience among students - Design involved 3rd-4th grade students in Turkey - Case studies, stories, in-class games meant to increase ability to act in forward-looking manner - Delivered by students' own teachers. (*Note:* measuring adherence is a nightmare and there could be teacher fixed-effects, so large sample is needed.) - Measure time preferences using gifts earlier or later and measure school performance --- class: MSU # Case study example (Alan and Ertac, 2018) present lots of simple case studies as teaching method. -- <img src="graphics/alanCaseStudy.png" width="200px" style="display: block; margin: auto;" /> > Zeynep is a girl who wants a bicycle for her next birthday but she needs to cover part of the cost herself. For this, she needs to save. After talking to the wise bird, she gets on a time machine and sees alternative birthday realizations (with and without the bicycle). Then she tries to infer what sort of actions might have led to each realization. --- class: MSU # Eliciting preferences ### MPL and CTB measurements <img src="graphics/alanCTB.png" width="720px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Significant increases in patience <img src="graphics/alanReg1.png" width="720px" style="display: block; margin: auto;" /> --- class: MSU # Results ### Effects persist for years(!) <img src="graphics/alanPersistence.png" width="650px" style="display: block; margin: auto;" /> --- class: inverseMSU # Evaluation **GREAT STUFF:** - First field-experimental evidence on persistent malleability of time preferences - Also show improvements in "behavior grade" in school (but **not** grades; this might be a useful fact for you all.) -- **INTRIGUING / OPEN STUFF:** - Effects appear to be on patience ( `\(\delta\)` ) not on present bias ( `\(\beta\)` ) - Not designed to get parameter estimates - Excellent follow-up paper by Alan, Boneva and Ertac (2019) on teaching "grit" with lovely experimental measure of sticking with tough task despite experiencing failure. - *Results in paper 2:* Treated students are more likely to exert effort to accumulate task-specific ability and more likely to succeed. --- class: MSU # Other Approachs One interpretation of present bias is that it results from excessive attention to current vs. future periods - Results in larger weight on current compared to future periods -- **Natural question:** Can we manipulate individuals' attention toward the future? -- - One approach: show individuals pictures of future selves - Attempt to increase *vividness of future self* (salience of future) -- (Hershfield, 2011) summarizes several studies. - Some promising results - So far only hypothetical choices <img src="graphics/Hershfield2011.png" width="400px" style="display: block; margin: auto;" />