Presented by Jonny Bravo · 7:57
Use the audio controls below to listen. Slides follow playback when JavaScript is enabled.
This transcript and its captions were generated automatically and may contain errors. Participant names may have been assigned by the Talk Owner.
Thanks for having me.
This is a deliberately synthetic talk about place effects and economic mobility.
Every number, sample, and result is fictional.
The point is to create a realistic seminar recording with enough structure to test whether a system can separate speakers and connect discussion to the right slide.
The motivating idea is that neighborhoods may shape children through schools, peers, safety, and access to labor markets.
But families select where to live, so a simple comparison across places confounds exposure with selection.
I want to isolate the effect of spending one additional year in a higher opportunity place.
Before you move on, what exactly do you mean by a higher opportunity neighborhood?
Is that defined using the outcomes of the same children in your sample?
Good question.
In the synthetic design, opportunity is an external destination score constructed from older cohorts, so it does not mechanically mechanically use the outcomes of the children whose exposure effects we estimate.
I will return to alternative destination measures on the robustness slide.
The main question is whether an extra year in a higher opportunity neighborhood raises adult earnings.
The second question is whether that effect declines with age at move.
A declining age profile would be consistent with cumulative childhood exposure rather than a one-time disruption around the move.
Q: Are you treating age at move as a continuous exposure measure, or are the age groups chosen after looking at the results?
A: The primary specification is continuous and fixed before estimation.
The age bins are only a visual summary.
That distinction matters because I do not want the central heterogeneity result to depend on a convenient breakpoint.
The fictional data contain 48,000 sibling pairs, whose families move across 220 commuting zones, Adult earnings are observed from ages 28 through 32.
The sibling structure gives us children from the same family who receive different durations of exposure after one common move.
How much sample selection comes from requiring two siblings and a move that you can observe?
I worry that these are unusually mobile families.
That is a real external validity concern.
The estimate applies to the synthetic population of observed mover families with siblings.
It is not automatically in effect for all children.
I will keep that limitation separate from the internal identification question.
The design compares siblings exposed to the destination for different lengths of childhood, after the same family move.
Family-by-move fixed effects absorb everything shared by the family and move.
Flexible age-at-move trends absorb smooth lifecycle differences.
The identifying variation is the sibling exposure difference within the same move.
But the older and younger sibling can respond differently to the same move.
Why should the difference be exposure rather than disruption at a sensitive age?
The design cannot rule that out by construction.
The empirical distinction comes from the shape of the age profile and from placebo outcomes.
A sharp one-time disruption should look different from a roughly linear dose response in years of exposure.
The synthetic headline estimate is that one additional year of childhood exposure raises adult earnings by 0.7%, with a fictional confidence interval from 0.3 to 1.1.
The magnitude is intentionally plausible, but it is not evidence about the real world.
Is that coefficient identified mostly by a small set of long-distance moves?
I would like to see the distribution of exposure changes, not just the average effect.
That is useful.
The current slide hides the support of the design.
In a real revision, I would add a distribution of destination score changes and report how the estimate moves after trimming the tails.
The key assumption is that conditional move timing is unrelated to child-specific potential outcomes.
A violation would occur if families time a move in response to a shock affecting one sibling differently.
The fixed effects do not solve that problem, because the shock varies within the family.
Is the assumption really about timing, destination choice, or both?
Right now, the slide seems to combine two different sources of endogeneity.
You are right that the slide is too compressed.
Timing determines the sibling exposure difference, while destination choice determines the treatment intensity.
The clean version should state the two assumptions separately and attach a distinct diagnostic to each.
Here is the central heterogeneity result: effects are largest for children who move before age 10 and become smaller with age at move.
The synthetic gradient is close to linear, which is what we would expect from cumulative exposure.
Could composition explain the age gradient rather than the mechanism?
Families moving with a 5-year-old may be very different from families moving with a 15-year-old.
That is the main concern.
The next robustness checks show age-specific balance tests and reweight the observable composition of mover families.
Those checks cannot prove the absence of selection, but they tell us whether observables move enough to explain the gradient.
Can you put the age-specific balance evidence beside this figure?
Sending us to the next slide makes it hard to judge the claim while looking at the gradient.
Yes, that is a presentation change I would make immediately.
Pair the key identifying diagnostic with the key heterogeneity figure instead of separating them.
The estimate survives alternative destination measures, family-specific trends, and placebo outcomes measured before the move.
It also remains similar after trimming the largest destination changes.
The checks address different threats, so a single robustness label is not very informative.
Which of these checks actually addresses endogenous mobility?
The list mixes sensitivity analysis with identification evidence.
The pre-move placebo and age-specific balance tests speak most directly to endogenous mobility.
The destination measures and trimming exercises are sensitivity checks.
I should organize the slide by threat rather than by specification name.
The fictional mechanism evidence points to school quality, peer networks, and labor market access.
These channels are correlated, so the exercise is descriptive.
It does not decompose the total place effect into causal shares.
Which result identifies the school channel rather than simply restating that better places have better schools?
None of the current results cleanly identifies that channel.
The honest claim is that school quality covaries with the estimated destination effect.
I would need a separate source of variation to make a mechanism claim.
The policy simulation asks when a hypothetical mobility program produces the largest gain.
Because the estimated effect accumulates with exposure, the simulated benefit is largest for moves before age 10.
The simulation holds family selection fixed by construction, so it should not be read as a forecast of program take-up.
Will the replication code and synthetic data generator be public?
That is not covered by the current DEC.
For this fictional exercise, the generator can be shared, but a real project would need a separate disclosure and reproducibility plan.
To conclude, the synthetic results say that duration of childhood exposure matters and that effects are larger for younger movers.
The seminar discussion sharpened the real burden of proof: separate timing from destination assumptions, put balance evidence beside the age gradient, and avoid presenting descriptive mechanisms as causal.
Thank you.