Características campesinas de la sociedad local

In this section, I provide further evidence of arms race-style preferences by examining how house- holds neighboring crash victims respond indirectly to the size of vehicles involved in the accidents. The idea is that a larger striking (or opposing) vehicle, and a smaller vehicle driven by a neighbor, produces more severe and salient outcomes (Anderson and Auffhammer, 2014). The household neighboring the victim then directly responds to these salient outcomes—a direct function of the crash weights.

Identification of indirect responses to the size of the vehicles involved in these accidents come from the as-good-as-random timing of the accidents. Neighboring households may directly respond to each other’s vehicle size by means of social norms. I will not be able to identify these baseline effects, but rather, identify the additional response to a victim’s vehicle size following the acci- dent. I estimate these indirect responses from a household to both the neighbor’s and the opposing vehicle’s weight in a two-car collision.

To identify the spillover effects of the vehicle sizes involved in an accident, I estimate the following equation:

weightit= αi+ λt+ γ · post_crashit+ β1 · weightnbri × post_crashit

+ β2· weightoppi × post_crashit+ εit

(2.3)

where αi and λtare household and month-of-sample fixed effects, respectively. Household fixed

effects will absorb any baseline sizes of the two vehicles involved in the accident that the house- hold may respond to prior to the accident. In some specifications I will simply control for these baseline weights, rather than household fixed effects (which absorb all one-purchase households in my data). post_crashitis an indicator for i’s neighbor incurring a vehicle accident at some time

s ≥ t, and εit is the unobserved error. β1 estimates the additional response to the weight of the

neighbor’s vehicle involved in the accident, weightnbr_i . β2estimates the additional response to the

weight of the opposing vehicle’s weight, weightopp_i , which hit the neighbor.

Note that β1 < 0 implies that a smaller neighbor’s vehicle involved in the accident on average

produces a more severe outcome, leading the household to purchase a heavier vehicle. An estimate of β2 > 0 implies that a larger vehicle hitting the neighbor’s car creates a more severe outcome,

2.5.1 Results and Robustness

The least squares estimates of Equation 2.3 are presented in Table A.4. These estimates report the effect of each 1,000 pounds of the crash vehicles’ weights. The first two columns estimate the effect of victims’ vehicle weights on neighbor fatality, using a linear probability model. This is simply the cross-section of vehicle-by-accidents. The results indicate that a 1,000 pound increase (decrease) in opposing (neighbor’s) vehicle weight increases the probability of neighbor fatality by about 6-7%. In the second column, I include census tract-level demographics, such as median income, median age, percent white, and percent black.

Because the weight effects not only risk of fatality, but also the severity of accident (and poten- tially other salient factors), using these crash weights as an instrument for neighbor fatality would violate the exclusion restriction. Therefore, inference should only be made from the reduced form effects of crash weights on neighbor of victim weight. These estimates are presented in Columns 3-7 of Table A.4.

There is a similarity in the estimates reported in Columns 1-2 of Table A.4 and the estimates re- ported by Anderson and Auffhammer (2014). They find that an additional 1,000 pounds of the striking (or opposing) vehicle generates about a 0.109 percentage point increase in risk of fatality. Note that I should not expect to get the same estimate, as my sample conditions on accidents in- volving at least one fatality, whereas Anderson and Auffhammer’s dataset also include accidents involving no fatalities.

The coefficients in Table A.4 were estimated using the two-neighbor sample. The results for the effect of opposing vehicle size indicate a statistically significant estimate of about 16-18 pounds. That is, a 1,000 pound increase in the opposing vehicle’s weight causes household i to increase their average fleet weight by 16-18 pounds.

The effect of the neighbor’s vehicle size represents the converse of the opposing vehicle’s size. That is, the smaller the neighbor’s car, the larger the response. Columns 3-4 illustrate a similar- in-magnitude response as the impact of oppsing vehicle. The inclusion of household fixed effects aborb much of this effect. Though estimates with and without household fixed effects are not statistically different from each other, conditioning on household fixed effects seems to eliminate statistical significance.

The reduced-form estimates of opposing vehicle weight offer corroborating evidence of arms race dynamics using another plausible source of quasi-random variation. The opposing vehicle is matched to the neighbor as-good-as-randomly, and the likelihood that these two victims were acquaintances is therefore small. I am also exploiting the as-good-as-random timing of the acci- dent for these estimates—the only source of (plausibly) random variation in estimating the effect of the neighbor’s fleet weight, post-crash.

I present the estimates for samples containing differing numbers of nearest next-door neighbors. These results are presented in Table A.5. Similar to previous results, these estimates become at- tenuated as I include neighbors further away. However, the estimates for opposing vehicle do not fully converge to zero in the 10 neighbor sample. This result differs from the estimates in Table A.3, where the estimates on weight converge to zero quickly.

In the event that these estimates remain positive for neighbors very far away, one should be con- cerned with the potential influence of confounders, such as common shocks. To mitigate these concerns, I could look at neighbors located on a street behind the household, as these households are less likely to be acquaintances with the crash victims. The process of finding neighbors a block over is very tedious. The simpler approach is to randomly assign neighbors to opposing vehicle weights from the sample. This is similar to the permutation approach in Section 2.4.4, except per- formed in one dimension. That is, I maintain the same crash time (the true crash time), which is a

source of random variation, but randomly re-assign a crash vehicle pair from the sample of two-car crashes. If I am picking up confounding factors common across all vehicles, I would expect to see effects not only from the opposing vehicle hitting the neighbor, but also other opposing vehicles in the sample.

I randomly draw, with replacement, victims of vehicle accidents and re-assign them to households. That is, I re-assign neighbors, collecting the weights from both vehicles in the accident. I make this re-assignment 1,000 times, estimating the coefficients in each permutation, and plot the dis- tribution of these placebo estimates. Using this approach resulted in a one-tailed p-value of 0.082 for the effect of neighbor’s weight, and 0.001 for the opposing vehicle. The distribution of the placebo estimates are presented in Figure A.5. These results suggest that it is highly unlikely that the estimates in Tables A.4 and A.5 are the outcome of common shocks.