Peer Effects and Voting Behavior

The results presented so far do not account for potential peer effects in voting behavior.

This section introduces a few types of potential peer effects and presents their associations with observed voting behaviors in reduced form models. The potential peer effects added here are a) the fraction of in favor votes at the meeting level, b) the fraction of in favor votes at the meeting level for all individuals except the reference individual (i.e. a number is calculated for each voter as the average of her panel without her vote included), and c) the fraction of in favor votes among conflicted panelists (only) at the meeting level. The results of regressions accounting for these potential sources of peer influence and interacting each with an indicator for financial conflict of interest are presented in Table 2.7.

An important note with respect to Table 2.7 is that the coefficients presented were recovered from linear probability models for ease of interpretation. So far all regression output tables have presented marginal effects from probit regressions. For the models

Table 2.7: Preliminary Analysis of Peer Effects in Voting Behavior

(A)

Any conflict 0.180** 0.104* -0.131*

(0.051) (0.045) (0.054)

Meeting Vote In Favor (Fraction) 0.983** 0.924**

(0.011) (0.018)

Conflict * Meeting Vote In Favor 0.378**

(0.076) Meeting Vote In Favor (Fraction) for All j6=ⁱ 0.607** 0.530**

(0.074) (0.077) Conflict * Meeting Vote In Favor for All j6=^{i 0.537**}

(0.143) Vote In Favor (Fraction) among Conflicted Panelists 0.609** 0.506**

(0.046) (0.053) Conflict * Vote In Favor among Conflicted Panelists 0.518**

(0.052)

Year Fixed Effects x x x

Panel Fixed Effects x x x

N 592 592 592

+<0.10, * p<0.05, ** p<0.01

Reported coefficients are from a linear probability model with standard errors clustered at the meeting level

presented in Table 2.7, results from probit regressions using the same dependent and independent variables are of the same sign and statistical significance in all cases, but for this exercise, I focus on the linear probability model results, for ease of comparability across specifications and exercises involving the summing of coefficients.

The models in Table 2.7 consider the relationships between each of the additional factors presented and an individual’s probability of voting in favor of a new product. Panel A considers the peer effects of a higher/lower fractional vote in favor among all panelists at a meeting. Panel B considers the fractional vote in favor among all panelists except the reference individual at a meeting; this is likely a better measure because it is not mechanically biased toward finding a positive result and the one I focus on in interpretation.

Panel C considers the average vote in favor among all conflicted panelists at a meeting.

While absolute probabilities are difficult to calculate without adding in averaged year and panel fixed effects as well as a constant, the differences between conflicted and unconflicted individuals ceteris paribus can be calculated easily algebraically. Coefficients in a linear probability model are interpreted as the percentage point change in y associated with a one unit change in x. Thus, the difference between conflicted and unconflicted individuals is estimated by summing the coefficient on “any conflict” and the coefficient on the interaction between “any conflict” and the specific meeting measure of interest, multiplied by, e.g., the meeting average vote, or a 1 standard deviation change in the meeting-level vote.

For example, in panel B, the difference in probabilities between conflicted and uncon-flicted individuals at the sample average for the fraction of in favor votes would be calculated as -0.187 (the coefficient on “any conflict”) + 0.537 (the coefficient on “any conflict * meeting average vote in favor for all j6=i”) * 0.601 (the average meeting vote in favor for all j6=^i).

That is, at the meeting level average of in favor votes, conflicted individuals are 13.6 percent-age points more likely to vote favorably for a new product than unconflicted individuals.

This is notable because the implied difference between conflicted and unconflicted panelists voting propensities with peer effects is actually smaller in magnitude than the naive estimate without peer effects in column 1 (13.6 vs. 18 percentage points).

However, a further implication of the positive coefficient on the interaction term is that in a setting with very high rates of other panelists voting in favor, the difference between conflicted and unconflicted individuals’ in favor voting propensities will increase. For example using the estimated coefficients from Panel B again, we can calculate that, at just 1 standard deviation above the mean fraction of co-panelists’ in favor votes, a conflicted individual is 28.1 percentage points more likely to vote favorably for a new product than an unconflicted individual. Thus allowing for peer effects suggests that the bias associated with a financial conflict of interest will vary positively with the fraction of other panelists voting favorably for a new product. A corollary to this, of course, is that the difference between conflicted and unconflicted individuals’ voting behaviors is expected to decrease as the fraction of in favor votes among other panelists shrinks; at just one standard deviation below the mean fraction of other panelists’ in favor votes, the estimated difference between conflicted and unconflicted individuals shrinks to roughly zero.

Another interesting fact implied by Table 2.7 is that the fraction of in favor votes among other panelists is also positively predictive of how unconflicted individuals vote (panel B).

This is consistent with peer effects that lead to “herding behavior” – even for individuals without a financial conflict of interest. Moreover, panel C implies that a greater fraction of in favor votes among conflicted panelists (who represent just over 16 percent of the total voters) is associated with a higher probability of in favor voting among unconflicted individuals as well.

The data suggest that financially conflicted individuals are likely to have a bias toward in favor voting and the size of that bias is statistically increasing in the fraction of in favor votes on the panel itself and separately, among conflicted individuals only. Moreover, although these are only reduced form estimates, they are generally supportive of a meaningful relationship between panel composition and individual voting behaviors, a framework that has not yet been introduced into research on conflict of interest and voting. Lastly, I note that Table 2.7 uses data on both temporary and permanent members, but coefficients of similar magnitude and statistical significance are estimated when using the (smaller) sample

of permanent members only.

A policy question raised by these estimation exercises is then: how often might bias impact meeting outcomes? Extensions of this preliminary work will focus on simulating meeting outcomes if the same individuals were to sit on panels with differing compositions of colleagues.