316 Personality and Social Psychology Review 18(4) disrogard that it to updat the e impr)Above.it was noted that the models impression)/(1 +exp(3 integration rule does not assume a negativity bias.An addi mpre the parame -1.0 is e prob reason for that specification 11s t a negativity bi on)it is 50 The fumn ein that mali y mak ould r is symmetric,so the probability of interaction for a positive impression of +0.5 is .82 and for +1.0,it is 953.Although easier for other observers to disregard them. nrell use the parameter the probability of interaction vary more linearly with the Evil Targets and Malicious Observers valence of the impression,whereas larger values make the To address the questions laid out in the introduction,we use patte more like a step function (in eract if and only if wo versions of the model.In o 34 ver ch decision the observers mp ression remains unchanged. the overall mean (4.5 rather than +0.5)and standard devia I use two versions of the model,one that allows this decision ion of 1.0.On all other occasions,their behaviors have sightly more p tive mean of +0./5 sc ave On sider thes Assumptions About Gossip opaths.who act unusually charming most of the time as mask for their tendency to produce rare,extremely nega stly borrowed from Smith an The key qu situation is wh :ta cific pr tion oftrials with g gets from the other.noma action with a target.So in a model run involving 100 trials Note that to form such differentiated impressions,the with Target Y,if Observer X gossips .4 c mpression-forn ation process has to produce nonlinea wil an expen neganv described)and on 40 trials,ObserverXwill gos sin instead of sitive behaviors.A linear rule (such a simple average interacting with Y. the valence of the behay viors)would allow the slightly posi used bers 01 sip can be used n the mode by t their ive trial Ohs s ofth ch erver i would othewise on thei curren mpres n proc me. at a single ega E: third narty's imnr sion is used to update Obse rver X's cu atly reduce the likelihood of further inter so it rent impression of Target Y using the same impression-for egative effect on the impres averaging formul ut mo on ity is of the target of whom holds the most eg f this eg tive that it has a high probability of being passed along to tive impression.It is as if X asks Z "tell me some juicy ther observers in gossip,spre ding the information through g gossip, targe social netw th of compensation by on of that ta using the same im sion fo n the ot ion of the model all ta nal tion ave ing formula producing behaviors with the same mean of +0.5.But one inally,in one version of the model. add an assum server fters high malicious gossip abou tion that may try to prote again ose targ ofasmplethicshoidfgospoOeinoamanthaif 1 egative-sO fers from the observer's current impression of the relevan mpression of those targets.The key question in this situa target by more than 1.0 in absolute value,the observer wil tion is whether the other observers are susceptible to this316 Personality and Social Psychology Review 18(4) occasion. Interaction is less likely the more negative the impression is, with interaction likelihood following a Luce choice function, p = exp(3 × impression) / (1 + exp(3 × impression)). With the parameters Denrell uses, the probabil￾ity of interaction for an impression of −1.0 is .047, for −0.5 it is .18, and for 0.0 (neutral impression) it is .50. The function is symmetric, so the probability of interaction for a positive impression of +0.5 is .82 and for +1.0, it is .953. Although Denrell uses 3 for the parameter value, it can be varied (see robustness analyses later in this article). Smaller values make the probability of interaction vary more linearly with the valence of the impression, whereas larger values make the pattern more like a step function (interact if and only if valence is above a threshold value). If the observer chooses not to interact with a target on a particular trial following this decision rule, the observer’s impression remains unchanged. I use two versions of the model, one that allows this decision process and one that does not (so interaction takes place on each trial, regardless of the observer’s impression). Assumptions About Gossip These assumptions are also mostly borrowed from Smith and Collins (2009). Each observer is assumed to gossip on a spe￾cific proportion of trials, with gossip replacing a direct inter￾action with a target. So in a model run involving 100 trials pairing Observer X with Target Y, if Observer X gossips .4 of the time, X will have 60 direct interaction opportunities (not all of which may take place, based on the decision rule just described) and on 40 trials, Observer X will gossip instead of interacting with Y. Two different versions of gossip can be used in the model. One (that used by Smith and Collins) we term directed gos￾sip: On a gossip trial, Observer X picks a different observer Z and obtains Z’s current impression of Target Y specifi￾cally—the target with which Observer X would otherwise have interacted with on this trial. It is as if, instead of inter￾acting with Y, X asks a friend “what do you think of Y?” The third party’s impression is used to update Observer X’s cur￾rent impression of Target Y using the same impression-for￾mation averaging formula. The alternative assumption is termed interesting gossip: Observer X still picks a third-party observer Z but obtains Z’s impression of the target of whom Z holds the most nega￾tive impression. It is as if X asks Z, “tell me some juicy, interesting gossip,” and hears about the target Z dislikes most. This impression is used to update Observer X’s current impression of that target, using the same impression-forma￾tion averaging formula. Finally, in one version of the model, we add an assump￾tion that observers may try to protect themselves against influence by biased or malicious gossip. This takes the form of a simple threshold: If gossip conveys information that dif￾fers from the observer’s current impression of the relevant target by more than 1.0 in absolute value, the observer will disregard that gossip (not using it to update the current impression). Above, it was noted that the model’s impression integration rule does not assume a negativity bias. An addi￾tional reason for that specification is that a negativity bias might be seen as artificially making this “disregard” rule more effective, in that malicious gossip would make observ￾ers’ impressions more extremely negative, making it even easier for other observers to disregard them. Evil Targets and Malicious Observers To address the questions laid out in the introduction, we use two versions of the model. In one, a small proportion of the targets (4 of 20) are assumed to be evil. On 5% of trials, they produce highly negative acts—with a mean 5.0 lower than the overall mean (−4.5 rather than +0.5) and standard devia￾tion of 1.0. On all other occasions, their behaviors have a slightly more positive mean of +0.75 so that overall, their behaviors have the same mean valence (+0.5) as all other targets. One could consider these targets as analogous to sociopaths, who act unusually charming most of the time as a mask for their tendency to produce rare, extremely nega￾tive behaviors. The key question in this situation is whether observers can form impressions that differentiate the evil tar￾gets from the other, normal targets. Note that to form such differentiated impressions, the impression-formation process has to produce nonlinear effects, so that that an experience of an extremely negative behavior cannot be compensated for by several subsequent positive behaviors. A linear rule (such as a simple average of the valence of the behaviors) would allow the slightly posi￾tive behaviors produced by the evil targets on 95% of trials to compensate for their 5% of strongly negative behaviors. Two aspects of the model can generate nonlinear effects. One is the version of the model in which observers actively decide whether to interact with a target based on their current impressions. This decision process means that a single nega￾tive action may make the impression sufficiently negative to greatly reduce the likelihood of further interactions, so its negative effect on the impression will not be compensated by future, more positive actions. The second possibility is the “interesting” gossip mode, where a single negative action may make this observer’s impression of this target so nega￾tive that it has a high probability of being passed along to other observers in gossip, spreading the information through the social network beyond the reach of compensation by future, more positive actions. In the other version of the model, all targets are normal, producing behaviors with the same mean of +0.5. But one observer offers highly negative, malicious gossip about four specific targets. When asked about one of those targets (in directed gossip) or when asked to give “interesting” gossip, this observer reports a strongly negative (−5.0) impression of those targets. The key question in this situa￾tion is whether the other observers are susceptible to this Downloaded from psr.sagepub.com at Remen University of China on September 6, 2015