* Collective memory
* Created: FdSS 2015/10/23
* Udated: FdSS 2015/11/13
capture cd c:\Users\KSS\Dropbox\!NetLogo\CollectiveMemory\
capture log close
log using cmAnalysis_2015-11-07, text replace
use cm05Clear, clear
* Constant deleted variables (their values):
* numberofevents (40) Number of events in each memory
* savefirstreligion (true) Whether graphical interface indicates first religion of agents
* numberofindividuals (84) Number of individual agents/memory spots
* connectivity (0.9) Coeficient influencing density of network of individuals
* step (1095) Number of simulation steps [1095 day - cca 3 years]
sum il-asr, detail
***
*** Analysis
***
*** Systematical analysis
* Simulated/experimented parameters, i.e. independent variables>
* nof {2, 3, 4}
* dt {1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0}
* nob {2, 4, 6, 8}
* il <1.928571, 32.83333> ~ not exactly set as parameter: just randomly generated
* ml <6.809524, 7.801974> ~ not exactly set as parameter: just randomly generated
* please note: ML was measured at the end of the simulation, so it is not "starting point",
* it is also partial result of the simulation, e.g. ML is dependent on following variable CM
* cm <74, 468> ~ not exactly set as parameter: just randomly generated
* Dependent variables:
* acs <0, 5503> all cases of dissonance which was solved during simulations
* ...
* nsm <0, 1159> unsolved cases of dissonance
* rc <0, 29> cases of religious attitude change
*************************
* Question 1: *
* Why mechanism starts? *
*************************
* the best predictor is dissonance treshold DT - more than 80% of explained variance!!!
reg lnacs dt
est store bestAlone
* but for better fit we need another variables: NOF, NOB, CM
reg lnacs dt cm nof nob
est store better
* interesting is that individually very poor variable IL
reg lnacs il
* is useful in combination with the others
reg lnacs dt cm nof nob il
est store best
* comparisons of models show that
lrtest bestAlone better
lrtest better best
est stat bestAlone better best
* But the IL effect is not very significant substantively:
* if we take median value of ACS (236), median value of IL (7),
* so, ceteris paribus, difference in IL of median (7) means
* lowering median ACS (236) by 9 (to 227)
* effect of IL is -0.006, by 7 it is -0.04, and than we should delogarithm it...
display(exp(ln(236) - 0.04))
* with some extreme values of ACS it would have bigger impact, e.g.
display(exp(ln(4000) - 0.04))
* here almost 160...
* So, IL plays a role in cases of other extreme values>
* low DT and high NOF and high NOB,
* than it is important whether the network is dense or not...
* Substantive significance of other variables>
* NOB - effect of one more boss is +0.17
display(exp(ln(236) + 0.17))
* so, one boss brings 43 cases more of processed dissonance
*
* NOF - effect of one more frame is +0.1
display(exp(ln(236) + 0.1))
* so, one frame brings 24 cases more of processed dissonance
*
* DT - effect of 0.1 higher treshold is -0.44
display(exp(ln(236) - 0.44))
* so, DT higher of 0.1 supress cases of processed dissonance by 84
*
* CM - effect of 20 more comemorations is 0.05
display(exp(ln(236) + 0.05))
* so, 20 comemorations bring 12 cases more of processed dissonance
*
* same as in case of IL, if we take as a benchmark case with higher ACS,
* the effects are sounder
* But according normal conditions it is evident, DT has the highest effect, NOB has very strong effect,
* NOF has strong effect, CM has reasonable effect, and IL has substantively poor effect
* So we can see, that crucial effect has DT
reg lnacs dt cm nof nob il
predict y
lab var y "Predicted all cases solving"
replace y = round(exp(y) - 24.217)
gen dtx = dt + 0.02
scatter y dt || scatter acs dtx
*scatter y dt
* Between 1.1 and 1.3 there is big variability, between 1.4 and 1.7 middle,
* and between 1.8 and 2.0 very low. It is evident that crucial DT is for high levels of DT.
* So, it sounds reasonable to reduce sample just to DT between 1.1 and 1.4 or 1.3,
* to find weight of other variables...
preserve
drop if dt > 1.3
*gen dtx = dt + 0.02
*scatter y dt || scatter acs dtx
* the best predictors comemorations CM and number of bosses NOB - both more than 30% of explained variance
reg lnacs nob
est store bestAlone1
reg lnacs cm
est store bestAlone2
* DT has still reasonable prediction power - about 25% of explained variance
reg lnacs dt
* but for better fit we need got together all reasonable variables: NOF, NOB, CM, DT
reg lnacs dt cm nof nob
est store better
* interesting is that individually very poor variable IL
reg lnacs il
* is useful in combination with the others
reg lnacs dt cm nof nob il
est store best
* comparisons of models show that
lrtest bestAlone1 better
lrtest bestAlone2 better
lrtest better best
est stat bestAlone1 bestAlone2 better best
* Substantive significance of variables>
* NOB - effect of one more boss is +0.15
display(exp(ln(1042) + 0.15))
* so, one boss brings 168 cases more of processed dissonance
*
* NOF - effect of one more frame is +0.073
display(exp(ln(1042) + 0.073))
* so, one frame brings 78 cases more of processed dissonance
*
* DT - effect of 0.1 higher treshold is -0.37
display(exp(ln(1042) - 0.37))
* so, DT higher of 0.1 supress cases of processed dissonance by 323
*
* CM - effect of 20 more comemorations is 0.05
display(exp(ln(1042) + 0.05))
* so, 20 comemorations bring 53 cases more of processed dissonance
*
* IL - effect of 5 more average links per group member is -0.02
display(exp(ln(1042) - 0.02))
* so, density of individual links higher of 5 more links per individual
* supress cases of processed dissonance by 21
*
* It is again evident, DT has the highest effect, NOB has very strong effect,
* NOF has strong effect, CM has reasonable effect, and IL has substantively poor effect
*restore
* Question is, why variables have their effect>
* NOB - more bosses stimulate more the rest of the group and it leads to higher ACS
* NOF - more frames are bigger potential for dissonance
* DT - higher treshold means that individual can stand case of actual dissonance,
* which fades away with Ebbinhause's forgeting curve, and because we have more cases
* of naturally standed dissonace, we face lower ACS
* CM - the more frequent comemorations the higher stimulation of events' saliency and
* higher levels of dissonance
* IL - IL is doble-edged: it should stimulate dissonance, because higher density means also
* more connection to bosses and more frequent stimulation during comemorations, which should lead
* to higher ACS; on the other hand, more dense network means that its members help each other to stand
* urgent cases of dissonance (counted in ACS but solved in first phase of mechanism in social network phase)
* so, these individuals change nothing in their minds and dissonance fades away with Ebbinhause's forgeting curve
********************************************************
* So, now answers to Question 1: Why mechanism starts? *
********************************************************
*
* {DT} indivuduals are not strong enough - they can't stand dissonance:
* high treshold of dissonance is best predictor of low ACS
* implication for model: focus more on DT - different tresholds for different individuals, different distributions over population
* implication for research: focus what are DT in real population... according my knowledge it is hard to measure it directly...
*
* {NOB} more bosses stimulates higher cognitive dissonance and higher ACS through their right choose event and comemorate it,
* they choose events consistent with their religious attitude, the others could face cognitive dissonance because inconsistent framing
* of their religious attitude and the event
* implication for model: rebuild comemoration routine to more realistic fashion - randomly choose event and let bosses to decide
* whether comemorate or not according consistency of framing (when consistent = comemorate), on the other hand this comemoration
* raises saliency of boss' inconsistently framed event anyway, so he could start suffering dissonance and start comunicating with his network neighbors
* implication for research: consult literature for the role bosses play in shaping collective memory
*
* {NOF} the more frames the bigger potential for dissonance, probably also because we have stil same number of events, so if they are divided to
* more frames, we could face extreme distributions easier and we face cases of cognitive dissonance more frequently
* implication for model: focus more on number of events and its interaction with NOF
* implication for research: focus on real number of religious attitudes frames and its connection with cognitive dissonance
************************************************
* Question 2: *
* Description of mechanism / success of phases *
************************************************
* For the low DT:
sum cm acs s1 s2 s3 s4 cof kml rc as nsm sm
recode nsm (11/max=10000)
tab nsm
* For whole sample
restore
preserve
sum cm acs s1 s2 s3 s4 cof kml rc as nsm sm
recode nsm (11/max=10000)
tab nsm
restore
* There are 244 comemoration on the average per one simulation, it is 3 comemoration per 2 weeks, a bit higher, may be...
* On the average it leads to 545 cases of cognitive dissonance under cure per simulation (2.2 per comemoration)
* On the average 22.5 % of these dissonant cases are postponed by network of individuals (122.5 per simulation)
* On the average 73 % of dissonant cases are postponed by wider picture (94 % of still unsolved dissonance, 399 per simulation)
* So, discusions and taking wider picture postpones 95.5 % of cognitive dissonance, in my view it sounds realistic,
* but it needs check to the experimental or other kind of data about coping with dissonance.
* On the average 1.2 still dissonant cases per simulation is solved by connecting to new same framed event
* (0.2% of all, 5% of still unsolved cases).
* On the average there are still 22.4 dissonant cases, 15.4 are reframed, in average 7 cases is applied killing of memory link between dissonant events.
* On the average 1.3 of 22.4 still dissonant cases per simulation is solved by re-framing of event or killing link between the events
* (0.2% of all, 6% of still unsolved cases).
* 21.2 cases are on the average unsolved per one simulation (3.9%).
* All cases of dissonance are solved in 46% of simulations, at least one case of unsolved dissonance is in 54% of simulations.
* At maximum 10 unsolved cases are in 76 % of simulation.
* Above I noted that on the average in 15.4 cases per simulation is reframed cognitive element. In 0.7 case it is religious attitude,
* it means, 4.5% of reframing is reframing of religious attitude. It means that religion is reframed more frequently than other
* cognitive elements (events) - even fraction would be 2.5%. It is because religious attitude is cinnected with all events at the start
* of simulation, so it is part of handling dissonance of all events which raises salincy of religion and in case the majority of events in the memory
* are framed differently, it could lead to change of religion.
* But in the subsample of DT < 1.4, there are a bit different results:
* First two phases postpone 94.7% cases - it is almost same, as unsolved cases (4.7%).
* But seemingly higher is number of solved cases by third and fourth phase
* (3.4 resp. 3.6 on average, 0.25% resp. 0.27% of all cases, 5% resp. 5%),
* but in relative figures it is same.
* So, with lower DT, there are higher number of cases, but solved fractions are same.
* But the only difference exists - completely all cases solved are in 4.8% of simulations,
* at maximum 10 unsolved cases are in 36% of simulations.
****************************************
* Question 3: *
* What influence success of mechanism? *
****************************************
* Our model
logit sm nof dt nob cm il
* with control of ACS
logit sm nof dt nob cm il acs
* Same for RC
logit rc nof dt nob cm il
* with control of ACS
logit rc nof dt nob cm il acs
exit