Creativity and Reason-Based Choice
11/3/17
Initial investigation on the relationship between creativity and reason-based choice. The idea is that because creative people can more easily justify things (they are not as cognitively entrenched/dependent on conventional justifications), they will be less susceptible context effects involving justification.
We measured creativity through a behavioral-based measure of the RAT at the beginning of the survey and a self-report at the end. The RAT measure was taken from Gino and Ariely (2012), but this measure is messy because it requires both divergent thinking (seeing words as multiple meanings, rather than just the dominant one), but also convergent thinking (discovering connections between disparate things). We generated the self-report measures from previous scales/our own intuition.
We had several versions of reason-based choice: 1) 3 compromise questions, 2) Shafir, Simonson, & Tversky 1993 accept/reject paradigm, 3) disjunction fallacy (which also included a reason listing procdure at the end).
Creativity Measures
The alpha for the 4 creativity items was 0.77.
Mean self-reported creativity was 4.79 out of 7.
SD self-reported creativity was 0.99.
Mean RAT score (# correct out of 16) was 3.47.
Mean # of RAT questions attempted was 5.56.
Distributions
qplot(create_total, data=d, xlab = "Self-reported Creativity")## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
qplot(rat_correct, data=d, xlab = "# of Correct RAT (out of 16)")## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
These measures do not significantly relate. We are going to have to examine them separately. (I’m not totally surprised by this finding– I think we are interested in divergent thinking and research suggests that the RAT might not capture that despite Ginao and Ariely using this as their manipulation check.)
cor.test(d$create_total, d$rat_correct)##
## Pearson's product-moment correlation
##
## data: d$create_total and d$rat_correct
## t = -0.85855, df = 68, p-value = 0.3936
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
## -0.3304875 0.1346964
## sample estimates:
## cor
## -0.1035552cor.test(d$create_total, d$rat_attempt) #this is number of RAT items attempted##
## Pearson's product-moment correlation
##
## data: d$create_total and d$rat_attempt
## t = 0.64314, df = 68, p-value = 0.5223
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
## -0.1601438 0.3071189
## sample estimates:
## cor
## 0.07775615Reason-based choice
Disjunctive
Only 26% said that they would not take the vacation. Looking back at the original materials, I just realized we made a mistake– we should have allowed a deferral option (i.e., “wait until I find out the outcome of test.”)
We also recorded the number of reasons listed for choosing to go on the vacation. People on average listed 4.26 reasons, with a SD of 1.97
Compromise
Out of the 3 choices, the average number of compromise options chosen was 1.07
Percent compromise for hotel: 51% Dogs: 29% Ticket: 27%
Accept/Reject
d$response[d$responseR=="1"] = "enriched"
d$response[d$responseR=="0"] = "impoverished"
table(d$response, d$cond)##
## accept reject
## enriched 22 15
## impoverished 16 17For the accept/reject question, we do not observe the expected effect. We should see that the enriched condition has a higher choice share for both accept and reject. This may just be due to small sample size.
Reason-based choice and Creativity
Disjunctive
summary(glm(disjunctionR ~ create_total, family="binomial", data=d))##
## Call:
## glm(formula = disjunctionR ~ create_total, family = "binomial",
## data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.0090 -0.7865 -0.7023 0.8601 1.8595
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.7433 1.4576 -1.882 0.0598 .
## create_total 0.3457 0.2894 1.195 0.2322
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 79.807 on 69 degrees of freedom
## Residual deviance: 78.323 on 68 degrees of freedom
## AIC: 82.323
##
## Number of Fisher Scoring iterations: 4This is opposite of what we predict (buy vacation = 0, don’t buy = 1). As creativity increases, their chance of not buying increases. But this may be due to the fact that we forgot to include the “delay” option.
We do get a promising effect for number of reasons generated:
summary(lm(reasons_total ~ create_total, data=d))##
## Call:
## lm(formula = reasons_total ~ create_total, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.2341 -1.3956 -0.4111 1.2100 6.3126
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.1660 1.1083 1.052 0.29654
## create_total 0.6459 0.2268 2.848 0.00582 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.874 on 68 degrees of freedom
## Multiple R-squared: 0.1066, Adjusted R-squared: 0.09342
## F-statistic: 8.11 on 1 and 68 DF, p-value: 0.005818ggplot(d, aes(x= create_total, y = reasons_total)) +
geom_point() +
geom_smooth(method="lm", formula = y ~ x)
summary(lm(reasons_total ~ create_total*disjunctionR, data=d))##
## Call:
## lm(formula = reasons_total ~ create_total * disjunctionR, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.6195 -1.3462 -0.3212 1.2001 5.4005
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.7091 1.2701 1.346 0.1830
## create_total 0.5118 0.2643 1.936 0.0571 .
## disjunctionR -1.9698 2.7262 -0.723 0.4725
## create_total:disjunctionR 0.4683 0.5406 0.866 0.3895
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.885 on 66 degrees of freedom
## Multiple R-squared: 0.1225, Adjusted R-squared: 0.08257
## F-statistic: 3.07 on 3 and 66 DF, p-value: 0.03377As creativity increases, the number of reasons listed increases. This is not moderated by the choice they made (to take or not take the vacation).
Now let’s look at the RAT:
summary(glm(disjunctionR ~ rat_correct, family="binomial", data=d))##
## Call:
## glm(formula = disjunctionR ~ rat_correct, family = "binomial",
## data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.9512 -0.7887 -0.6918 0.7668 1.8923
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -1.6079 0.5333 -3.015 0.00257 **
## rat_correct 0.1499 0.1195 1.254 0.20979
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 79.807 on 69 degrees of freedom
## Residual deviance: 78.224 on 68 degrees of freedom
## AIC: 82.224
##
## Number of Fisher Scoring iterations: 4This is opposite of what we predict (buy vacation = 0, don’t buy = 1). As creativity increases (measured by RAT), they are more likely to select “dont buy.”
summary(lm(reasons_total ~ rat_attempt, data=d))##
## Call:
## lm(formula = reasons_total ~ rat_attempt, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.6010 -1.3148 -0.1244 0.9948 6.2076
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 2.93293 0.51733 5.669 3.18e-07 ***
## rat_attempt 0.23829 0.08391 2.840 0.00595 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.875 on 68 degrees of freedom
## Multiple R-squared: 0.106, Adjusted R-squared: 0.09288
## F-statistic: 8.065 on 1 and 68 DF, p-value: 0.005947summary(lm(reasons_total ~ rat_correct + rat_attempt, data=d))##
## Call:
## lm(formula = reasons_total ~ rat_correct + rat_attempt, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.671 -1.233 -0.078 1.042 5.877
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 2.96322 0.52492 5.645 3.63e-07 ***
## rat_correct -0.05637 0.12766 -0.442 0.6603
## rat_attempt 0.26805 0.10802 2.482 0.0156 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.886 on 67 degrees of freedom
## Multiple R-squared: 0.1086, Adjusted R-squared: 0.08202
## F-statistic: 4.082 on 2 and 67 DF, p-value: 0.02123It looks like the number of reasons listed does not relate to the RAT except for when looking at number attempted (which could be a proxy for effort). This suggests our self-report measure may be better.
Summary: no effect of creativity on choice in the disjunction paradigm, but our question wasn’t identical to past work. We do get an effect of self-reported creativity on number of reasons people generated, which is in line with our thinking. This also suggests that self-reported creativity might be a better measure than RAT.
Compromise
summary(lm(comp_total ~ create_total ,data=d))##
## Call:
## lm(formula = comp_total ~ create_total, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.33368 -0.74725 -0.05322 0.65357 2.17626
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 2.04759 0.48532 4.219 7.43e-05 ***
## create_total -0.20397 0.09932 -2.054 0.0438 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8207 on 68 degrees of freedom
## Multiple R-squared: 0.05841, Adjusted R-squared: 0.04456
## F-statistic: 4.218 on 1 and 68 DF, p-value: 0.04384I’m not sure if a linear regression is the right way to test the compromise data (ranging from 0 = no compromise options chosen to 3 = all compromise options chosen), but we do get a significant effect of self-reported creativity on compromise p =.04. As creativity increases, the number of compromise options chosen decreases.
Here are the results for each individual item. Directional for each item, only significant (p=.05) for dogs.
#hotel
summary(glm(comp_1 ~ create_total, family="binomial",data=d))##
## Call:
## glm(formula = comp_1 ~ create_total, family = "binomial", data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.576 -1.172 0.901 1.113 1.475
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 1.6490 1.2268 1.344 0.179
## create_total -0.3322 0.2506 -1.326 0.185
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 96.983 on 69 degrees of freedom
## Residual deviance: 95.166 on 68 degrees of freedom
## AIC: 99.166
##
## Number of Fisher Scoring iterations: 4#dogs
summary(glm(comp_2 ~ create_total, family="binomial",data=d))##
## Call:
## glm(formula = comp_2 ~ create_total, family = "binomial", data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.2855 -0.8543 -0.6774 1.2421 1.8990
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 1.5889 1.3323 1.193 0.2330
## create_total -0.5353 0.2845 -1.881 0.0599 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 83.758 on 69 degrees of freedom
## Residual deviance: 79.973 on 68 degrees of freedom
## AIC: 83.973
##
## Number of Fisher Scoring iterations: 4#tickets
summary(glm(comp_3 ~ create_total, family="binomial",data=d))##
## Call:
## glm(formula = comp_3 ~ create_total, family = "binomial", data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.8651 -0.8038 -0.7742 1.5529 1.7026
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -0.5722 1.3178 -0.434 0.664
## create_total -0.0871 0.2718 -0.320 0.749
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 81.854 on 69 degrees of freedom
## Residual deviance: 81.752 on 68 degrees of freedom
## AIC: 85.752
##
## Number of Fisher Scoring iterations: 4Now let’s look at the RAT
summary(lm(comp_total ~ rat_correct ,data=d))##
## Call:
## lm(formula = comp_total ~ rat_correct, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.08461 -1.06088 -0.07322 0.91919 1.94197
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.084609 0.185290 5.854 1.53e-07 ***
## rat_correct -0.003797 0.044733 -0.085 0.933
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8457 on 68 degrees of freedom
## Multiple R-squared: 0.0001059, Adjusted R-squared: -0.0146
## F-statistic: 0.007204 on 1 and 68 DF, p-value: 0.9326summary(lm(comp_total ~ rat_attempt ,data=d))##
## Call:
## lm(formula = comp_total ~ rat_attempt, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.08467 -1.06278 -0.07007 0.91777 1.93965
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 1.057913 0.233371 4.533 2.42e-05 ***
## rat_attempt 0.002432 0.037851 0.064 0.949
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8457 on 68 degrees of freedom
## Multiple R-squared: 6.071e-05, Adjusted R-squared: -0.01464
## F-statistic: 0.004129 on 1 and 68 DF, p-value: 0.949No relationship between RAT and compromise.
Here are the results for each individual item.
#hotel
summary(glm(comp_1 ~ rat_correct, family="binomial",data=d))##
## Call:
## glm(formula = comp_1 ~ rat_correct, family = "binomial", data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.356 -1.212 1.009 1.133 1.443
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) 0.4103 0.4447 0.923 0.356
## rat_correct -0.1017 0.1079 -0.943 0.346
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 96.983 on 69 degrees of freedom
## Residual deviance: 96.077 on 68 degrees of freedom
## AIC: 100.08
##
## Number of Fisher Scoring iterations: 4#dogs
summary(glm(comp_2 ~ rat_correct, family="binomial",data=d))##
## Call:
## glm(formula = comp_2 ~ rat_correct, family = "binomial", data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.8786 -0.8273 -0.8107 1.5091 1.6602
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -0.75276 0.48105 -1.565 0.118
## rat_correct -0.04784 0.11932 -0.401 0.688
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 83.758 on 69 degrees of freedom
## Residual deviance: 83.595 on 68 degrees of freedom
## AIC: 87.595
##
## Number of Fisher Scoring iterations: 4#tickets
summary(glm(comp_3 ~ rat_correct, family="binomial",data=d))##
## Call:
## glm(formula = comp_3 ~ rat_correct, family = "binomial", data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.9837 -0.8147 -0.7137 1.1843 1.8638
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -1.5431 0.5241 -2.945 0.00323 **
## rat_correct 0.1527 0.1180 1.294 0.19579
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 81.854 on 69 degrees of freedom
## Residual deviance: 80.164 on 68 degrees of freedom
## AIC: 84.164
##
## Number of Fisher Scoring iterations: 4Just for completeness, looking at both RAT and self-report in the same model:
summary(lm(comp_total ~ rat_correct + create_total ,data=d))##
## Call:
## lm(formula = comp_total ~ rat_correct + create_total, data = d)
##
## Residuals:
## Min 1Q Median 3Q Max
## -1.29122 -0.72875 -0.03326 0.64906 2.18701
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 2.10825 0.52884 3.987 0.000168 ***
## rat_correct -0.01317 0.04394 -0.300 0.765302
## create_total -0.20710 0.10053 -2.060 0.043280 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.8263 on 67 degrees of freedom
## Multiple R-squared: 0.05967, Adjusted R-squared: 0.0316
## F-statistic: 2.126 on 2 and 67 DF, p-value: 0.1273Conclusion: This is promising! It looks like, for at least the self-reported creativity, compromising is reduced for more creative individuals.
Accept/Reject
summary(glm(responseR ~ cond*create_total, family="binomial", data=d))##
## Call:
## glm(formula = responseR ~ cond * create_total, family = "binomial",
## data = d)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -1.514 -1.137 0.875 1.177 1.414
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -1.3245 1.6971 -0.780 0.435
## condreject 1.0773 2.4093 0.447 0.655
## create_total 0.3478 0.3535 0.984 0.325
## condreject:create_total -0.3226 0.4945 -0.652 0.514
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 96.812 on 69 degrees of freedom
## Residual deviance: 94.961 on 66 degrees of freedom
## AIC: 102.96
##
## Number of Fisher Scoring iterations: 4ggplot(d, aes(x= create_total, y = responseR, group=cond, color=cond)) +
geom_point() +
geom_smooth(method="glm", formula = y ~ x)
Here we expect an interaction: the proportion choosing the enriched option should not differ for people low in creativity, but the proportion should differ for people high in creativity. although not significant, the graph shows some evidence for this relationship. Some problems: 1) we don’t conceptually replicate the main effect of accept/reject having the same proportion of the enriched option. 2) the sample size is very small to observe an interaction.
Summary: I think this inconclusive and deserves a second test