Results from Bayesian models. - see [[220906_192146 taskus replication lucid - india and philippines|lucid replication results]] - [prereg](https://docs.google.com/document/d/1MzvJf3dtW9oGKcD3Feyisthh4O0pGf6E42--QXi-NV4/edit) - coding - condition (0/1): control-interest vs treatment-accuracy - repetition (-0.5/0.5): novel vs repeated - veracity (-0.5/0.5): false vs true - models with covariates (11 in total) - numeric covariates are z-scored - categorical covariates have many levels and are dummies - see exploratory [[220729_114133 forest - doubly robust estimators|causal forest doubly robust estimator results]] # bayesian models Including all the categorical control variables and interactions led to joint multicollinearity. Have to determine which features were collinear and then set tighter priors on those to constrain the sampler. # hypothesis 2: false headlines only model: `accuracy_rating ~ repetition * condition` (false headlines only) ```r param b low high p_direction 1: b_Intercept 3.200 2.931 3.472 1.000 2: b_repetitionc 0.077 -0.061 0.214 0.862 # rep effect for control group 3: b_conditiond 0.137 -0.139 0.408 0.840 4: b_repetitionc:conditiond -0.126 -0.300 0.048 0.922 # almost significant ``` ![[1659982188.png]] After controlling for all covariates, interaction effect became significant. Marginal effects of repetition and condition also became significant. ```r param b low high p_direction 1: b_repetitionc:conditiond -0.256 -0.445 -0.066 0.996 # marginal effects term estimate conf.low conf.high 1: repetitionc -0.003 -0.075 0.072 # no effect of repetition 1: conditiond -0.223 -0.445 0.000 # treatment group rated headlines as less true ``` # hypothesis 1: all headlines - only CONTROL group, because H1 interaction effect is significant model: `accuracy_rating ~ veracity * repetition` (control group only) ```r param b low high p_direction 1: b_Intercept 3.512 3.309 3.710 1.000 2: b_veracityc 0.623 0.373 0.863 1.000 # true headline srated more true 3: b_repetitionc 0.042 -0.059 0.143 0.791 # no repetition effect 4: b_veracityc:repetitionc -0.028 -0.223 0.173 0.609 # no interaction ``` ![[1659980775.png]] After controlling for covariates, no change in interaction estimate ```r param b low high p_direction 1: b_veracityc:repetitionc -0.03 -0.221 0.171 0.627 # no interaction # marginal effects contrast estimate conf.low conf.high 1: veracityc 0.621 0.251 0.986 # true headlines rated more true 2: repetitionc 0.042 -0.060 0.142 # no repetition effect # only false headlines contrast estimate conf.low conf.high 2: repetitionc 0.057 -0.089 0.202 # slightly stronger effect than true headlines (below) # only true headlines contrast estimate conf.low conf.high 2: repetitionc 0.027 -0.110 0.167 # slightly weaker effect than false headlines ``` --- # extra analyses below # hypothesis 1: all headlines - CONTROL & TREATMENT groups model: `accuracy_rating ~ veracity * repetition` (control and treatment groups) ```r param b low high p_direction 1: b_Intercept 3.624 3.475 3.773 1.000 2: b_veracityc 0.650 0.431 0.863 1.000 # true headlines rated as more true 3: b_repetitionc 0.016 -0.042 0.072 0.704 # no repetition 4: b_veracityc:repetitionc 0.026 -0.085 0.135 0.686 # no interaction ``` After controlling covariates, no change in interaction estimate ```r param b low high p_direction 1: b_veracityc:repetitionc 0.027 -0.086 0.14 0.689 # marginal effects contrast estimate conf.low conf.high 1: veracityc 0.642 0.382 0.892 # true headlines rated as more true 2: repetitionc 0.015 -0.044 0.071 # no repetition effect # false headlines contrast estimate conf.low conf.high 2: repetitionc 0.001 -0.084 0.086 # true headlines contrast estimate conf.low conf.high 2: repetitionc 0.028 -0.046 0.102 ```