- see [[220808_172005 distribution of dwells - buy vs no buy|distribution of dwells before and after adjusting for motor movement]]
# PCA on headline features
```r
> pca$rotation # weights/eigenvectors (pc1: informational, pc2: sketchiness)
PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8
familiarity 0.434 -0.166 0.279 0.257 0.683 -0.269 0.319 -0.018
favorability 0.116 0.311 0.811 -0.063 -0.420 -0.043 0.205 -0.090
impactful 0.429 0.285 -0.245 -0.196 0.052 0.655 0.361 -0.274
informative 0.451 0.074 -0.348 -0.399 -0.295 -0.635 0.147 -0.002
provocative 0.170 0.523 0.120 -0.291 0.381 -0.034 -0.673 0.015
sharing 0.346 0.285 -0.189 0.794 -0.282 -0.038 -0.227 -0.004
surprising -0.261 0.544 -0.129 0.053 0.137 -0.057 0.405 0.656
truth 0.441 -0.374 0.132 -0.119 -0.155 0.297 -0.192 0.697
```
## top 3 PC1 headlines
![[077_o9n_jXuPe.png]]
![[273_no_bJ5QB.png]]
![[010_no_bbcxR.png]]
## top 3 PC2 headlines
![[015_o9n_5OuMZ.png]]
![[055_o9n_dXijO.png]]
![[120_o9n_WAVTh.png]]
# predict motor-artifact-adjusted dwell time (amount of trying)
- linear regression
```r
> feols(ldwelladj ~ engagedc * ( scale(pc1) * scale(pc2) ) | user_id,
+ data = d0cov2[dwelladj %between% c(1000, 15000)], cluster = c("user_id", "entity_id"))
OLS estimation, Dep. Var.: ldwelladj
Observations: 45,758
Fixed-effects: user_id: 524
Standard-errors: Clustered (user_id & entity_id)
Estimate Std. Error t value Pr(>|t|)
engagedc 0.122676 0.010624 11.54757 < 2.2e-16 ***
scale(pc1) -0.007363 0.004774 -1.54229 1.2415e-01
scale(pc2) 0.021536 0.005160 4.17407 4.0184e-05 ***
scale(pc1):scale(pc2) -0.005308 0.004953 -1.07157 2.8485e-01
engagedc:scale(pc1) 0.010640 0.008812 1.20753 2.2827e-01
engagedc:scale(pc2) 0.034854 0.009865 3.53318 4.8144e-04 ***
engagedc:scale(pc1):scale(pc2) -0.012671 0.010450 -1.21261 2.2632e-01
```
## predict binary engagement (yes/no) (buy vs nobuy)
- logistic regression
```r
> feglm(engaged ~ scale(ldwelladj) * ( scale(pc1) * scale(pc2) ) | user_id,
+ data = d0cov2[dwelladj %between% c(1000, 15000)], cluster = c("user_id", "entity_id"), family = 'binomial') |> summary()
NOTE: 8 fixed-effects (247 observations) removed because of only 0 (or only 1) outcomes.
GLM estimation, family = binomial, Dep. Var.: engaged
Observations: 45,511
Fixed-effects: user_id: 516
Standard-errors: Clustered (user_id & entity_id)
Estimate Std. Error t value Pr(>|t|)
scale(ldwelladj) 0.288546 0.024313 11.86806 < 2.2e-16 ***
scale(pc1) 0.214525 0.050491 4.24873 2.1498e-05 ***
scale(pc2) -0.257264 0.048612 -5.29223 1.2084e-07 ***
scale(pc1):scale(pc2) -0.124139 0.044917 -2.76377 5.7138e-03 **
scale(ldwelladj):scale(pc1) 0.063588 0.020624 3.08320 2.0479e-03 **
scale(ldwelladj):scale(pc2) 0.036866 0.022937 1.60727 1.0800e-01
scale(ldwelladj):scale(pc1):scale(pc2) -0.054512 0.021176 -2.57426 1.0045e-02 *
```