Table 3.
Results of regression models.
| Variables | Model | ||||||||
| Poisson | Negative binomial | Zero-inflated negative binomial | Hurdle negative binomial | ||||||
| Coefficienta | P value | Coefficienta | P value | Coefficienta | P value | Coefficienta | P value | ||
| Independent variables | |||||||||
|
|
Beverage | –0.65 | .01 | –0.11 | .29 | 0.66 | .02 | –0.23 | .12 |
|
|
Coffee | –0.85 | <.001 | –0.02 | .73 | 0.14 | .34 | –0.15 | .21 |
|
|
Sweet | –0.11 | .34 | –0.19 | .006 | –0.60 | <.001 | –0.34 | .001 |
|
|
Dessert & bakery | –0.49 | .003 | –0.28 | <.001 | –0.70 | <.001 | –0.52 | <.001 |
|
|
Fruits | –0.29 | <.001 | –0.16 | <.001 | –0.13 | .008 | –0.24 | <.001 |
|
|
Herbs & spices | –0.07 | .87 | –0.57 | <.001 | –0.64 | <.001 | –0.60 | <.001 |
|
|
Menthol & mint | –0.35 |
|
–0.20 | .003 | –0.12 | .21 | –0.08 | .46 |
|
|
Nutty | 0.11 | .71 | 0.15 | .17 | –0.27 | .11 | 0.002 | .99 |
|
|
Cream | 0.09 | .49 | –0.13 | .13 | –0.79 | <.001 | –0.15 | .27 |
|
|
Tobacco | –1.22 | <.001 | –0.44 | <.001 | 0.33 | .14 | –0.55 | .001 |
|
|
Chocolate | –0.09 | .59 | 0.03 | .82 | –0.36 | .11 | 0.09 | .65 |
| Control variables | |||||||||
|
|
Promotion | 3.12 | <.001 | 3.06 | <.001 | 2.01 | <.001 | 3.16 | <.001 |
|
|
Posts (Coefficient x 10–4) | 3.46 | <.001 | 2.78 | <.001 | 1.18 | <.001 | 3.58 | <.001 |
|
|
Average share (Coefficient x 10–2) | .48 | <.001 | 6.67 | <.001 | 4.18 | <.001 | 10.17 | <.001 |
|
|
Topic 1 | 6.15 | <.001 | 1.61 | <.001 | 0.79 | <.001 | 1.39 | <.001 |
|
|
Topic 2 | 7.75 | <.001 | 1.11 | <.001 | 1.42 | <.001 | 1.17 | <.001 |
|
|
URL mention | 0.09 | .12 | 0.25 | <.001 | 0.27 | <.001 | 0.19 | <.001 |
|
|
Hashtag | –0.39 | <.001 | –0.05 | .25 | 0.01 | .75 | –0.02 | .72 |
| Intercept | –3.73 | <.001 | –0.97 | <.001 | –0.53 | <.001 | –18.79 | <.001 | |
| Model zero | No |
|
No |
|
Yes |
|
Yes |
|
|
| Model dispersion | No |
|
Yes |
|
Yes |
|
Yes |
|
|
| AIC/nb | 33.74 |
|
3.20 |
|
3.24 |
|
3.14 |
|
|
aEstimate of coefficient for each variable in the model.
bAIC: Akaike information criterion. The hurdle negative binomial regression model had the best performance as indicated by the lowest AIC/n (AIC value divided by number of observation).