Edchat HLM

Loading, setting up

library(lme4)
library(tidyverse)
library(sjPlot)

g_tweet_file <<- "data/tweet_annotated.csv"
g_profile_file <<- "data/profile_data.csv"


safe_log <- function(x) {
    ifelse(x == 0, 0, log(x))
}


integer_NAs_to_zeros <- function (x)  {
    
    if (is.na(x) && is.integer(x)) {
        0
    } else {
        x
    }
}

readData <- function (f) {
    message("reading source  file at ", f)
    myDF <- read.csv(f, stringsAsFactors = F, na.strings = c("NA","NaN", " ")) # changed to read_csv from the readr package, as it has nice defaults (stringsAsFactors = F) and is faster
    
    myDF[] <- lapply(myDF, integer_NAs_to_zeros)
    myDF
}

if (!exists('TweetDF')) TweetDF <- readData(g_tweet_file)

if (!exists('ProfileDF')) ProfileDF <- readData(g_profile_file)

MergedDF <- merge(x=TweetDF, y=ProfileDF, by.x="Screen_Name", by.y="screen_name")

MergedDF$impressions_l <- safe_log(MergedDF$impressions)
MergedDF$profile_follower_count_l <- safe_log(MergedDF$profile_follower_count)
MergedDF$tweets_l <- safe_log(MergedDF$tweets)

# this seems to be giving some problems

# MergedDF$hour_category<-recode(MergedDF$hour_of_day,"0:3='12-3am';4:7='4-7am';8:11='8-11am';12:15='12-3pm';16:19='4-7pm';20:23='8-11pm'")

MergedDF <- as_tibble(MergedDF) %>% distinct() # this removes a few hundred observations


LIWC <- read_csv("data/tweet_annotated_liwc.csv")

LIWC <- select(LIWC,Tweet_URL = `Source (P)`, WC:OtherP)

df <- left_join(MergedDF, LIWC, by = "Tweet_URL")

Models

m1 <- lmer(Total_Impressions ~ 
               1 +
               (1 | Screen_Name),
           data = df)

sjstats::icc(m1) # 76.42 % variability at tweeter level

## Linear mixed model
##  Family: gaussian (identity)
## Formula: Total_Impressions ~ 1 + (1 | Screen_Name)
## 
##   ICC (Screen_Name): 0.764196

# Let's add tweeter vars

m2 <- lmer(Total_Impressions ~ 
               1 +
               scale(profile_follower_count) +
               scale(profile_following_count) +
               scale(tweets) +
               scale(retweets) + 
               (1 | Screen_Name),
           data = df)

sjt.lmer(m2, show.icc = TRUE, show.r2 = TRUE, p.kr = FALSE, show.re.var = F)

## Computing p-values via Wald-statistics approximation (treating t as Wald z).

		Total_Impressions
		B	CI	p
Fixed Parts
(Intercept)		7564.41	7144.88 – 7983.95	<.001
scale(profile_follower_count)		31197.69	31097.91 – 31297.46	<.001
scale(profile_following_count)		-975.34	-1066.41 – -884.27	<.001
scale(tweets)		-691.92	-2052.50 – 668.66	.319
scale(retweets)		4908.14	4704.89 – 5111.40	<.001
Random Parts
NScreen_Name		54027
ICCScreen_Name		0.096
Observations		191794
R2 / Ω02		.895 / .895

# Let's add tweet vars

m3 <- lmer(Total_Impressions ~ 
               1 +
               scale(profile_follower_count) +
               scale(profile_following_count) +
               scale(tweets) +
               scale(retweets) + 
               
               scale(n_chars) +
               Type + # consider re-leveling these 
               day_of_week +
               as.factor(hour_of_day) + # need to recode this
               (1 | Screen_Name),
           data = df)

sjPlot::sjt.lmer(m3, show.icc = TRUE, show.r2 = TRUE, p.kr = FALSE, show.re.var = F)

## Computing p-values via Wald-statistics approximation (treating t as Wald z).

		Total_Impressions
		B	CI	p
Fixed Parts
(Intercept)		7356.51	6714.17 – 7998.85	<.001
scale(profile_follower_count)		31116.95	31017.93 – 31215.98	<.001
scale(profile_following_count)		-1069.88	-1160.28 – -979.48	<.001
scale(tweets)		-2702.43	-4058.58 – -1346.27	<.001
scale(retweets)		4497.96	4295.33 – 4700.60	<.001
scale(n_chars)		698.57	638.53 – 758.61	<.001
Type (Regular)		3330.99	3023.66 – 3638.32	<.001
Type (reply)		-6498.92	-7206.24 – -5791.60	<.001
Type (retweet)		-1614.46	-1902.44 – -1326.47	<.001
day_of_week (Monday)		332.62	133.80 – 531.44	.001
day_of_week (Saturday)		48.05	-149.64 – 245.73	.634
day_of_week (Sunday)		381.77	182.38 – 581.16	<.001
day_of_week (Thursday)		-115.22	-301.20 – 70.76	.225
day_of_week (Tuesday)		-124.53	-314.16 – 65.10	.198
day_of_week (Wednesday)		-20.82	-200.81 – 159.16	.821
as.factor(hour_of_day) (1)		-251.43	-785.88 – 283.02	.356
as.factor(hour_of_day) (2)		-43.69	-560.43 – 473.06	.868
as.factor(hour_of_day) (3)		-103.80	-585.94 – 378.34	.673
as.factor(hour_of_day) (4)		20.33	-434.65 – 475.31	.930
as.factor(hour_of_day) (5)		236.32	-210.17 – 682.81	.300
as.factor(hour_of_day) (6)		-87.78	-529.70 – 354.14	.697
as.factor(hour_of_day) (7)		-73.34	-518.06 – 371.39	.747
as.factor(hour_of_day) (8)		-84.52	-529.88 – 360.83	.710
as.factor(hour_of_day) (9)		-324.28	-768.95 – 120.39	.153
as.factor(hour_of_day) (10)		-272.61	-721.28 – 176.06	.234
as.factor(hour_of_day) (11)		-487.27	-940.98 – -33.56	.035
as.factor(hour_of_day) (12)		-319.42	-773.15 – 134.30	.168
as.factor(hour_of_day) (13)		86.89	-367.69 – 541.47	.708
as.factor(hour_of_day) (14)		-42.53	-499.44 – 414.37	.855
as.factor(hour_of_day) (15)		45.58	-410.06 – 501.23	.845
as.factor(hour_of_day) (16)		65.01	-384.93 – 514.94	.777
as.factor(hour_of_day) (17)		-38.47	-491.90 – 414.96	.868
as.factor(hour_of_day) (18)		34.90	-417.59 – 487.40	.880
as.factor(hour_of_day) (19)		238.65	-218.14 – 695.43	.306
as.factor(hour_of_day) (20)		282.73	-185.93 – 751.39	.237
as.factor(hour_of_day) (21)		192.13	-305.12 – 689.38	.449
as.factor(hour_of_day) (22)		-12.39	-535.05 – 510.27	.963
as.factor(hour_of_day) (23)		148.56	-379.35 – 676.46	.581
Random Parts
NScreen_Name		54027
ICCScreen_Name		0.097
Observations		191794
R2 / Ω02		.897 / .897

# adding a few LIWC variables

m4 <- lmer(Total_Impressions ~ 
               1 +
               scale(profile_follower_count) +
               scale(profile_following_count) +
               scale(tweets) +
               scale(retweets) + 
               
               scale(n_chars) +
               Type + # consider re-leveling these 
               day_of_week +
               as.factor(hour_of_day) + # need to recode this
               
               scale(cogproc) +
               scale(social) + 
               scale(affiliation) + 
               scale(posemo) + 
               (1 | Screen_Name),
           data = df)

sjPlot::sjt.lmer(m4, show.icc = TRUE, show.r2 = TRUE, p.kr = FALSE, show.re.var = F)

## Computing p-values via Wald-statistics approximation (treating t as Wald z).

		Total_Impressions
		B	CI	p
Fixed Parts
(Intercept)		7380.10	6737.71 – 8022.49	<.001
scale(profile_follower_count)		31115.74	31016.75 – 31214.74	<.001
scale(profile_following_count)		-1067.86	-1158.25 – -977.48	<.001
scale(tweets)		-2689.40	-4044.92 – -1333.89	<.001
scale(retweets)		4491.12	4288.57 – 4693.68	<.001
scale(n_chars)		688.20	627.67 – 748.72	<.001
Type (Regular)		3390.32	3079.63 – 3701.01	<.001
Type (reply)		-6460.37	-7168.47 – -5752.27	<.001
Type (retweet)		-1692.06	-1981.89 – -1402.24	<.001
day_of_week (Monday)		337.45	138.60 – 536.30	<.001
day_of_week (Saturday)		56.43	-141.26 – 254.12	.576
day_of_week (Sunday)		386.39	186.82 – 585.96	<.001
day_of_week (Thursday)		-114.26	-300.25 – 71.74	.229
day_of_week (Tuesday)		-114.43	-304.08 – 75.23	.237
day_of_week (Wednesday)		-19.98	-199.99 – 160.04	.828
as.factor(hour_of_day) (1)		-231.32	-765.76 – 303.12	.396
as.factor(hour_of_day) (2)		-35.63	-552.35 – 481.08	.892
as.factor(hour_of_day) (3)		-102.45	-584.56 – 379.65	.677
as.factor(hour_of_day) (4)		18.34	-436.62 – 473.31	.937
as.factor(hour_of_day) (5)		234.73	-211.72 – 681.19	.303
as.factor(hour_of_day) (6)		-90.47	-532.37 – 351.42	.688
as.factor(hour_of_day) (7)		-77.37	-522.07 – 367.32	.733
as.factor(hour_of_day) (8)		-91.13	-536.47 – 354.20	.688
as.factor(hour_of_day) (9)		-326.50	-771.13 – 118.13	.150
as.factor(hour_of_day) (10)		-278.45	-727.08 – 170.19	.224
as.factor(hour_of_day) (11)		-492.70	-946.37 – -39.03	.033
as.factor(hour_of_day) (12)		-322.16	-775.85 – 131.53	.164
as.factor(hour_of_day) (13)		88.73	-365.81 – 543.27	.702
as.factor(hour_of_day) (14)		-45.99	-502.87 – 410.90	.844
as.factor(hour_of_day) (15)		37.62	-418.00 – 493.23	.871
as.factor(hour_of_day) (16)		55.75	-394.16 – 505.67	.808
as.factor(hour_of_day) (17)		-44.55	-497.98 – 408.89	.847
as.factor(hour_of_day) (18)		29.98	-422.50 – 482.46	.897
as.factor(hour_of_day) (19)		229.03	-227.76 – 685.81	.326
as.factor(hour_of_day) (20)		272.44	-196.22 – 741.09	.255
as.factor(hour_of_day) (21)		185.45	-311.77 – 682.67	.465
as.factor(hour_of_day) (22)		-9.29	-531.91 – 513.33	.972
as.factor(hour_of_day) (23)		155.94	-371.94 – 683.83	.563
scale(cogproc)		44.09	-11.20 – 99.39	.118
scale(social)		94.30	19.87 – 168.73	.013
scale(affiliation)		37.85	-24.96 – 100.66	.238
scale(posemo)		-120.74	-173.27 – -68.22	<.001
Random Parts
NScreen_Name		54027
ICCScreen_Name		0.097
Observations		191794
R2 / Ω02		.897 / .897