# Analysis script for "Effects of face masks on speech recognition" 
# J.C. Toscano and C.M. Toscano, Villanova University

library(lme4)
library(plyr)

##################################################################################################
# Set up data frame for analysis

# Load data
data <- as.data.frame(read.delim('data.csv', sep=',', header=T))

# Omit practice trials
data2 <- subset(data, hintlist!='NA')

# Omit trials with RT > 60 seconds
data2 <- subset(data2, rt<60)

# Convert subject and sentence columns to factors
data2$subject <- as.factor(data2$subject)
data2$hintlist <- as.factor(data2$hintlist)
data2$hintsentence <- as.factor(data2$hintsentence)

# Create HINT 'list_sentence' column
data2$hint <- as.factor(paste0(data2$hintlist, '_', data2$hintsentence))

# Create SNR and talker variables
data2$snrNum <- ifelse(data2$snr==3,-0.5,0.5)
data2$talkerNum <- ifelse(data2$talker=='jt',-0.5,0.5)

# Create effect codes for mask type
# Key:
#   dispose = disposable surgical mask
#   fabric1 = fitted cloth mask
#   fabric2 = pleated cloth mask
#   n95mask = N95 respirator
data2$maskDispose <- ifelse(data2$mask=='dispose',0.8,-0.2)
data2$maskFabric1 <- ifelse(data2$mask=='fabric1',0.8,-0.2)
data2$maskFabric2 <- ifelse(data2$mask=='fabric2',0.8,-0.2)
data2$maskN95Mask <- ifelse(data2$mask=='n95mask',0.8,-0.2)

# Create proportion correct column
data2$propcorrect <- data2$wordscorrect / data2$numwords

# Set glmer options
glmeropt = glmerControl(optimizer='bobyqa', optCtrl = list(maxfun = 200000))

##################################################################################################
# Descriptive Statistics

# Total number of trials
nrow(data2)
# [1] 3564

# Total number of subjects
length(unique(data2$subject))
# [1] 181

# SNR
ddply(data2, .(snr), summarize, mean=mean(propcorrect), sd=sd(propcorrect))
  # snr      mean        sd
# 1   3 0.3688941 0.3673171
# 2  13 0.9229677 0.1988491

# Talker
ddply(data2, .(talker), summarize, mean=mean(propcorrect), sd=sd(propcorrect))
  # talker      mean        sd
# 1     ct 0.6834107 0.3914347
# 2     jt 0.6126850 0.4142413

# Talker x SNR
ddply(data2, .(talker,snr), summarize, mean=mean(propcorrect), sd=sd(propcorrect))
  # talker snr      mean        sd
# 1     ct   3 0.4138122 0.3733912
# 2     ct  13 0.9485159 0.1582753
# 3     jt   3 0.3238742 0.3556851
# 4     jt  13 0.8973055 0.2298350

# Mask type 
ddply(data2, .(mask,snr), summarize, mean=mean(propcorrect), sd=sd(propcorrect))
      # mask snr      mean        sd
# 1  dispose   3 0.4238229 0.3751347
# 2  dispose  13 0.9350444 0.1839943
# 3  fabric1   3 0.2698052 0.3014067
# 4  fabric1  13 0.8878891 0.2323401
# 5  fabric2   3 0.3513878 0.3543970
# 6  fabric2  13 0.9181812 0.2045042
# 7  n95mask   3 0.3460698 0.3636381
# 8  n95mask  13 0.9309524 0.1838518
# 9  no_mask   3 0.4522263 0.4070690
# 10 no_mask  13 0.9426648 0.1813404


##################################################################################################
# Mixed-effects Models

model <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ 
	maskDispose + maskFabric1 + 	maskFabric2 + maskN95Mask + 
	snrNum + talkerNum + snrNum:talkerNum +
	maskDispose:snrNum + maskFabric1:snrNum + maskFabric2:snrNum + maskN95Mask:snrNum + 
	maskDispose:talkerNum + maskFabric1:talkerNum + maskFabric2:talkerNum + maskN95Mask:talkerNum + 
	maskDispose:snrNum:talkerNum + maskFabric1:snrNum:talkerNum + maskFabric2:snrNum:talkerNum + maskN95Mask:snrNum:talkerNum + 
	(1|subject) + 
	(1|hint) +
	(0+snrNum|subject) + 
	(0+snrNum|hint) +
	(0+talkerNum|subject) + 
	(0+talkerNum|hint) +
	(0+snrNum:talkerNum|subject) + 
	(0+snrNum:talkerNum|hint) +
	(0+maskDispose|subject) + (0+maskFabric1|subject) + (0+maskFabric2|subject) + (0+maskN95Mask|subject) +
	(0+maskDispose:snrNum|subject) + (0+maskFabric1:snrNum|subject) + (0+maskFabric2:snrNum|subject) + (0+maskN95Mask:snrNum|subject) +
	(0+maskDispose:talkerNum|subject) + (0+maskFabric1:talkerNum|subject) + (0+maskFabric2:talkerNum|subject) + (0+maskN95Mask:talkerNum|subject) +
	(0+maskDispose|hint) + (0+maskFabric1|hint) + (0+maskFabric2|hint) + (0+maskN95Mask|hint) +
	(0+maskDispose:snrNum|hint) + (0+maskFabric1:snrNum|hint) + (0+maskFabric2:snrNum|hint) + (0+maskN95Mask:snrNum|hint) +
	(0+maskDispose:talkerNum|hint) + (0+maskFabric1:talkerNum|hint) + (0+maskFabric2:talkerNum|hint) + (0+maskN95Mask:talkerNum|hint), 
	data=data2, family=binomial, control=glmeropt)

summary(model)

                              # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)                   1.728076   0.179320   9.637  < 2e-16 ***
# maskDispose                   0.002214   0.123851   0.018  0.98573    
# maskFabric1                  -1.070332   0.146447  -7.309 2.70e-13 ***		Fitted
# maskFabric2                  -0.542202   0.176781  -3.067  0.00216 ** 		Pleated
# maskN95Mask                  -0.380902   0.167531  -2.274  0.02299 *  		N95
# snrNum                        5.479648   0.212302  25.811  < 2e-16 ***		SNR
# talkerNum                     1.008158   0.196543   5.129 2.91e-07 ***		Talker
# snrNum:talkerNum              0.505536   0.252203   2.004  0.04502 *  		SNR x Talker
# maskDispose:snrNum            0.285615   0.319579   0.894  0.37147    
# maskFabric1:snrNum            0.597503   0.302610   1.974  0.04833 *  		Fitted x SNR
# maskFabric2:snrNum            0.572854   0.275046   2.083  0.03727 *  		Pleated x SNR
# maskN95Mask:snrNum            0.829566   0.356007   2.330  0.01980 *  		N95 x SNR
# maskDispose:talkerNum         0.221199   0.356517   0.620  0.53496    
# maskFabric1:talkerNum        -0.424661   0.335666  -1.265  0.20582    
# maskFabric2:talkerNum         0.793616   0.326684   2.429  0.01513 *  		Pleated x Talker
# maskN95Mask:talkerNum        -0.184253   0.253929  -0.726  0.46808    
# maskDispose:snrNum:talkerNum  0.029926   0.456605   0.066  0.94774    
# maskFabric1:snrNum:talkerNum  0.377307   0.426782   0.884  0.37666    
# maskFabric2:snrNum:talkerNum  0.164940   0.449775   0.367  0.71383    
# maskN95Mask:snrNum:talkerNum  1.262623   0.459155   2.750  0.00596 ** 		N95 x SNR x Talker


# SNR x talker follow-up

data3 <- subset(data2, snrNum < 0)
modelA <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ talkerNum +
	(1|subject) + (1|hint) + 
	(0+talkerNum|subject) + (0+talkerNum|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelA)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)  -0.8293     0.1703  -4.869 1.12e-06 ***
# talkerNum     0.5961     0.1703   3.501 0.000464 ***		Talker (low SNR)

data3 <- subset(data2, snrNum > 0)
modelB <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ talkerNum +
	(1|subject) + (1|hint) + 
	(0+talkerNum|subject) + (0+talkerNum|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelB)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)   3.8677     0.1966  19.669  < 2e-16 ***
# talkerNum     1.0736     0.2534   4.236 2.27e-05 ***		Talker (high SNR)


# Pleated mask x talker follow-up

data3 <- subset(data2, (mask=='fabric2' | mask=='no_mask') & talkerNum < 0)
data3$maskFabric2 <- ifelse(data3$mask=='fabric2', 0.5, -0.5)
modelA <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskFabric2 +
	(1|subject) + (1|hint) + 
	(0+maskFabric2|subject) + (0+maskFabric2|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelA)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)   0.6015     0.1107   5.436 5.45e-08 ***
# maskFabric2  -0.4905     0.1491  -3.291 0.000999 ***		Pleated (male talker)

data3 <- subset(data2, (mask=='fabric2' | mask=='no_mask') & talkerNum > 0)
data3$maskFabric2 <- ifelse(data3$mask=='fabric2', 0.5, -0.5)
modelB <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskFabric2 +
	(1|subject) + (1|hint) + 
	(0+maskFabric2|subject) + (0+maskFabric2|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelB)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)  1.25877    0.13621   9.242   <2e-16 ***
# maskFabric2 -0.20541    0.09985  -2.057   0.0397 *  		Pleated (female talker)


# Fitted mask x SNR follow-up

data3 <- subset(data2, (mask=='fabric1' | mask=='no_mask') & snrNum < 0)
data3$maskFabric1 <- ifelse(data3$mask=='fabric1', 0.5, -0.5)
modelA <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskFabric1 +
	(1|subject) + (1|hint) + 
	(0+maskFabric1|subject) + (0+maskFabric1|hint), 
	data=data3, family=binomial, control=glmeropt)
summary(modelA)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)  -0.9468     0.2166  -4.371 1.24e-05 ***
# maskFabric1  -1.1816     0.2276  -5.192 2.08e-07 ***		Fitted (low SNR)

data3 <- subset(data2, (mask=='fabric1' | mask=='no_mask') & snrNum > 0)
data3$maskFabric1 <- ifelse(data3$mask=='fabric1', 0.5, -0.5)
modelB <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskFabric1 + 
	(1|subject) + (1|hint) + 
	(0+maskFabric1|subject) + (0+maskFabric1|hint), 
	data=data3, family=binomial, control=glmeropt)
summary(modelB)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)   4.4351     0.3300  13.438  < 2e-16 ***
# maskFabric1  -1.2474     0.3502  -3.562 0.000369 ***		Fitted (high SNR)


# Pleated mask x SNR follow-up

data3 <- subset(data2, (mask=='fabric2' | mask=='no_mask') & snrNum < 0)
data3$maskFabric2 <- ifelse(data3$mask=='fabric2', 0.5, -0.5)
modelA <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskFabric2 + 
	(1|subject) + (1|hint) + 
	(0+maskFabric2|subject) + (0+maskFabric2|hint), 
	data=data3, family=binomial, control=glmeropt)
summary(modelA)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)  -0.7091     0.2122  -3.342 0.000832 ***
# maskFabric2  -0.6683     0.2070  -3.228 0.001248 ** 		Pleated (low SNR)

data3 <- subset(data2, (mask=='fabric2' | mask=='no_mask') & snrNum > 0)
data3$maskFabric2 <- ifelse(data3$mask=='fabric2', 0.5, -0.5)
modelB <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskFabric2 + 
	(1|subject) + (1|hint) + 
	(0+maskFabric2|subject) + (0+maskFabric2|hint), 
	data=data3, family=binomial, control=glmeropt)
summary(modelB)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)   4.8290     0.3688  13.094   <2e-16 ***
# maskFabric2  -0.7763     0.4147  -1.872   0.0612 .  		Pleated (high SNR, marginal)


# N95 x SNR x talker follow-up

data3 <- subset(data2, (mask=='n95mask' | mask=='no_mask') & snrNum < 0 & talkerNum < 0)
data3$maskN95Mask <- ifelse(data3$mask=='n95mask', 0.5, -0.5)
modelA <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskN95Mask +
	(1|subject) + (1|hint) + 
	(0+maskN95Mask|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelA)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)  -1.0837     0.3160  -3.430 0.000604 ***
# maskN95Mask  -0.4408     0.2170  -2.032 0.042186 *  		N95 (low SNR, male talker)

data3 <- subset(data2, (mask=='n95mask' | mask=='no_mask') & snrNum < 0 & talkerNum > 0)
data3$maskN95Mask <- ifelse(data3$mask=='n95mask', 0.5, -0.5)
modelB <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskN95Mask +
	(1|subject) + (1|hint) + 
	(0+maskN95Mask|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelB)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)  -0.5169     0.2550  -2.027   0.0426 *  
# maskN95Mask  -1.2121     0.2086  -5.812 6.18e-09 ***		N95 (low SNR, female talker)

data3 <- subset(data2, (mask=='n95mask' | mask=='no_mask') & snrNum > 0 & talkerNum < 0)
data3$maskN95Mask <- ifelse(data3$mask=='n95mask', 0.5, -0.5)
modelC <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskN95Mask +
	(1|subject) + (1|hint) + 
	(0+maskN95Mask|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelC)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)   4.3892     0.4161  10.549   <2e-16 ***
# maskN95Mask  -0.4306     0.4795  -0.898    0.369    

data3 <- subset(data2, (mask=='n95mask' | mask=='no_mask') & snrNum > 0 & talkerNum > 0)
data3$maskN95Mask <- ifelse(data3$mask=='n95mask', 0.5, -0.5)
modelD <- glmer(cbind(wordscorrect, numwords-wordscorrect) ~ maskN95Mask +
	(1|subject) + (1|hint) + 
	(0+maskN95Mask|hint),
	data=data3, family=binomial, control=glmeropt)
summary(modelD)
            # Estimate Std. Error z value Pr(>|z|)    
# (Intercept)   8.4950     0.9670   8.785   <2e-16 ***
# maskN95Mask  -0.3394     0.6695  -0.507    0.612