ConversationAlign_Step3_Summarize

This is the final step where ConversationAlign will compute summary statistics including main effects and alignment statistics for the vectorized dataframe you produced using prep_dyads(). Users have several options for how to output their data, and these choices should be guided by your analysis strategy. For example, a linear mixed effects approach might involve modeling the rise and fall of values across turns. In contrast, a standard ANOVA would work on grouped summary data.

Main effects for each dimension of interest aggregated by Conversation (Event_ID) and Person (Participant_ID). For example, main effects of concreteness would involve aggregated means of all the concreteness values for words produced by Mary vs words produced by Dave in an individual conversation.
dAUC_raw: difference area under the curve reflecting the difference between interlocutors at each turn on each dimension uncorrected for conversation length. For example, if Mary’s concreteness at Exchange 1 was 8 (on a scale of 0-9), and Dave’s concreteness on Exchange 1 was 4, the difference between Mary and Dave in this one-turn conversation would be 4. dAUC reflects area across all turns estimated using the trapezoidal rule.
dAUC_scaled100: normalized AUC value to 100 turns using proportional scaling. e.g., (Observed AUC/Turns Raw) = (Normalized AUC)/100).
Lead_Corr2: Pearson or Spearman lagged correlation reflecting turn-by-turn covariance across partners for each specified dimension
Lag_Corr2: Lead correlation
Who_Talked_First: Interlocutor who started the conversation (needed for interepreting lead/lag stats)

Arguments to summarize_dyads() include:
1) df_prep= dataframe created by prep_dyads()function
2) custom_lags= default is NULL, any additional user-specified lagged correlations. will automatically produce lead of 2 turns, immediate response, lag of 2 turns for each dimension of interest.
3) sumdat_only= boolean default is TRUE, produces grouped summary dataframe with averages by conversation and participant for each alignment dimension, FALSE retrains all of the original rows, filling down empty rows of summary statistics for the conversation (e.g., AUC)
4) corr_type= default=‘Pearson’, other option ‘Spearman’ for computing turn-by-turn correlations across interlocutors for each dimension of interest.

MarySumDat <- summarize_dyads(df_prep = NurseryRhymes_Prepped, custom_lags=NULL, sumdat_only = TRUE, corr_type='Pearson') 
colnames(MarySumDat)
#>  [1] "Event_ID"                "Participant_ID"         
#>  [3] "Talked_First"            "Dimension"              
#>  [5] "Dimension_Mean"          "AUC_raw_Immediate"      
#>  [7] "AUC_scaled100_Immediate" "AUC_raw_Lag1"           
#>  [9] "AUC_scaled100_Lag1"      "TurnCorr_Lead2"         
#> [11] "TurnCorr_Immediate"      "TurnCorr_Lag2"
knitr::kable(head(MarySumDat, 15), format = "simple", digits = 3)

Event_ID	Participant_ID	Talked_First	Dimension	Dimension_Mean	AUC_raw_Immediate	AUC_scaled100_Immediate	AUC_raw_Lag1	AUC_scaled100_Lag1	TurnCorr_Lead2	TurnCorr_Immediate	TurnCorr_Lag2
ItsySpider	Maya	Yin	emo_anger	0.001	0.783	1.630	0.749	1.560	-1	-1	-1
ItsySpider	Yin	Yin	emo_anger	-0.033	0.783	1.630	0.749	1.560	-1	-1	-1
JackJill	Ana	Franklin	emo_anger	-0.066	3.729	4.662	3.634	4.542	1	1	1
JackJill	Franklin	Franklin	emo_anger	0.030	3.729	4.662	3.634	4.542	1	1	1
LittleLamb	Dave	Mary	emo_anger	-0.001	1.486	1.486	1.456	1.456	NA	NA	NA
LittleLamb	Mary	Mary	emo_anger	-0.031	1.486	1.486	1.456	1.456	NA	NA	NA

Generating sham conversations

Some research questions would benefit from the use of conversations that control for some temporal effects. The function generate_shams() accepts the output of prep_dyads() and returns a data frame in the same structure with each interlocutor’s time series randomly shuffled. Since the output has the same format as prep_dyads() output, it can easily be supplied to summarize_dyads() and compared to the real conversations.

Arguments to generate_shams() include:
1) df_prep= dataframe created by prep_dyads()function
2) seed= numeric, a number to supply as a seed. This allows for reproducible results.

MaryShams <- generate_shams(df_prep = NurseryRhymes_Prepped, seed = 10)
MarySumDatShams <- summarize_dyads(df_prep = MaryShams, custom_lags=NULL, sumdat_only = TRUE, corr_type='Pearson') 
knitr::kable(head(MarySumDatShams, 15), format = "simple", digits = 3)

Event_ID	Participant_ID	Talked_First	Dimension	Dimension_Mean	AUC_raw_Immediate	AUC_scaled100_Immediate	AUC_raw_Lag1	AUC_scaled100_Lag1	TurnCorr_Lead2	TurnCorr_Immediate	TurnCorr_Lag2
ItsySpider	Maya	Maya	emo_anger	0.002	0.759	1.580	0.701	1.459	0.100	-1	0.100
ItsySpider	Yin	Maya	emo_anger	-0.032	0.759	1.580	0.701	1.459	0.100	-1	0.100
JackJill	Ana	Ana	emo_anger	-0.066	3.729	4.662	3.634	4.542	-0.158	1	-0.158
JackJill	Franklin	Ana	emo_anger	0.030	3.729	4.662	3.634	4.542	-0.158	1	-0.158
LittleLamb	Dave	Dave	emo_anger	-0.001	1.486	1.486	1.456	1.456	NA	NA	NA
LittleLamb	Mary	Dave	emo_anger	-0.031	1.486	1.486	1.456	1.456	NA	NA	NA

summarize_dyads()

Jamie Reilly, Ben Sacks, Ginny Ulichney, Gus Cooney, Chelsea Helion

October 14, 2025

Generating sham conversations