Getting the most out of interactive and developmental data Daniel Messinger 5.2001

Basics Know your phenomena Know your data Watch the tapes Know your data Look at descriptives and frequencies Look for patterns that you can see in individuals or dyads Not just in the group as a whole

Topics Types of codes and coding Frequency, duration, and combined approaches Agreement and reliability Measures of association between two behaviors Development

Coding Types of codes Type of coding Agreement/Reliability Objective vs. socially recognizable Type of coding Frequency, duration, combined approaches Agreement/Reliability Percent agreement, Cohen’s Kappa, intra-class correlation

Types of codes Objectively describable codes E.g., Facial Action Coding System Socially recognizable codes E.g., Emotionally positive moments Measurement and analysis more geared to objectively describable but there is no qualitative difference between different types of codes

Type of coding Duration Frequency Combined approaches Agreement/Reliability Agreeing on what we saw

Agreeing on what we saw Agreement Reliability Whether same thing was observed at the same time The focus of percent agreement and Kappa Reliability Did we see the same number of things in a given interactive period Summary measure of codes over session It all depends on the research question

Duration How long does behavior last Code onsets and offsets E.g., Total time smiling Code onsets and offsets Smile begins, smile ends Or exclusive categories Smile, neutral, smile

Duration

Agreement on duration Agreement = Total duration of time both coders indicated same event was occurring Disagreement = Total duration of time coders indicated different events were occurring Its ok to collapse codes

Duration agreement statistics Proportion agreement Agreement / (Agreement + Disagreement) Observed agreement Cohen’s Kappa = (Observed agreement – Expected agreement) / (1 – Expected agreement) Expected agreement = Sum of product of marginals Bakeman & Gottman

Frequency How often does behavior occur Code onsets only Always express as onsets per unit of time E.g., number of smiles per minute You can calculate frequencies from duration codes Just count the onsets But onset and offset coding is more difficult E.g., vocalizations

Frequency

Frequency agreement Percent agreement Number of times the same code is recorded by independent coders within a given time window (e.g., 2 seconds) Divided by one half the total number of codes No Kappa for frequency There is no expected measure of agreement

Frequency and Duration: Pros and cons Duration is a relatively stable measure of what’s going on But not how they occurred E..g., Total gazing at mother Frequency tells you about discrete activities But not how long they lasted E.g., number of speech acts Mixed approaches

Mixed approaches Duration of smiling in a given period of interaction initiated by infant Duration in that it’s a total time measure But frequency in that its onset of infant action Calculate both types of agreement

Reliability (intra-class correlation) Summary measure of codes over session Variance attributable to differences between subjects expressed as a proportion of all variance Including variance between coders – want to minimize Type of ANOVA Better than a simple correlation

Measures of association between two behaviors Group level analysis Individual level analysis Duration and frequency approaches

Example Infant gaze direction Mother smile Coded continuously in time At mother’s face or away Mother smile Yes or no Coded continuously in time

SPSS (10.0) VALUE LABELS m12 1.00000000000000 "Mother Not Smiling" 2.00000000000000 "Mother Smiling"

General duration variable create /leadweeK=lead(week 1) /LEADSECs=LEAD(SECS 1). EXECUTE. IF (week=leadweek) Duration=LEADSECS-SECS. IF SYSMIS(DUR) DUR=0.

Analysis of entire group Weight by duration, then . . .

Analysis of entire group Doesn’t tell you if any given infant/dyad shows the association Or the strength of association for a given infant/dyad Use when necessary E.g. small amounts of data for individual infants/dyads

Analysis of individuals Determine frequency and duration of variables for whole group Then aggregate By subject/participant for general analysis By time period for developmental analyses Construct variables Analysis

Specific variable duration How much time do infants spend gazing at mother And away from mother How much time do mothers spend smiling And not smiling

Specific variable duration IF (infgaze=1) infgazeM=Duration. IF (infgaze=2) infgazeA=Duration. IF (momsmile=1) M1=Duration. IF (momsmile=2) M2=Duration. Execute.

Duration of two co-occurring behaviors How long do infants gaze at mother when she is smiling?

Duration of two co-occurring behaviors IF (infgaze=1 & momsmile=1) GMM1=Duration. IF (infgaze=1 & momsmile=2) GMM2=Duration. IF (infgaze=2 & momsmile=1) GAM1=Duration. IF (infgaze=2 & momsmile=2) GAM2=Duration. EXECUTE.

Combined duration and frequency approach Duration of two co-occurring behaviors given that one has just occurred How long do infants gaze at smiling mother having just gazed at her Not when they were gazing and she smiled Attention: The following technique assumes a dataset created so that only the variables of interest (in this case, two variables of interest) exist and that cases (rows) exist only when one of these variables changes (or there is a new session).

Combined duration and frequency approach CREATE /momsml_1=LAG(momsml 1) /infgaz_1=LAG(infgaz 1). IF (infgaz=2 & infgaz_1~=2 & momsml=2) GAM2F=duration. IF (infgaz=1 & infgaz_1~=1 & momsml=2) GMM2F=duration. IF (infgaz=2 & infgaz_1~=2 & momsml=1) GAM1F=duration. IF (infgaz=1 & infgaz_1~=1 & momsml=1) GMM1F=duration. EXECUTE.

Aggregate: Summarizing data for analysis Aggregate over subject for overall effects Aggregate over subject and time period for developmental analyses

Aggregate creates new variables

Summarizes over time-linked cases Summary measures Number of cases for frequency Sum of values for duration

Ouch! AGGREGATE /OUTFILE=* /BREAK=newsub / infgazm2= N(infgazem) /infgaza2 = N(infgazea) /gmm1_3 = SUM(gmm1) /gmm2_3 = SUM(gmm2) /gam1_3 = SUM(gam1) /gam2_3 = SUM(gam2) /gam2f_2 = N(gam2f) /gmm2f_2 = N(gmm2f) /gam1f_2 = N(gam1f) /gmm1f_2 = N(gmm1f) /gam2f_3 = SUM(gam2f) /gmm2f_3 = SUM(gmm2f) /gam1f_3 = SUM(gam1f) /gmm1f_3 = SUM(gmm1f).

Voile’

Constructing variables New duration and frequency dependent measures are calculated per subject in new file Same dependent measures will be calculated per time period within subject (in a different file) for developmental analyses

Creating durational proportions COMPUTE GMM2P=Gmm2_3/M2_3. Note: M2_3 (total time mother is smiling) can be created during aggregation or computed as = GMM2_3+GAM2_3. Number of seconds of gazing at mother while mother is smiling divided by total time gazing at mother Do the same for gazes at mother while mother is not smiling These variables are calculated for each subject in new aggregated file

Analyses, finally

Results Look at results subject by subject by graphing

Duration and frequency Can tell you the same thing about an interaction Or different things

Creating frequency per minute Example: COMPUTE GAM2PM= (GAM2F_2/M2_3)*60. This is calculated for each subject in new aggregated file Number of gazes away while mother is smiling divided by total time mother is smiling per minute Do the same for gazes away while mother is not smiling

Duration and frequency together Infant gazes at mother, mother smiles, infant then gazes away Combined (frequency and duration) approach might have shed light on this directly

Development – Conceptual How do individuals change over time? Unit of analysis is individuals (or individual dyads) - singoli Can developmental effects be seen in each individual’s data? Or in a significant proportion of each individual’s data? Keep it simple General trends preferred over particular periods Linear versus curvilinear effects

Development - Practical Choices for developmental analyses Hierarchical linear modeling Individual growth (Linear and curvilinear models) T-tests, binomial tests, Graph individual and group data

Development - How to Go back to the original file Create a case for every week Or other age category E.g.: AGGREGATE /OUTFILE=* /BREAK=newsub WEEK / infgazm2= N(infgazem) /infgaza2 = N(infgazea) / etc.

Voile’ - Development Each case now summarizes the variables of interest for a given age period for a given subject Create summary proportional duration and frequency per minute variables as before

Individual growth modeling Conduct developmental analyses within individuals Regression analyses (within individuals) Typically linear effects

Analysis by individual

Analysis techniques Is mean slope different from 0 T-test Data can be copied from regression output to make a new developmental data file

Graphing and binomial tests Do a significant proportion of subjects show an increase (or decrease) with age? Binomial test (Hays) Consistency within variability

Consistency across individuals

Hierarchical linear modeling Age is nested within individual Calculates “mean” slope for group T-test Estimates variability around that mean Chi-square Assumes normal distribution of slope parameters Should have many subjects Level 2 units

Specialized programs – WHLM Ask for OLE output for each subject 1 0.54269 -0.02004 2 0.41946 -0.01507 3 0.50880 -0.02033 4 0.45630 -0.01529 5 0.99627 -0.04153 6 0.67704 -0.02057 7 0.34434 -0.00026 8 0.73655 -0.02059 9 0.38579 -0.00123 10 0.51995 -0.01611 Need to get your data into the program. Can be tricky

Table: Gaze away

Review Types of codes and coding Frequency, duration, and combined approaches Agreement and reliability Measures of association between two behaviors Complete picture may involve both duration and frequency (or combined approaches) Development Occurs and should be studied as it occurs in individuals (or individual dyads)

Duchenne Smiles (M = 20.95%) Non-Duchenne Smiles (M = 39.29%) 59.8% 46.3% 53.7% 40.3% 22.7% 30.2% 27.5% 20.5% Non-Duchenne Smiles (M = 39.29%) 72.5% 69.8% Non-Smiles (M = 39.76%) 77.3% 79.5%

Transitional probabilities Likelihood of one behavior following another Typically within a modality of actiojn One code following another within a category