Associations between psychometrically assessed life history strategy and daily behavior: data from the Electronically Activated Recorder (EAR)

The present study

As a preliminary inquiry into the ethology of LHS variation, I address a simple question: how well does a widely-used psychometric LHS instrument, the Arizona Life History Battery (ALHB: Figueredo, 2007) predict the everyday behavior of a sample of US college students? The ALHB, or its short form the Mini-K, has been used in many studies (Figueredo et al., 2015), but almost all of this work describes associations with other self-report instruments. A possible weakness of the ALHB/Mini-K is its undercoverage of aspects of extraversion (dominance striving, sensation-seeking) and openness to experience (imagination, unconventionality) that are theoretically expected (Del Giudice, 2014) to be associated with a faster LHS. In earlier analyses of the data set used in this paper (Manson, 2017a), self-reported ALHB scores were found to be only weakly (non-significantly) associated with other-reported LHS as indicated by similarity to the slow life history template (Dunkel et al., 2015; Sherman, Figueredo & Funder, 2013) of the California Adult Q-Sort (CAQ: Block, 1978). Analysis of this template indicates that it taps the fast LHS-associated facets of extraversion and openness (Manson, 2017a). Furthermore, the slow LHS template of the CAQ has been validated by correlating it with biodemographic LH indicators such as age at sexual debut and number of sex partners (Dunkel et al., 2015; see also Kubinski, Chopik & Grimm, 2017).

In the present study, participants’ everyday behavior was sampled using the Electronically Activated Recorder (EAR: Mehl et al., 2001; Mehl & Robbins, 2012). The EAR unobtrusively collects periodic brief audio snippets of people’s daily lives, using a portable recording device that participants wear attached to their clothes during their waking hours. Behavioral variables coded from EAR recordings (e.g., location and general activity) have shown substantial inter-rater reliability and within-participant temporal stability (Mehl & Pennebaker, 2003), and have provided novel insights into processes such as parent–child transmission of negative emotionality (Slatcher & Trentacosta, 2012) and the role of family environment in children’s recovery from trauma (Alisic et al., 2015). Self-reported major personality dimensions are correlated in predictable ways with some EAR-measured variables (e.g., extraversion negatively correlated with proportion of time spent alone; Mehl, Gosling & Pennebaker, 2006).

Table 1 lists the behaviors measured in the present study, and their predicted associations with slow life history strategy as measured by the ALHB, based on findings from psychometric studies (Brumbach, Figueredo & Ellis, 2009; Figueredo et al., 2007; Figueredo et al., 2005; Figueredo et al., 2014a; Sherman, Figueredo & Funder, 2013). College students pursuing a slower LHS were predicted to engage more frequently in behavior indicative of (1) future time perspective, the ability to learn from past experience, and investment in embodied capital (Kaplan et al., 2000) (class attendance, talking about long-term plans and talking about past experiences), (2) altruistic dispositions toward kin and the wider community (talking to kin; talking about kin; volunteering); and (3) religiosity (attending religious services and study groups). A slower LHS was predicted to be associated with lower rates of (1) entertainment consumption (TV and movie watching; videogame playing; being at amusement venues such as bars, concerts, and sports events), (2) talk endorsing alcohol or recreational drug use, (3) indicators of negative affect (general complaining; sighing; see Robbins et al., 2011; Roth, 2005), and (4) indicators of antagonistic social schemata (Patch & Figueredo, 2017) or an alienated social orientation (arguing; complaining about authority figures). A final prediction, that a slower LHS would be negatively related to amount of daytime sleeping, requires additional explanation. Research suggests that LHS is related to chronotype (morningness/eveningness, Adan et al., 2012), such that people pursuing a slower LHS are more active in the morning, whereas those pursuing a faster LHS are more active in the evening (Ponzi et al., 2015). Because the audio recording protocol in the present study included an 00:00–06:00 blackout period (see Materials and Methods), and because participants rarely went to sleep before the blackout period began (see Results), sleeping time mostly consisted of sleeping between 06:00 and 18:00. Sleep time during this time interval (corresponding to late rising and afternoon napping) was isolated for analysis, and I tested the prediction that it would be negatively associated with a slower LHS.

Finally, because the ALHB/Mini-K is strongly positively related to extraversion (Figueredo et al., 2014a; Gladden, Figueredo & Jacobs, 2009; Manson, 2015; Strouts, Brase & Dillon, 2016), including among the present study’s participants (Manson, 2017a), ALHB scores were predicted to be positively associated with two indicators of sociality (engagement in social interaction; proportion of social interactions that involve more than one interlocutor).

Ethics statement

All procedures described here were approved by UCLA’s Institutional Review Board (Approval #12-001128-AM-00001). Informed consent was obtained from all participants in accordance with the terms of this approval.

Relation to previous research

The dataset used in the present study is the same as that used by Manson (2017a), Manson (2017b) and Manson & Robbins (2017), but the analyses reported here are novel.

Participants

Ninety-two students (55.4% female, M ± SD = 20.0 ± 3.1 years old), enrolled at the University of California, Los Angeles, were recruited via posted flyers and classroom announcements to participate in a study bearing the public title “Audio Sampling of Daily Life.” Based on self-reported ethnic identity, 53.2% of the sample was Asian or Asian-American, 16.3% White, 12.0% Latino/a, 3.2% Middle Eastern, and 15.3% mixed or “other.” Although unrepresentative of college-aged Americans generally, the sample’s ethnic composition was fairly representative of UCLA’s undergraduate student body. Data were collected between October 2013 and January 2015.

Procedure

Data preparation and analysis

Modeling relationships between predictor variables and behavior frequencies

To assess relationships between the ALHB and behavior frequencies, I carried out Bayesian multi-level aggregated binomial regressions, using inferential procedures and notation described by McElreath (2015). Previous EAR-based research (e.g., Baddeley, Pennebaker & Beevers, 2013; Mehl, Gosling & Pennebaker, 2006) has scored behavioral variables as proportions, e.g., the number of a participant’s audio clips containing dyadic interactions divided by the total number of that participant’s clips containing interactions (dyadic plus larger conversational groups). However, converting count data to proportions or percentages discards the information contained in the magnitude of the denominator; e.g., the observation that one participant’s multi-interlocutor social interactions comprised 10 out of 20 clips containing any social interaction is less reliable (contains less information) than the observation that another participant’s multi-interlocutor interactions comprised 60 out of 120 clips containing any interaction. A binomial regression procedure models the observed (count) form of the data as a binary outcome (here, whether a particular behavior does or does not occur in a particular audio clip). A multi-level binomial regression, using participants as clusters, estimates both an overall intercept (here, the probability of a behavior occurring in a clip when the predictor variable[s] is at its mean value) and a specific intercept for each participant. Thus, the model takes into account (1) unmeasured differences between participants that are related to the occurrence of the designated behavior, (2) the additional uncertainty introduced by rarely occurring behaviors, and (3) skewed distributions of behavior frequencies. Estimates of participant-specific intercepts that are farther from the overall intercept, and/or based on a smaller denominator (i.e., the number of clips in which the behavior could possibly have occurred), are “shrunk” farther toward the overall intercept. These intercepts are then incorporated into a linear model of the relationship between the predictor (here, the LHS or personality measure) and the probability of a behavior occurring during a given clip. A logit link maps this probability (which must lie between 0 and 1) onto a linear model (in which the outcome variable is expressed as log-odds). Rather than a p-value, each Bayesian model yields a posterior distribution of parameter value combinations, i.e., the relative probability of having generated the observed data associated with each combination of intercept, “slope” (i.e., the relationship between the predictor variable and the probability of engaging in that behavior), and standard deviation of the outcome variable. As an example, the relationship between ALHB score and the frequency of multi-interlocutor social interactions was modeled as follows: (1)${\displaystyle {\text{Count-of-multi-interlocutor-interactions}}_i\sim \text{Binomial}\left(n_i,p_i\right)}$ (2)${\displaystyle \text{logit}\left(p_i\right)=\alpha +{\alpha }_{\mathrm{PARTICIPANT[i]}}+\left(\beta \ast {\text{ALHB}}_i\right)}$ (3)${\displaystyle \alpha \sim \text{Normal}\left(0,2\right)}$ (4)${\displaystyle \beta \sim \text{Normal}\left(0,2\right)}$ (5)${\displaystyle {\alpha }_{\mathrm{PARTICIPANT[i]}}\sim \text{Normal}\left(0,{\sigma }_{\mathrm{PARTICIPANT}}\right)}$ (6)${\displaystyle {\alpha }_{\mathrm{PARTICIPANT}}\sim \text{HalfCauchy}\left(0,1\right)}$ Line (1), the likelihood, states that the number of clips containing multi-interlocutor interactions involving participant i is distributed binomially, with n_i as the total number of i’s clips containing any social interactions, and p_i as the probability that a clip containing social interaction by participant i contains a multi-interlocutor interaction. Line (2) is the logit link function, incorporating the overall intercept, the participant-specific intercept offset, and the slope (β). Lines (3) through (6) are the prior distributions (described as mean, standard deviation) for the overall intercept, the slope, the participant-specific intercept, and the participant-specific standard deviation.

Before modeling relationships between ALHB scores and behavior frequencies, I modeled relationships between behavior frequencies and (1) whether or not a clip was recorded during a weekend (17:00 Friday to 24:00 Sunday, or during a school holiday) and (2) the participant’s sex. These were treated as dummy variables, with weekend and female coded as 1, while weekday and male were coded as 0. Class attendance, alone among all analyzed behaviors, was assumed to be possible only during weekdays (see Table 1). Weekend is a nuisance variable in the context of the present study. Sex, however, is a theoretically noteworthy variable, in that women generally pursue slower life history strategies than men (Figueredo et al., 2005; Kubinski, Chopik & Grimm, 2017). Because of the rarity (see Table 2) of two of the recorded behaviors, volunteering and religious service/study, and consequent problems with model performance, clips containing either of these behaviors were summed into a single variable, predicted to be positively associated with the ALHB.

For 61 of the 91 participants (67%), n_i (the denominator, i.e., the number of clips in which the behavior could have been observed) comprised all clips in which the participant was compliant and awake for the behaviors social interaction, presence at amusement venues, TV and movie watching, videogame playing and the aggregated religious services/volunteering category. For the remaining 30 participants, n_i varied across these behavior categories because one or more of these behaviors could not be coded in one or more clips (e.g., the participant was at a bar in which the music was so loud that it was impossible to determine whether he or she was engaged in social interaction). The maximum percentage of a participant’s awake/compliant clips that was subtracted from the total for such an adjustment was 6.6%. See Data S1 for more details.

Each model generated by these procedures was then compared, using the Widely Applicable Information Criterion, to a null model (i.e., a model of the same behavior containing no predictor variable, only the intercept). Information criteria reward models for explaining more variance in the dependent variable, but penalize models for additional parameters, thus reducing overfitting. The Akaike weights of the compared models sum to 1. A model’s weight is an estimate of the probability that the model will make the best predictions on new data, conditional on the set of models considered. For each behavior, all models with Akaike weights greater than that of the null model were then compared, as a single set of models, to each other as well as to the null model. Interpretations took into account both (1) the distribution of a model’s estimated value of β (i.e., the extent to which its credible interval overlapped zero) and (2) the model’s Akaike weight relative to other models predicting that behavior.

Models were fitted using the Markov chain Monte Carlo estimation engine Stan, via the R Rethinking package (McElreath, 2016; R Core Team, 2016; Stan Development Team, 2017).

Because I examined the relationships of many behaviors to the ALHB, correction for multiple tests is necessary. For those behaviors that, based on the procedures described above, appeared to be predicted by the ALHB, I randomly re-assigned the ALHB scores (but no other variables) among the participants and re-ran the models 100 times. I then compared the estimated slopes and Akaike weights of the models based on the real data to the distributions of slopes and Akaike weights generated from these resampled data sets.

Descriptive statistics

As reported by Manson (2017a), participants deleted an average of 1.1% of their audio clips (SD = 2.4%, median = 0, range = 0–15.0%). The mean number of compliant clips per participant was 232 (SD = 25.2, range = 148–261). For 43 of the 91 participants (47.3%), at least one clip was recorded during a weekend. Of 21,112 compliant clips, 5679 (26.9%) were recorded during weekends.

Table 1 shows Cohen’s kappa for each of the 18 behaviors analyzed. Six behaviors (sighing, talking to kin, and four talk topics) were excluded from further analysis because of inadequate (κ < .41) inter-rater reliability. Table 1 also describes, for each behavior, the relevant denominator, i.e., the set of audio clips in which the behavior could possibly have been observed (n_i in the binomial models). Table 2 presents data on the distributions of observed frequencies of each of the 12 behaviors with adequate inter-rater reliability. Five of the behaviors were observed at least once in >80% of participants, whereas three behaviors (volunteering, attending religious services or study, and talk endorsing alcohol or recreational drug use) were observed in <20% of participants.

Participants rarely went to sleep before the overnight non-recording period began at 24:00. Sixty-seven of 91 (73.6%) participants were never observed to do so. Among 243 observation days in which participant compliance continued until 24:00, participants were still awake at 24:00 on 209 (86.0%) days. One participant chose to delete all his early morning clips, reducing the sample size to 90 for analyses of daytime sleeping. From these 90 participants, of 4,306 audio clips in which sleeping was recorded, only 338 (7.8%) were recorded between 18:00 and 24:00.

Four participants, who participated during summer 2014, were dropped from the analysis of predictors of classroom attendance, because inspection of their audio clips and event diaries indicated that there were not enrolled in any classes during the recording period.

As reported in by Manson (2017a), the ALHB’s internal reliability (i.e., across its seven scales) was .68.

Model tests

Seven behaviors were more frequent during weekends than weekdays. This was evident both in the estimates of β (positive, with the 95% credible interval not including zero), and in comparisons of models using weekend as a predictor to null models of the same behaviors (the former having 100% of the Akaike weight in all seven cases). These seven behaviors were daytime sleeping, social interaction, proportion of social interactions that involve more than one interlocutor, presence at amusement venues, TV and movie watching, videogame playing, and the aggregated religious services/volunteering category. For these behaviors, models of their associations with ALHB were compared to models including weekend as the only predictor (hereafter, weekend-only models). Only one behavior showed a sex difference, such that the 95% CI of the estimate of β did not include zero, and the model with sex as a predictor had more Akaike weight than the null model. This behavior was videogame playing, which was engaged in more by men than by women (β =  − 2.76, 93% CI = [−4.23, −1.28], Akaike weight of predictor model = 84%). The model of the relationship between the ALHB and videogame playing were compared to a model in which weekend and sex were the only predictors.

Table 3 shows the results of tests of models using ALHB as a predictor of behavior frequencies. Only one behavior, presence at amusement venues, was associated with ALHB scores (Akaike weight compared to the weekend-only model: 1.00). This association was in the opposite of the predicted direction: people whose ALHB score indicated a slower life history strategy spent more time at amusement venues than those whose ALHB score indicated a faster life history. However, random reassignment of ALHB scores among participants casts doubt on the strength of the relationship. Of the 100 runs of the model with randomly reassigned ALHB scores, six generated positive relationships between ALHB and presence at amusement venues with an Akaike weight of 1.00 compared to the intercept-only model (median estimate of β: 1.47). Five other runs generated negative relationships between ALHB and presence at amusement venues with an Akaike weight of 1.00 compared to the intercept-only model (median estimate of β: −1.63).