Briefer Assessment of Social Network Drinking: A Test of the Important People Instrument-5 (IP-5)

Abstract

Introduction

The Important People instrument (IP; Longabaugh et al., 2010) is one of the most commonly-used measures of social network drinking. Although its reliability and validity are well-supported, the length of the instrument may limit its use in many settings. The present study evaluated whether a briefer, 5-person version of the IP (IP-5) adequately reproduces scores from the full IP.

Method

College freshmen (N=1053) reported their own past-month drinking, alcohol-related consequences, and information about drinking in their close social networks at baseline and one year later. From this we derived network members’ drinking frequency, percentage of drinkers, and percentage of heavy drinkers, assessed for up to 10 (full IP) or 5 (IP-5) network members. We first modeled the expected concordance between full-IP scores and scores from simulated shorter IP instruments by sampling smaller subsets of network members from full IP data. Then, using quasi-experimental methods, we administered the full IP and IP-5 and compared the two instruments’ score distributions and concurrent and year-lagged associations with participants’ alcohol consumption and consequences.

Results

Most of the full-IP variance was reproduced from simulated shorter versions of the IP (ICCs≥0.80). The full IP and IP-5 yielded similar score distributions, concurrent associations with drinking (r=0.22 to 0.52), and year-lagged associations with drinking.

Conclusion

The IP-5 retains most of the information about social network drinking from the full IP. The shorter instrument may be useful in clinical and research settings that require frequent measure administration, yielding greater temporal resolution for monitoring social network drinking.

Social environments are integrally involved in the initiation, maintenance, and cessation of problem alcohol use (Hunter-Reel, McCrady, & Hildebrandt, 2009; McCrady, 2004). For example, a heavy-drinking network increases the likelihood of future heavy drinking among adolescents and young adults (social influence), and inversely, individuals who drink heavily are likely to develop social networks that drink more heavily (i.e., social selection; Bullers, Cooper, & Russel, 2001; Parra, Krull, Sher, & Jackson, 2007; Reifman, Watson, & McCourt, 2006). Reductions in numbers of drinking- and heavy-drinking network members and increases in numbers of non-drinking and non-heavy drinking network members are hypothesized to be causal mechanisms of behavior change in many interventions (Longabaugh, Magill, Morgenstern, & Huebner, 2013), and results from several studies are consistent with this hypothesis (DeMartini, Prince, & Carey, 2013; Kelly, Stout, Magill, & Tonigan, 2011; Litt, Kadden, Kabela-Cormier, & Petry, 2009).

The Important People instrument (IP, Clifford & Longabaugh, 1991) is one of the most commonly-used measures for assessing social network drinking. In its usual format (e.g., Longabaugh, Wirtz, Zywiak, & O’Malley, 2010), participants identify up to 10 people in their social network who are considered important and with whom they have had recent contact. Participants then answer a set of questions about each individual, including their relationship to the participant, their general supportiveness to the participant, their reactions to the participant’s drinking, and their typical drinking behavior. Although originally developed as an interview, the IP has been validated in paper-and-pencil and computer-based formats (Hallgren, Ladd, & Greenfield, 2013) and extended to assess other network behaviors related to HIV risk (Zywiak, Stout, Braciszewski, & Wray, 2014), smoking (Roberts, Nargiso, Gaitonde, Stanton, & Colby, 2015) and drug use (Owens & McCrady, 2014; Zywiak et al., 2009). The instrument has been validated in a variety of populations, including college students (e.g., DeMartini et al., 2013; Hallgren et al., 2013; Orford, Krishnan, Balaam, Everitt, & Van der Graaf, 2004), adults in outpatient treatment (e.g., Hunter-Reel, McCrady, Hildebrandt, & Epstien, 2010; Longabaugh et al. 2010; Zywiak, Longabaugh, & Wirtz, 2002), and adults in criminal justice settings (e.g., Nargiso, Kuo, Zlotnick, & Johnson, 2014; Owens & McCrady, 2014).

Psychometric studies have generally supported the reliability, concurrent validity, and predictive validity of IP indices (Hallgren et al., 2013; Longabaugh, Wirtz, Zweben, & Stout, 1998). However, factor analytic studies have yielded limited support for a consistent factor structure, suggesting anywhere from 2 to 6 factors representing different network constructs (Groh, Olson, Jason, Davis, & Ferrari, 2007; Hallgren et al., 2013; Longabaugh, Beattie, Noel, Stout, & Malloy, 1993; Longabaugh et al., 2010; Zywiak et al., 2009). Although a number of network-related summary variables can be computed from the IP (e.g., Longabaugh et al. 2010 identify 46), indices that reflect the drinking behavior of network members have consistently demonstrated the strongest concurrent and predictive validity (DeMartini et al., 2013; Groh et al., 2011; Hallgren et al., 2013; Owens & McCrady, 2014; Zywiak et al., 2002) and have been recommended as the best candidates for predicting subsequent drinking in participants (Groh, Jason, Ferrari, & Halpert, 2011). In contrast, constructs that reflect network members’ general supportiveness or encouragement for drinking/treatment have had limited and mixed associations with concurrent and future drinking, respectively (Groh et al., 2007, 2011; Zywiak et al., 2002).

Despite research supporting its reliability and validity in multiple populations, the length of the IP may still limit its feasibility for use in many settings. For example, its typical format can consist of up to 80 questions (eight questions per network member, up to 10 network members) and is estimated to take an average of 12 minutes to complete (Longabaugh et al., 2010), though in our anecdotal experience can take up to 20 minutes for some participants. This may result in administrators limiting the frequency that network drinking is assessed, despite the potential benefit of frequent, routine assessments in many clinical and research settings. There is currently no existing research that systematically evaluates shorter IP instruments.

Efforts to shorten the IP may have been limited by the instrument’s unusual structure. Specifically, the instrument asks individuals to nominate up to 10 important people and then asks a number of questions about each of these people. IP indices are computed by summing or averaging the responses provided about each network member (e.g., percentage of heavy-drinking network members), making each person listed analogous to an “item” or question on a traditional instrument. However, unlike many traditional instruments, the IP allows each participant to identify different numbers of network members, which leads to different numbers of “items” completed per participant. Additionally, the IP does not instruct participants to identify network members in any particular order, and therefore the order of items (network members) is exchangeable rather than fixed as in traditional instruments. These unique features limit the use of many traditional psychometric approaches for developing short forms (e.g., item-response theory, Spearman-Brown prediction formula).

In the present study, we utilize a combination of simulation and quasi-experimental methods to test whether a briefer version of the IP can provide similar information about social network drinking as the full IP while retaining its concurrent and time-lagged associations with alcohol use. This testing took place in two stages, both of which used data from a longitudinal investigation that administered the IP on two occasions one year apart. In the first stage of testing, scores from shorter IP instruments were simulated by sampling subsets of network members listed in the full IP. This method allowed us to simulate scores from hypothetical versions of shorter IP instruments, drawing from the full IP data while accounting for different ways that a smaller network might be identified. Scores from the hypothetical, shorter IP could then be compared with scores from the full IP, allowing us to identify how shorter versions would likely perform when different numbers of network members could be listed. In the second stage of testing, we compared distributions of IP scores and concurrent and year-lagged associations between participant and network drinking indices using quasi-experimental methods in which the full IP (10-person) or IP-5 (5-person) was administered to multiple cohorts of participants. We hypothesized that the simulation results would indicate that a shorter IP could capture most of the variance and concurrent associations between network drinking and participant alcohol use. We also hypothesized that the full IP and IP-5 tested in the quasi-experimental study would obtain similar means and distributions of IP drinking indices and would return similar concurrent and time-lagged associations with alcohol use.

Method

Participants

Participants were 1053 college freshmen recruited prior to their first year of college in a study of naturalistic changes in alcohol use (Barnett, Orchowski, Read, & Kahler, 2013). Participants were recruited from three cohorts in 2004 (n=245), 2005 (n=395), and 2006 (n=413) at three Northeastern U.S. colleges and universities. Participants completed two online surveys approximately one year apart. Year-1 measures were completed after high school but before the first year of college (in July or August); Year-2 measures were completed after the first year and before the second year of college (in May or June).

Eligibility criteria included attending high school in the United States and planning to live on campus at college. Mean age was 18.36 (SD=0.49) and 606 participants (57.5%) were female. Latino/Hispanic ethnicity was reported by 125 participants (11.9%). For race, 691 (65.6%) were Caucasian, 135 (12.8%) were Asian, 76 (7.2%) were Black/African American, 62 (5.9%) were multiracial, and 89 (8.4%) reported other races or did not report a race (most of these participants reported Latino ethnicity).

Measures

Network Drinking Survey

The IP (Clifford & Longabaugh, 1991) assessed participants’ social networks via a computerized questionnaire. Participants were first instructed to list people who had been important to them in the past year, including family members, friends, people from work, or anyone that was seen as having a significant impact on their lives (Longabaugh et al., 2010). Participants then answered 13 questions about each network member including demographic information (age, gender, race); the duration and type of relationship; frequency of contact; general supportiveness to the participant; the network members’ drinking frequency, maximum number of drinks per day, and general drinking status (e.g., abstainer, occasional drinker, frequent or heavy drinker); and frequency that participants drank with the identified network member. The instrument was adapted from the original IP to be more relevant to a general population of college students (e.g., support for treatment items were removed); however, the items used to compute network drinking indices in the present analyses were identical to those used elsewhere (e.g., Longabaugh et al., 2010).

Administration of the full IP and the IP-5 was stratified by cohort. Cohort 1 completed the full IP in Years 1 and 2; cohort 2 completed the full IP in year 1 and the IP-5 in Year 2; cohort 3 completed the IP-5 in Years 1 and 2. The full IP and IP-5 allowed participants to nominate up to ten and five network members, respectively, but participants could nominate fewer network member in both versions.

IP indices in the present study focused specifically on network members’ drinking behavior, as these indices have generally had the strongest associations with participant drinking (Groh et al., 2011; Hallgren et al., 2013). IP drinking indices included the frequency of network drinking (mean number of drinking days in the past year per network member), the percentage of drinking network members, and the percentage of heavy-drinking network members (i.e., network members identified as frequent, problem, or heavy social drinkers or alcoholic compared to those identified as abstainers, occasional/light, or moderate/average social drinkers). IP index computations were based on Longabaugh et al. (2010). The full instructions, content, and scoring procedures of the IP measure used in the present study are included in online supplementary materials.

Alcohol consumption

Typical alcohol consumption was assessed over the past month in Year 1 (post-high school, pre-college) and Year 2 (post-first year of college, pre-second year of college). Participants were provided with information about standard drink sizes, then indicated the number of drinks they consume in a typical day of drinking. Participants also entered their weight and the duration of their typical drinking episode to allow computation of estimated blood alcohol concentration (eBAC; Matthews & Miller, 1979).

Negative alcohol-related consequences

Negative consequences due to drinking were assessed using the Young Adult Alcohol Problems Screening Test (YAAPST; Hurlbut & Sher, 1992). The instrument contains 27 items in which participants report the frequency of experiencing various consequences during the past year. Nineteen items were rated on a scale from 0 (did not occur) to 3 (occurred three or more times); eight items were rated dichotomously as either 0 (did not occur) or 1 (occurred). Responses to these items were summed to indicate the severity of past-year negative alcohol-related consequences. The instrument was developed and validated specifically for use with young adult college students.

Procedure and Analytic Plan

Simulation procedure

The concordance between full-IP scores and hypothetical scores were obtained by simulating shorter versions of the instrument using data from the first cohort of participants, who completed the full IP. Simulated scores were produced by sampling subsets of IP responses from up to k network members per person, where k ranged from 1 network member to 9. For example, when k=5, data from up to five network members would be sampled from the network members listed, and IP scores were computed based only on these five network members. This yielded data sets with simulated IP indices that mimicked the potential responses that could have been obtained if participants had completed shorter versions of the IP.

When fewer than k network members were listed by a participant, the original number of network members listed was used to compute IP indices. When more than k network members were listed by a participant, the subset of network members that participants would identify on a shorter IP would be unknown, and we therefore simulated different strategies that could plausibly reflect different methods that participants might use to prioritize the naming of network members if the instructions limited the number they could identify. We simulated four strategies that reflect how participants might select a smaller subset of k network members. These included prioritizing network members based on (a) their order of importance, (b) their order of supportiveness, (c) their order of who was known longest, and (d) randomly selecting network members (without replacement) via a computerized random number generator. An illustrative example of these strategies is included in the online supplementary materials corresponding to this paper.

The concordance between network drinking indices obtained from the full (10-person) IP and indices obtained from simulated shorter versions was assessed using two-way absolute-agreement intra-class correlations (ICCs). Although there are no universal cutoffs for acceptable concordance levels, we used Cicchetti’s (1994) rules of thumb for interpreting ICCs as poor (<0.40), fair (0.40 to 0.59), good (0.60 to 0.74) or excellent (0.75 to 1.00). Concurrent associations between simulated IP drinking variables and participant drinking indices were assessed using Pearson correlations. These correlations were then compared to the Pearson correlations obtained from the full (10-person) IP to assess whether the shorter and full IP scores yielded similar correlations between participant and network members’ drinking.

Quasi-experimental full IP and IP-5 administration

A quasi-experimental design was then used to assess whether the IP-5 provides similar information as the full (10-person) IP. Different versions of the IP were administered based on the cohort (calendar year) that participants were recruited. Since the full IP was administered in both years for cohort 1, the IP-5 was administered in both years for cohort 3, and the two versions were administered in Year 1 and Year 2, respectively for cohort 2, we were able to evaluate the similarities and differences between samples that received different lengths of the instrument. Means and variances of network-drinking indices were compared between each cohort to test whether similar distributions were obtained between the full IP and IP-5. Associations between Year-1 IP network drinking indices and Year-1 participant drinking, Year-1 IP network drinking indices and Year-2 participant drinking, and Year-1 participant drinking and Year-2 IP network drinking indices were compared between cohorts to evaluate whether concurrent and predictive associations of IP indices and drinking were similar for the full IP and IP-5. Concurrent associations were evaluated using Pearson correlations. Time-lagged associations were evaluated using standardized regression coefficients to facilitate effect size comparison while controlling for Year-1 values of the Year-2 dependent variable.

Results

Sample Description

Descriptive statistics for the full sample and each cohort are presented in Table 1. None of the participant drinking indices differed by cohort, and each significantly increased from Year 1 to Year 2 (all p < .001).

Table 1.

Descriptive Statistics for Study Sample

	Full Sample N=1053 M (SD)	Cohort 1 n=245 M (SD)	Cohort 2 n=395 M (SD)	Cohort 3 n=413 M (SD)	p-value
Age (Year 1)	18.36 (0.49)	18.39 (0.54)	18.37 (0.51)	18.33 (0.42)	0.14
Female, % (N)	57.55 (606)	62.04 (152)	54.94 (217)	57.38 (237)	0.21
Racial/Ethnic Minority, % (N)	34.38 (362)	33.47 (82)	34.43 (136)	34.87 (144)	0.94
Typical Drinks per Drinking Day, Year 1	2.55 (3.03)	2.58 (3.08)	2.48 (3.08)	2.59 (2.95)	0.88
Typical Drinks per Drinking Day, Year 2	3.21 (3.25)	3.22 (3.20)	3.26 (3.16)	3.15 (3.38)	0.76
Typical eBAC, Year 1	0.04 (0.06)	0.05 (0.07)	0.04 (0.06)	0.04 (0.06)	0.48
Typical eBAC, Year 2	0.07 (0.07)	0.08 (0.07)	0.07 (0.06)	0.07 (0.07)	0.25
Alcohol Consequences, Year 1	3.31 (4.97)	3.60 (5.43)	3.16 (4.69)	3.29 (4.94)	0.52
Alcohol Consequences, Year 2	5.57 (6.54)	5.45 (6.46)	5.77 (6.81)	5.45 (6.34)	0.89
Number of Network Members, Year 1	6.20 (2.71)	7.55 (2.86)	7.28 (2.74)	4.38 (1.13)	0.00
Number of Network Members, Year 2	4.11 (2.52)	5.42 (3.76)	3.68 (1.82)	3.75 (1.83)	0.00
Network Drinking Frequency, Year 1	50.76 (56.91)	51.19 (55.56)	52.15 (56.34)	49.17 (58.33)	0.60
Network Drinking Frequency, Year 2	65.08 (60.02)	62.55 (53.90)	63.28 (58.49)	68.16 (64.48)	0.25
Percent Drinkers in Networks, Year 1	69.25 (30.10)	69.44 (28.27)	68.24 (30.30)	70.11 (30.97)	0.68
Percent Drinkers in Networks, Year 2	80.05 (27.68)	80.33 (25.98)	80.48 (27.32)	79.49 (28.96)	0.69
Percent Heavy Drinkers in Networks, Year 1	12.41 (18.72)	14.56 (20.16)	11.42 (15.82)	12.10 (20.30)	0.17
Percent Heavy Drinkers in Networks, Year 2	15.09 (21.86)	15.38 (20.19)	14.30 (22.04)	15.68 (22.58)	0.76

Note. p-values reflect tests for differences in means/proportions between cohorts, computed using ANOVA for continuous measures and chi-square tests for categorical measures. Network drinking frequency reflects the mean number of drinking days per network member per year as reported by participants. Percent drinkers in networks indicates the percentage of network members who were described as having a non-abstinent drinking status. Percent heavy drinkers in networks indicates the percentage of network members who were identified as frequent, problem, or heavy social drinkers or alcoholic. eBAC = estimated blood alcohol concentration. Year 1 network indices were assessed using the full IP (10-person) for cohorts 1-2 and the IP-5 (5-person) for cohort 3. Year 2 network indices were assessed using the full IP for cohort 1 and the IP-5 for cohorts 2-3.

As expected, the number of network members listed was higher for participants who completed the full IP (cohorts 1–2 in Year 1; cohort 1 in Year 2) compared to the IP-5 (cohort 3 in Year 1; cohorts 2–3 in Year 2). Also, regardless of measure administered, all cohorts reported significantly fewer network members in Year 2 than in Year 1 (all p < .001).

In Year 1, participants reported that about 69% of their network members drank, and about 12% were heavy drinkers, with a mean network drinking frequency of about 51 drinking days per year per network member. There were no mean differences between cohorts in either year, but all of these indices increased from Year 1 to Year 2 in the direction of greater network drinking (all p < .001).

Simulated IP-5 Results

Score concordance

Figure 1 displays intra-class correlations (ICCs, y-axis) that indicate the concordance between full-IP scores and simulated scores from shorter IPs. Different values for the maximum numbers of network members (k) used in the simulated instrument are presented on the x-axis; different strategies for how subsets of network members were chosen are presented as separate lines.

Figure 1.

Concordance (ICC, y-axis) between indices from the full IP and brief IPs of different lengths (k, x-axis). Estimates from simulated instruments with a maximum of k=5 network members are circled. See online version of this article for color figure.

Although there was no universal level of k where concordance suddenly increased, concordance was often in the fair or good range when k ≤ 3, and generally within the excellent range when k ≥ 5, indicating a high level of agreement between scores from the full IP and simulated shorter IP versions. With one exception, the selection strategies yielded similar ICCs, suggesting that concordance between full-IP and simulated shorter IP scores was similar regardless of how participants chose a smaller set of network members to name. However, when the most supportive network member subsets were retained, agreement was slightly lower for the index reflecting percentage of heavy drinkers in the network (right panel). When k=5 (circled points in Figure 1), the ICC for this index and selection strategy was 0.80, while ICCs for all other indices and strategies were ≥ 0.87, indicating high concordance between full and simulated shorter IP scores for the other indices and strategies.

Concurrent associations with drinking

Figure 2 displays Pearson correlations for concurrent associations between Year 1 participant drinking measures and simulated shorter IP indices. The 3 × 3 layout of Figure 2 displays each drinking measure by row and each IP index by column. As with Figure 1, estimates obtained by using each selection strategy are presented as separate lines. Pearson correlations obtained using the full IP are presented as dashed horizontal lines, and 95% CIs for these correlations are presented as gray regions.

Figure 2.

Concurrent associations (r, y-axis) between indices from the full IP and brief IPs of different lengths (k, x-axis). The 3 × 3 layout displays Pearson correlations for each selection strategy as a separate line for each drinking measure (separate rows) and each IP index (separate columns). Pearson correlations obtained using the full IP are presented as dashed horizontal lines, and 95% CIs for these correlations are presented as gray regions. Estimates from simulated instruments with a maximum of k=5 network members are circled. See online version of this article for color figure.

All concurrent associations between drinking and full-IP indices were significant and positive. Correlations from the full instrument were similar to correlations obtained from simulated scores of shorter IP instruments when fewer people (k) were nominated. As should be expected, correlations closely approximated the full-IP correlations as k increased. When k=5 (circled points in Figure 2), the difference in Pearson correlations using the simulated versus the full IP scores differed by no more than 0.06 for all combinations of drinking indices, IP indices, and selection strategy; the mean difference in correlations at k=5 was 0.02. In summary, the indices from simulated shorter IP instruments yielded correlations with participant drinking that were highly similar, albeit slightly lower, than correlations that used full IP data.

Quasi-Experimental Comparison of Full IP and IP-5

Distributional Comparison

Year-1 network drinking indices were compared by cohort to evaluate the consistency of distributions obtained from the full, 10-person IP (cohorts 1–2) and the 5-person IP-5 (cohort 3). As described earlier, there were no differences in means between cohorts on drinking or IP-drinking indices in Years 1 or 2 (see Table 1). Bartlett tests for equality of variances, which provide one test for similarity in score distributions, also indicated no differences in variances across the three cohorts for network drinking indices in both years, with the exception of the percentage of heavy drinkers in networks in Year 1, χ²(2)=28.1, p < .001, and network drinking frequency in Year 2, χ²(2)=8.3, p = .01 (standard deviations of IP indices are presented by cohort and year in Table 1; graphical displays of full IP and IP-5 score distributions are included in online supplementary materials). Follow-up contrasts revealed significantly lower variance for cohort 2 compared to the other cohorts for Year 1 percentage of heavy drinkers in networks, F(644,391) = 1.64, p < .001 but no difference between cohorts 1 and 3, F(237, 406) = 0.99, p = 0.91, who were administered the full IP and the IP-5, respectively. Likewise, cohort 3 had higher variance than cohorts 1 and 2 for Year 2 network drinking frequency, F(352,525) = 1.29, p = .001, but cohorts 1 and 2 did not differ, F(191,333) = 0.85, p = .21. In summary, of the 18 combinations in which IP index variances were contrasted between cohorts, there were two occasions when cohorts that received different versions of the IP had significantly different variances and two occasions in which cohorts that received identical versions of the IP had significantly different variances.

Concurrent and predictive associations

Effect size indices were computed for each cohort to compare concurrent and year-lagged associations between participant drinking and network drinking indices (see Figure 3). In the top panel of Figure 3, each three-line grouping indicates the Pearson correlation and 95% CI for concurrent (Year 1) associations between each IP network drinking index and participant drinking variable. All concurrent associations (top panel) were positive, significant, and in the medium to large range (0.22 ≤ r ≤ 0.52). Point estimates for cohort 3, which completed the short IP, were always within the 95% CIs of cohort 1 or cohort 2. Overall, these results indicate that there were minimal systematic differences in the concurrent associations obtained between the full-IP and IP-5 cohorts.

Figure 3.

Associations between participant alcohol use and IP network drinking indices with full IP (10-person) and IP-5 (5-person) versions. Colors of lines (online version only) also indicate which version of the IP was used in each cohort. The bottom panel displays the drinking and network variables in the same order as the other panels to facilitate comparison; note, however, that the dependent and independent variables are reversed in this analysis compared to the middle panel. Regression coefficients in the middle and bottom panels are estimated after controlling for baseline values of their respective outcome variables.

The middle panel of Figure 3 displays standardized regression coefficients and 95% CIs for time-lagged associations of Year-1 network drinking indices predicting Year-2 participant drinking (controlling for Year-1 participant drinking). The results indicate mixed associations between Year-1 network drinking and Year-2 participant drinking. The most consistently significant associations were for the percentage of drinkers in networks, which predicted typical drinks per drinking day and typical eBAC (cohorts 2–3) and alcohol consequences (all cohorts). No other associations were significant in more than one cohort, and all standardized effect sizes were small (B ≤ 0.20). Point estimates for cohort 3, which received the IP-5, were almost always within the 95% CIs for the other two cohorts, indicating that similar results were obtained using the IP-5 and the full IP. However, on one occasion the point estimate for cohort 3 (who completed the IP-5) was outside of the CI from cohort 2 (who completed the full IP) for the association between the percentage of heavy drinkers in networks and subsequent participant alcohol consequences. Moderation-by-cohort analyses indicated that these coefficients were significantly different, B=−0.17, SE=0.06, t=−2.62, p = .01. Overall, these results indicate minimal systematic differences in full-IP and IP-5 scores predicting next-year drinking.

The bottom panel of Figure 3 displays standardized regression coefficients and 95% CIs for Year-1 participant drinking predicting Year-2 network drinking (full IP for cohort 1, IP-5 for cohorts 2–3), controlling for Year-1 network drinking (full IP for cohorts 1–2, IP-5 for cohort 3). Coefficients in this panel were all positive and significant, and in the small-to-medium effect size range (0.13 ≤ B ≤ 0.40) with the exception of Year-1 typical eBAC predicting Year-2 percentage of drinkers in networks, which was positive but non-significant for cohort 1, B=0.11, SE=0.07, t=1.62, p = .11. Point estimates for one cohort often, but not always, fell within 95% CIs of the other cohorts; however, moderation-by-cohort analyses indicated that these strengths of association were never significantly different between cohorts (all p ≥ .08).

Discussion

The primary purpose of the present study was to determine whether a briefer version of the IP would obtain similar information as longer IP versions used in current practice. Our simulation results indicated that IP scores and concurrent associations with drinking could likely be reproduced by substantially reducing the length of the instrument to assess as few as five network members. Moreover, the simulation findings suggested that the manner in which the network members were selected for the shorter instrument had only small effects on the results obtained. Next, a quasi-experimental comparison of the full IP (10-person) and IP-5 (5-person) administered to different college student cohorts indicated that both versions yielded similar distributions of network indices, similar concurrent associations between network indices and participant drinking, and similar year-lagged associations of IP indices predicting participant drinking and participant drinking predicting IP indices. These results indicate that the IP-5 can recover most of the information obtained from the full IP among college students.

Using the IP-5 may yield several practical advantages. For example, reduced time and effort to complete assessments may make the IP-5 more feasible to administer in clinical settings where administration and scoring of instruments reduces time for other clinical activities or is not always feasible or billable. The IP-5 may be more useful for routine administration throughout the course of alcohol interventions, particularly when changes in social networks are directly targeted or hypothesized to facilitate change or indicate treatment progress. In research settings, the IP-5 may also facilitate more frequent assessment of networks, providing greater temporal resolution of network changes (e.g., weekly or real-time assessments of social network drinking) in contrast to the common practice of assessing network drinking less frequently (e.g., yearly, pre- and post-intervention only, or baseline only). When paired with previous findings suggesting that network drinking behavior is more strongly linked with participant drinking than network general supportiveness or network encouragement for treatment/drinking (Groh et al., 2011; Hallgren et al., 2013), it may be reasonable for investigators and clinicians to further shorten the instrument by including only the network drinking behavior questions (3 items) for up to five network members, reducing the instrument to 15 items from 80.

Participants significantly decreased the number of network members reported from Year 1 to Year 2 regardless of whether the IP-5 (mean decrease = 0.63 network members) or the full IP (mean decrease = 2.09 network members) were administered. This was contrary to the expectation that network sizes would increase over the first year of college (Brissette, Scheier, & Carver, 2002). One interpretation of this finding could be that participants were hesitant to identify more network members during the second assessment, particularly when the full IP was used, knowing that doing so would substantially increase the length of the assessment. Despite reporting fewer network members in Year 2, social network drinking indices increased from Year 1 to Year 2. This could be due in part to increases in drinking from existing social network members as well as changes in social network composition during the first year of college, where young adults move toward greater affiliation with friends and less affiliation with family members (Meisel, Kenney, & Barnett, 2016). Participant drinking also increased during this time, reflecting the expected increase in alcohol consumption during the first year of college.

Although it is unknown how participants decide on whom to name in their social networks when they are given a restrictive upper limit (or how they decide in what order to name them), our simulation results determined there was very little difference between methods of selecting network members (i.e., first listed, longest known, most supportive, or random subset). The IP instructions are vague as to which network members participants should identify, only stating that they should be important and have had contact with the participant in recent months (Longabaugh et al., 2010). Our simulation findings suggest that results from the full IP or the IP-5 are likely robust across strategies, and there is likely little benefit to refining these instructions.

Effect sizes of associations between participant drinking and network drinking were typically medium-to-large when measured concurrently, small-to-medium when Year-1 drinking predicted Year-2 networks, and small (and often non-significant) when Year-1 networks predicted Year-2 drinking. Although the present study was not designed to draw causal conclusions about associations between drinking and social networks, this pattern may indicate some degree of social selection, where college students select and maintain relationships over the first year of college with people who drink similarly to the level that participants drank at the end of high school. There was less consistent evidence for social influence (i.e., pre-college network drinking predicting participants’ drinking one year later); however, it should be noted that the first-year IP assessed network drinking at a point prior to which substantial change in networks is likely to occur (i.e., before transitioning into college). It is therefore possible that greater support for social influence would have been found if additional, temporally-closer assessments of drinking and social networks had occurred (e.g., within the first year of college). The relative importance of social influence versus social selection is still not well differentiated; some studies with adolescents and young adults have found stronger social selection than social influence (Bullers et al., 2001; Mundt, Mercken, & Zakletskaia, 2012; Reifman et al., 2006), while others have found evidence for both effects simultaneously (e.g., Cheadle, Walsemann, & Goosby, 2015; Huang et al., 2014; Parra et al., 2007) or both effects operating at different levels depending on the age group being studied (Burk, van der Vorst, Kerr, & Stattin, 2012; Mercken, Steglich, Knibbe, & Vries, 2012; Parra et al., 2007). More frequent assessments of drinking and social network changes and statistical advances in social network analysis (e.g., Latkin & Knowlton, 2015) may help untangle selection and influence effects across age groups and populations.

Strengths and Limitations

This study had several strengths, including its use of data from a large sample of students from three universities with many participants who were racial and ethnic minorities. Although traditional analytic methods were not always able to be used given the measure’s design, the methods adhered to the spirit of short form development (e.g., Smith, McCarthy, Anderson, 2000) by comparing score distributions, concordance of full IP and IP-5 scores, and concurrent and predictive associations with participant drinking and drinking-related consequences. The use of simulation methods allowed us to control and test variables that could not be directly manipulated, such as the manner in which network members were selected, while also allowing for comparison of full IP and simulated IP-5 scores within the same sample of subjects at the same point in time. The quasi-experimental study loosened several of these controls and provided greater external validity by allowing participants to freely name network members and allowing empirical comparisons of results obtained from the full IP and IP-5 when administered to different cohorts of participants.

There are also several limitations of the present study. Our use of simulation and quasi-experimental methods each contained tradeoffs that could limit the conclusions. The quasi-experimental design stratified full IP and IP-5 by cohort, and it is possible that comparisons between assessment types could be confounded by cohort differences. Although the present study evaluated time-lagged associations between individual drinking and network drinking, the timing between administration points limited the ability to test stronger causal hypotheses related to social influence and social selection; this could be better addressed in future research by more frequent assessment and formal social network analysis (e.g., Cheadle et al., 2012; Mundt et al., 2012). We did not assess whether the IP-5 was perceived as face valid to clinicians and patients, which could affect its use in clinical encounters. For example, it is possible that clients in alcohol treatment settings may view results from a social network assessment of 10 individuals as having greater credibility or representativeness of their social networks than a briefer measure that assesses only 5 individuals. Finally, the sample consisted of college students in their first and second year and did not include other individuals for whom the instrument is also appropriate (e.g., adults in alcohol treatment or in criminal justice settings). The study also focused on indices reflecting drinking behavior of network members and it is possible that other IP indices (e.g., encouragement for treatment or drinking, general supportiveness) or other content areas (e.g., HIV risk, drug use) may be differently affected by shortening the instrument.

Conclusion

The IP-5 appears to adequately reproduce full-IP scores and concurrent and year-lagged associations with drinking and alcohol consequences in college student freshmen. By assessing five network members, the IP-5 could reduce the maximum length of the full IP by half with minimal information loss. Clinical and research settings that prioritize assessment of network drinking behavior (rather than general supportiveness or encouragement of drinking/treatment) could further shorten instrument by assessing only the network drinking questions for each network member and omitting other items. The IP-5 may facilitate more frequent measurements of network drinking in clinical and research settings.