Tracking Changes in the Endorsement of Injunctive Drinking Norms in Response to the COVID-19 Pandemic Using Longitudinal Alignment Analysis

Abstract

The onset of the pandemic saw shifts in messaging around the acceptability of alcohol consumption at different times and contexts. A psychometric analysis of responses to injunctive norms may reveal important differences in specific aspects of norms that were influenced by the pandemic. Study 1 used alignment analysis to evaluate measurement invariance in low- and high-risk injunctive norms across samples of Midwestern college students from 2019 to 2021. Study 2 used an alignment-within-confirmatory factor analysis (CFA) approach to replicate the solution from Study 1 in an independent longitudinal sample (N = 1,148) who responded between 2019 and 2021. For Study 1, the latent mean for high-risk norms was significantly higher in 2021, and the endorsement of four specific norms also differed. In Study 2, increases in latent means for low- and high-risk norms were observed across 2020 and 2021, and differential endorsement emerged for one high-risk norm item. Examining scale-level changes in injunctive drinking norms provides insight into how college students’ perceptions changed in response to the COVID-19 pandemic.

Young people and more specifically college students are major consumers of alcohol (Karam et al., 2007). This is concerning because college students are prone to heavy drinking behaviors (Borsari & Carey, 2003; Reid et al., 2021). Within the United States, 60% of college students are considered to be drinkers (Balestrieri et al., 2018) and 2 (44.4%) out of five college students admit to participating in binge drinking (Wechsler et al., 2002). Heavy drinking can lead to a multitude of problems such as failure to attend classes, low grades, and sexual assault (A. White & Hingson, 2013). Furthermore, heavy drinking behaviors may be associated with other problems, such as depression and financial issues (Fearnow-Kenny et al., 2001).

The Impact of COVID-19 on College Students’ Drinking

The COVID-19 pandemic may have influenced young peoples’ drinking (Lechner et al., 2020); however, the literature is mixed as to whether young peoples’ consumption increased or decreased. These mixed findings could be due to the pandemic having a dramatic impact on students’ social environments and living situations. As college campuses closed, students returned home or found alternative living situations. Students who moved back in with their parents during the pandemic exhibited a decrease in frequency and quantity of drinking in comparison to those that remained in their living situations following campus closures (Rosansky & Rosenberg, 2021; H. R. White et al., 2020). Students who returned to living at home potentially decreased their drinking due to their parents’ disapproval (Epler et al., 2009). On the contrary, students who remained living on their own reported that their frequency of consumption increased (Rosansky & Rosenberg, 2021; H. R.White et al., 2020).

Other literature suggests that the increases or decreases in students’ alcohol consumption may have been due to individual differences. Overall, a meta-analysis examined how individual difference factors affected alcohol use during the pandemic and uncovered that the majority of studies among young adults (aged 18–25) found decreases in consumption (Acuff et al., 2022). However, college students who experienced greater symptoms of depression and anxiety during the pandemic tended to drink more. Conversely, another study found that those who indicate higher levels of perceived social support reported lower levels of alcohol use during quarantine (Lechner et al., 2020).

Given the inconsistencies within the literature, it is still largely unclear what might have been the driving force behind changes in students’ alcohol consumption during the pandemic. One important factor that might explain fluctuations in students’ drinking is changes in their perceptions of other students’ views around alcohol during the pandemic (e.g., changes in drinking norms); students perceived drinking norms are one of the most robust predictors of students’ consumption (Dibello et al., 2018; Neighbors et al., 2007).

Norms and Drinking

The two types of drinking norms most often studied within the alcohol literature with respect to college student drinking norms are descriptive and injunctive norms. Descriptive norms refer to the quantity in which students perceive other students are drinking (e.g., the number of drinks a college student believes similar college peers are consuming; Zou & Savani, 2019). Conversely, injunctive norms are indicative of students’ perception of what their peers deem as being socially acceptable or socially unacceptable with respect to drinking (Cialdini et al., 1990).

A few studies have demonstrated the impact of changes in college students’ drinking norms during the pandemic on students’ own alcohol use (Bonar et al., 2021; Graupensperger et al., 2021; Litt et al., 2021; Rosansky & Rosenberg, 2021). For instance, Graupensperger and colleagues (2021) found that the majority of students who perceived decreases in their fellow classmates’ alcohol use likewise reported reductions in their own alcohol consumption. Another study that focused on binge drinking frequency examined students’ drinking precampus and postcampus closures and found that students who reported higher perceived drinking norms postclosure also tended to report greater binge drinking and were more likely to be in Greek Life (Bonar et al., 2021). Furthermore, perceiving peers to be frequently posting on social media about their alcohol use to cope with and/or alleviate their boredom due to the pandemic was both associated with students’ own likelihood for posting similar alcohol-related posts and increases in their own drinking (Litt et al., 2021).

The current literature on college students’ drinking during the pandemic has exclusively focused on the role of descriptive drinking norms; however, prior work suggests that perceptions of injunctive norms also play an important role in shaping the drinking patterns of young adults (Borsari & Carey, 2003; Lewis et al., 2010). As such, the current study bridges the gap in the literature on norms by examining changes in students’ endorsement of injunctive norms during the pandemic. According to the theory of planned behavior (Ajzen, 1991), injunctive norms in combination with attitudes and perceived behavioral control influence behavioral intentions, which in turn, can impact health behaviors. Moreover, injunctive norms are malleable depending on contextual factors. For instance, a study conducted prior to the pandemic provided college students with feedback regarding how other students from their university negatively perceived heavy drinking behaviors and found a reduction in participants’ estimates of their university peers’ approval of heavy drinking behaviors (Prince & Carey, 2010).

Students’ social environments with respect to drinking changed dramatically due to campus closures and social distancing mandates. Due to these shifts in the contexts in which students drank, injunctive drinking norms surrounding behaviors that would normally be considered indicators of risky drinking, such as drinking alone, may have changed as well. This underscores the importance of gaining a more nuanced understanding of how different aspects of injunctive norms may have been deemed more or less acceptable in response to the pandemic. Consequently, if students believe that their university peers are more approving of drinking alone, this drinking behavior may become more normalized.

Current Study

The current study utilized a novel psychometric approach to understand differences in college students’ endorsement of low- and high-risk injunctive drinking norms before, during, and after the onset of the COVID-19 pandemic in relation to their own drinking in both cross-sectional and longitudinal samples. Specifically, we applied the alignment method for evaluating measurement invariance of an established injunctive drinking norms questionnaire (Lewis et al., 2010) across independent samples of students who completed the measure during the spring of 2019, 2020, and 2021. This allowed us to evaluate differences across time in the endorsement of specific injunctive norms as well as latent mean differences. We then examined the tenability of this aligned solution in a longitudinal sample of students from the same institution who completed the measure across two or three of the assessment points (2019 & 2020; 2019 & 2021; 2020 & 2021; 2019–2021). These complementary psychometric analyses provide precise and (statistically) powerful insights into potential changes in participants’ approval of injunctive norms in response to COVID-19.

Method

Procedures

The current project uses three waves (2019, 2020, and 2021) of the Student Health Survey (SHS). The SHS was administered to all full-time students at a midsized Midwestern school (approximately 15,000 students were enrolled each year) during March. The students received an email invitation to participate in the online survey, which was administered via Qualtrics. The participants viewed the questionnaires in a random order to increase validity. In addition, to reduce the length of the survey, participants only received a subset of the questionnaires. Response rates for each year were 23.4% (2019), 14.8% (2020), and 17.4% (2021). Upon completion of the survey, students received a coupon to a local coffee shop for a free cup of coffee.

Participants

Demographic characteristics for the serial cross-sectional and longitudinal samples are provided in Table 1. Among the N = 6,691 (M_Age = 20.71) students included in the cross-sectional sample, 59.69% identified as women, 31.68% were freshmen, 5.09% identified as Hispanic or Latinx, and 78.57% indicated their race as White. In addition, 45.63% of students in the cross-sectional sample reported living On-campus, and 34.17% were members of a Greek organization. Turning to the N = 1,148 individuals in the longitudinal sample (M_Age = 20.82), 61.29% identified as women, 30.66% were freshmen, 4.49% identified as Hispanic or Latinx, and 79.09% indicated their race as White. Finally, 43.82% of students in the longitudinal sample reported living On-campus, and 37.11% were members of a Greek organization. A series of χ² and independent samples t tests found no evidence of differences in demographic characteristics across the samples.

Table 1.

Demographic Characteristics for Cross-sectional and Longitudinal Samples.

Characteristic	Cross-sectional (only available)	Longitudinal (first available)	χ2obs or tobs (ϕ or g)
N	6,691	1,148	—
% Womena	59.69	61.29	1.05 (.01)
% Freshmanb	31.68	30.66	0.47 (−.01)
% Hispanic/Latinxc	5.09	4.49	0.73 (.01)
Race			—d
% Asian	11.25	10.02
% Black	2.99	2.87
% Endorse Multiple	4.75	5.75
% Hawaiian or Pacific Islander	0.07	0.09
% Native American or Alaskan Native	0.13	0.26
% No Race Indicated	0.49	0.61
% Other	1.73	1.31
% White	78.57	79.09	0.16 (.00)
% On-campus	45.63	43.82	1.30 (−.01)
% Greek	34.17	37.11	3.75† (.02)
MAge (SD)	20.71 (4.15)	20.82 (4.69)	−0.78 (−.02)
M#Days Drinking (SD)	1.30 (1.37)	1.37 (1.39)	−1.49 (−.05)
MLow-Risk Norms (SD)	5.24 (1.48)	5.48 (1.18)	−6.11** (−.17)e
MHigh-Risk Norms (SD)	3.06 (1.18)	3.21 (0.99)	−4.54** (−.13)e

Note. ^an_missing = 4. ^bn_missing = 3. ^cn_missing = 87. ^d Test statistic omitted due to expected cell counts < 5. ^e Satterthwaite adjusted t-test.

^**p < .01. †p < .10.

Measures

Demographics

The researchers used a basic demographic questionnaire to collect the participants’ gender, ethnicity, age, and year in school.

Alcohol Consumption

The researchers provided the participants with the definition of a standard drink (i.e., 12 ounce bottle of beer, 5 ounces of wine, 1.5 ounce shot of distilled spirits). Using this information, participants indicated if they had previous experience consuming alcohol, the number of days per week they typically consumed at least one drink, the typical number of drinks consumed on a typical day on one drinking occasion, and the highest number of drinks they consumed on one alcohol consumption event in the past 30 days.

Injunctive Norms

Injunctive norms were measured (Baer, 1994; Lewis et al., 2010) with the following directions “How acceptable (or unacceptable) do you think the typical [local university] student finds each of the following behaviors? The participants responded with a 7-point Likert-type scale rating from “strong disapproval” (1) to “strong approval” (7). The administered measure was comprised of 17 items, including (a) Drinking alcohol every weekend; (b) Drinking alcohol daily; (c) Driving a car after drinking; (d) Drinking enough alcohol to pass out; (e) Playing drinking games; (f) Drinking to have fun; (g) Drinking shots; (h) Drinking to meet people; (i) Drinking to get drunk; (j) Drinking with friends; (k) Drinking under the age of 21; (l) Drinking alcohol; (m) Drinking to blow off steam; (n) Drinking alone; (o) Never drinking; (p) Drinking to blackout; and (q) Drinking 21 shots on your 21st birthday. Item 17 was not included in the measure developed by Lewis et al. (2010) but was included in the present administration based on prior research investigating social norms surrounding high-risk drinking behaviors (Neighbors et al., 2005). Initial reliability analyses revealed that Item 15 (Never drinking) was weakly and inconsistently correlated with other items, and it was removed from further analysis. These remaining items were grouped into low-risk and high-risk injunctive norms (Lewis et al., 2010). The internal reliability of low- and high-risk injunctive norms was .97 and .67 respectively.

Analysis Strategy

The present study used the alignment approach (Asparouhov & Muthén, 2014) to evaluate measurement invariance (MI) across independent groups of individuals who provided responses to the injunctive drinking norms measure in the serial cross-sectional sample (2019 only, 2020 only, 2021 only), as well as complementary Alignment-within-confirmatory factor analysis (CFA [AwC]; Marsh et al., 2018) model to examine MI in the longitudinal sample comprised of individuals who provided responses across multiple study years (2019/2020, 2019/2021, 2020/2021, and 2019/2020/2021). Generally, the alignment approach provides several advantages over traditional approaches to MI testing (e.g., Stark et al., 2006), most notably, the ability to provide a multigroup factor solution in which the scale of latent variables are “aligned” across groups on a common metric, which allows for the comparison of latent factor means under partial scalar invariance (Asparouhov & Muthén, 2014). In the present study, all factor models were specified using continuous indicators and estimated using robust maximum likelihood to correct standard errors and fit indices for violations of conditional multivariate normality.

Cross-sectional Alignment Model

The standard alignment analysis begins with a minimally constrained baseline model in which the scales of the latent variable(s) in a reference group, 2019 only in our analysis, are fixed (M = 0; variance = 1). In addition, the factor means (α) and variance (ψ) parameters of the latent variables in the other groups (i.e., 2020 only, 2021 only) are freely estimated relative to the reference group. Solutions for the factor loadings (λ), intercepts (ν), and residual variances (θ) are obtained by optimizing a component loss function (similar to an exploratory factor analysis rotation) that attempts to minimize cross-group differences for parameters that are more similar across groups and maximize differences for parameters that are less similar. Once the convergence criterion is satisfied, the aligned model provides pooled estimates for invariant parameters, along with group-specific estimates for noninvariant parameters. Pairwise tests (corrected for multiple comparisons, α = .001; Asparouhov & Muthén, 2014) provide insight into the meaningfulness of cross-group differences for the model parameters involved in evaluating scalar invariance (i.e., approximately equivalent λ and ν across groups). Finally, because the scales of the latent factors are aligned on a common metric, we can meaningfully evaluate differences in latent factor means across groups, even when individual item parameters differ across groups.

Longitudinal AwC Model

Marsh and colleagues (2018) demonstrated that the aligned solution provided by Mplus (Muthén & Muthén, 1998–2018) can be embedded within a general structural equation modeling (SEM) framework as a traditional multigroup CFA model. More specifically, the measurement parameter estimates obtained in an alignment analysis can be used as starting values in a minimally constrained multigroup CFA, and the resulting AwC model will have the same likelihood (and estimated values) as the original alignment model. The present study employed a novel extension of the alignment and AwC analysis by applying the aligned solution from our serial cross-sectional sample to an AwC model in an independent longitudinal sample collected over the same time period. More specifically, the measurement parameters from the aligned cross-sectional sample for each study year (i.e., 2019 only, 2020 only, 2021 only) were used as starting values and ultimately fixed parameter values in a series of increasingly restrictive longitudinal AwC (LAwC) measurement models (covering the same study years) using a separate sample of individuals. We argue that this approach for evaluating longitudinal invariance is superior to traditional alternatives because it allows us to perform an independent, but identically scaled, replication of the invariance analysis from a large sample cross-sectional invariance analysis.

The LAwC analysis applied here is comprised of a series of increasingly restrictive models, mirroring the widely applied progression for the invariance testing approach outlined by Merideth and Horn (2001). The nonindependence introduced by repeated measures was accommodated by allowing scores on the latent factors to covary across constructs and over time. In addition, error covariances were estimated for the same items across time points (i.e., error for each indicator in 2019 covaries with error for the corresponding indicators in 2020 and 2021). Otherwise, our initial configural invariance model (LAwC M1) is identical to the baseline model described by Merideth and Horn (2001) and is identified by constraining the latent factor means for the low- and high-risk norms at 0 and the factor variances at 1, for the latent factors representing each of the T = 3 study years (2019, 2020, and 2021). The remaining measurement parameters (i.e., λ, ν, θ) in LAwC M1 are then freely estimated. The LAwC M2 involves constraining all factor loadings (λ) identified as invariant in the cross-sectional analysis to their estimated values from that model, resulting in T*k_{invariant λ} fewer λ parameters. These λ constraints also allow for the free estimation of latent factor variances for study years 2020 and 2021 (i.e., ψ_Low-Risk2020, ψ_{High-Risk2020}, ψ_Low-Risk2021, ψ_{High-Risk2021}), resulting in T*k_{invariant λ} – 4 fewer model parameters relative to the baseline model (LAwC M1). Finally, the LAwC M3 involves constraining all invariant λ and measurement intercepts (ν) to the values obtained from the cross-sectional analysis, leading to an additional reduction of T*k_invariant intercept parameters. In addition, these intercept constraints allow for the free estimation of latent factor means for the final two study years (i.e., α_Low-Risk2020, α_{High-Risk2020}, α_Low-Risk2021, α_{High-Risk2021}), resulting in T*k_{invariant λ} + T*k_{invariant ν} – 8 fewer model parameters relative to the baseline model.

Overall model fit will be evaluated using the traditional SEM fit indices (i.e., comparative fit index [CFI], Tucker–Lewis index [TLI], root mean square error of approximation [RMSEA], and standardized root mean square residual [SRMR]) based on the (robust) Satorra–Bentler corrected model χ² statistic. Given the large sample size and the number of parameters involved, the tenability of more restrictive models will be evaluated by examining the change in these fit metrics using the benchmarks recommended by Chen (2007), along with changes in the information criteria indices (i.e., Akaike information criterion [AIC], Bayesian information criterion [BIC], and Schwarz Bayesian Information Criterion [SBIC]). More generally, if the λ constraints in our LAwC M2, and the additional ν constraints in our LAwC M3 do not result in a notable deterioration in model fit, we can conclude that the aligned factor solution holds in an independent sample of individuals measured longitudinally. Assuming LAwC M2 or M3 provides a reasonable fit to the data, follow-up analyses comparing the unconstrained λ or ν parameters for the noninvariant study year to a pooled estimate of the invariant study years will be conducted. These analyses will evaluate whether the differential functioning observed in the initial cross-sectional study year alignment analysis replicates in a longitudinal sample.

Results

Sample Descriptive Analysis

Recall that both the cross-sectional and the longitudinal samples described in the present study were independent (i.e., comprised of different individuals) and differ only in the individuals in the cross-sectional sample (N_crossec = 6,691) providing responses to the injunctive norms questionnaire during a single year of the study (i.e., n_2019only = 2,916, n_2020only = 1,763, n_2021only = 2,012), whereas individuals in the longitudinal sample (N_longit = 1,148) provided responses during two (i.e., n_{2019/2020only} = 388, n_{2019/2021only} = 180, n_{2020/2021only} = 429) or all three study years (n_all = 151).¹Table 1 also provides descriptive statistics for the participant demographic characteristics, as well as observed sample means for the high- and low-risk injunctive drinking norm composite variables. Generally, demographic characteristics were comparable for both samples (all ps > .05). Although members of the cross-sectional and longitudinal samples reported comparable frequency of days drinking (M_Diff = −0.07), t(7,836)= −1.49, p = .14, Hedges’ g = −.05, means for low-risk (M_Diff = −0.23), t(1,834.7) = −6.11, p < .01, Hedges’ g = −.17, and high-risk (M_Diff = −0.15), t(1,757.3) = −4.54, p < .01, Hedges’ g = −.13, norms were higher on the first available report from individuals in the longitudinal sample. However, it is important to note that the standardized effect sizes of these differences were small (Cohen, 1992).

Sample 1: Cross-sectional Alignment Results

Pooled and group-specific factor loadings (λ) and intercept (ν) parameters, along with latent factor means (α), for the two-factor (low-risk norms and high-risk norms) alignment analysis are provided in Table 2.² Analysis revealed that λ parameters for all items were invariant across all groups, indicating that the pattern of covariances among low- and high-risk normative approval remained stable across study years. In addition, ν parameters for all indicators were invariant across the 2019 and 2020 study years, suggesting that the conditional endorsement rate of indicators also remained stable over that time frame, after accounting for any latent mean differences. However, four of the 16 measurement intercepts (ν) differed for the 2021 study year. More specifically, participants in 2021 provided more approving (higher) ratings of Item 13 (ν_13-2021 = 4.776; “Drinking to blow off steam”) on the low-risk factor, relative to those in 2020 or 2019 (ν_13-Pooled = 4.660). Turning to the high-risk norms factor, participants in 2021 were more approving of Item 14 (ν_14-2021 = 4.446; “Drinking alone”) relative to 2020/2019 (ν_14-Pooled = 3.312) but provided less approving (lower) responses to Item 4 (ν_4-2021 = 2.737; “Drinking enough alcohol to pass out”) and Item 16 (ν_16-2021 = 3.011; “Drinking to blackout”) relative to earlier groups (ν_4-Pooled = 2.902; ν_16-Pooled = 3.149).

Table 2.

Alignment Analysis Parameters for Sample 1 Analysis.

Item	2019		2020		2021		Pooled
	λ	ν	λ	ν	λ	ν	λ	ν
Low risk
1. Drinking alcohol every weekend	1.389	5.105	1.407	5.091	1.421	5.075	1.403	5.092
5. Playing drinking games	1.546	5.390	1.531	5.370	1.567	5.380	1.549	5.382
6. Drinking to have fun	1.521	5.678	1.522	5.670	1.506	5.627	1.516	5.660
7. Drinking shots	1.567	5.292	1.533	5.287	1.541	5.330	1.550	5.302
8. Drinking to meet people	1.489	5.255	1.473	5.260	1.498	5.281	1.488	5.264
9. Drinking to get drunk	1.578	4.891	1.626	4.939	1.595	4.973	1.596	4.928
10. Drinking with friends	1.427	5.748	1.440	5.720	1.411	5.689	1.426	5.723
11. Drinking under the age of 21	1.500	4.799	1.530	4.843	1.490	4.753	1.505	4.797
12. Drinking alcohol	1.506	5.471	1.508	5.471	1.511	5.443	1.508	5.463
13. Drinking to blow off steam	1.387	4.666	1.345	4.650	1.374	4.776	1.372	4.660
High risk
2. Drinking alcohol daily	1.076	3.089	1.111	3.162	1.115	3.140	1.097	3.123
3. Driving a car after drinking	0.692	1.760	0.654	1.687	0.722	1.766	0.691	1.742
4. Drinking enough alcohol to pass out	1.259	2.908	1.241	2.893	1.258	2.737	1.254	2.902
14. Drinking alone	0.889	3.293	0.826	3.344	0.781	3.446	0.840	3.312
16. Drinking to blackout	1.308	3.151	1.331	3.146	1.314	3.011	1.316	3.149
17. Drinking 21 shots on your 21st birthday	1.284	3.754	1.360	3.727	1.291	3.766	1.306	3.751
Low-risk factor variance (ψLow-Risk)	1.000		0.865		0.981		—
High-risk factor variance (ψHigh-Risk)	1.000		0.961		1.093		—
Low-risk factor mean (αLow-Risk)	0.000 a		0.043a		−0.011a		—
High-risk factor mean (αHigh-Risk)	0.000 a		0.039a		0.194 b		—
nobs	2,916		1,763		2,012		6,691

Note. The right-most column contains pooled estimates for the invariant groups. Bolded parameters can be interpreted as globally invariant across all groups; conversely, bolded elements of the country-specific columns indicate differentially functioning parameters. Italicized values represent fixed parameters. Cohort-specific coefficients in bold are noninvariant (differentially functioning) for that group. Pooled coefficients in bold are globally invariant (DIF-free). Factor mean and variance were fixed at 0 and 1 (respectively) for the reference cohort (2019), and cohort-specific means listed on the same row with different superscripts differ from each other.

No cross-group differences emerged for the means of the low-risk factor, indicating that latent levels of low-risk normative approval remained stable across study years 2019 (α_{Low-Risk 2019} = fixed@0), 2020 (α_{Low-Risk 2020} = 0.043), and 2021 (α_{Low-Risk 2021} = −0.011). Similarly, no difference emerged between 2019 (α_{High-Risk 2019} = fixed@0) and 2020 (α_{High-Risk 2020} = 0.039) across the high-risk factors; however, the latent mean for the 2021 group (α_{High-Risk2 021} = 0.194) was significantly higher than both prior years, which suggested that approval of high-risk norms was elevated during the final study year.³

Sample 2: LAwC Results

Fit statistics for all LAwC models are provided in Table 3. The baseline LAwC M1 converged successfully and did a good job of explaining the observed data (CFI_M1 = .923, RMSEA_M1 = .047 90% confidence interval [CI]: [.045, .047], SRMR_M1 = .052). Fit statistics for the more restrictive LAwC M2 suggest that fixing the 48 factor loadings (λ) to the values obtained from the cross-sectional alignment model while allowing the latent factor variances for 2020 and 2021 the study years to be freely estimated did not result in a notable deterioration in model fit (CFI_M2 =.922, RMSEA_M2 =.046 [.045, .048], SRMR_M2 =.065). Indeed, the change in CFI (Δ_{CFI M2-M1} = −.001), RMSEA (Δ_{RMSEA M2-M1} = −.001), and SRMR (Δ_{SRMR M2-M1} = +.013) from M1 to M2 were well below the meaningful change thresholds of .010, .015, and .030, respectively, recommended by Chen (2007).⁴ In addition, all information criteria indices were smaller for LAwC M2, relative to M1 (Δ_{AIC M2-M1} = −4, Δ_{BIC M2-M1} = −225, Δ_{SBIC M2-M1} = −76), suggesting that any degradation in model fit was reasonable, given the added parsimony of the factor loading constraints. Fit statistics for LAwC M3 (CFI_M3 = .919, RMSEA_M3 = .047 [.045, .048], SRMR_M3 = .067) suggest that imposing constraints on the item threshold (ν) parameters identified as invariant in the initial alignment analysis, and allowing the latent factor means for 2020 and 2021 to be freely estimated, did not result in a meaningful deterioration in model fit. As before, the change in CFI (Δ_{CFI M3-M2} = −.003), RMSEA (Δ_{RMSEA M2-M1} = +.001) and SRMR (Δ_{SRMR M2-M1} = +.002) from M2 to M3 were well below the meaningful change thresholds of .010, .015, and .010, respectively, for when evaluating threshold invariance across groups (Chen, 2007). The AIC estimate for LAwC M3 (Δ_{AIC M3} = 106,651) increased relative to M2 (Δ_{AIC M3-M2} = +50), suggesting a meaningful decrease in model fit; however, both the BIC and the SBIC values for M3 were lower than M2 (Δ_{BIC M3-M2} = −12, Δ_{SBIC M3-M2} = −10), suggesting that the intercept constraints were reasonable given the added parsimony.

Table 3.

Model Fit Statistics for the Sample 2 Longitudinal Alignment-Within-CFA (LAwC) Analysis.

Model	(df) S-B χ2	CFI	TLI	RMSEA [90% CI]	SRMR	AIC	BIC	SBIC
LAwC M1—Configural	(1,017) 3,603.28	.923	.915	.047 [.045, .049]	.052	106,605	107,649	106,992
LAwC M2—Fix All λ	(1,061) 3,693.39	.922	.917	.046 [.045, .048]	.065	106,601	107,424	106,906
LAwC M3—Fix All λ+Inv. ν	(1,093) 3,808.84	.919	.917	.047 [.045, .048]	.067	106,651	107,312	106,896

Note. Item residuals were allowed to covary over time in all models. S-B χ² = Sattorra–Bentler chi-square statistic. LAwC M1 was identified by fixing all latent factor means to 0 and variances to 1, and freely estimating all other model parameters. LAwC M2 fixed factor loadings (λ) for all items to the estimated values from the cross-sectional cohort analysis (Table 2), which allowed the free estimation of latent factor variances (ψ_Low-Risk, ψ_High-Risk). LAwC M3 added fixed measurement intercepts (ν) to the estimated values from the cross-sectional cohort analysis (Table 2) for invariant items (1-3, 5-12, 17), which allowed the free estimation of latent factor variances (ψ_Low-Risk, ψ_High-Risk) and means (α_Low-Risk, α_High-Risk). CFA = confirmatory factor analysis; LAwC = longitudinal alignment-within-CFA; CI = confidence interval; AIC = Akaike information criterion; BIC = Bayesian information criterion; SBIC = Schwarz Bayesian Information Criterion; CFI = comparative fit index; TLI = Tucker–Lewis index; RMSEA = root mean square error of approximation; SRMR = standardized root mean square residual.

Overall, the pattern of model fit statistics across model configurations indicates that the factor solution obtained from the initial alignment analysis provided a good fit of the data obtained from the longitudinal sample. As a result, we conducted follow-up analyses examining the difference between the intercepts of Items 13, 4, 14, and 16 (Table 2) from study year 2021 and the pooled intercept estimates from study years 2019 and 2020. Table 4 provides these unconstrained ν parameters for all study years (ν₂₀₁₉, ν₂₀₂₀, ν₂₀₂₁), the pooled (ν_19/20) estimates, the difference estimates (ν₂₀₂₁ – ν_19/20), and 95% CIs for the difference, which are based on 1,000 bootstrapped resamples. Although the direction of the difference between the 2021 and 2019/2020 pooled intercept (ν₂₀₂₁ – ν_19/20) for Item 13 (“Drinking to blow off steam”) was in the expected direction, the confidence interval suggested that this difference was not statistically meaningful (ν_13-Diff = .020, 95% CI_boot[−.055, .092]).⁵ Turning to the indicators for the High-Risk factor, the intercept difference for item 4 (“Drinking enough alcohol to pass out”) suggests a significantly lower rate of endorsement among 2021 respondents (ν_4-Diff = −.099, 95% CI_boot[−.185, −.005]). The intercept difference for item 14 (“Drinking alone”) suggested comparable rates of endorsement across study years (ν_14-Diff = −.007, 95% CI_boot [−.100, .089]). Finally, the intercept difference for Item 16 (“Drinking to blackout”) emerged in the expected direction (ν_16-Diff = −.093), although the upper-bound of the 95% confidence interval was near 0 (95% CI_boot [−.187, .003]).

Table 4.

Intercept Estimates and Differences, Latent Factor Means, and Variances for Longitudinal the Sample 2 AwC M3 Analysis.

Parameter	2019	2020	2021	Pooled, 19/20	2021—Pooled 19/20, [95% CI]
ν13	4.663	4.726	4.715	4.695	.020 [−.055, .092]
ν4	2.944	2.908	2.827	2.926	−.099 [−.185, −.005]
ν14	3.420	3.412	3.409	3.416	−.007 [−.100, .089]
ν16	3.122	3.169	3.052	3.146	−.093 [−.187, .003]
ψLow-Risk	1.000	0.740	0.705	—	—
ψHigh-Risk	1.000	0.893	1.021	—	—
αLow-Risk	0.000 a	0.113 b	0.285 b	—	—
αHigh-Risk	0.000 a	0.139 b	0.431 c	—	—
nobs	719	967	760	—	—

Note. Estimated measurement intercepts (ν) for items identified as noninvariant in the cross -sectional cohort analysis (iItems 13, 4, 14, 16). Confidence intervals (CI) for intercept differences (ν2021 – ν19/20) and factor mean comparisons were based on 1,000 bootstrapped resamples. Estimates in bold are different from 0 at p < .05. Cohort-specific factor means listed on the same row with different superscripts differ from each other.

Finally, the means for the low-risk factor were significantly higher in 2020 (α_{Low-Risk 2020} = .113) and 2021 (α_{Low-Risk 2021} = .258), relative to 2019 (α_{Low-Risk 2019} = fixed@1), reflecting a generalized increase in students’ beliefs about the acceptability of these behaviors across the final two measurement occasions. In addition, latent means for the high-risk factor exhibited a significant increase between 2019 (α_{High-Risk 2019} = fixed@1) and 2020 (α_{High-Risk 2020} = .139), as well as a significant change between 2020 and 2021 (α_{High-Risk 2021} = .431).

Discussion

The current study suggests that college students’ endorsement of injunctive norms changed during the pandemic such that more students perceived others as being more approving of high- and low-risk drinking behaviors. A pandemic-induced shift of injunctive norms has the potential to lead to the reversal of the downward trend in college student drinking (Schulenberg et al., 2020) as norms are considered one of the strongest predictors of alcohol consumption (Neighbors et al., 2007). Globally it appears that participants perceived others to be more approving of drinking during the COVID-19 pandemic. Nuances of the findings are discussed subsequently.

In the serial cross-sectional sample, the data showed a consistent pattern. In the spring of 2021, students reported a global increase in perceived approval by their peers for high-risk drinking norms (i.e., higher latent mean), but no general increase in perceived approval for low-risk injunctive norms. However, students in the 2021 cross-sectional sample were more likely to endorse the low-risk item “drinking to blow off steam”; this aligns with the Dumas et al. (2020) study which found a relationship between drinking and COVID-19 fear and depression symptoms. This may be because “drinking to blow off steam” often accompanies a specific stressful event rather than being suggestive of drinking to cope. Although the pandemic may have contributed to students’ long-term stress, students may not have considered it a one-off stressor. In addition to the general increase in the endorsement of high-risk behaviors observed in the 2021 sample, students were even more likely to endorse “drinking alone,” which is consistent with findings reported by Dumas and colleagues (2020) indicating that the prevalence of solitary drinking increased in Canadian high school students at the very beginning of the pandemic.

The intercept terms for “drinking enough alcohol to pass out” and “drinking to blackout” were lower among 2021 participants in the serial cross-sectional sample, suggesting that after accounting for the high-risk latent mean increase, these participants did not exhibit much of a shift in their belief of perceived approval of drinking enough to pass out and drinking to blackout. Drinking to blackout or with the intention to blackout is considered a marker of high-risk alcohol consumption (e.g., Ward & Guo, 2020; Yuen et al., 2021). According to social media data, discussions of drinking to blackout continued to be prominent during the pandemic (Ward et al., 2021). However, we did not see a corresponding shift in these beliefs in our sample. This finding is particularly interesting as drinking to pass out and drinking to blackout are behaviors that are often related to parties. Given the social isolation during this period and the relative lack of party environments, a decrease in the perceived approval of these behaviors might have been anticipated.

The longitudinal analyses partially replicated the pattern of findings observed in the serial cross-sectional analyses. Specifically, there was a global (latent mean) increase in approval of high-risk norms between 2020 and 2021, but a lower intercept parameter for “drinking enough to pass out,” suggesting that the increases in approval of high-risk norms did not generalize to this item. Similarly, there was evidence that the intercept parameter for the high-risk item “drinking to blackout” may have decreased in 2021; however, the confidence interval for the difference test captured zero (–.187, .003). The increases in the intercept parameters for “drinking to blow off steam” and “drinking alone” observed in the serial cross-sectional analysis did not replicate in the longitudinal analysis. However, the latent mean of the low- and high-risk factors did increase in the longitudinal sample, suggesting that overall endorsement (approval) of these items did increase over time. Further research is needed to explore these effects.

Over several waves of data, there seems to be a shift in students’ perceptions of their peers’ approval of high-risk alcohol consumption. This increase in approval is particularly surprising given the parameters of the pandemic (i.e., limits on social gatherings). Given the corresponding increase in alcohol purchasing (Hu et al., 2021), these findings warrant further investigation. The approval shift in drinking alone may be due to the social distancing policies and changes in access to parties (see Miech et al., 2021 for further discussion of the impact of these policies on drinking). Thus, it may have been temporary rather than indicative of long-term increases in drinking due to the fact that students may not have been drinking alone by choice but rather they were drinking alone in response to social distancing mandates. Future research should investigate whether students continue to perceive their peers as being more approving of drinking alone and whether holding these beliefs map onto increases in students’ drinking prospectively.

Limitations

A strength of this current study is that, to our knowledge, it is the first to examine how shifts in injunctive norms during the pandemic. However, there are several limitations to note. One major limitation is our study’s sole focus on injunctive norms. As previously mentioned, most of the literature during the pandemic has spotlighted the effects of descriptive norms on college students’ drinking (Bonar et al., 2021; Graupensperger et al., 2021; Litt et al., 2021; Rosansky & Rosenberg, 2021). Incorporating both descriptive and injunctive norms might have helped to disentangle the unique effects of each type of norm on students’ consumption during the pandemic. Furthermore, our results may not be generalizable at the population level. The sample was primarily students who identify as women, White, and first-year students. Additional samples with more gender, race, ethnicity, and age diversity are needed.

The present study provided a unique opportunity to compare the results of cross-sectional and longitudinal measurement invariance analysis of injunctive norms in independent samples drawn from the same population over the same time interval. To the authors’ knowledge, the present study was the first to apply alignment analysis to independent samples across time, and then attempt to replicate the aligned solution (using an alignment-within-CFA model) in a true longitudinal sample over the same period. Although the pattern of findings across studies differed slightly, they were not contradictory, and the incomplete replication across samples may be attributable to differences in sample size. More specifically, the cross-sectional sample was comprised of substantially more observations at each time point (n₂₀₁₉ = 2,916; n₂₀₁₉ = 1,763; n₂₀₁₉ = 2,012) than the longitudinal sample (n₂₀₁₉ = 719; n₂₀₁₉ = 967; n₂₀₁₉ = 760), which likely translates into superior power to detect smaller differences in the serial cross-sectional analysis. Future work should further evaluate the sample size requirements necessary to ensure sufficient power to replicate findings using this approach.

Methodological Implications

The practice of attempting to replicate factor solutions across independent samples is well-established, particularly when applying exploratory approaches in the context of scale development (Fabrigar et al., 1999). However, to the best of our knowledge, the present study was the first to explore the replicability of an alignment solution using a serial cross-sectional sample and a separate longitudinal sample obtained over the same period. Because the alignment module available in version 8.8 of Mplus only supports independent groups, it is not capable of directly evaluating measurement invariance in true longitudinal designs featuring repeated assessments from the same individuals. However, the present work illustrates an intuitive and practical solution to this problem when a serial cross-sectional sample is available to provide an initial aligned solution, which can then be tested using a LAwC applied to a sample comprised of repeated measures over the same time interval. We suspect that cohort sequential monitoring designs such as the one featured here are quite common and we encourage future researchers to consider applying the sample division procedure illustrated here (i.e., identifying participants who respond only once for inclusion in a serial cross-sectional subsample, and those who responded more than once for inclusion in a longitudinal subsample) to take full advantage of the data generated by this form of monitoring protocol.

Conclusion

This application of the alignment approach shows that injunctive norms shifted during the COVID-19 pandemic at the global level and among certain specific types of norms. Given that substance use during the pandemic increases the likelihood of contracting COVID-19 (Dumas et al., 2020) and that injunctive norms relate to alcohol consumption levels (Neighbors et al., 2008), these shifts in norms perceptions may have lasting consequences.