Psychometric Research · CFA · Measurement Invariance

Multigroup CFA of the PWI-SC

9–11 Years vs. 12–18 Years: Testing Age Extension of the Scale
Lindsay Tunbridge  ·  Australian School Children, 2018–2023
8,137
Original Participants
4,824
Cleaned Sample
9–18
Age Range (years)
4
Invariance Levels
1

Introduction

🎯
Research Objective The current PWI-SC manual restricts recruitment to children aged ≥ 12 years (Cummins & Lau, 2023, p. 6). This study seeks evidence to lower that age cut-off to at least 9 years, applying Confirmatory Factor Analysis and Multigroup CFA to Australian school students aged 9–18 years.

Data were collected using the PWI-SC from 8,137 Australian Primary and Secondary school students between 2018 and 2023. After cleaning, two roughly equal-sized groups were randomly extracted for comparison:

🏫
Group 1 — Secondary School
12–18 years · n = 485 (matched subsample)
Reference group for testing measurement invariance. A random subsample of 485 was drawn to create roughly equal group sizes for comparison.
🏡
Group 2 — Primary School
9–11 years · n = 583
The focal group whose validity with the PWI-SC is under examination — extending the age range one year younger than the previous 10–11 year study.
Previous study scope
10–11 years
Primary school group in earlier analysis (n = 480)
Current study scope
9–11 years
Extended to include 9-year-olds — n = 583

Confirmatory Factor Analysis (CFA; SPSS AMOS Version 29) first examined whether the PWI-SC fits the hypothesised one-factor model in the 9–11 year group. A Multigroup CFA then tested whether the scale functions identically (invariantly) across both age groups at four increasingly stringent measurement levels.

ℹ️
Why Multigroup CFA? Unlike multiple regression — which assumes explanatory variable errors are zero — Multigroup CFA incorporates error into its computations. SWB research with students necessarily involves developmental differences that can introduce common method bias, making direct comparisons invalid without demonstrated measurement invariance (Cummins, 2018a).
2

MGCFA Methods: Establishing Measurement Invariance

Measurement invariance is established by comparing changes in model fit across four increasingly stringent levels of constraint. Each level builds on the previous, testing a more restrictive set of assumptions about how the two age groups relate to the underlying subjective wellbeing construct.

The Four Measurement Levels

1. Configural
Same structure
→
2. Metric
Same loadings
→
3. Scalar
Same intercepts
→
4. Strict
Same residuals
🔷 Level 1 · Configural Invariance
Research question: Do items load onto the same factor across groups?
The least rigid assumption. Tests whether the instrument represents the same factor structure in each group (Gregorich, 2006). Items must load onto the same factor(s) in both the 9–11 and 12–18 year groups. If they do, similar (though not necessarily identical) latent variables exist across age groups (Chen et al., 2005). All parameters are freely estimated within each group.
🔵 Level 2 · Metric Invariance
Research question: Are factor loadings the same across groups?
Constrains factor loadings to equality. If achieved, the PWI-SC factor carries the same meaning — indicating that younger primary school children and older secondary students interpret the wellbeing domains similarly (Kamody et al., 2014). Failure here would suggest at least some items are understood differently across ages.
🟡 Level 3 · Scalar Invariance
Research question: Are item intercepts the same across groups?
Constrains item intercepts in addition to loadings. Tests whether observed group means are associated with the same latent wellbeing (Gregorich, 2006). If established, a respondent in either age group would obtain the same observed score for an equal level of latent wellbeing, making latent mean comparisons valid (Milfont & Fischer, 2010).
🟢 Level 4 · Strict Invariance
Research question: Are item residuals the same across groups?
The most stringent level. Constrains all residual variances (specific + error) to equality. Confirms that direct comparisons of observed item means are valid, and that group mean differences reflect true differences in wellbeing rather than differences in measurement error (Fuller-Tyszkiewicz, 2018).

Why Strict Invariance?

The required level of invariance for validly comparing group means is debated. Milfont and Fischer (2010) argue scalar invariance suffices; Stokes et al. (2018) and Van De Schoot et al. (2015) assert strict invariance is required. This study favours strict invariance: without it, group mean comparisons may reflect differences in carelessness or indecision rather than genuine SWB differences. Van De Schoot et al. (2015) state that comparing latent factor means validly is impossible without strict invariance.

📐
Sequential Constraint Progression Each model is compared to its predecessor: Metric vs. Configural → Scalar vs. Metric → Strict vs. Scalar. Changes in fit statistics at each step determine whether invariance holds. Exceeding the preset thresholds on two or more indices signals non-invariance and stops the sequence.
3

Measurement & Comparative Fit Statistics

Following Chen (2007), χ² is not used as the primary fit measure due to its sensitivity to large sample sizes. Three preferential fit indices are employed:

CFI
Good fit: ≥ .95
Comparative Fit Index. Compares the hypothesised one-factor model to a baseline (no-covariance) model. Higher values reflect greater improvement over the baseline (Hu & Bentler, 1999).
RMSEA
Good fit: < .06
Root Mean Square Error of Approximation. Estimates the population covariance matrix, producing p-values and confidence intervals. A narrow CI (lower limit ≈ 0, upper ≤ .08) indicates stable population-level fit (Browne & Cudeck, 1993).
SRMR
Good fit: ≤ .05
Standardised Root Mean Square Residual. The average standardised residual across all model residuals. Values near zero indicate excellent average model-data correspondence (Hu & Bentler, 1999).

Invariance Change Thresholds — Chen (2007)

Comparison TypeΔCFI limitΔRMSEA limitΔSRMR limit
Factor loadings (Metric vs. Configural)≥ .005≥ .015≥ .025
Intercepts & residuals (Scalar/Strict vs. prior)≥ .005≥ .010≥ .005
⚠️
Two-Statistic Rule Non-invariance is flagged only when at least two change statistics simultaneously exceed their thresholds. A single threshold violation without corroboration from a second index is not sufficient evidence of non-invariance.
4

Data Cleaning

A rigorous multi-step cleaning procedure was applied. Crucially, this study also explicitly compares rates of data quality issues between both age groups as an early test of cognitive equivalence.

Cleaning Steps Applied

  1. Homeostatic defeat exclusion: Participants scoring ≤ 50%SM were removed (n = 267). Empirically, a one-factor CFA fits poorly for respondents below this threshold, domain intercorrelations are lower, and Rasch modelling reveals haphazard score variation (Richardson et al., 2015).
  2. Missing data: SPSS multiple imputation identified 416 participants with missing values who were deleted. A further 1,179 students without a year level were also removed (Pallant, 2005).
  3. Response sets: 173 participants at the ceiling (10) and 49 at the floor (0) across all PWI-SC domains were removed, indicating acquiescent responding or poor scale comprehension (International Wellbeing Group, 2013).
  4. Scale competency failures: 917 participants who failed the scale's competency check questions were deleted.
  5. Outliers: Z-score screening identified univariate outliers; 5% trimmed means showed no significant influence, so all were retained. A χ² criterion of 22.46 (df = 6, p < .001) identified 312 multivariate outliers; independent-samples t-tests confirmed no significant mean differences before and after removal.
  6. Skew & kurtosis: All variables within the acceptable range of −7.0 to 7.0 (Pallant, 2005). No transformations required.
  7. Multicollinearity: SPSS 26 collinearity diagnostics yielded Tolerance values all > .10 and VIF values all < 10 — no multicollinearity or singularity detected.

Data Quality Comparison Across Age Groups

An explicit comparison of data quality rates between the 9–11 and 12–18 year groups challenges the assumption that younger children are prone to greater response bias or acquiescent responding (Huebner, 2004):

Response Sets (%)
9–11 years3.3%
12–18 years2.7%
Outliers (%)
9–11 years4.1%
12–18 years5.1%
Non-Normality (%)
9–11 years10.7%
12–18 years8.5%
✅
Comparable Data Quality Rates Primary and secondary school students showed broadly similar percentages of students removed during cleaning. This comparability challenges the assumption that 9- to 11-year-olds are prone to greater response bias than adolescents (Huebner, 2004).

Homeostatic Normative Range

🧮
Calculating the Normative Range (Capic et al., 2017) 4,824 students returned M = 73.89%SM, SD = 10.97. Formula: Lower bound = M − 2SDs; Upper bound = M + 2SDs → normative range 51.95–95.83%SM. Scores ≤ 50%SM indicate homeostatic defeat (Cummins, 2018b) and were excluded.

Sample Size Adequacy

Wolf et al.'s (2013) Monte Carlo benchmarks for one-factor CFA models indicate a minimum sample of 100 for the PWI-SC factor loadings. Both groups substantially exceed this threshold.

5

Measures: Personal Wellbeing Index (School Children)

The Personal Wellbeing Index (PWI), described by the International Wellbeing Group (2013) as a measure of subjective wellbeing following homeostasis theory, contains seven satisfaction items spanning distinct quality-of-life domains. The PWI-SC is adapted for school-aged children.

The Seven Life Domain Items

🏠
Standard of Living
❤️
Health
🏆
Achieving in Life
👥
Personal Relationships
🛡️
Personal Safety
🌍
Community Connectedness
🔮
Future Security

Scores across all seven domains are averaged and converted to a percentage point scale from 0 to 100 (%SM). Available from the Australian Centre on Quality of Life at acqol.com.au.

6

Results

Descriptive Statistics

Table 1 · Participants Before and After Data Cleaning

PopulationOriginal NCleaned N (%)MaleFemaleOtherNot Specified
All Students8,1374,824 (59.28%)2,210 (45.8%)2,276 (47.2%)29 (0.6%)309 (6.4%)
Students 12–18 years5,5864,241 (87.9%)1,900 (44.8%)2,018 (47.6%)14 (0.3%)309 (7.2%)
Students 9–11 years771583 (75.62%)309 (53.0%)259 (44.2%)15 (0.17%)—
ᵃ Gender statistics after data cleaning. ᵇ % of original participants remaining after cleaning.

Table 2 · PWI Mean, Standard Deviation, and Group Data Quality Comparison

PopulationnPWI Mean (%SM)SDResponse SetsOutliersNon-Normality
Original uncleaned data8,13773.8916.29———
Original cleaned data4,82478.1310.97———
Students 12–18 years4,24177.9810.952.7%5.1%8.5%
Students 9–11 years58379.2811.193.3%4.1%10.7%

CFA Results: PWI-SC for 9–11 Year Olds

Before testing invariance, a single-group CFA confirmed the PWI-SC's one-factor structure in primary school students aged 9–11 years.

24.63
χ² (p = .04, df = 14)
.96
CFI (≥ .95 ✓)
.04
RMSEA (< .06 ✓)
.03
SRMR (≤ .05 ✓)
✅
Good Fitting Model Confirmed All three fit indices met or exceeded their cut-offs (CFI ≥ .95; RMSEA < .06; SRMR ≤ .05), confirming a good-fitting one-factor model for the PWI-SC in the 9–11 year group. This qualified the data to proceed to Multigroup CFA.

Multigroup CFA: Measurement Invariance Results

The Multigroup CFA compared PWI-SC performance between 9–11 years (n = 583) and 12–18 years (n = 4,241). Model fit and change statistics at each invariance level are presented below.

Table 4 · Evaluations of Measurement Invariance — PWI-SC Groups (9–11 vs. 12–18 years)

Modelχ²dfΔCFIΔRMSEAΔSRMRCFIRMSEASRMRVerdict
Configural383.789***28 ——— .953.051.036 Baseline
Metric405.152***34 .002.004.0001 .951.048.036 ✓ Invariant
Scalar445.995***35 .004.002.0000 .947.045.036 ✓ Invariant
Strict474.000***42 .002.002.0001 .944.043.036 ✓ Invariant
Groups compared: PWI-SC (12–18 years, n = 4,241) and PWI-SC (9–11 years, n = 583).
Thresholds for excessive change — Factor loadings (any two of): ΔCFI ≥ .005, ΔRMSEA ≥ .015, ΔSRMR ≥ .025; Intercepts/residuals (any two of): ΔCFI ≥ .005, ΔRMSEA ≥ .010, ΔSRMR ≥ .005.
*p = .05, **p = .01, ***p = .001

Interpreting Each Invariance Step

🔷 Configural — Baseline Model
CFI = .953, RMSEA = .051, SRMR = .036. Both groups show the same one-factor structure. The seven wellbeing domains load onto a single latent factor in both the 9–11 and 12–18 age groups, establishing the structural foundation for further testing. Notably the configural CFI already meets the ≥ .95 threshold.
🔵 Metric — Factor Loadings Invariant ✓
ΔCFI = .002, ΔRMSEA = .004, ΔSRMR = .0001. All three change statistics remained well below all thresholds. Strong confirmation that factor loadings are statistically equivalent across groups — the seven wellbeing domains carry the same psychometric meaning for 9- to 11-year-olds and 12- to 18-year-olds.
🟡 Scalar — Item Intercepts Invariant ✓
ΔCFI = .004, ΔRMSEA = .002, ΔSRMR = .0000. All changes far below thresholds — including a ΔSRMR of exactly zero. Item intercepts are equivalent across age groups, confirming that observed scores reflect the same latent wellbeing level regardless of age group membership.
🟢 Strict — Residuals Invariant ✓
ΔCFI = .002, ΔRMSEA = .002, ΔSRMR = .0001. All change values comfortably below all thresholds. Strict invariance confirmed: residual variances are equivalent across both age groups, fully validating direct comparisons of observed item means.
📊
Notably Clean Invariance Results Unlike the 10–11 year comparison (where ΔCFI reached exactly .005 at the metric and strict levels), this 9–11 year comparison produced change values comfortably below all thresholds at every stage — including ΔSRMR values of .0001 or .0000. This is strong, unambiguous evidence of full strict measurement invariance extending to 9-year-olds.
7

Discussion

🎉

Strict Measurement Invariance Achieved: Ages 9–18 Years

The Multigroup CFA demonstrated full measurement invariance between the 9–11 year and 12–18 year groups across all four levels — configural, metric, scalar, and strict — confirming that direct mean comparisons between these age groups are valid and the PWI-SC can be extended to students aged 9–18 years.

Data Quality as Supporting Evidence

Results from Table 2 reveal that primary school students (9–11 years) and secondary students (12–18 years) showed similar rates of response sets (3.3% vs. 2.7%), outliers (4.1% vs. 5.1%), and non-normality (10.7% vs. 8.5%). This directly challenges the belief articulated by Huebner (2004) that younger children lack the cognitive maturity for reliable self-evaluative responding and are prone to acquiescent response biases.

Implications for Practice

📋 Key Findings for the PWI-SC Manual
1 The current manual (Cummins & Lau, 2023) restricts PWI-SC use to students aged ≥ 12 years.
2 Strict measurement invariance was achieved across all four levels, confirming the scale performs equivalently across both age groups.
3 The PWI-SC can validly be extended to students aged 9–18 years, encompassing both primary and secondary school cohorts.
4 Future editions of the manual should consider revising the minimum age criterion from 12 to 9 years.

This conclusion is consistent with Van De Schoot et al. (2015), who assert that validly comparing latent factor means between groups requires strict invariance. The change statistics here were comfortably below all thresholds at every level — including ΔSRMR = .0000 at the scalar step — providing unusually robust psychometric support for the age extension.

The convergence of multiple lines of evidence — comparable data quality rates between age groups, a well-fitting one-factor CFA in the 9–11 year group, and full strict measurement invariance across all four levels — supports both the structural validity and measurement equivalence of the PWI-SC when applied to primary school children from 9 years of age in an Australian context.

8

References