Title: Comparing Reflective and Formative Measurement Models on the Same Indicators
1Comparing Reflective and Formative Measurement
Models on the Same Indicators
- George R. Franke, University of Alabama
- Nick Lee, Aston University
2Formative Principles
- The indicators cause the construct
- Indicators need not be correlated
- Omitted indicators omit part of the construct
- Error term is structural (effect of omitted
causes), not measurement - The construct depends on its outcomes, not just
its indicators
F/R
3Reflective Principles
- The construct causes the indicators
- Indicators should be internally consistent
- Equally-reliable indicators are substitutable
- Error terms are measurement error
- Different antecedents and outcomes of the
construct have limited effects on the indicators
R
R
4Measurement (mis?)specification
- One common view only one model is right or
wrong for each construct - Wrong model causes major problems
- Diamantopoulos Siguaw (2006) Type I and
Type II errors from misspecification - Various estimates of misspecification rates
- 29 in marketing
- 31 in information systems
- 47 in leadership
- 69 in strategy
5Measurement (mis?)specification
- Purported consequences
- misspecification of even one formatively
measured construct within a typical structural
equation model can have very serious consequences
for the theoretical conclusions drawn from that
model (Jarvis, MacKenzie Podsakoff 2003, p.
212) - misspecification can inflate unstandardized
structural parameter estimates by as much as 400
or deflate them by as much as 80 (MacKenzie,
Podsakoff Jarvis 2005, p. 728)
6Alternative Views
- A given set of indicators may (possibly) be
modeled equally well as formative or reflective - A given construct may (possibly) be measured
formatively or reflectively - Model fit and relationships between constructs
are not necessarily fundamentally affected by the
modeling approach
7Formative model, R2 .30, n400
t 1.98
t 7.74
8Reflective specification, R2 .36
t 15.03
t 8.16
t 7.74
9Hybrid specification, R2 .30 (reduced R2
.25 indirect effect of KSI on ETA, t 7.90)
t 1.98
t 14.92
X1
.14
.71
X2
Y1
.14
.71
.71
.14
.71
KSI
ETA
X3
.71
.14
.71
Y2
.14
t 7.74
X4
.71
X5
Chi-Square0.00, df9, P-value1.00000,
RMSEA0.000
10Comparison
11Empirical comparison, MPJ 2005
Case A Exogenous and Endogenous Formative
Correlations 0.5
1
1
1
1
1
1
1
1
1
1
V1
V2
V4
V5
V3
V8
V9
V11
V12
V10
1
1
1
1
1
1
1
1
1
1
0.3
Construct A
Construct B
1
1
D1
1
1
1
1
D2
0.50
0.50
V6
V7
V13
V14
1
1
1
1
E7
E13
E14
E6
0.32
0.32
0.32
0.32
12Empirical comparison, MPJ 2005
Case A Standardized values
Correlations 0.5
1
1
1
1
1
1
1
1
1
1
V1
V2
V4
V5
V3
V8
V9
V11
V12
V10
.25
.25
.25
.25
.25
.24
.24
.24
.24
.24
.29
Construct A
Construct B
1
1
D1
.99
.99
.99
.99
D2
.03
.03
V6
V7
V13
V14
1
1
1
1
E7
E13
E14
E6
.02
.02
.02
.02
13Empirical comparison, MPJ 2005
Case B Exogenous Reflective
E1
E2
E3
E4
E5
1
1
1
1
1
1
1
1
1
1
V1
V2
V4
V5
V3
V8
V9
V11
V12
V10
1
1
?
Construct A
Construct B
1
1
D1
D2
V6
V7
V13
V14
1
1
1
1
E7
E13
E14
E6
14Empirical comparison, MPJ 2005
Case C Endogenous Reflective
E8
E9
E10
E11
E12
1
1
1
1
1
1
1
1
1
1
V1
V2
V4
V5
V3
V8
V9
V11
V12
V10
1
1
?
Construct A
Construct B
1
1
D1
D2
V6
V7
V13
V14
1
1
1
1
E7
E13
E14
E6
15Empirical comparison, MPJ 2005
Case D Exogenous and Endogenous Reflective
E1
E2
E3
E4
E5
E8
E9
E10
E11
E12
1
1
1
1
1
1
1
1
1
1
V1
V2
V4
V5
V3
V8
V9
V11
V12
V10
1
1
1
?
Construct A
Construct B
1
1
D1
D2
V6
V7
V13
V14
1
1
1
1
E7
E13
E14
E6
16Population data, n500
v1
v2
v4
v5
v3
v8
v9
v11
v12
v10
1.0 (t13.78)
1.0 (tna)
1.0 (t13.64)
1.0 (tna)
.30 (t13.65)
A
B
standardized .29
1.0 (t21.25)
1.0 (t21.03)
v6
v7
v13
v14
Chi-sq 0, df 66, RMSEA 0
17Population data, another scaling
v1
v2
v4
v5
v3
v8
v9
v11
v12
v10
.30 (t17.82)
1.0 (t21.03)
1.0 (tna)
A
B
standardized .29
3.33 (t31.35)
1.0 (tna)
1.0 (t113.02)
3.33 (t31.35)
v6
v7
v13
v14
Chi-sq 0, df 66, RMSEA 0
18Population data, final scaling
v1
v2
v4
v5
v3
v8
v9
v11
v12
v10
1.0 (t21.25)
1.0 (t21.03)
.30 (t31.35)
A
B
standardized .29
1.0 (t108.26)
1.0 (tna)
1.0 (tna)
1.0 (t113.02)
v6
v7
v13
v14
Chi-sq 0, df 66, RMSEA 0
19Comparison
20Conclusions from Analysis
- It is not necessarily a case of right/wrong, but
either/or - Do you focus on individual (formative) or common
(reflective) effects? - Evidence for one model does not necessarily
preclude the other - Choice of model may have minimal impact on
estimated structural relationships
21Theoretical Implications
- Formative and reflective models imply different
upfront measure development activities - These may result in different sets of measurement
indicators - formative variety of moderately-correlated
indicators to maximize variance explained while
limiting collinearity - reflective fewer highly-correlated indicators to
maximize reliability and interpretability
22Practical Implications
- It is becoming more common to recommend formative
treatment of reflective indicators - Our results suggest that for a given set of
indicators, choice of model may not be of major
importance to structural relationships
23Conclusion
- Our results support calls for the field as a
whole to think more carefully about measurement
model relationships and do a better job of making
sure that the measurement models used match that
conceptualization (Jarvis, MacKenzie and
Podsakoff 2003, p. 216)