Data Analysis and Findings
Evaluation of the Measurement Model
Table
1 presents the descriptive statistics for the latent constructs. The
results of the measurement model reveal that the model meets all of the
minimum requirements. Firstly, all of the first- and second
order-constructs with reflective items suggest a good indicator
reliability with few indicator loadings below 0.70 (Table 2). However,
since the criteria for reliability and convergent validity were met, we
decided to retain all original items, as suggested by Hair et al.
Secondly, Cronbach's alpha and composite reliability values range from
0.70 to 0.91, thus meeting the commonly accepted threshold level.
Thirdly, all of the average variance extracted (AVE) values of the
first- and second-order constructs are above 0.50, supporting the
construct measures' convergent validity.
Table 1. Descriptive statistics, correlations for the study constructs.
|
M |
S.D |
ASC |
ESC |
RHSC |
RSC |
SCU |
DU |
SU |
SCS |
FLEX |
SPD |
OP |
ASC |
3.26 |
0.75 |
0.85 |
|
|
|
|
|
|
|
|
|
|
ESC |
3.91 |
0.70 |
0.32 ** |
0.79 |
|
|
|
|
|
|
|
|
|
RHSC |
3.25 |
0.79 |
0.27 ** |
0.31 ** |
0.84 |
|
|
|
|
|
|
|
|
RSC |
3.76 |
0.62 |
0.60 ** |
0.11 ** |
0.52 ** |
0.71 |
|
|
|
|
|
|
|
SCS |
3.14 |
0.51 |
0.50 ** |
0.50 ** |
0.27 ** |
0.47 ** |
0.60 |
|
|
|
|
|
|
DU |
2.81 |
0.74 |
0.12 |
0.19 * |
0.13 |
0.18 * |
0.18 * |
0.81 |
|
|
|
|
|
SU |
3.47 |
0.65 |
0.10 |
0.32 ** |
0.16 * |
0.23 ** |
0.23 ** |
0.077 |
0.84 |
|
|
|
|
SCU |
3.54 |
0.59 |
0.15 |
0.36 ** |
0.20 * |
0.28 ** |
0.28 ** |
0.600 ** |
0.50 ** |
0.66 |
|
|
|
FLEX |
3.61 |
0.59 |
0.23 ** |
0.29 ** |
0.23 ** |
0.31 ** |
0.31 ** |
0.327 ** |
0.19 * |
0.34 ** |
0.77 |
|
|
SPD |
3.89 |
0.53 |
0.17 * |
0.17 * |
0.21 ** |
0.14 |
0.21 ** |
0.112 |
0.19 * |
0.18 * |
0.46 ** |
0.78 |
|
OP |
3.75 |
0.48 |
0.24 ** |
0.28 ** |
0.26 ** |
0.27 ** |
0.31 ** |
0.26 ** |
0.22 ** |
0.31 ** |
0.57 ** |
0.64 ** |
0.66 |
M
= Mean, S.D = Standard deviation. ASC = Agile supply chain. ESC =
Efficient supply chain. RHSC = Risk-Hedging supply chain. RSC =
Responsive supply chain. SCS = Supply chain strategy. DU = Demand
uncertainty. SU = Agile supply chain. SCU = Supply chain uncertainty.
FLEX = Flexibility. SPD = Speed. OP = Operational performance. Diagonal
and italicized bold elements are the square roots of the AVE (average
variance extracted). Off-diagonal elements are the correlation between
constructs' variables. * |t| ≥ 1.65 at p = 0.10 level. ** |t| ≥ 1.96 at p = 0.05 level.
Table 2. Measurement Model.
Construct |
Code |
Items |
FOL |
SOL |
Measurement model at first-order factors (First-stage) |
ASC (reflective)
|
C.R = 0.89; α = 0.83; AVE = 0.67; VIF = 2.99 |
SSA1 |
Our supply chain always faces the volatile customer demand. |
0.81 |
0.74 |
SSA2 |
Our supply chain needs to maintain a higher capacity buffer in response to the volatile market. |
0.85 |
0.75 |
SSA3 |
Our supply chain provides the customer with personalized products. |
0.81 |
0.69 |
SSA4 |
Our supply chain structure often changes to cope with a volatile market. |
0.80 |
0.76 |
ESC (reflective)
|
|
|
C.R = 0.84; α = 0.71; AVE = 0.64; VIF = 1.12 |
SSE1 |
Our supply chain supplies predictable products |
0.83 |
0.57 |
SSE2 |
Our supply chain reduces any waste as much as possible |
0.83 |
0.64 |
SSE3 |
Our supply chain reduces costs through mass production. |
0.72 |
0.47 |
RHSC (reflective)
|
|
|
C.R = 0.88; α = 0.80; AVE = 0.71; VIF = 1.78 |
SSH1 |
Our supply chain partners are ready to share resources whenever necessary |
0.89 |
0.75 |
SSH2 |
Our supply chain reduces costs through sharing capacities/resources |
0.86 |
0.69 |
SSH3 |
Our supply chain partners are always ready to support and cooperate |
0.78 |
0.76 |
RSC (reflective)
|
|
|
C.R = 0.80; α = 0.70; AVE = 0.50; VIF = 1.81 |
SSR1 |
Our supply offers wider product range |
0.76 |
0.57 |
SSR2 |
Our supply chain offers new products more frequently |
0.77 |
0.64 |
SSR3 |
Our supply chain offer more innovative products |
0.64 |
0.47 |
SSR4 |
Our supply chain provides fast deliveries |
0.66 |
0.74 |
DU
|
|
|
C.R = 0.88; α = 0.81; AVE = 0.65; VIF = 1.56 |
SUD1 |
Our master production schedule has a high degree of variation in demand over time. |
0.86 |
0.54 |
SUD2 |
Our demand fluctuates drastically from week to week. |
0.86 |
0.49 |
SUD3 |
Our requirements for raw materials supply vary drastically from week to week. |
0.84 |
0.51 |
SUD4 |
Customer requirements/services for products change dramatically. |
0.63 |
0.42 |
SU
|
|
|
C.R = 0.91; α = 0.87; AVE = 0.71; VIF = 2.84 |
SUS1 |
Our suppliers always provide us a correct lead time estimation |
0.82 |
0.62 |
SUS2 |
Our suppliers consistently meet our delivery requirements |
0.87 |
0.73 |
SUS3 |
Our suppliers provide us the input with consistent quality |
0.88 |
0.77 |
SUS4 |
Our suppliers consistently meet specified volume requirements |
0.81 |
0.65 |
FLEX (reflective)
|
|
|
C.R = 0.86; α = 0.78; AVE = 0.60; VIF = 3.14 |
FPF1 |
Ability to customize products/services |
0.79 |
0.60 |
FPF2 |
Ability to respond to changes in delivery requirement |
0.74 |
0.50 |
FPF3 |
Ability to adjust production volumes |
0.81 |
0.73 |
FPF4 |
Ability to produce a range of products/services |
0.75 |
0.72 |
SPD (reflective)
|
|
|
C.R = 0.86; α = 0.79; AVE = 0.61; VIF = 1.28 |
FPS1 |
On time delivery |
0.80 |
0.71 |
FPS2 |
Delivery Dependability |
0.81 |
0.68 |
FPS3 |
Delivery Speed |
0.70 |
0.60 |
FPS4 |
Time to Market |
0.82 |
0.73 |
Measurement model at second-order factors (Second-stage) * |
SCS |
|
Supply Chain Strategy |
|
|
C.R = 0.82; α = 0.74; AVE = 0.92 |
ASC |
Agile supply chain |
0.80 |
|
ESC |
Efficient supply chain |
0.85 |
|
RHSC |
Risk Hedging supply chain |
0.77 |
|
RSC |
Responsive supply chain |
0.95 |
|
SCU |
|
Supply Chain Uncertainty |
|
|
C.R = 0.92; α = 0.90; AVE = 0.87 |
DU |
Demand uncertainty |
0.75 |
|
SU |
Supply uncertainty |
0.76 |
|
OP |
|
Operational performance |
|
|
C.R = 0.86; α = 0.82; AVE = 0.89 |
FLEX |
Flexibility |
0.64 |
|
SPD |
Speed |
0.96 |
|
Note:
All loadings and weights are significant at 0.001 level (2-tailed). *
Second order construct, two-stage approach. FOL = First-order loadings.
SOL = Second-order loadings. C.R = Composite reliability. α = Cronbachs
Alpha. AVE = Average variance extracted. VIF = Variance inflation
factor.
To
examine the discriminant validity, this study uses two approaches:
First, the results of cross loading show that all of the items load is
higher on their respective constructs than on the other constructs, and
that the differences between loadings and cross loading are much higher
than the suggested threshold of 0.1. Second, Table 1 shows that the
square root of AVE is greater than the corresponding construct
correlation on the diagonal.
Finally, the constructs in this
study are operationalized in a reflective-reflective type 1 model based
on theoretical considerations. This study applies a two-stage
approach to evaluate the hierarchical second-order latent constructs. In
the first stage, the repeated indicator approach is used to obtain the
latent variable scores for all the first-order constructs, which, in the
second stage, serve as manifest variables in the measurement model of
second-order constructs.
Evaluation of the Structural Model
First,
this study examines the model for collinearity. The results show minimal
collinearity with the variance inflation factor (VIF) for two sets of
(predictor) constructs because they are much less than the common
threshold of 5. Therefore, collinearity among the predictor's
constructs in the structural model is not a problem. Second, the model
predictability is assessed using R2 values for the dependent latent
variables. The R2 values of SCS (0.16) and OP (0.24) are in line with
prior research, supporting the PLS-SEM model's in-sample
explanatory power.
Third, the sizes and significance of the path
coefficients that represent the derived hypotheses were examined. To
obtain the significance levels, the bootstrapping procedure (with some
5000 bootstrap samples and 146 bootstrap cases, using no sign changes)
was run. As shown in Table 3, the results indicate that SCU has a
significant effect on SCS (β = 0.30 ***; t = 3.91; CI0.90: (0.22, 0.27)
and OP (β = 0.27 **; t = 3.02; CI0.90: (0.26, 0.31). In addition, SCS
has a significant effect on OP (β = 0.25 **; t = 3.20; CI0.90: (0.22,
0.27). These results confirmed that all three hypotheses were accepted.
Table 3. Mediation Analysis Results.
Mediation Model |
Standardized Coefficient |
t-Value |
95% CI |
VAF (%) |
f2 † |
q2 † |
Conclusion |
Direct effect |
SCU on OP |
0.27 ** |
3.02 |
(0.37, 0.43) |
|
0.17 |
|
|
SCU on SCS |
0.30 *** |
3.91 |
(0.38, 0.44) |
|
0.21 |
0.04 |
H1 supported |
SCS on OP |
0.25 ** |
3.20 |
(0.15, 0.19) |
|
0.17 |
0.04 |
H2 supported |
Indirect effect |
SCU on OP (mediated by SCS) |
0.07 *** |
13.31 |
(0.06, 0.08) |
22.15 |
|
|
H3 supported |
Note: CI = Confidence interval; VAF (%) = Variance represented percentage. † = Effect size. The values of f2 and q2; 0.02, 0.15, 0.35 for weak, moderate, strong effects. * |t| ≥ 1.96 at p = 0.05 level; ** |t| ≥ 2.58 at p = 0.01 level; *** |t| ≥ 3.29 at p = 0.001 level.
Fourth,
the values of the effect size f2 and q2 of 0.02, 0.15, and 0.35 are
regarded as small, medium, or large, respectively. The results of
the f2 and q2 effect sizes concerning all of the relations in the model
are provided in Table 3.
Finally, a blindfolding procedure was run to
evaluate the model's predictive relevance. All of the Q2 values are
considerably above zero, thus providing support for the model's
predictive relevance regarding out-of-sample prediction as shown in
Figure 1.
Mediation Analysis
This study follows Zhao et
al.'s recommendations to examine mediation effects in the model.
Zhao et al. posit the key condition in showing mediation is that the
indirect effect is significant. The bootstrapping procedure
facilitates the exploration of the SCS (mediator) simultaneously in the
association between SCU (independent variable) and OP (dependent
variable). Based on Zhao et al., this study applies the recommended
5000 bootstrap samples at the 90% confidence level. The structural
equation model is examined to determine whether SCS mediates the effect
of SCU on OP. The results show the existence of a significant indirect
effect of SCU on OP with the SCS mediator (β = 0.07 ****; t = 13.31;
CI0.90: (0.06, 0.07), thus suggesting that SCS mediates the association
between SCU and OP. However, the direct effect of SCU on OP shows
partial mediation. The variance was (VAF) 22.15%; that is, there was a
partial mediation or complementary mediation. Further, the
meditation testing, using the Baron and Kenny's procedure, produced
highly similar results to those attained using the Zhao et al.'s method, thus suggesting SCS mediates the association between SCU and OP.
fsQCA Results
fsQCA
analysis requires a calibration of the conventional variables measured
using Likert scales. First, this study measures each latent variable by
calculating the average of the values of their items. This process
results in six conditions and two variables reflecting the outcome.
Second, the eight variables are then calibrated into fuzzy sets using
the direct calibration method. Accordingly, the decision was made
for the threshold of full membership in the sets (fuzzy score = 0.95).
The threshold for full non-membership (fuzzy score = 0.05) and the
cross-over point, which indicates a point of maximum ambiguity where
respondents are not in or out of the sets (fuzzy-score = 0.50) (pp.
104–105). In line with extant studies, such as Fiss, this study
employs the 75th percentile of each variable as an anchor for the full
membership. For the full non-membership, the 25th percentile and the
50th percentile for the crossover point were used. This study does not
use absolute anchors for two reasons. Firstly, it is the first
exploration of a research incorporating SCU and an SCS in a single
empirical frame. Secondly, concept measurements are based on subjective
scales. However, due to the asymmetric nature of the distribution of
responses associated with variable supply uncertainty, the percentile
for the calibration process was not used; instead, the six-value fuzzy
set was used. Because the outcome is originally measured by the
flexibility and the speed of the supply chain, a so-called
macro-variable was created by combining them. Ragin qualifies the
macro-variable as a 'higher-order construct', which benefits from using
the maximum of the values of the measures of which it is composed.
Table
4
exhibits the results of the fsQCA analysis of high OP. The notation
in Table 4 is based on the extant literature.
Table 4. Solutions table indicating the configurations needed to achieve high OP.
Table
4 shows the existence of five causal paths leading to high OP. They
represent various combinations of SCS and SCU. Each part of the overall
solution has a consistency greater than 0.80, which can be considered
acceptable. These five solutions demonstrate the existence of a
first-order equifinality (e.g., Fiss) in the analysis of the OP of the
supply chain. In Solution 1 (1a and 1b), the SCS consists of a
conjunction of lack of ASC, efficiency, responsiveness, and a lack of
risk hedging, while the SCU is only based on a demand uncertainty. In
Solution 1a, efficiency and a lack of risk hedging constitute the core
conditions regarding the SCS. These conditions have a strong causal
relation with high OP (the outcome). Conversely, responsiveness and lack
of agility are peripheral or contributing conditions, which imply that
they have a low causal relation with the high OP. Regarding SCU; demand
uncertainty maintains a strong causal relationship with the high OP of
the supply chain, regardless of the presence or absence of supply
uncertainty. Solutions 1a and 1b provide evidence of a second-order
equifinality (e.g., Fiss). Indeed, comparing Solutions 1a and 1b
indicates that a high demand uncertainty and a high uncertainty in
supply can be treated as substitutes. These results also demonstrate
that companies can achieve a high OP even though they do not hedge
against risks and do not encounter agility in their supply chain.
Solution 2 indicates that certain companies may engage an SCS to a
lesser extent, although they face an uncertainty context marked by high
demand uncertainty and low supply uncertainty. In contrast, Solution 3
emphasizes that companies may deploy other massive supply chain
strategies in the absence of high uncertainty in their supply chain.
Solutions 4 and 5 suggest that companies encountering a high level of
uncertainty may implement differentiated supply chain strategies.
To
additionally examine the causal asymmetry assumption underlying fsQCA, a
new outcome variable was created that represents the absence of the
high OP of the supply chain. The new outcome is computed as the negation
of the high OP of the supply chain examined in Table 4, and
investigated with the same six conditions reflecting the SCS and
uncertainty. The analysis was not conclusive in emphasizing the manner
in which the asymmetrical nature of the causal relations led to either a
high or low OP. Indeed, we obtain a solution only when lowering the
cut-off of the truth table algorithm from "0.8" down to "0.75".
Testing for Predictive Validity
Previous
PLS-SEM and fsQCA analyses demonstrate how well the model investigated
fits the data. They are not indicative of how independent variables
(conditions) predict dependent variables (result). Thus, a predictive
analysis is performed in this study. This study follows the
procedure in Cepeda-Carrión et al. to report the predictive validity of
the PLS-SEM model as follows: first, two-thirds of the sample is
randomly chosen, which is composed of 96 samples as the training set;
the remaining 50 samples represent the holdout sample. The training set
is used to estimate the parameters in the model. Using the holdout
sample, each indicator is standardized, and the construct scores are
formed as linear combinations of the respective indicators using the
weights obtained from the training sample. The construct scores are
standardized. For the OP of the endogenous construct in the holdout
sample, the predictive scores are created by using the path coefficients
obtained from the training sample. The correlation between the
predictive scores and construct scores is 0.63 (p < 0.01), which
suggests that the PLS-SEM model has acceptable predictive validity.
Furthermore, the predictive validity of fsQCA analysis is made following
recent empirical studies. Table 5 highlights that the
patterns of the complex combination of conditions are causally
consistent indicators of a high level of supply chain OP. Furthermore,
Figure 2 shows that the first part of the solution of the modeling
subsample is causally relevant in predicting the high OP of the supply
chain with a consistency higher than "0.80" (0.91). Predictive tests for
the remaining four parts of the solution of the modeling subsample show
the high consistency of the outcome under investigation. All of the
results will be provided at the request of any interested readers.
Figure 2. Test of Part 1 of the solution from the modeling subsample using data from the holdout subsample.
Table 5. Complex configurations of SCU and strategy dimensions, which indicate a high level of OP for the modeling subsample.
Solution from the Modeling Subsample |
Raw Coverage |
Unique Coverage |
Consistency |
1. ~su*du*~esc*~asc*~rsc*~rhsc |
0.12 |
0.05 |
0.84 |
2. su*~du*esc*~asc*rsc*~rhsc |
0.14 |
0.08 |
0.95 |
3. ~su*du*esc*~asc*rsc*~rhsc |
0.07 |
0.01 |
0.98 |
4. su*du*~esc*asc*~rsc*rhsc |
0.15 |
0.08 |
0.89 |
5. su*du*esc*asc*rsc*rhsc |
0.19 |
0.12 |
0.87 |
Overall solution consistency: |
0.87 |
Overall solution coverage: |
0.47 |
Note:
SU: Supply uncertainty, DU: Demand uncertainty, ASC: Agile supply
chain, ESC: Efficient supply chain, RSC: Responsive supply chain, RHSC:
Risk-hedging supply chain, and "~": the negation of the condition.