Networks can facilitate the search for resources, knowledge, and information to foster innovation (Beuren & Dal Vesco, 2022), including the execution of fundamental activities by partners (Thomas et al., 2017). The absence of networks hinders access to critical opportunities and resources (Aldrich & Ruef, 2006). Diverse networks can provide information about markets, business locations, innovation alternatives, sources of capital, and potential investors (Aldrich & Ruef, 2006). The affective strength of networks increases trust and proximity and enables long-lasting relationships, which provide protection against opportunism and uncertainty (Aldrich & Ruef, 2006).

Networks require trust, rapid communication, rules, and self-regulation mechanisms for effective management (Mouritsen & Thrane, 2006). Cooperation can be frustrated by divergent interests among participants ( y suggest a lack of trust in the network and inhibit creativity, whereas more flexible controls foster information symmetry with partners (Chenhall et al., 2011). In uncertain environments, strict formal controls may fail, necessitating adaptable controls (Kherrazi, 2020).

Informal controls encompass norms, values, culture, and the internalization of goals to encourage desired behaviors and results (Das & Teng, 2001). They are flexible, data-rich, and minimally standardized (e.g., involving subordinates in budget planning). They aim to reduce divergences between individual preferences to promote cooperation and commitment among stakeholders (Eisenhardt, 1985). Informal controls include selection processes, training (Eisenhardt, 1985), and the socialization of technical skills, integrity, and values (Beuren & Dal Vesco, 2022).

Formal controls primarily focus on the short term by evaluating the achievement of results and monitoring business partners’ behavior more immediately to correct deviations and track performance (Stouthuysen et al., 2017). Informal controls provide stability and security by positioning the organization and focusing on the long term (Henri & Wouters, 2020). Informal controls help humanize and socialize experiences, strengthen relationships and loyalty, and enhance effectiveness (Stouthuysen et al., 2017). Chenhall et al. (2011) analyzed the effects of social networks and organic innovation culture on innovation and found an indirect and positive effect of social networks on innovation through organic innovation culture.

An appropriate control design (Chenhall, 2003) can contribute to the management of resources available in the network (Xu et al., 2022) and monitor the innovation process (Mouritsen & Thrane, 2006). Kherrazi (2020) investigated the role of a research and development ecosystem in the design of managerial controls and found that a strong ecosystem enhances managerial controls and that effective controls improve performance. However, the study did not confirm whether their combination yields better results.

Stouthuysen et al. (2017) examined the relationship between exploration, exploitation, and ambidextrous alliances and alliance performance, moderated by formal and informal controls. Their findings indicated stronger relationships between formal result controls and exploitation alliances, behavioral controls and exploration alliances, combined with ambidextrous alliances. Informal controls suggest an increase in the effectiveness of behavioral controls in exploration alliances, whereas the benefits of result controls diminished in exploitation alliances.

In the context of collaborative innovation, Guo et al. (2019) found that entry controls promote process innovation, while behavioral and output controls positively impact both process and product innovation. Pistoni et al. (2022) investigated the relationship between mechanisms of coordination and partnership control and innovation performance. They concluded that trust serves as a more effective informal control for innovation than direct contacts. Their study also highlighted that excessive rules can be ineffective and suggested that weak formal controls may prevent the inhibition of initiatives and creativity. Additionally, they emphasized that combining formal and informal controls leads to innovation.

Micheli et al. (2020) found that having a diverse network comprising companies of various sizes, along with changes in network ties, fosters BMI. Anwar and Ali Shah (2020) observed that financial, business, and political networks facilitate resource acquisition, generate new ideas and knowledge, provide access to distinct resources and knowledge, and impact BMI. Xu et al. (2022) investigated the effect of entrepreneurial networks on BMI in Chinese startups and found that the size and strength of networks have a positive impact on BMI.

Based on the above, the following hypothesis is proposed:

H1: Formal controls (H1a), informal controls (H1b), and the combination of formal and informal controls (H1c) positively influence the relationship between cooperation networks and BMI.

Figure 1 presents the theoretical model of the research and highlights the formulated hypotheses.

Figure 1 Source: Prepared by the authors. Note. The dotted arrows refer to the mediation hypotheses.

The theoretical model predicts a mediating effect of formal controls, informal controls, and the combination of formal and informal controls on the relationship between cooperation networks and BMI. Control variables were also included in the model.

The research population consists of managers of companies located in Brazilian technology parks listed on the Ministry of Science, Technology, and Innovation (Ministério da Ciência, Tecnologia e Inovações [MSTI]) platform and InovaDatabr. Technology parks integrate educational institutions, research centers, government bodies, and new businesses to foster cooperation (Pazetto & Beuren, 2022). In Brazil, there are 2,657 resident companies (MSTI, 2023) distributed across 103 technology parks, of which 43 are operational and 60 are in the implementation phase (Associação Nacional de Entidades Promotoras de Empreendimentos Inovadores [Anprotec], 2023). The parks in the implementation phase were included as they offer valuable resources, partnerships with incubators, and an environment for knowledge and experience exchanges.

After mapping the 2,657 companies on the MSTI InovaDatabr platform, the company profiles were verified on the social network LinkedIn. This process led to the removal of 171 profiles that were not companies, 43 duplicates, 758 not found, and 3 requiring an email connection, resulting in 1,683 companies. A letter introducing the research, explaining ethical procedures, confidentiality, and participant anonymity was sent to these companies. Those who agreed to participate received the survey link via the QuestionPro platform.

Data collection took place from January to March 2024, yielding 94 valid responses from managers of companies located in technology parks. The sample was determined using G*Power 3.1.9.7 software, following Faul et al.’s (2009) guidelines, with the model having three predictors, a mean effect size of 0.15, a significance level of 0.05 (α = 0.05), and a sample power of 0.8 (β = 0.8), resulting in a minimum sample size of 77 responses, which aligns with the valid responses obtained. The profiles of the respondents and the companies surveyed are consistent with the research objectives.

The respondents are predominantly male (85%), aged between 31 and 50 years (68%). Most held postgraduate degrees, with 33% having specialization, 17% a master’s degree, and 27% a doctorate. The dominant field of study was Engineering (37%), followed by Business Administration (18%), Biological Sciences (17%), and Information Systems (15%). The majority indicated they are partners, co-founders, or directors (86%) and have held their position for less than 6 years (61%). Most companies were relatively young: 27% were 4 years old or less, 26% were 5 to 8 years old, and 24% were 8 to 16 years old. Over half of the companies operate in the technology sector (53%), followed by medicine, health, and chemistry (19%). Most were small companies, with 75% employing fewer than 20 people. The majority have been located in the park for 6 years or less (79%).

The theoretical model consists of four constructs: cooperation networks, business model innovation, formal controls, and informal controls. The constructs and their corresponding items (see http://doi.org/10.17632/47gv4vwb4h.1) were validated in previous research, and the research instruments were translated from English into Portuguese. The construct ‘cooperation networks’ is composed of two instruments: cooperation networks between the company and other companies in the technology park, and cooperation networks between the companies and the park.

Cooperation networks between the company and other companies in the park were measured using three items adapted from Wu et al. (2020), originally from Collins and Clark (2003). It is a first-order latent variable (e.g., CNP1) that captures information about the volume, frequency, and proximity to other companies in the park. The seven-point scale was modified to five points, ranging from strongly disagree (1) to strongly agree (5), to standardize the scale of the instruments. Adaptations to the items were made in relation to the original context of business incubators.

Cooperation networks between companies and the park were measured with two items adapted from Bocquet et al. (2016), originally from Lynn et al. (1996). It is a first-order latent variable (e.g., NP1) that captures information about exposure to and use of cooperation network services provided by the park over the last three years. The four-point scale was changed to five points, ranging from strongly disagree (1) to strongly agree (5). The items were adapted due to the original context of network management organizations.

Business model innovation was adapted from the reflective-formative model by Spieth and Schneider (2016). It is a second-order variable composed of three first-order elements, capturing changes in the company’s business model over the last three years. The construct is composed of three variables (value proposition, value delivery, value capture) (e.g., VP1, VD1, VC1). A five-point scale was used, ranging from not at all (1) to completely (5). Items were adapted to reduce the length of statements.

Formal controls consist of two first-order unidimensional latent variables, and informal controls consist of one unidimensional latent variable, all adapted from Stouthuysen et al. (2017). Formal controls consist of six items capturing results and behavior controls, grouped into a second-order reflective variable. Informal controls consist of five items related to culture, values, and information sharing (e.g., RC1, BC1, and IC1). The seven-point scale was adjusted to five points, ranging from strongly disagree (1) to strongly agree (5). Items were adapted due to the context.

Control variables were included in the model, specifically: cooperation networks of the technology park with other institutions outside the park, the number of cooperation networks, and the intensity of the cooperation networks.

Cooperation networks between the technology park and other institutions outside the park were measured with two items adapted from Wu et al. (2020), originally from Collins and Clark (2003) It is a first-order latent variable (e.g., PEN1) that captured information about the variety and closeness of relationships with other institutions in the park’s networks. The seven-point scale was changed to five points, ranging from strongly disagree (1) to strongly agree (5). Items were adapted due to the context of business incubators.

The number of cooperation networks was measured with items adapted from Wu et al. (2020). This instrument captured three individual items: the number of networks between companies in the park, with companies outside the park, and with other institutions. The measurement of each item was done through its natural logarithm. The number of networks is relevant because the more networks available, the more resources are accessible, and innovation performance depends on the abundance of external resources (Dongling et al., 2022).

The intensity of cooperation networks assumes that greater intimacy, trust, reciprocity, and emotional engagement lead to a higher transfer of resources between partners (Dongling et al., 2022). This variable was measured with items adapted from Wu et al. (2020) on a five-point Likert scale, ranging from low intensity (1) to high intensity (5).

To test the hypotheses, partial least squares structural equation modeling (PLS-SEM) was used, as it allows the analysis of relationships between dependent and independent variables, with intervening variables, using a causal-predictive approach and seeking to explain the variance in the dependent variables (Hair et al., 2022). Path coefficients and other references are estimated to maximize the explained variance of dependent constructs in various partial regressions (Hair et al., 2022).

In mediation analysis, both the absence of mediation (or no mediation) and the presence of mediation were considered (Hair et al., 2022). In the absence of mediation, only direct non-mediation was evaluated, where the direct effect is significant and the indirect effect is not, as well as non-mediation without effect, where both direct and indirect effects are not significant (Hair et al., 2022). The presence of the mediator effect was evaluated by mediation: complementary (significant direct and indirect effects in the same direction), competitive (significant direct and indirect effects in opposite directions), and indirect (only the indirect effect is significant) (Hair et al., 2022). The combined mediator effect of formal and informal controls was analyzed based on the total indirect effects from SmartPLS, which represents the sum of the indirect effects of each mediator (Hair et al., 2022).

Biases due to common method, non-response, and early and late respondents were also assessed. To mitigate common method bias, the guidelines of Podsakoff et al. (2003) were followed. Respondents were asked to answer the items honestly and were informed that there were no right or wrong answers and that their anonymity was guaranteed. The Harman single-factor test was applied to assess whether one factor explained more than 50% of the variance in the data (Podsakoff et al., 2003). The result from the SPSS software, version 25.0.0.2, showed that the first factor explained 25.82% of the total variance, suggesting the absence of this bias.

The non-response bias arises because only a portion of the invited participants responded to the questionnaire. Assuming that non-respondents resemble late respondents, the first-last method was applied, which involves a t-test for independent samples comparing the early and late respondents. If no significant differences are found at the 5% level, the non-response bias is considered non-representative (Pazetto & Beuren, 2022). Comparing the means of the first 10 and last 10 respondents revealed no significant differences in the constructs (p-value between 0.128 and 1).

The evaluation of the measurement model shows that the convergent validity, after the exclusion of items (BC5 and VD4) with factor loadings below 0.708, resulted in an increase in the average variance extracted (AVE). The reliability and AVE of the constructs were above 0.5. The internal consistency reliability, evaluated by Cronbach’s alpha (α) and composite reliability (CR), was above 0.7. Although the construct ‘value capture’ had α lower than 0.7 (0.612), it was kept in the model as it captures essential elements of the BMI construct, and the other indicators met the assumptions. The discriminant validity of the constructs had HTMT lower than 0.85. The evaluation of the AVE and rho_c of the reflective-reflective (second-order) measurement models were run separately (Hair et al., 2022), all of which were above the recommended thresholds of 0.5 and 0.7, respectively. Table 1 presents the results.

Table 1

Variables	Mean	Standard deviation	Factor loading	Reliability and convergent validity
				α	rho_a	rho_c	AVE
CN	3.229	1.369		0.730	0.793	0.878	0.782
CNP	3.128	1.352		0.834	0.839	0.901	0.752
CNP1	3.404	1.315	0.855
CNP2	2.979	1.271	0.922
CNP3	3.000	1.414	0.822
NP	3.383	1.384		0.818	0.820	0.916	0.846
NP1	3.436	1.426	0.925
NP2	3.330	1.332	0.915
FC	3.054	1.313		0.745	0.889	0.886	0.795
BC	3.333	1.209		0.789	0.793	0.877	0.705
BC1	3.638	1.184	0.797
BC2	3.298	1.192	0.900
BC3	3.064	1.174	0.819
RC	2.777	1.356		0.885	0.889	0.929	0.813
RC 1	2.968	1.356	0.880
RC 2	2.617	1.354	0.878
RC 3	2.745	1.328	0.946
IC	3.391	1.366		0.796	0.817	0.864	0.614
IC1	3.447	1.217	0.665
IC2	3.436	1.242	0.758
IC3	3.691	0.990	0.828
IC4	2.989	1.171	0.754
IC5	2.553	1.234	0.607
BMI
VP	3.146	1.236		0.771	0.778	0.867	0.685
VP1	2.862	1.268	0.831
VP2	3.181	1.194	0.860
VP3	3.394	1.178	0.793
VD	3.213	1.162		1.743	0.758	0.853	0.660
VD1	3.160	1.170	0.727
VD2	3.436	1.135	0.797
VD3	3.043	1.138	0.807
VD4	2.777	1.281	0.675
VC	2.989	1.254		0.612	0.633	0.836	0.719
VC1	2.872	1.265	0.838
VC2	3.106	1.224	0.860

Note. CN = cooperation networks; CNP = cooperation networks with companies located in the park; NP = cooperation networks with the park; FC = formal controls; BC = behavioral controls; RC = result controls; IC = informal controls; BMI = business models innovation; VP= value proposition; VD = value delivery; VC= value capture; AVE = average variance extracted; α = Cronbach’s alpha; rho_c = composite reliability.Source: Prepared by the authors.

The measurement model meets the parameters of discriminant validity for all constructs. The square root of the AVE is greater than the correlations shown below the diagonal, and the HTMT above the diagonal is less than 0.85.

The analysis of redundancy of the reflective-formative higher-order construct resulted in a global evaluation above 0.7 for the three items, with loadings between 0.638 and 0.898, for a 95% confidence interval. The VIF demonstrated a result below 3. The weight of the variables showed that value capture was significant at the 5% level and had the highest weight. The loadings of the constructs showed values higher than 0.5 and significance at the 5% level, which allows keeping all the constructs.

Robustness tests complement the analysis, including tests for linearity, endogeneity, full collinearity test, and additional sensitivity analysis to the size of the companies (Hair et al., 2022; Kock, 2015). The linearity test indicates that the relationships are linear due to non-significance at the 5% level and low f² power. Endogeneity is not an issue for the model, as the results of the Gaussian copula test showed no significance at the 5% level. The full collinearity test confirmed the absence of common variance by showing VIF below the suggested threshold for all variables (VIF < 3.3). Finally, the sensitivity to company size was not significant at the 5% level for all variables.

Table 3 presents the results of the structural model, highlighting the hypothesis tests, effects of the control variables, and quality criteria of the model.

Table 2

Panel A: Main and higher-order latent constructs
		1	2	3	4	5	6
CN		0.885	0.431	0.531	0.126	0.543	0.794
FC		0.338	0.888	0.698	0.176	0.451	0.181
IC		0.426	0.505	0.784	0.177	0.484	0.294
NCN		0.077	0.179	0.116	1.000	0.360	0.097
ICN		0.369	0.317	0.339	0.298	0.857	0.466
PEN		0.610	0.138	0.254	-0.087	0.332	0.931
Panel B: Lower-order latent constructs
	1	2	3	4	5	6	7
CNP	0.867	0.694	0.374	0.432	0.219	0.326	0.272
NP	0.575	0.920	0.213	0.193	0.083	0.139	0.099
BC	0.303	0.168	0.840	0.711	0.265	0.208	0.263
RC	0.370	0.166	0.594	0.902	0.279	0.407	0.547
VP	0.177	0.032	0.196	0.230	0.828	0.698	0.644
VD	0.254	-0.036	0.077	0.333	0.528	0.813	0.757
VC	0.188	-0.028	0.137	0.399	0.443	0.523	0.849

Note. CN = cooperation networks; CNP = cooperation networks with companies residing in the park; NP = cooperation networks with the park; FC = formal controls; BC = behavioral controls; RC = result controls; IC = informal controls; BMI = business model innovation; VP = value proposition; VD = value delivery; VC = value capture; NCN = number of cooperation networks; ICN = intensity of cooperation in the networks; PEN = park external networks; AVE = average variance extracted; HTMT = heterotrait-monotrait. The values in bold indicate the square root of the AVE; the lower diagonal shows the correlations; the upper diagonal shows the HTMT values.Source: Prepared by the authors.

Table 3

Hypotheses	Relationships	β	P-value	CI [5%; 95%]	Decision
Panel A: Direct relations
	CN → BMI	-0.261	0.115	[-0.52; 0.014]
	CN → FC	0.338	0.001***	[0.170; 0.501]
	CN → IC	0.426	0.000***	[0.313; 0.569]
	FC → BMI	0.317	0.033**	[0.062; 0.547]
	IC → BMI	0.029	0.854	[-0.21; 0.308]
Control variables
	NCN → BMI	0.167	0.158	[-0.035; 0.354]
	ICN → BMI	0.062	0.662	[-0.164; 0.300]
	PEN → BMI	0.397	0.004***	[0.164; 0.597]
Endogenous variable		R²	Q²	VIF
FC		0.114	0.086	1.545
IC		0.182	0.145	2.255
BMI		0.247	0.033	1.621
Panel B - Indirect relations
Hypotheses	Relations	β	p-value	CI [5%; 95%]
H1a	CN → FC → BMI	0.107	0.087*	[0.014; 0.219]	Accept
H1b	CN → IC → BMI	0.013	0.863	[-0.096; 0.138]	Reject
H1c	CN → FC+IC → BMI	0.12	0.088*	[0.015; 0.242]	Accept

Note 1. CN = cooperation networks; BMI = business model innovation; FC = formal controls; IC = informal controls; NCN = number of cooperation networks; ICN = intensity of cooperation in the networks; PEN = park external networks; CI = confidence interval; R² = coefficient of determination; Q² = Stone-Geisser indicator. * p < 0.10; ** p < 0.05; *** p < 0.01. CI calculated by accelerated bootstrap and bias-corrected, with 5,000 re-samples and a two-tailed test. Source: Prepared by the authors.

Panel A presents the results of the direct relationships, despite the absence of hypotheses. Panel B presents the results for the indirect relationships, where H1a and H1c were supported, contrasting with H1b.

The quality criteria of the model demonstrate an acceptable VIF adjustment for the latent variable (less than 3). R² is below 0.25, indicating that significant relationships have weak adjustments but are aligned with previous research (Stouthuysen et al., 2017). The evaluation of Q² presents positive values, demonstrating the model’s accuracy.

The discussion of the results is guided by the research model, based on the hypotheses. Figure 2 presents the hypotheses that were not rejected.

Figure 2 Source: Significant relationships at the level of * p < 0.10; ** p < 0.05; *** p < 0.01. The dashed arrows correspond to mediation hypotheses. Source: Prepared by the authors.

Prior theoretical support is required to explore the effect of variables (Memon et al., 2018). The absence of hypotheses for direct relationships stems from the lack of literature to support them. Chenhall et al. (2011) did not find a direct relationship between networks and innovation.Chan and Lau (2005) inferred, in research on startups in a Chinese technology park, that these do not form a cooperative network due to fear of losing information. Faccin et al. (2020) observed in their research the need for the orchestration of the network, even its reconfiguration, for specific collaborative practices.

According to Sant’Anna et al. (2019), despite the increase in the number of parks in Brazil and the willingness to collaborate, there are still controversies between perceptions and collaborative practices. Schmidt and Balestrin (2015) emphasize that these environments affect collaborative innovation but not with infrastructure services. For Clauss (2017), because interactions with external suppliers/partners and customers are the main types of networks that impact BMI, companies seem to interact more with this environment to capture market needs. Furthermore, internal cooperation networks may not provide the necessary resources for BMI. Despite previous studies recognizing the importance of network formation to drive BMI (Anwar & Ali Shah, 2020), other variables may indirectly influence this relationship.

The mediating effect of formal controls in the relationship between cooperation networks and BMI was confirmed, contrasting with informal controls, which were significant only when combined with formal controls. Although prolonged contacts or close relationships may not access resources for BMI, companies use controls in their cooperation networks. Using controls in managing networks influences behaviors, directs efforts, restricts opportunism, minimizes failure risks, and brings trust to partners (Das & Teng, 1998, 2001). Formal controls demonstrated support for managing networks in companies located in technology parks. According to the findings of Pistoni et al. (2022), there is greater strength of formal controls compared to informal controls in managing collaborative networks. This suggests a greater emphasis on short-term maintenance brought by formal controls for sustaining a partnership, as opposed to the long-term view provided by informal controls, which reinforce trust in the network (Chenhall et al., 2011).

Another important finding is the positive relationship between formal controls and BMI. This finding aligns with the results of Ganguly and Das (2020) and Guo et al. (2019). Although formal controls may restrict innovation in achieving goals, they enhance creativity by imposing limits on the use of company resources and allow the identification of opportunities within the company’s domains (Ganguly & Das, 2020). Formal controls assist in the innovation planning process and result evaluation through indicators (Guo et al., 2019).

H1a has support for acceptance. It aligns with Beuren and Dal Vesco (2022), who state that a performance evaluation system applied to a network ensures the fair distribution of information among participants and enables knowledge sharing; with Chenhall et al. (2011), who argue that formal controls stand out in organizations more prone to learning, leading to innovation; with Pistoni et al. (2022), who claim that formal contracts promote relational governance and stability in interorganizational exchanges; and with Stouthuysen et al. (2017), who suggest that the synergistic configuration of results and behavior controls generates dynamic tensions between innovation and goals, which helps balance strategies.

H1b does not present statistical significance, leading to its rejection. This aligns with the research of Pistoni et al. (2022), who found that direct and personal contacts, status-based rewards, and trust-based coordination together do not have an effect on innovation. According to Pistoni et al. (2022), trust-based coordination is an informal mechanism that supports the relationship between collaborative partnerships and innovation. Additionally, personal contacts and rewards are less effective, which encourages future research to evaluate the issue of trust.

H1c has support for acceptance. This result contrasts with Chenhall et al. (2011), who state that the combination of formal controls with an organic innovative culture (informal controls) does not add benefits to innovation. However, it converges with Pistoni et al. (2022), who argue that companies can combine formal control mechanisms with information sharing. This combination promotes information symmetry between the parties and leads to innovation as an alternative to informal controls (Pistoni et al., 2022). This suggests that controls have complementary roles in BMI.

The analysis of the results of the structural model allows some assumptions. First, the non-significance of H1b suggests that social interactions between the research participants may not occur in the search for resources for BMI. Second, the significant relationship of H1a and the non-significant relationship of H1b suggest that formal controls are used more to control the flow of information and govern cooperation networks than informal controls. Third, the non-significance of H1b may indicate a lack of trust among the participants of the cooperation network and, consequently, reluctance to provide information. Fourth, even though H1c is confirmed, the inclusion of informal controls raised statistical significance by only 0.01. This suggests that informal controls are not crucial and that formal controls are the determinants of the model.

The control variables, number of cooperation networks, and intensity of cooperation in the networks, associated with BMI, did not present statistical significance. This suggests that interacting more or less and having more or fewer networks may not result in obtaining resources. This analysis seems to complement the examination of the hypotheses, as it indicates that cooperation for BMI may not occur between the companies located in the park or with the technology park itself.

Companies located in technology parks may not use the resources from cooperation networks for BMI, contrasting with the RBV, perhaps due to the fear of resource exchange among the companies located in parks. The effect of external park networks on BMI was significant. This suggests that companies located in parks make greater use of the resources provided by external park networks than by the resources of internal networks. This result aligns with Bocquet et al. (2016), who found that external resources are more important than internal ones in fostering innovation. It is argued that external park connections provide more resources for companies located in parks.