Effects of Structured Support from Leader on Employee Satisfaction, Intention to Quit and Store Performance: Experimental Evidence from Grocery Stores

All participating stores receive access to an online application providing several leadership-training topics. The app lists two employees that leaders should work with while using the app. The treatment version of the app includes two additional features to facilitate structured support from leaders: First, it includes a tool to encourage the leaders to set up weekly development meetings with the two employees. This tool helps leaders to organize meeting notes, set goals, and prompts the leader to schedule follow-up meetings. Second, to strengthen leaders’ beliefs in development opportunities, it includes a module designed to help leaders develop a growth mindset for their own abilities as a leader and the growth potential of their employees.

2023-09-11

2023-10-29

Update after pilot but prior to implementation: removed engagement as pilot showed little variation in this variable, updated dates


Primary outcomes :
Employee level outcomes:
Short term:
intention to quit, work satisfaction

Long term:
employee turnover

Store-level outcomes:
Index on store performance both Short-term (October-November) and Long-term (December-April)

Primary outcomes:
Employee level outcomes (from the two targeted employees):
-Short term:
Intention to quit: measured using Michaels and Spector’s (1982) 3-Item Turnover Intentions Scales which uses a 6-point Likert scale. We will build an index using equal weighting.

Satisfaction: measured using the 18-item basic need satisfaction scale by Van Broek et al. (2010) aggregated with equal weights.

-Long term:
employee turnover: binary variable indicating whether the employee is still working in the store 6 months after the intervention

Store level outcomes:
store performance: index consisting of the following administrative measures provided by the cooperation partner: percent satisfied customers per month, contribution margin per month, and shrinkage per month relative to revenue. We are using a principal component analysis to generate the index.

leader’s need support
frequency of leader-employee meetings
leader mindset

leader’s need support (employee reported): 12-item survey measure developed by Tafvelin and Stenling (2019).

frequency of leader employee meetings (employee reported):
Question: How frequently do you have a work-related conversations with your leader
(1)almost every day (2) 2-3 times per week (3) once a week (4) 2-3 times a month (5) once a month or less

We intend to validate this by checking consistency with frequency of meetings reported by the leader for the 2 chosen employees.

leader mindset: measured using a 13 item survey measure developed by Kangas et al (2022)

Update after pilot but prior to implementation: removed engagement as an outcome variable and reduced the number of selected employees to two instead of three

We conduct a field experiment in a large grocery store chain in Norway to examine the effects of structured support from leaders on employees’ satisfaction and, subsequently, core KPIs of the stores. In the experiment, leaders will work with a leadership training app for 7-weeks. In addition, the leaders are encouraged to target two employees they work with closely while using the app. The content and features of the app differ by treatment arms.

Update after pilot but prior to implementation: removed engagement as an outcome variable and reduced the number of selected employees to two instead of three

We conduct a field experiment in a large grocery store chain in Norway to examine the effects of structured support from leaders on employees’ satisfaction, intention to quit and, subsequently, core KPIs of the stores. In the experiment, leaders in both the treatment and control group will work with a leadership training app for 7-weeks. In addition, the store managers are encouraged to target two employees they work with closely while using the app. These employees will be selected by the research team among those employees who consented to participate in the research project. We will follow the following selection procedure: If the store has an assistant manager, s/he will be the first of the two employees. For the rest of the selection, we only consider employees who consented to participate. Only consider employees who started after 2017 and consented. Out of those, choose the ones who work the most. If there is less than one employee (plus the assistant manager) who fulfill these requirements, choose the ones who work the most and who most recently started. If the store has no assistant manager, two employees are selected based on the procedure described above.

Using block design, stores get randomized to either receive the treatment or the control version of the app. The treatment version of the app includes two additional features to facilitate structured support from leaders: First, it includes a tool to encourage the leaders to set up weekly development meetings with the two employees. This tool helps leaders to organize meeting notes, set goals, and schedule follow-up meetings. Second, to strengthen leaders’ beliefs in development opportunities, it includes a module designed to help leaders develop a growth mindset for their own abilities as a leader and the growth potential of their employees. Leaders and employees are blind to treatment.

We investigate the treatment impact on work satisfaction, intention to quit, turn-over, and store performance (see definitions under primary outcomes). For satisfaction and intention to quit, we focus on the two targeted employees in our main test of the hypotheses. We will also test for spill-over effects using employees who were not targeted by the leader but have consented to participate.

When estimating treatment impacts, we use several model specifications investigating robustness to adding control variables. All models control for blocks. We will also include additional controls, which are selected based on model fit tested without the inclusion of the treatment variable. We will also run a model where we control for the baseline value of the outcome variable.

We investigate differential treatment effects based on baseline measures of the following variables implementation readiness (median split), employee tenure (median split), perceived leadership support (median split), intention to quit (at least agree or higher). We will also test the effect of the treatment on the individual components of the store performance index.

Stratified randomization blocked on the contributions margin per month, Intention to quit at store level, and the share of satisfied customers, all median split based on baseline data. In addition, we have generated a block including only two stores in which the targeted employees have a 0% fulltime employment position. The fact that the position is set to 0% is likely an error in the data, but we expect that these employees might be different from those with a fixed positive employment rate. We randomized using the stratarand command in STATA on the resulting 9 strata. In the extended trial, we will also block on corporation (4-5 different corporations) and might not need to have a block for the 0% fulltime employment position. In the extended trial, we will also block on corporation (4-5 different corporations) and might not need to have a block for the 0% fulltime employment position.


Update on extended trial randomization after pilot but prior to implementation:

In the extended trial, we use stratified randomization blocked on corporation, Intention to quit at store level and for large corporations the contributions margin per month based on baseline data. For small corporations (less than 10 stores) we stratify using a median split on intention to quit. For medium corporations (more than 10 but less than 20 stores) we use intention to quit terciles, for large corporations (more than 20 stores) we use intention to quit terciles and median split on the contributions margin to generate strata. We randomized using the stratarand command in STATA on the resulting 31 strata.

store

Yes

First trial: 42.
Extension:If the trial goes well, we hope to expand the number of participating stores to 150. If we do not have to make major changes after the first trial, we will stack the data of both trials.

Update Extension: We were able to recruit 169 stores. Due to poor take-up in the first trial, we do not intend to stack the data.

First trial: 42 stores. For each store, we observe at least the manager, the assistant manager, and 2 two additional employees. For individual-level employee outcomes, this gives us a sample of 3*42=126. For store-level outcomes, we have 42 observations. Extension: 150 stores. For each store, we observe at least the manager, the assistant manager, and 2 two additional employees. For individual-level employee outcomes, this gives us a sample of 3*150=450. For store-level outcomes, we have 150 observations. Update Extension:169 stores. For each store, we observe at least the manager, the assistant manager, and one additional employee. For individual-level employee outcomes, this gives us a sample of 2*169=338. For store-level outcomes, we have 169 observations.

Observations will be split equally between treatment and control. For the first trial, this means 21 stores in treatment and 21 stores in control, or 63 employees in treatment and 63 employees in control. For the extension, this means 75 stores in treatment and 75 stores in control, or 225 employees in treatment and 225 employees in control.

Update extension: We plan to have 85 stores in treatment and 84 stores in control, or 170 employees in treatment and 168 in control.

We calculate the MDES using PowerUP! Software Dong, N. and Maynard, R. A. (2013). First trial: Individual level outcomes: We have 9 blocks and, on average, 4.67 clusters per block with 3 observations in each cluster. Assuming constant treatment effects, an ICC of 0.01, an alpha of 0.05, and a power of 80%, we achieve a MDES of 0.58 standard deviations. This MDES can be further reduced by taking into account the explanatory power of the blocking variable and the inclusion of control variables, especially baseline measures for the outcome variable. If we, for example, assume that the blocking variable can explain 20% of the variance and an R^2 of 0.8, the MDES reduces to 0.27 standard deviations. Store level outcomes: We have 9 blocks with, on average, 4.67 stores per block. Assuming constant treatment effects, an alpha of 0.05, and a power of 80%, we achieve a MDES of 0.99. This MDES can be reduced by taking into account the explanatory power of the blocking variable and the inclusion of control variables, especially baseline measures for the outcome variable. If we, for example, assume the proportion of variance in Level 1 outcome explained by Block and Level 1 covariates is 0.8, the MDES reduces to 0.44 standard deviations. Extension: We are assuming that we have additional 4 corporations which are of similar size as the one participating in the first trial. Individual level outcomes: We have 32 blocks and, on average, 4.67 clusters per block with 3 observations in each cluster. Assuming constant treatment effects, an ICC of 0.01, an alpha of 0.05, and a power of 80%, we achieve a MDES of 0.29. standard deviations. This MDES can be further reduced by taking into account the explanatory power of the blocking variable and the inclusion of control variables, especially baseline measures for the outcome variable. If we, for example, assume that the blocking variable can explain 20% of the variance and an R^2 of 0.8, the MDES reduces to 0.14 standard deviations. Store level outcomes: We have 32 blocks with, on average, 4.67 stores per block. Assuming constant treatment effects, an alpha of 0.05, and a power of 80%, we achieve a MDES of 0.53. This MDES can be further reduced by taking into account the explanatory power of the blocking variable and the inclusion of control variables, especially baseline measures for the outcome variable. If we, for example, assume that the blocking variable can explain 20% of the variance and an R^2 of 0.8, the MDES reduces to 0.23 standard deviations. Update Extension: Individual level outcomes: We have 31 blocks and, on average, 5.45 clusters per block with 2 observations in each cluster. Assuming constant treatment effects, an ICC of 0.01, an alpha of 0.05, and a power of 80%, we achieve a MDES of 0.32. standard deviations. This MDES can be further reduced by taking into account the explanatory power of the blocking variable and the inclusion of control variables, especially baseline measures for the outcome variable. If we, for example, assume that the blocking variable can explain 20% of the variance and an R^2 of 0.8, the MDES reduces to 0.15 standard deviations. Store level outcomes: We have 31 blocks with, on average, 5.45 stores per block. Assuming constant treatment effects, an alpha of 0.05, and a power of 80%, we achieve a MDES of 0.45. This MDES can be further reduced by taking into account the explanatory power of the blocking variable and the inclusion of control variables, especially baseline measures for the outcome variable. If we, for example, assume that the blocking variable can explain 20% of the variance and an R^2 of 0.8, the MDES reduces to 0.20 standard deviations.