The impact of peers’ behaviour on firms’ training decisions

The local employment agency in the province of Trento provides important training subsidies to craft firms, financed by the Autonomous Province of Trento. Despite a high coverage rate of the cost of training up to 100% of the total expenses, take-up is relatively low, with only 5% of craft firms having benefited from this financing channel in the past 3 years. This is consistent with the overall evidence from the literature that craft firms underinvest in training despite low costs and potentially high benefits. We run an information treatment experiment providing craft firms with information about the training activities done by their peers through the local Employment Agency. The aim is to evaluate the peer effects on individual firms’ training decisions and test whether providing information on peers can be used as a lever to induce more training. Through administrative databases, we obtain the universe of training financed by the local employment agency to craft firms. We construct the average share of craft firms in the province that have benefited from this financing channel in the past 3 years in each sector, the average amount of funding, the average number of hours, as well as the generic type of training (technical vs. Soft skills) and the most common titles of courses taken, namely the top 5 courses. We construct a dashboard that easily allows users to browse this information by selecting the sector of interest and potentially filtering by firms with and without employees (many craft firms are sole proprietorships with no employees).
For the primary analysis in this study, we mainly leverage two sources of data: i) firm-level demographic information (e.g., sector, class size, number of employees) on the universe of craft firms in the province derived from administrative archives; ii) information about all the training activities financed by the local employment agency to craft firms. In November 2024, a baseline survey (CAWI) was conducted on the population of craft firms in the Trento province to obtain information about the attitudes and propensity towards training in general (i.e., not specifically financed by the local employment agency). This baseline survey was administered to all craft firms in the province, achieving a 21% response rate. Aside from socio-demographic characteristics, the survey also collected detailed information on all training activities carried out (including those not funded by the Employment Agency), as well as on the type of training, funding sources, decision-making processes, and future investment plans in training. Similarly, a follow-up survey is planned to be conducted after the implementation of the intervention and before the end of June 2026, which marks the end of the training "season” for the current year (firms can sign up for training financed by the local employment agency roughly from September to June of a given year). The survey is meant to complement the main analysis on administrative data and allow for exploratory analysis of secondary outcomes, as detailed below. For this reason, sample stratification also includes a stratum for baseline survey respondents.

2025-09-15

2026-06-30

The primary outcomes are:
1. Request for additional information via email (binary, 0/1 – intermediate outcome ).
2. a. Extensive margin: Firm participation in training funded by the local employment agency (binary, 0/1) in the 2025-26 cycle (and potentially onwards); 2b) Intensive margin: number of hours of training funded by the local employment agency (continuous)
3. Qualitative margin: The type of training the firm participates in (technical vs. Soft skills, as well as other potential categorisations to be decided afterwards); whether a firm is more likely to participate in a course that falls within the category of the top 5 courses other firms have benefited from, indicated in the dashboard; 0 otherwise.

All primary outcomes are obtained through administrative data: the intermediate outcome (contact via e-mail to the employment agency to obtain information about the financing channel) will be provided by the local employment agency; similarly, information about firm extensive and intensive participation in the training financed by local employment agency, as well as the type of training, will be obtained from administrative datasets operated by the local employment agency.

The secondary outcomes are:
1. General attitudes toward training: This outcome is captured through a composite index based on seven survey items measuring firms’ perceptions and attitudes toward training. The items cover perceived benefits, costs, and barriers to training (continuous).
2. Total hours of training conducted in the year 2025-2026, besides the one financed through the local employment agency (continuous).
3. Percentage of firms that are aware of at least one incentive and/or subsidy provided by the Employment Agency or directly by the province. In the baseline survey, more than 50% of responding firms reported not being aware of the existence of such funding opportunities.
4. Average training expenditure within the firm, excluding contributions received from the Province or the Employment Agency.
5. Planned increase in training effort (hours and budget): Plans to increase training hours over the next 12–24 months (binary 1/0).

All secondary outcomes are collected by the baseline and follow-up survey, which include measures of general attitudes toward training, perceived benefits from training in terms of productivity and experience with training in general. On top of the main analysis of administrative data for the universe of craft firms in the province, we will conduct a separate analysis on the sample of firms that respond to both the baseline (21%) and follow-up survey:

Our experimental design follows a stratified four-arm randomised controlled trial (three treatment groups + one control).
Experimental details will be available as soon as the trial is over.

Our experimental design follows a stratified four-arm randomised controlled trial (three treatment groups + one control).
All firms (including control firms) receive a generic email containing information about training opportunities through the financing channel of the local employment agency dedicated to craft firms in the province. Within the text of the email, an option is mentioned to request further details via a dedicated email address. Treated firms receive the same email, which, at the end, also includes a link to the dashboard mentioned above. The dashboard contains various information on training funded by ADL over the 2022–2024 period: in addition to the number of firms that participated in training initiatives, it reports the average training hours, the average funding provided, and the corresponding totals by sector. When the sector is locked (see below), the values refer to all sectors. Linked to these data are details on the breakdown by type of course (i.e., technical and transversal skills), as well as the list of the most frequently attended course cycles for these two categories.
The treatment group is divided into three arms, which differ depending on the type of training information accessible through the dashboard:
• Treatment 1 - receives information limited to the sector in which the firm operates; we obtain the sector from the administrative database containing the universe of craft firms in the province.
• Treatment 2 - receives the "full” dashboard, allowing navigation across all sectors, but pre-set on the sector in which the firms operate; this implies that when opening the dashboard, firms in this treatment arm see the same page as treatment 1, but are allowed to further browse the dashboard across sectors. They can also see the total for all sectors combined.
• Treatment 3 - receives only information for all sectors combined, without the possibility to browse across sectors or to access information about its own specific sectors. This information is also available for firms in treatment 2 by selecting “All sectors”.
The receipt of the email can be monitored to ensure that generic information about the financing channel is delivered in equal proportion to both treatment and control groups. The email subject line will be identical across groups. Treatment take-up will be measured by the number of clicks on the section of the email linking to the dashboard. A pilot analysis on a small sample of firms predicted a take-up rate of approximately 20%.


Protocol deviations

Amendment note
Implementation deviations: after randomisation, two implementation issues occurred (i) firms assigned to Treatment 2 (T2) did not receive the planned email or dashboard; they did not received anything (no-contact); and (ii) firms assigned to Treatment 1 (T1), received instead T1 plus a second email 1.5 hours later containing the T2 dashboard link (cross-sector navigation enabled). The email body was otherwise identical. We retain the original four-arm intention-to-treat (ITT) analysis by assignment and pre-specify the comparisons that remain valid under these deviations (see Analysis plan).
To sum up, treatment assignment remained randomised, but assignment to the treatment arms deviated from the initial setting. The new configuration is the following:
• Control (C): Generic email only (no dashboard), [no change];
• T3: Dashboard with aggregate/all-sectors only (no sector page, no navigation), [no change];
• T1: Two identical mails but different links to the dashboard (sector-locked plus cross-sector);
• T2: No contact (Super Controls).

From the first summary statistics provided by the Employment Agency, almost three weeks after the emails were sent (01.10.2025), delivery rates among contacted arms were similar (Controls: 4,251/4,579; T1: 2,117/2,291; T3: 2,191/2,369). General take-up (dashboard access) was 10.1% in T1 vs 9.2% in T3, indicating limited behavioural impact of the duplicate email. Among the T1 group (n = 232), it emerges that 40.9% opened only the T1 link, 45.7% only the T2 link, and 13.4% opened both.

Given the results on the take-up, in the reminders scheduled from now on, we will deliver to the T1 group the "cross-sector navigation enabled" dashboard, pre-set on their sector-specific information, but allowing navigation across all sectors.
For these reasons, we keep the four-arm stratified RCT (Control, T1, T2, T3) and we analyse the pre-specified outcomes, considering the following comparisons:
• Any contact vs no contact: (Control + T1 + T3) vs T2 (now Super Controls).
• Dashboard (any) vs generic email: (T1 + T3) vs Control (restricted to contacted arms).
• cross-sector (pre-set on their sector-specific information) vs aggregate: T1 vs T3.

To maintain a consistent definition of T1 exposure, subsequent reminders to T1 will include the cross-sector dashboard (pre-set on the firm’s own sector, with navigation enabled). C and T3 reminders remain as originally defined; T2 remains a no-contact group for the whole intervention period.

Computer random number generator.

The randomisation unit is the individual craft firm operating in the province of Trento.

No

No clustering.

Indicatively 12,000 craft firms (the full universe of craft firms in the province of Trento) obtained from administrative data.

3 treatment groups (20% of firms each) and 1 control group (40% of firms).

Primary analysis on Administrative Data From the administrative data, we know that the baseline participation rate in financed training is equal to 5 % and the average number of hours is equal to 78.8. Considering an alpha level of 0.05 (two-sided), 80% power, allocation 60/40 (T/C), full compliance, full response rate (N = 12,000), the ITT minimum detectable effects (MDEs) would be: • Participation (binary): 1.21 pp • Hours of training (continuous, SD = 26.78 h): 1.40 h If analyses are restricted to 20% responders (20% response rate, N ≈ 2,400), the ITT MDEs worsen: 2.89 pp (participation) and 3.13 h (hours). But because only a fraction of treated firms may comply with the intervention (e.g. clicking the dashboard, engaging with the information), the ITT effect is further diluted according to the compliance rate. Below, we report the implied LATE MDEs under alternative compliance rates: Compliance rate Outcome N=12,000 (MDE) N≈2,400 (MDE) 10% Participation (pp) 12.1 28.9 10% Hours (h) 14 31.3 20% Participation (pp) 6.05 14.45 20% Hours (h) 7 7.1 15.65 30% Participation (pp) 4.03 9.63 30% Hours (h) 4.67 10.43 50% Participation (pp) 2.42 5.78 50% Hours (h) 2.8 6.26