Cultivating Connectivity: Measuring the Impact of a Digital Agricultural Extension Platform on Millennial Farmers and Extension Workers in Indonesia (LenteraDigiEx)

The intervention will consist of two treatment groups. In each treatment group, we will implement two types of training interventions: One for MFs and one for EWs.
The one for MFs will include information on digital literacy, the LenteraDESA platform, and its new feature for farm recording (in Bahasa, Catatan Tani). The training for EWs will include information on how to integrate digital platforms and digital content into their activities in addition to the other topics.
In the first treatment group, the 'light' training, the recipients will be offered an online session, while the second treatment group, the `intensive' training, will replicate the same session in-person(offline).
Both types of treatment will be followed by a self-paced online training on LenteraDESA, covering various topics related to digitization in agriculture and digital solutions for agriculture.
The training will be concluded with an online presentation by the participants who will obtain a certificate of completion at the end of the course.
MFs will also be able to use the digital farm record for loan requests.

2025-05-02

2025-05-31

Shared Outcomes (MFs and EWs)

Digital Literacy and platform uptake
• What is the digital literacy level of the respondent? (composite index - primary outcome)
– What is the level of digital information and data literacy of the respondent? (composite
index)
– What is the level of digital communication and collaboration literacy of the respondent?
(composite index)
– What is the level of digital content creation literacy of the respondent? (composite index)
– What is the level of digital safety literacy of the respondent? (composite index)
– What is the level of digital problem solving literacy of the respondent? (composite index)
– What is the level of digital knowledge of the respondent? (composite index)
– What is the level of digital attitude of the respondent? (composite index)
– What is the level of digital skills of the respondent? (composite index)
• Does the respondent know about LenteraDESA? (primary outcome)
• Does the respondent use LenteraDESA? (primary outcome)


Outcomes only related to EWs

Performance and capability of extension workers
•Has the respondent integrated digital tools into activities in the last 12 months?
(primary outcome)
• How many extension activities did the respondent perform during the last 12 months? (in total primary outcome)
•Did the intervention improve the respondent’s capability? (primary outcome)
– How many farmers does the respondent reach?
– How many MFs does the respondent reach?
– How many acres does the respondent supervise?


Outcomes only related to MFs
Welfare of MFs
• How much income did the respondent get from their agricultural activities in the
last 12 months? (primary outcome)

The digital literacy index is based on the DigiComp 2.0 framework, adjusted for the agricultural context in Indonesia.

Shared outcomes

Training uptake behavior Within these outcomes, a special focus will be on the training topic of the intervention (digitization in agriculture).
• Has the respondent attended any training in the last six months?
• Has the respondent obtained any certificate for a training in the last six months?
• How many types of training did the respondent attend in the last six months?
• Has the respondent attended any training that covers the use of digital technology in supporting
agricultural extension in the last six months?
• How many types of training on digital technology use did the respondent attend in the last six
months?

Working life satisfaction
• What is the level of working life satisfaction of the respondent?


Outcomes only related to EWs

Performance and capability of extension workers
• Have the training sessions helped improve the respondents’ ability to use digital technology for
extension services?
• What is the respondent’s level of acceptance of digital technology in extension activities?
• What is the level of personal digital innovation of the respondent?
• Has the respondent integrated LD into their activities in the last 12 months?
• How many extension activities did the respondent perform during the last 12 months? (by type of activity)


Outcomes only related to MFs

Welfare of MFs

• How many agricultural and business practices does the respondent know?
• How many agricultural and business practices does the respondent use?
• How many sources of agricultural and business practices does the respondent know?
• Does the business of the respondent have any level of formalization?
• How many certification types does the business of the respondent have?
• Is the respondent a member of a farmer group?
• Is any (other) household member part of a farmer group?
• Did the respondent have any loan in the last 12 months?
• How many loans did the respondent have in the last 12 months?
• What is the total loan amount of the(se) loans?
• Is there a loan from a formal institution?
• How many sources of loans does the respondent have?
• What is/are the purpose(s) of the(se) loan(s)?
• How many Rupiees has the respondent invested in their business in the last 12 months?
• Does the respondent currently employ any worker?
• How many people does the respondent currently employ?
• How many family members that are not paid for work does the respondent currently have?
Within the agricultural and business practices outcomes, a special focus will be given to the topics
of the intervention (agricultural digitization, social media marketing, e-commerce, and digital farm
recording).

Sampling

Our sample of respondents is drawn from 51 subdistricts that are comparable to agricultural extension
service offices1, located in five regencies in the Yogyakarta special region of Indonesia. The 51 subdis-
tricts cover almost the entire area under analysis. The covered area is incomplete as five offices have
been excluded from the analysis since they reach at most two MFs. The reasons for the exclusion are
implementation efficiency and cost-benefit.
The planned proportional random sample consists of 839 MFs and 171 EWs drawn from a sampling
frame of 1,624 MFs and 303 EWs. The sampling frame comes from administrative data collected by
the agriculture and food security office of the local government2 and consists of governmental EWs and
registered MFs. We selected the sample to have a similar number of respondents across treatment arms
to increase statistical power. Therefore, we first considered 50% of the individuals to be distributed
across the three treatment arms (two treatment groups and one control), such that
_______________________________
where RT is the ideal number of respondent per treatment group, and N is the total number of
individuals in the sampling frame. In a second step, we select the number of respondent per subdistrict
as explained in the following equation:
_______________________________
In particular, the number of selected respondents R in a subdistrict S within a treatment arm T is
rounded up integer of the proportion of individuals in a subdistrict of a given treatment arm NT S
over the total population of that treatment arm NT , times the total number of respondents in a given
treatment arm RT . We rounded up all decimal points to select the number of respondents within each
subdistrict, to ensure at least the ideal number of respondents per treatment arm to be reached.
During the baseline data collection (February 2025), we realized that the reliability of the MFs’
data was low, as many the potential respondent only attended one or more events held by BPSDMP,
therefore, we requested additional data from the local agricultural department (DINAS). After ran-
domizing the potential replacements from this list, we obtained a final number of respondent of 784
MFs and 170 EWs. As we plan two additional waves, we aim at obtaining 2,862 observations from
954 respondents.

11.2 Data collection
Data collection is planned in three phases: Baseline, midline, and endline. These three phases will
occur at the same time for both respondent types. Baseline data collection ended in March 2025.
Midline data collection is planned for October 2025, and endline is planned for February and March
2026.


Randomization procedure:
We applied a stratified random assignment with misfit corrections based on the distribution of EWs
across subdistricts. First, we considered the distribution of EWs in our sampling frame and created
strata based on whether the number of EWs in a subdistrict is above the median number of EWs per
subdistrict. Then, we randomly assigned one-third of the subdistricts (17) to each treatment arm (two
treatment groups and one control), making sure that misfit corrections were globally applied based on
the distribution of EWs in the subdistricts.

Spillovers
We do not expect major spillovers in terms of training participation, as the intervention is based on
subdistrict level and is only offered to the selected respondents (due to budget constraints), preventing
EWs from sharing their freshly acquired knowledge with MFs outside of their working subdistrict area.
Moreover, access to self-paced training will be first offered only to selected respondents4. Still, given
that digital platforms are nonexclusive5, the respondents in the control group could access the platform
themselves. Therefore, the control group can become aware of the platform and access its general
content, learning about those agricultural and business practices that might affect their behavior. As
we plan to analyze intent-to-treat effects, the impact of the training effect on these outcomes will only
be underestimated.

Multiple hypothesis testing
In our study, there is a high probability of falsely rejecting the null hypothesis due to the large number
of outcomes. Therefore, to ensure that our findings are robust to p-value adjustments for multiple
hypothesis testing, we will control for the False Discovery Rates as in Benjamini and Hochberg (1995).
Specifically, we will calculate sharpened q-values within each family of outcomes.

Cost-benefit analysis
We plan to perform a cost-benefit analysis to inform the local government about the value of a potential
scale-up of the project.

Not available

Randomization done in office by computer

We applied a stratified random assignment with misfit corrections based on the distribution of EWs
across subdistricts. First, we considered the distribution of EWs in our sampling frame and created
strata based on whether the number of EWs in a subdistrict is above the median number of EWs per
subdistrict. Then, we randomly assigned one-third of the subdistricts (17) to each treatment arm (two
treatment groups and one control), making sure that misfit corrections were globally applied based on
the distribution of EWs in the subdistricts.

Cluster: Sub-district:
Cluster sub-unit: Individual

Yes

51 sub-districts

The planned total of 2,862 observations (2,352 for MFs and 510 for EWs) from three survey waves from 170 EWs and 784 MFs. If attrition rates are greater than 10% and we find evidence of differential attrition by treatment status, we will estimate the pairwise Lee bounds for our treatment effects.

17 sub-districts control, 17 sub-districts online only treatment , 17 sub-districts blended learning treatment

Power analysis We updated our power calculations using our baseline data collection for our composite digital literacy indicator (scale 0-4) based on a 95% confidence interval and a power of 80%. The baseline data shows a mean of digital literacy of 1.74 scale points, with a standard deviation of 0.49 and an intra-cluster correlation (ICC) of 0.069. Given our average cluster size of 18 respondents per cluster and 17 clusters per treatment arm, we are powered to detect a minimum effect of 0.165. For a separate analysis, we can detect an effect of 0.189 for MFs and 0.205 for EWs.

PAP_LenteraDigiEx