001
002
003
004
005
006
007
008
009
010
011
012
013
014
015
016
017
018
019
020
021
022
023
024
025
026
027
028
029
030
031
032
033
034
035
036
037
038
039
040
041
042
043
044
045
046

Heterogeneous multidimensional efficiency effects

of agro-ecological pest management transition and

intensification in smallholder systems–Evidence

from mango fruit fly control in Kenya

Sulman Olieko Owili1,2*, David Jakinda Otieno1,
Evans Ligare Chimoita1, Frederick Philbert Baijukya2

1*Department of Agricultural Economics, University of Nairobi,
Nairobi, Kenya.

2Natural Resource Program, International Institute of Tropical
Agriculture, Nairobi, Kenya.

*Corresponding author(s). E-mail(s): oliekosulman@gmail.com;
Contributing authors: david.jakinda@uonbi.ac.ke;

echimoita@uonbi.ac.ke; f.baijukya@cgiar.org;

Abstract

Agro-ecological transition is an important step towards sustainable and resilient
food systems in the face of systemic threats from climate-change-induced dis-
turbances. In smallholder systems, the transition towards agro-ecological pest
management (APM) offers the prospect of reconciling agronomic performance
with environmental and social imperatives by replacing indiscriminate chemi-
cal applications with locally-derived biorational options. However, the efficiency
implications of APM transitions remain insufficiently documented, particularly
in smallholder systems and in relation to invasive alien pests that are prone to
resurgence and reinfestation under suboptimal management. This paper evalu-
ates whether the adoption and intensification of APM improve both technical
and eco-efficiency in smallholder settings, with a focus on the Oriental fruit fly
(Bactrocera dorsalis L.) in mango (Mangifera indica L.) orchards. We apply a
latent class stochastic metafrontier model to a sample of 418 orchard managers
from Makueni County, Kenya, selected through a multistage sampling procedure.
This approach enables us to classify orchard managers into non-adopters, non-
intensive adopters, and intensive adopters, and to compute meta-technical and
meta-eco-efficiency scores, from which we derive an environmentally adjusted

1



047
048
049
050
051
052
053
054
055
056
057
058
059
060
061
062
063
064
065
066
067
068
069
070
071
072
073
074
075
076
077
078
079
080
081
082
083
084
085
086
087
088
089
090
091
092

efficiency measure. We find no significant sample selection bias and treatment
effects are estimated using a doubly robust Inverse-probability-weighted regres-
sion adjustment estimator. Intensive adoption had a positive average treatment
effect (ATE) and average treatment effect on the treated (ATT) of 8.1% and
5.6%, respectively, whereas non-intensive adoption showed no significant effect
(ATE = –1.1%, ATT = –1.5%). Efficiency effects were heterogeneous and
inefficiency varied with orchard manager’s APM adoption intensity, education
level, orchard prospects, group membership, and participation in knowledge co-
creation activities. Policymakers and development practitioners should support
farmers by institutionalising continuous learning and establishing multi-pronged
participatory training platforms that use existing social networks.

Keywords: agroecology, eco-efficiency, environmentally adjusted efficiency, invasive
alien pests, latent class stochastic metafrontier, sustainable pest management,
technical efficiency

JEL Classification: C38 , D24 , Q12 , Q16 , Q57

1 Introduction

Agriculture faces the dual challenge of simultaneously decoupling productivity from
environmental footprints while improving the food security for a growing global pop-
ulation [1]. Historically, attempts to improve productivity have relied on conventional
intensification approaches, increasingly relying on external inputs to bolster food pro-
duction and stabilise yields. The Green Revolution epitomised this paradigm, where
the intensive use of synthetic pesticides, fertilisers, and improved cultivars drastically
increased global food production by more than 50% [2, 3]. Although this intensifi-
cation mitigated the need for additional land conversions, reducing the pressure on
marginal lands, forests, and riparian areas to meet increasing food demand, the cumu-
lative and pervasive reliance on synthetic damage-control inputs such as pesticides
has led to a series of ecological, economic, and health-related concerns [4–7]. This
has prompted calls for sustainable transitions that leverage ecological processes to
maintain or increase yields.

Recently, a growing stream of literature has examined the extent to which sus-
tainable intensification can match conventional yields in practice [8–11]. This debate
is particularly relevant for smallholder systems, where resource constraints, as well
as biotic and abiotic pressures, pose increasing threats to agricultural sustainability.
Unlike organic systems that have been shown to achieve lower yields between 19–25%
so that an increase of 23–33% in land size is required to meet the current output levels
under conventional systems [12], agro-ecological systems have been found to improve
yields as well as land and labour productivity in smallholder systems by countering
local constraints [8, 13].

Mango (Mangifera indica L.) is Kenya’s second most important fruit crop after
banana [14]. However, tephritid fruit fly, particularly Bactrocera dorsalis (Diptera:
Tephritidae), poses a major constraint to mango productivity and marketing, causing

2



093
094
095
096
097
098
099
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138

fruit and quality loss if improperly managed. This polyphagous pest inflicts exten-
sive damage, with yield losses reported to range from 30% to as high as 90% [15, 16].
Female flies oviposit within the fruit, and subsequent larval feeding not only renders
the fruit unmarketable, but also predisposes it to secondary bacterial infections, com-
pounding yield losses [15]. Conventional fruit fly control typically relies on chemical
insecticides, which often act quickly to reduce infestations but are often associated with
high external costs, including pesticide residues, human health risks, loss of beneficial
insects, and pest resistance [17]. In contrast, agro-ecological pest management (APM)
offers a holistic alternative through a suite of locally available eco-friendly practices
[18–26], including orchard sanitation, the release of natural enemies such as ovivorous
ants and parasitoid wasps, the application of biopesticides, food baits, male annihila-
tion technique [27], and other cultural and indigenous controls [28–31]. Transitioning
to effective APM requires farmers to synchronise and intensify pest-control efforts
not through blanket chemical applications but by adopting ecologically grounded,
indigenous, locally available, knowledge-intensive biorational practices.

It has been argued that the conscious adoption of agro-ecological practices can
potentially close yield gaps, maintain ecological integrity, reduce reliance on synthetic
pesticides, and improve resource-use efficiency at the farm level [1, 17, 32, 33]. Farm-
level analyses in Kenya found that APM-adopting mango smallholders obtained higher
yields [29, 34] and higher net income [29, 34–37] while using significantly less inorganic
pesticides [29, 38]. It has also been found that APM uptake increases inclusivity in
decision making by enhancing women empowerment [39]. These findings support the
optimistic view that APM can improve both productivity, social and environmental
performance of farming systems, ultimately enhancing eco-efficiency.

Eco-efficiency (EE) refers to the ratio of economic value added to the associated
deleterious environmental impacts [40, 41]. The EE index requires optimising the
ratio of agricultural outputs to environmental impacts and has been promoted as a
strategy to quantify the benefits of sustainable pest management, such as APM [42–
44]. By lowering the environmental footprint of food production without sacrificing
yield, APM can increase the output gained per unit of environmental cost, a key
requirement for sustainability. Current literature propose constructing an EE index
as the ratio of crop yield to total applied toxicity (or the risk quotient) [see, e.g.,
38, 43, 44]. To provide a more comprehensive picture of sustainability, this study
also included additional environmental impact categories, including carbon footprint,
nutrient balance, biodiversity loss, and energy balance, in the analysis.

Recently, Weltin and Hüttel [9] proposed a directional meta-frontier approach with
matching to measure EE of farms under various sustainable intensification regimes
and found that farmers under higher intensification were more eco-efficient than less
intensified ones. However, directional distance functions are sensitive to the choice
of directional vector and pressure-output normalisation. Additionally, their analysis
is not crop-specific. This matters for EE because key biophysical and management
drivers are often crop-dependent. For example, crop phenology (e.g., flowering and
fruit-set timing), water and nutrient responses, and crop-specific pest/disease pressures
influence both outputs and environmental pressures in ways generic farm-level studies
cannot capture. Therefore, such non-targeted analyses risk smoothing over technology

3



139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184

choices that are uniquely relevant for a specific crop (e.g., pest management regimes)
and mismeasuring pressures where crop-specific units and scaling are critical.

Extant studies on the EE of mango production vary considerably in scope, context
and methodology. Basset-Mens et al. [45] used a cradle-to-farm-gate life cycle anal-
ysis to compare locally grown and imported mango, apple, peach, and clementine in
French markets. Their results showed that mango generally performed well in various
environmental impact categories such as eco-toxicity and eutrophication. However, the
authors acknowledged that fruit fly was not as problematic in these contexts as it is in
Africa. Additionally, due to the broader scope of the study, there were uncertainties
in obtaining representative data for the individual crops, which could have increased
uncertainty in the results. Similarly, in southern Iran, Rasoolizadeh et al. [46] evalu-
ated the EE of five tropical fruits (guava, mango, banana, jujube, and sapodilla) using
life cycle analysis approach. Mango emerged with the highest EE score, outperform-
ing the other fruits. Although the life cycle analysis procedure is a useful approach
in aggregation of environmental impacts, it is a subjective weighting method that
relies on expert judgement in the assignment of pressure weights. This can potentially
bias estimates as the resulting EE index could be a function of the expert’s values
and beliefs [47]. To overcome the pitfalls of life cycle analysis, Heidenreich et al. [48]
employed an input-oriented order-m approach with Data Envelopment Analysis to
measure EE among mango, macadamia, coffee and cocoa farms in Kenya and Ghana.
Their findings indicated substantial variability, with mango farms ranking as least
efficient and requiring a 25% cut in environmental pressures to reach optimal perfor-
mance. Although the study accounted for regional heterogeneities in the production
environment, the authors presupposed uniform production technologies and ecological
conditions at the orchard level. The assumption of uniform production technologies is
often restrictive and is rarely observed in smallholder systems.

Existing studies on the efficiency effects of pest management strategies have
predominantly focused on economic performance, particularly technical efficiency
(TE), while largely overlooking the environmental implications of such practices.
For instance, Yi et al. [49] investigated productivity and TE disparities among
shallot farmers in Indonesia, distinguishing growers by their compliance to alterna-
tive pest management protocols. Similarly, Rahman and Norton [50] examined the
impacts of integrated pest management adoption on the TE of eggplant growers in
Bangladesh. More recently, Rodrigues et al. [51] evaluated the TE of biological pest
control adoption in Brazil, accounting for heterogeneity between intensive and non-
intensive user subgroups. In this paper, we contribute to this discourse by evaluating
whether the transition to and intensive adoption of APM can enhance multidimen-
sional farm-level efficiency. Following Andrieu et al. [52], we implement a multi-criteria
approach to explicitly accommodate both economic and ecological performance of
APM by analysing EE alongside the traditional TE. We then derive an environmen-
tally adjusted efficiency score that reflects the dual objectives of optimising input
conversion and minimising environmental externalities, providing a better picture from
a sustainability perspective. To the best of our knowledge, this is the first attempt
at examining the multidimensional efficiency effects of agro-ecological transitions in

4



185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230

smallholder systems, particularly in relation to invasive alien pests that are prone to
resurgence and reinfestation under suboptimal management.

The rest of the paper is structured as follows. Section 2 describes the materials and
methods including the study area, sampling procedure, data collection, and analytical
framework. In Section 3, we present the empirical results and discuss the findings. We
then conclude in Section 4 by outlining avenues for future research and prospects for
scaling up agro-ecological transitions in smallholder systems, with reference to broader
policy implications.

2 Materials and methods

2.1 Study area

We utilised a subset of observational data from a household survey conducted in
Makueni county, Kenya. Makueni County spans an area of 8,214 square kilome-
tres, located between latitudes 1◦35′ and 2◦59′S and longitudes 37◦10′ to 38◦30′E
(Figure 1), with a population of approximately 1,098,584 [53]. Administratively, the
county has six sub-counties, each subdivided into wards, sub-wards, and villages.
The county’s main economic activities include subsistence agriculture, apiculture, and
small-scale trade [54]. The county majorly has a low-lying terrain, with diverse agro-
ecological zones (see Figure 1). The hillier sections receive about 800–1200 mm of
rainfall annually and involves cultivation of a wide variety of crops, including maize,
pulses, fruit crops and coffee, as well as livestock production such as dairy and bee
keeping. The lower plains receive as low annual rainfall as 250–400 mm and is mainly
associated with livestock keeping and tourism [54]. The county’s average annual rain-
fall is estimated at 500–750 mm making it ideal for growing most crops. Mean air
temperatures range from 20.2◦C to 35.8◦C, with the hills remaining noticeably cooler
[55]. This warm climate favours production of fruit crops such as mango and citrus.

2.2 Sampling technique and data collection

To determine the required sample size, the Yamane [56] formula was applied, using a
known population of 28,696 mango farmers in the county [54]. A multistage sampling
strategy was implemented. First, Makueni County was purposively selected because
it is Kenya’s main mango producing region, facilitated, in part, by its proximity to
important export hubs, such as Nairobi and Mombasa, as well as favourable climatic
conditions for mango production. In the second stage, the Makueni, Mbooni, and Kaiti
sub-counties were chosen due to their prominence in mango production. In the third
stage, six wards (Kako-Waia, Kee, Kiteta-Kisau, Kitise, Nzaui, and Ukia) and 12 sub-
wards (2 per ward) were randomly selected within these sub-counties. Finally, since
almost all households in the selected areas grow mango, systematic random sampling
was employed within these areas to select every third household.

Mango orchard managers were identified as key respondents due to their direct
control and awareness of most orchard-level activities. The interviews were conducted
between August and September 2023 by trained enumerators, with informed consent
obtained at the beginning of each interview. Of 434 orchard managers interviewed, nine

5



231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276

Fig. 1: Map of Makueni county showing its agro-ecological zones and location of sampled
households.

responses were excluded after controlling for non-exposure bias, and seven were dis-
carded due to incomplete responses. This resulted in 418 valid responses for subsequent
analyses.

2.3 Analytical framework

2.3.1 System boundary and life cycle inventory

We adopted a farm gate approach as the system boundary, so that production extends
only to the point where materials leave the orchard, assuming no value addition occurs
within the orchard. This delimitation ensures that all input quantities and ecological
pressures are under direct control of the orchard manager. To allow global comparison,
the functional unit chosen was one hectare (ha) of mango orchard and all inputs,
outputs, and environmental pressures were normalised per ha. Typically, smallholder
mango production involves farm activities such as tillage, fertilisation, control of pests,
diseases, and weeds, and harvesting.

The selection of an eco-efficiency indicator depends on the availability of data, the
interest of the policymaker, and the intended use of the resulting scores [11, 41], and
remains largely an empirical matter [57]. We considered six environmental pressures
from mango production following the relevant literature (Table 1). Water resource
pressures were excluded because smallholder mango systems in SSA are predominantly
rainfed.

6



277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322

T
a
b
le

1
:
S
u
m
m
a
ry

o
f
in
d
ic
a
to
rs

a
n
d
eq
u
a
ti
o
n
s
u
se
d
to

m
ea
su
re

en
v
ir
o
n
m
en
ta
l
im

p
a
ct
s
in

m
a
n
g
o
p
ro
d
u
ct
io
n

In
d
ic
a
to

r
D
e
sc

r
ip

ti
o
n

C
o
m

p
u
ta

ti
o
n

R
e
fe
r
e
n
c
e
s
a
n
d

n
o
te

s

C
a
rb

o
n

fo
o
t-

p
ri
n
t
(C

F
)

E
st
im

a
te
s
to
ta
l
g
re
en

h
o
u
se

g
a
s
(G

H
G
)
em

is
si
o
n
s
fr
o
m

o
rc
h
a
rd

i,
a
cc
o
u
n
ti
n
g

fo
r

ex
te
rn

a
l

in
p
u
ts
,

d
ir
ec
t

fi
el
d
em

is
si
o
n
s,

a
n
d
ca

rb
o
n

se
q
u
es
tr
a
ti
o
n
p
o
te
n
ti
a
l.

C
F
i
=

(G
H
G

fr
o
m

in
p
u
ts
i
+

G
H
G

fr
o
m

fi
el
d
i
)−

G
H
G

se
q
u
es
te
re
d
i
,

G
H
G
i
ψ

=
∑ Ψ ψ=

1
(A

D
i
ψ
·E

F
ψ
)

M
a
n
u
re

p
ro
d
u
ct
io
n

(e
n
te
ri
c

fe
rm

en
ta
-

ti
o
n
)

em
is
si
o
n
s

ex
cl
u
d
ed

;
4

k
g

C
E

h
a
−

1
fo
r

fi
el
d

cu
lt
iv
a
ti
o
n

[5
8
].

A
te
n
-

y
ea

r-
o
ld

m
a
n
g
o

o
rc
h
a
rd

se
q
u
es
tr
a
ti
o
n

p
o
te
n
ti
a
l:

1
1
.0
4
tC

O
2
eq

h
a
−

1
(a
p
p
ro
x
.

1
.1
tC

O
2
eq

h
a
−

1
y
r−

1
)
[5
9
].
E
m
is
si
o
n
fa
c-

to
rs

o
b
ta
in
ed

fr
o
m

IP
C
C

[6
0
],
W

a
ll
in
g
a
n
d

V
a
n
ee
ck

h
a
u
te

[6
1
],
N
a
y
a
k
et

a
l.
[6
2
],
C
ec
h

et
a
l.
[6
3
].

T
o
x
ic
it
y

(T
O
X
)

Q
u
a
n
ti
fi
es

to
x
ic
it
y

fr
o
m

h
a
za

rd
o
u
s
p
es
ti
ci
d
es

b
a
se
d

o
n

a
ct
iv
e

in
g
re
d
ie
n
t

(A
I)

q
u
a
n
ti
ti
es

1
.

B
a
se
d
o
n
to
ta
l
q
u
a
n
ti
ty

(i
n
k
g
A
I
h
a
−

1
y
r−

1
)

o
f
h
a
za

rd
o
u
s
a
ct
iv
e
in
g
re
d
ie
n
ts

a
p
p
li
ed

.
H
a
za

rd
o
u
s
cl
a
ss
ifi
ca

ti
o
n
p
er

P
A
N

In
te
rn

a
-

ti
o
n
a
l
L
is
t
2
0
2
4
[6
4
];
cr
o
ss
-c
h
ec
k
ed

a
g
a
in
st

K
en

y
a
P
es
t
C
o
n
tr
o
l
P
ro
d
u
ct
s
B
o
a
rd

d
a
ta
.

N
u
tr
ie
n
t

b
a
l-

a
n
ce

(N
,
P
)

M
ea

su
re
s

n
u
tr
ie
n
t

(ζ
i
)

in
p
u
t-
o
u
tp

u
t
g
a
p
to

a
ss
es
s

d
efi

ci
ts

o
r

su
rp

lu
se
s

in
n
it
ro
g
en

(N
)
a
n
d
p
h
o
sp

h
o
-

ru
s
(P

)
(i
n
k
g
h
a

−
1
y
r−

1
)

ζ i
b
a
la
n
ce

=
a
′
·X

ζ
i
−

b
′
·Y
ζ
i
ζ i

d
efi

ci
t
=

{ |ζ
i
b
a
la
n
ce

|,
if
ζ i

b
a
la
n
ce

<
0

0
,o

th
er
w
is
e

ζ i
su

rp
lu
s
=

{ ζ i
b
a
la
n
ce
,i
f
ζ i

b
a
la
n
ce

>
0

0
,o

th
er
w
is
e

N
u
tr
ie
n
t
fl
o
w
s
es
ti
m
a
te
d
fr
o
m

in
p
u
t
la
b
el
s

a
n
d

fr
u
it

co
m
p
o
si
ti
o
n

d
a
ta
.
O
n
ly

d
efi

ci
t

v
a
ri
a
b
le
s
re
ta
in
ed

d
u
e
to

li
m
it
ed

su
rp

lu
s

o
b
se
rv
a
ti
o
n
s2
.
A
p
p
ro
a
ch

b
a
se
d
o
n
m
a
te
ri
-

a
ls

b
a
la
n
ce

p
ri
n
ci
p
le

[6
5
,
6
6
].

E
n
er
g
y
b
a
la
n
ce

(E
N
)

A
ss
es
se
s
en

er
g
y

in
p
u
ts

v
s

o
u
tp

u
ts

(i
n

M
J

h
a

−
1

y
r−

1
).

In
p
u
t
=

en
er
g
y
fr
o
m

la
b
o
u
r,

m
a
n
u
re
,
fe
rt
il
is
-

er
s,

p
es
ti
ci
d
es
;
O
u
tp

u
t

=
en

er
g
y

fr
o
m

h
a
r-

v
es
te
d
m
a
n
g
o
fr
u
it
s.

E
n
er
g
y

co
n
te
n
t:

0
.2
5
–
0
.7
9
5

M
J

1
0
0
g
−

1

d
ri
ed

m
a
n
g
o
[6
7
].
F
re
sh

fr
u
it

co
m
p
o
si
ti
o
n
:

8
3
.4
%

w
a
te
r
[6
8
].
E
m
is
si
o
n
/
en

er
g
y
fa
ct
o
rs

fr
o
m

H
ei
m
p
el

et
a
l.
[6
9
].

S
p
ec
ia
li
sa
ti
o
n

(S
P
)

P
ro
x
y
fo
r
b
io
d
iv
er
si
ty

lo
ss

d
u
e

to
m
o
n
o
cu

lt
u
re

te
n
-

d
en

cy
.

S
P
i
=

M
a
n
g
o

a
r
e
a
i

T
o
t
a
l
o
r
c
h
a
r
d

a
r
e
a
i

H
ig
h
er

v
a
lu
es

in
d
ic
a
te

in
cr
ea

se
d
m
o
n
o
cu

l-
tu

re
a
n
d
p
o
te
n
ti
a
l
b
io
d
iv
er
si
ty

lo
ss
.
B
a
se
d

o
n
H
ei
d
en

re
ic
h
et

a
l.
[4
8
].

N
o
te
s:

A
b
b
re
v
ia
ti
o
n
s:
—

A
D
,
a
c
ti
v
it
y

d
a
ta

;
A
I,

a
c
ti
v
e
in
g
re
d
ie
n
t;

C
E
,
c
a
rb

o
n

e
q
u
iv
a
le
n
t;

C
O

2
e
q
,
c
a
rb

o
n

d
io
x
id
e
e
q
u
iv
a
le
n
t;

E
F
,
e
m
is
si
o
n

fa
c
to

r;
h
g
,

h
e
c
to

g
ra

m
;
M

J
,
m
e
g
a
jo
u
le
s.

1
W

e
re
p
o
rt

ra
w

q
u
a
n
ti
ti
e
s
o
f
p
e
st
ic
id
e
p
ro

d
u
c
ts

a
p
p
li
e
d

b
e
fo
re

m
ix
in
g
w
it
h

w
a
te
r.

A
s
m
o
st

p
ro

d
u
c
ts

w
e
re

li
q
u
id

fo
rm

u
la
ti
o
n
s,

a
p
p
li
c
a
ti
o
n

ra
te
s
a
re

p
re
se
n
te
d
a
s
v
o
lu
m
e
s
(L

h
a
−

1
y
r−

1
).

T
o
a
c
c
o
m
m
o
d
a
te

th
e
fe
w

in
st
a
n
c
e
s
w
h
e
re

so
li
d
fo
rm

u
la
ti
o
n
s
(s
u
ch

a
s
g
ra

n
u
la
r
o
r
w
e
tt
a
b
le

p
o
w
d
e
rs
)
w
e
re

a
p
p
li
e
d
,

w
e
u
se
d

a
1
:1

m
a
ss
–
v
o
lu
m
e
c
o
n
v
e
rs
io
n
,
a
ss
u
m
in
g

u
n
it

d
e
n
si
ty
.
T
h
is

a
p
p
ro

a
ch

e
n
su

re
s
c
ro

ss
-i
n
d
ic
a
to

r
c
o
m
p
a
ra

b
il
it
y

a
n
d

d
e
li
v
e
rs

a
si
n
g
le
,
c
o
n
si
st
e
n
t

m
e
a
su

re
o
f
in
se
c
ti
c
id
e
q
u
a
n
ti
ty

fo
r
m
o
d
e
ll
in
g
.
In

li
n
e
w
it
h

st
a
n
d
a
rd

p
ra

c
ti
c
e
,
a
c
ti
v
e
-i
n
g
re
d
ie
n
t
q
u
a
n
ti
ti
e
s
a
re

e
x
p
re
ss
e
d

a
s
m
a
ss

(k
g
A
I
h
a
−

1
y
r−

1
).

2
O
n
ly

tw
o
o
rc
h
a
rd

s
e
x
h
ib
it
e
d

n
it
ro

g
e
n

su
rp

lu
s
a
n
d

tw
e
lv
e
p
h
o
sp

h
o
ru

s
su

rp
lu
s;

h
e
n
c
e
,
su

rp
lu
s
v
a
ri
a
b
le
s
w
e
re

o
m
it
te
d

fr
o
m

th
e
m
a
in

a
n
a
ly
si
s
m
o
d
e
ls
.

Id
e
a
ll
y
,
n
u
tr
ie
n
t
b
a
la
n
c
e
w
o
u
ld

in
c
lu
d
e
so

il
st
o
ck

s
a
n
d

b
e
e
st
im

a
te
d

d
y
n
a
m
ic
a
ll
y
[7
0
,
7
1
],

b
u
t
su

ch
d
a
ta

w
e
re

n
o
t
a
v
a
il
a
b
le
.

7



323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368

We employed the value-added approach for the desirable output, net value added
(NVA), which permits a natural interpretation of the EE scores. Since a typical mango
farming household consumes part of its production, the NVA from mango was cal-
culated incorporating the value of the fruit consumed. Following Kuosmanen and
Kortelainen [47], labour costs were not deducted from the NVA because they represent
wages and rents circulating within society, rather than costs of production. In contrast
to regions where collusive practices among orchard owners and financiers intensify
labour exploitation, thereby diminishing the overall societal benefits from farming,
such as those documented by Sacramento and Cañete [72] in the context of Philippine
mango fruit farming, the relatively competitive labour market in Makueni discourages
such collusion behaviours, ultimately ensuring better protection for orchard labourers.
Machinery depreciation and maintenance were not considered because smallholders in
developing countries such as Kenya typically use negligible amounts of machinery in
mango production.

All orchards were assumed to have a uniform selling price to ensure that variations
in economic performance arise from technical management rather than price differ-
ences. Based on insights gathered from key informants, the market price of mango
typically ranges between KES 15–30 kg−1 of fresh fruit, with fluctuations largely influ-
enced by seasonality. Given that all yield data reported in this study pertain to the
main (on-) season, when market prices are generally lower due to peak production, we
adopted a conservative pricing approach by using the lower bound of the price range
(KES 15 kg−1, corresponding to ≈ USD 0.117 kg−1 based on the exchange rate at the
time of the survey).

2.4 Empirical framework

2.4.1 Latent class stochastic metafrontier

To assess the efficiency with which orchard managers convert resources into desir-
able outputs while limiting undesirable by-products, we employ a production-frontier
approach rooted in classical efficiency theory [73]. In this framework, it is assumed
that each orchard manager uses the inputs optimally to achieve the maximum possible
output with minimum deleterious environmental impacts such as pollution or resource
depletion. We adopt an output-oriented perspective to evaluate how much orchard
managers can increase their desirable outputs without consuming additional inputs
or increasing deleterious by-products. From a sustainability perspective, an orchard
can be considered output inefficient while still having room to reduce environmen-
tal impacts. Pareto–Koopmans efficiency is achieved when no further improvements
in yield or net value added are possible without increasing input or environmental
degradation, thus balancing productivity, resource conservation, and broader societal
goals [74, 75]. The production frontier represents the highest possible production level
under similar technological and environmental conditions, and any deviation from this
frontier is attributed to inefficiency [76].

Smallholders rarely use homogeneous technologies due to resource and socio-
economic constraints, requiring a mechanism for accounting for potential technological

8



369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414

heterogeneity in their production. However, the qualitative classification of individu-
als into technology-specific groups based on observed and unobserved characteristics
is often a complex process that risks introducing bias, particularly when threshold
identification levels are non-trivial. To address this challenge, we employ a latent
class stochastic frontier (LCSF) procedure formalised in Greene [77] to estimate the
posterior probability, Π(i, j), of an orchard manager i’s membership in class j, con-
ditioned on observed (separating) variables si. The class assignment is usually based
on the largest Π(i, j) obtained for class j using Bayes rule and parametrised using a
multinomial logit function as:

Π(i, j) =
exp(si; Ωj)∑C
c=1 exp(si; Ωj)

(1)

The probability of membership in class j, is computed as:

Pr(i, j) =
2√

σ2
v|j + σ2

u|j

ϕ

 ξi|j√
σ2
v|j + σ2

u|j

Φ

−
(
σu|j

σv|j

)
ξi|j√

σ2
v|j + σ2

u|j

 , (2)

and the associated loglikelihood function is the weighted sum of the j-class likelihood
functions:

logL =

N1,2∑
i=1

log

 2∑
j=1

Π(i, j) · Pr(i | j)

 (3)

where N1,2 is the number of observations not in class 0 (i.e., those for whom j ∈
{1, 2}). The determination of the optimal number of classes is guided by the model’s
information criterion. The Bayesian Information Criterion is often preferred and the
model with the lowest absolute value is selected.

We first classified farmers as adopters or non-adopters of APM, according to the
criteria described in Owili et al. [30]. However, without an objective APM inten-
sity threshold to define cut-off levels, the number of distinct adopter subgroups and
the heterogeneity within the adopter class remained indeterminate. Consequently, we
applied the LCSF model exclusively to the subsample of APM adopters. The approach
identified two classes of APM adopters, which we classified as non-intensive and inten-
sive adopters. We continue this discussion in Section 3.3.1. Thus, together with the
non-adopter category, our sample consisted of three classes of orchard managers.

The ideal approach to controlling for selection bias in observational studies is to
use a randomised experiment so that all individuals have equal chances of assignment
to each treatment class. However, this approach was not feasible in this study due
to cost implications. To test for the presence of selectivity bias in the presence of
heterogeneous technologies, we employ the bias-corrected LCSF procedure proposed
by Dakpo et al. [78]. The approach uses a more efficient quadrature method within
the LCSF framework as an alternative to the maximum simulated likelihood approach
derived by Greene [79], and mitigates confounding bias from unobservables. In our
case, the coefficient of the selectivity variable RHO (ρ) was not significantly different

9



415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460

from zero at 5%, in the pooled and class-specific frontiers, indicating that our sample
does not suffer from selection bias. This implies that the standard LCSF suffices to
estimate regime-specific frontiers and efficiencies.

Following Aigner et al. [80] and Meeusen and van Den Broeck [81], we estimate
two output-oriented stochastic frontiers for each of the three classes of farmers: (i) a
production frontier that captures technical efficiency and (ii) a damage or pressure-
generating technology (PGT) function that assesses eco-efficiency, as:

lnYi|j = ln f j
(
Xik|j ,β

)
+ vi|j − ui|j , TEi|j = exp

(
−ui|j

)
,

lnNVAi|j = lnDj
(
Pik|j , τ

)
+ νi|j − µi|j , EEi|j = exp

(
−µi|j

)
PGT = {(NVA,P) ∈

RN+1
+ | NVA can be produced with P pressures},

(4)

where lnYi|j is the logarithm of mango yield per hectare for the ith orchard managed

by a farmer in the jth class, Xik|j is a 1 × K vector of normalised positive inputs,
and β is the coefficient vector of interest. Similarly, lnNVAi|j denotes the logarithm
of net value added, Pik|j a vector of normalised environmental pressures, and τ the
parameter vector. For each orchard i = 1, . . . , N and class j = 0, 1, 2, vi|j ∼ N(0, σ2

v)
and νi|j ∼ N(0, σ2

ν) are iid white-noise errors, independent of the inefficiency terms.

ui|j ≥ 0 ∼ N+
(
0, σ2

u

)
and µi|j ≥ 0 ∼ N+(0, σ2

µ) are half-normal inefficiency terms.
TEi|j ∈ [0, 1] and EEi|j ∈ [0, 1] denote technical and eco-efficiency, respectively.

The zero-observation problem is a common productivity analysis pitfall, particu-
larly in smallholder systems where farmers may not apply some inputs, resulting in a
large proportion of genuine zeros in the dataset. To accommodate this, we apply an
inverse hyperbolic sine (IHS) transformation, which, unlike the logarithm, is defined
for zero and negative values and naturally approximates ln(x) for a sufficiently large
x [82]. Since the IHS transformation can be adversely affected by the chosen unit
of measurement [83], following Aihounton and Henningsen [82], we perform several
tests to determine the appropriateness of the chosen units of measurement and the
corresponding scaling factors. The results are displayed in Table A2 of Appendix A.

We perform likelihood ratio (LR) tests for the deterministic kernels ln f j (•) and
lnDj (•) of Equation (4). In both cases, the test strongly rejects the Cobb-Douglas
specification in favour of the log-linear Translog functional form (see Table A1 of the
Appendix A). Although often associated with multicollinearity, the Translog specifi-
cation is flexible and has the ability to capture non-linearities in the regressors, allows
for potential substitutions among inputs, and places no constraints on returns to scale
[84]. All terms of Translog are normalised by their geometric means to allow the first-
order coefficients of the stochastic frontier to be interpreted as partial elasticities with
respect to the mango yields at the sample mean [85]. To satisfy the regularity condi-
tions, we impose monotonicity on all inputs using a three-step procedure developed
by Henningsen and Henning [86].

To account for technological heterogeneity across the three adoption classes and
enable meaningful benchmarking, we embed the LCSF within a metafrontier frame-
work. A common misconception is that the standard LCSF automatically provides
a common benchmark for all technology regimes, making efficiency scores directly

10



461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506

comparable. Many recent studies fall prey to this error by reporting cross-technology
efficiency comparisons from LCSF results. Since the estimation of LCSF implies tech-
nology heterogeneity, it follows, therefore, that a metafrontier estimation is required.
While LCSF is robust at uncovering unobserved “technology regimes”, it does not pro-
duce an enveloping frontier for benchmarking. Only by nesting class-specific frontiers
under a shared metafrontier can one validly compare efficiency scores across regimes
[87].

To justify the metafrontier estimation, we conduct generalised likelihood-ratio
(GLR) tests. The GLR test statistic follows a Chi-square distribution under the null
hypothesis and, in our case, is obtained as:

−2

(
log L

(
ln fM (•)

)
−

log L(ln f0(•)
)
+

N1,2∑
i=1

log

 2∑
j=1

Π(i, j) · Pr(i, j)

 ) (5)

where log L
(
ln f0(•)

)
is the loglikelihood of the non-adopters frontier model and

log L
(
ln fM (•)

)
is the loglikelihood of the metafrontier. The GLR tests strongly

rejected the null hypotheses, confirming that the technologies used among the
three adoption groups are heterogeneous and differ systematically (Table A1 of the
Appendix A). In other words, farmers in the three adoption groups have different
production possibility frontiers, making them directly incomparable. Empirically, this
suggests that estimating a metafrontier provides a better fit compared to the three
separate class-specific frontiers. By definition, a metafrontier is an overarching technol-
ogy that encompasses multiple class-specific frontiers, forming a common benchmark
technology available to the whole industry and is similar for all farmers [87, 88].

In a two-stage procedure, we first estimate the class-specific frontiers as in
Equation (4). In the second stage, the predicted fitted values ln f̂ j(•) and ln D̂j(•)
from the three groups are pooled to construct metafrontiers, following Huang et al.
[89]:

ln f̂ j(Xi|j ,β) = ln fM (Xi|j ,β)− uM
i|j + vMi|j , MTEi|j = e−uMi|j × e−ui|j ,

ln D̂j(Pi|j , τ ) = lnDM (Pi|j , τ )− µM
i|j + νMi|j , MEEi|j = e−µMi|j × e−µi|j ,

(6)

where Equation (6) has the usual properties of the frontiers given in Equation (4);
however, uM

i|j and µM
i|j denote the non-negative technology gap ratio (TGR) and the

PGT gap ratio (PTGR) component for production and eco-efficiency frontiers, respec-
tively, and are distributed as uM

i ≥ 0 ∼ N+(0, σ2
u) and µM

i|j ≥ 0 ∼ N+(0, σ2
µ). In

this case, vMi|j ∼ N(0, σ2
v) and νMi|j ∼ N(0, σ2

ν) may not be iid and are therefore
assumed to be asymptotically normally distributed. MTE and MEE refer to the meta-
technical efficiency and meta-eco-efficiency scores, respectively. The MTE and MEE
scores are directly comparable between the various technology classes relative to the
metafrontier.

11



507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552

In mango production, TGR represents the ratio of predicted mango yield of a class-
specific production frontier to the potential mango yield given by the metafrontier,
and thus shows the position of the class-specific production frontier relative to the
frontier achievable by the industry as a whole. Thus, the metafrontier allows for the
segregation of production inefficiencies into those caused by poor agronomic practices
and those caused by technology gaps within the industry. In this study, technology gaps
quantify the extent to which various pest management technologies deviate from global
best practice and therefore can inform policy interventions aimed at the promotion
of best-performing pest management strategies. These gaps arise from the choice of a
particular pest management technology from the various technologies available to the
industry as a whole, depending on their accessibility to the individual farmers and the
rates of technology adoption.

2.4.2 Environmental adjustment procedure

To create a multidimensional efficiency score, we compute environmentally adjusted
composite measures for both efficiency and technology gaps to create adjusted meta-
technical efficiency MTEadj

i|j and adjusted technology gap ratios TGRadj
i|j , respectively

as:

MTEadj
i|j =

∏
q∈{uM

i|j , ui|j , µ
M
i|j , µi|j}

e−q = exp
[
−
(
uM
i|j + ui|j + µM

i|j + µi|j
)]
,

TGRadj
i|j =

∏
r∈{µM

i|j , u
M
i|j}

e−r = exp
[
−
(
µM
i|j + uM

i|j
)]
.

(7)

In essence, we make MTEadj
i|j depend on MTEi|j and MEEi|j , whereas TGRadj

i|j
depends on TGRi|j and PTGRi|j . The multiplicative operator endows both com-
posites with several desirable properties. First, strict monotonicity holds in every
argument so that an improvement in either MTEi|j or MEEi|j raises MTEadj

i|j ,

and an improvement in either TGRi|j or PTGRi|j raises TGRadj
i|j , ceteris paribus.

Second, the products are bilinear, symmetric and homogeneous of degree two, ensur-
ing that no single component is given precedence over its counterpart. Third, the
aggregation rule is non-compensatory and effectively penalises poor performance so
that a deficiency in one component cannot be masked by superiority in the other.
As any component approaches zero, the corresponding composite measure is driven
sharply downwards, constraining overall performance by the weakest link. Conse-
quently, the formulation is consistent with the notion that sustainable performance
enforces holistic progress both in technical efficiency and eco-efficiency, while aiming
at closing the technology gaps as well. Finally, since each component index satis-
fies: MTEi|j ,MEEi|j ,TGRi|j ,PTGRi|j ∈ [0, 1]; hence, the adjusted scores MTEadj

i|j

and TGRadj
i|j are properly bounded within the unit interval, rendering interpretation

straightforward.

12



553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598

For policy purposes, we model drivers of environmentally adjusted inefficiency,
MTIEadj

i|j , using a fractional probit as:

MTIEadj
i|j = G(δ;Zi|j) = δ0 + δ1Z1i|j + · · ·+ δnZni|j

MTIEadj
i|j = 1−MTEadj

i|j

(8)

where G(·) is the Bernoulli specification of the quasi-maximum likelihood estimator
of the standard normal cumulative density function with a probit link, (δ0|j , · · · , δn|j)
are the parameters of interest, and (Z1i|j + · · · + Zni|j) are exogenous variables that
are hypothesised to influence orchard-level adjusted inefficiency. Since it is widely
recognised that multi-step estimation procedures are generally associated with biased
standard errors [90, 91], we apply a bootstrap procedure following Simar and Wilson
[92] and Simar and Wilson [93] with 1000 replications to correct the standard errors.

2.4.3 Inverse-probability-weighted regression adjustment for
multi-valued treatment effect estimation

To determine the effect of APM transition and intensification regimes, we subject
the MTEadj

i|j scores to an inverse-probability-weighted regression adjustment (IPWRA)

procedure. In this framework, for each adoption level A ∈ {0, 1, 2} where 0, 1 and 2
denote non-adopters, non-intensive adopters and intensive adopters, respectively, the
average potential outcome (POMean) is estimated as:

Θ̂(a) = E

θ̂(a, Zi) +
1 [Ai = a]

(
MTEadj

i|j − θ̂(a, Zi)
)

P̂ ra(Zi)

 . (9)

Here, θ̂(a, Zi) denotes the predicted outcome under treatment level a and is mod-

elled using a fractional probit specification such that θ̂(a, Zi) = Φ
(
Ziδ̂a

)
, with Φ(·)

representing the standard normal cumulative distribution function. The probability of
receiving treatment level a is estimated using a multinomial logit model. The inclusion
of both an outcome regression and inverse probability weighting guarantees a consis-
tent estimation of treatment effects provided that either of the models is correctly
specified and hence doubly robust. The treatment effects are then determined by com-
paring the estimated POMeans at different treatment levels, to obtain the average
treatment effect (ATE) and the average treatment effect on the treated (ATT) as.

ATE1,0 = E[MTEadj
i|1 −MTEadj

i|0 ] ATT1,0 = E[MTEadj
i|1 −MTEadj

i|0 | Ai = 1]

ATE2,0 = E[MTEadj
i|2 −MTEadj

i|0 ] ATT2,0 = E[MTEadj
i|2 −MTEadj

i|0 | Ai = 2]

ATE2,1 = E[MTEadj
i|2 −MTEadj

i|1 ] ATT2,1 = E[MTEadj
i|2 −MTEadj

i|1 | Ai = 2]

(10)

As a robustness check for selectivity bias from observables, we compare the IPWRA
estimates with those obtained from a propensity score matching (PSM) and regression

13



599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644

adjustment (RA). Whereas RA depends on a correct specification of the outcome

model θ̂(a, Zi), the PSM relies on propensity scores P̂ ra(Zi) to match an orchard
manager in a treatment class with an orchard manager from the control group based
on similar characteristics. In this case, a propensity score is a conditional probability
of an orchard manager being assigned to a treatment class based on a vector of their
observed covariates. We impose both covariate balance and common support to ensure
that for every combination of covariates, there is a non-zero probability of being treated
and untreated. To determine the potential influence of unobserved confounding in
both models, we perform a sensitivity analysis to assess the stability and validity of
our treatment effect estimates using Oster’s δ [94]. Oster’s δ indicates how strong
the influence of unobservables would need to be, relative to the influence of observed
covariates, to reduce the estimated treatment effect to zero.

It has been shown that the benefits of adopting APM do not apply uniformly
in all contexts [see 95]. To uncover heterogeneities in treatment effects, we use the
doubly robust conditional average treatment effect (DRCATE) visualisation procedure
proposed by Lee et al. [96]. This procedure models the conditional average treatment
effect (CATE) function using an augmented inverse probability weighting estimator
of a covariate of interest by combining a propensity score model with a local linear
regression for the POMeans and ATEs.

3 Results and discussion

3.1 Adoption of agro-ecological fruit fly management options

Figure 2 illustrates the distribution of adopters across three broad categories of APM
practices, including habitat management, orchard sanitation, and reactive options.
Adoption rates were moderate, with habitat management being the most adopted
category, followed by orchard sanitation and reactive options. The most common prac-
tices were regular scouting and monitoring, the management of alternate hosts, and
male annihilation, which were adopted by at least half of the respondents.

Within habitat management category, adoption was highest for regular scouting
and monitoring (53.5%), management of alternate hosts (50.2%), and inter-tree raking
(43.3%), suggesting a preference for ecologically grounded practices that are relatively
straightforward to implement. In contrast, adoption of more specialised practices such
as intercropping with non-host crops and trap-cropping with passion fruit remained
limited. For orchard sanitation, feeding infested fruit to livestock (45.6%) and deep
burying of infested fruits (35.2%) were the most widely adopted, whereas technical
measures like solarisation and the use of an augmentorium were negligible. The pattern
was even more pronounced in reactive control strategies in which male annihilation
(50.2%) was the only widely adopted method, while all other approaches, including
the use of biopesticides and botanical sprays, showed limited uptake. These findings
indicate the uneven diffusion of APM options, with adoption strongly skewed toward
practices that are familiar, locally adaptable, and possibly less resource-intensive.

14



645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690

2.1

13.0 13.4

43.3
50.2 53.5

0
10

20
30

40
50

60
Sh

ar
e 

of
 a

do
pt

er
s 

(%
)

Tr
ap

-c
ro

pp
in

g 
wi

th
 p

as
sio

n 
fru

it

Ea
rly

 h
ar

ve
st

in
g

In
te

rc
ro

pp
in

g 
wi

th
 n

on
-h

os
t c

ro
p

In
te

r-t
re

e 
ra

kin
g

M
an

ag
em

en
t o

f a
lte

rn
at

e 
ho

st
s

Re
gu

la
r s

co
ut

in
g 

an
d 

m
on

ito
rin

g

(a) Habitat management

0.2 3.2
6.9

17.1

35.2

45.6

Us
e 

of
 a

n 
au

gm
en

to
riu

m

So
la

ris
at

io
n 

wi
th

 s
pe

cia
l b

ag
s

Bu
rn

in
g 

in
fe

st
ed

 fr
ui

ts

Co
m

po
st

in
g 

in
fe

st
ed

 fr
ui

ts

De
ep

 b
ur

yin
g 

in
fe

st
ed

 fr
ui

ts

Fe
ed

in
g 

in
fe

st
ed

 fr
ui

t t
o 

liv
es

to
ck

(b) Orchard sanitation

0.5 0.5 1.6 4.2

14.4

50.2

So
il i

no
cu

la
tio

n 
wi

th
 b

io
pe

st
ici

de

As
h/

ba
kin

g 
po

wd
er

 a
nd

 to
ba

cc
o

Sp
ot

 s
pr

ay
 o

f f
oo

d 
ba

its

Sp
ra

yin
g 

bo
ta

ni
ca

l p
es

tic
id

es

Sm
ok

in
g 

he
rb

s 
an

d 
du

ng

M
al

e 
an

ni
hi

la
tio

n

(c) Reactive options

Fig. 2: APM adoption rates by category of practice. Source: Survey Data (2023).

3.2 Characteristics of adopters and non-adopters of
agro-ecological pest management

Table 2 shows the characteristics of the surveyed orchard managers, disaggregated
by their APM adoption classes. The average orchard size was 0.56 ha, confirming
that most of the producers in the sample are smallholders. On average, non-intensive
adopters used slightly more land than intensive adopters. As expected, labour con-
sumption was higher among intensive adopters than among non-intensive adopters.
Intensive adopters implemented more labour-intensive practices and also applied
slightly more inputs such as fertilisers and manure, which could have consumed
additional labour due to input application.

The intensive class exhibited greater specialisation, as indicated in their average
tree density of 164 trees ha−1, compared to 108 and 136 trees ha−1 among non-
adopters and non-intensive adopters, respectively, on average. On average, adopters
applied 0.95 L ha−1 more insecticides (both organic and inorganic combined) than
conventional farmers. This was expected, as organic pesticides are typically applied
in larger quantities than their inorganic counterparts used by non-adopters. Among
adopters, non-intensive users applied 1.13 L ha−1 of organic insecticides on average,
compared to 0.94 L ha−1 by intensive adopters, which explains the higher overall
insecticide use among adopters. As expexted a priori, intensive adopters applied 0.23
L ha−1 less inorganic insecticide than non-adopters. This finding is consistent with
the results of Midingoyi et al. [29] and Mwungu et al. [38], who found that adopters
of integrated pest management used significantly less synthetic pesticides than non-
adopters. These results suggest that integrating multiple pest control strategies can
effectively reduce reliance on the often-costly chemical pesticides. In contrast, non-
intensive adopters used slightly more inorganic insecticides than non-adopters, with
an average difference of 0.06 L ha−1.

15



691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736

T
a
b
le

2
:
C
h
a
ra
ct
er
is
ti
cs

o
f
p
o
o
le
d
,
n
o
n
-a
d
o
p
te
rs
,
n
o
n
-i
n
te
n
si
v
e,

a
n
d
in
te
n
si
v
e
a
d
o
p
te
rs

o
f
A
P
M

P
o
o
le
d

N
o
n
-a

d
o
p
te

r
s

N
o
n
-i
n
te

n
si
v
e

In
te

n
si
v
e

V
a
r
ia
b
le

P
a
r
a
m

e
te

r
M
ea

n
(S

D
)

M
ea

n
(S

D
)

M
ea

n
(S

D
)

M
ea

n
(S

D
)

D
e
p
e
n
d
e
n
t
v
a
ri
a
b
le
s

M
a
n
g
o
y
ie
ld

(k
g
h
a
−
1
y
r−

1
)

7
5
1
9
.5
7
(8
3
8
7
.3
0
)

6
2
1
2
.1
3
(5
6
6
1
.2
7
)

6
7
0
9
.1
1
(7
3
5
8
.0
6
)

1
0
2
4
8
.6
8
(1
1
5
3
6
.1
6
)

N
et

v
a
lu
e
a
d
d
ed

(’
0
0
K
E
S
h
a
−
1
y
r−

1
)

1
0
9
9
.4
2
(1
2
5
8
.4
8
)

9
0
8
.1
5
(8
5
1
.0
9
)

9
7
4
.6
9
(1
1
0
2
.3
0
)

1
5
0
5
.1
0
(1
7
3
2
.4
2
)

In
p
u
ts

(T
ec
h
n
ic
a
l
effi

ci
en

cy
)

L
a
n
d
(h

a
)

β
L
A

0
.5
6
(0
.5
1
)

0
.5
5
(0
.4
7
)

0
.6
1
(0
.6
4
)

0
.5
2
(0
.4
0
)

L
a
b
o
u
r
(m

a
n
d
a
y
s
h
a
−
1
y
r−

1
)

β
L
B

1
5
.8
3
(2
3
.4
6
)

1
4
.2
4
(2
0
.5
5
)

1
3
.7
5
(1
4
.5
0
)

2
0
.3
1
(3
2
.7
4
)

F
er
ti
li
se
rs

(k
g
h
a
−
1
y
r−

1
)†

β
F
E

3
.3
2
6
(8
.2
2
6
)

3
.7
5
0
(9
.2
8
0
)

2
.0
9
0
(3
.2
8
5
)

4
.0
0
4
(9
.9
5
4
)

In
se
ct
ic
id
es

(k
g
h
a
−
1
y
r−

1
)†

‡
β
IN

2
.2
9
3
(4
.7
5
1
)

1
.7
3
4
(2
.2
3
7
)

2
.9
2
2
(7
.1
3
9
)

2
.4
4
3
(4
.2
2
6
)

F
u
n
g
ic
id
es

(k
g
h
a
−
1
y
r−

1
)†

β
F
U

2
.0
2
3
(2
.7
8
9
)

1
.8
7
3
(2
.2
9
3
)

2
.0
5
5
(2
.6
8
5
)

2
.2
0
7
(3
.4
7
4
)

M
a
n
u
re

(k
g
h
a
−
1
y
r−

1
)†

β
M

A
2
8
7
9
.9
8
9
(6
8
2
.2
9
7
)

2
6
6
4
.7
5
4
(6
5
5
.0
9
2
)

1
9
7
2
.9
3
0
(3
3
0
.5
2
0
)

4
1
3
5
.1
3
8
(9
3
3
.8
0
8
)

In
p
u
ts

(E
co

-e
ffi
ci
en

cy
)

C
F

(k
g
C
O

2
eq

h
a
−
1
y
r−

1
)

τ G
H
G

1
3
5
5
3
.9
1
(4
2
0
8
4
.3
0
)

1
2
2
7
0
.5
8
(4
0
3
9
5
.3
6
)

7
6
4
1
.3
9
(2
1
0
1
8
.9
2
)

2
1
5
6
2
.9
1
(5
7
3
0
2
.2
5
)

N
d
efi

ci
ts

(k
g
h
a
−
1
y
r−

1
)†

τ N
D

4
3
7
.6
3
0
(8
2
7
.0
3
1
)

4
0
1
.8
2
0
(7
9
2
.6
5
1
)

3
1
5
.1
4
6
(4
2
1
.3
6
5
)

6
1
6
.8
4
4
(1
1
1
9
.6
7
0
)

P
d
efi

ci
ts

(k
g
h
a
−
1
y
r−

1
)†

τ P
D

6
6
.3
6
2
(1
3
1
.1
4
2
)

6
4
.2
7
4
(1
4
3
.1
6
8
)

4
6
.8
8
3
(6
2
.6
3
1
)

8
9
.6
6
3
(1
6
0
.0
0
2
)

T
o
x
ic
it
y
(k
g
A
I
h
a
−
1
y
r−

1
)†

τ T
O
X

3
.0
4
1
(6
.0
3
0
)

2
.7
1
0
(5
.7
6
8
)

3
.1
8
6
(4
.7
4
9
)

3
.3
6
6
(7
.4
5
4
)

E
n
er
g
y
d
efi

ci
ts

(M
J
h
a
−
1
y
r−

1
)

τ E
N

1
4
3
8
.6
4
(3
2
2
6
.2
9
)

1
3
2
9
.7
5
(3
0
9
0
.0
8
)

1
0
0
8
.7
3
(1
5
7
0
.5
7
)

2
0
4
3
.4
4
(4
4
1
8
.5
1
)

S
p
ec
ia
li
sa
ti
o
n
(p

ro
p
o
rt
io
n
)

τ S
P

0
.3
3
(0
.2
8
)

0
.2
7
(0
.2
4
)

0
.3
4
(0
.3
1
)

0
.4
1
(0
.3
0
)

O
b
se
rv
a
ti
o
n
s

4
1
8

1
7
3

1
2
5

1
2
0

N
o
te
s:

*
,
*
*
a
n
d

*
*
*
d
e
n
o
te

si
g
n
ifi
c
a
n
c
e
a
t
th

e
1
0
,
5
a
n
d

1
%

le
v
e
ls
,
re
sp

e
c
ti
v
e
ly
.
V
a
lu
e
s
in

p
a
re
n
th

e
se
s
a
re

st
a
n
d
a
rd

d
e
v
ia
ti
o
n
s.

A
b
b
re
v
ia
ti
o
n
s:
—

A
I,

a
c
ti
v
e
in
g
re
d
ie
n
t;

C
F
,
c
a
rb

o
n

fo
o
tp

ri
n
t;

C
O

2
e
q
,
c
a
rb

o
n

d
io
x
id
e
e
q
u
iv
a
le
n
t;

M
J
,
m
e
g
a
jo
u
le
s.

†
V
a
ri
a
b
le
s
m
a
rk

e
d

w
it
h

th
is

sy
m
b
o
l
w
e
re

tr
a
n
sf
o
rm

e
d

u
si
n
g
th

e
in
v
e
rs
e
h
y
p
e
rb

o
li
c
si
n
e
(I
H
S
)
to

p
re
se
rv

e
o
b
se
rv
a
ti
o
n
s
w
it
h

g
e
n
u
in
e
z
e
ro

v
a
lu
e
s
a
n
d

su
b
se
q
u
e
n
tl
y
re
-s
c
a
le
d

fo
r
a
n
a
ly
si
s.

In
o
rd

e
r
to

m
it
ig
a
te

a
n
y
p
o
te
n
ti
a
l
d
is
to

rt
io
n
a
ry

e
ff
e
c
ts

o
f
th

e
IH

S
tr
a
n
sf
o
rm

a
ti
o
n
,
in
p
u
ts

u
se
d

in
th

e
te
ch

n
ic
a
l
e
ffi
c
ie
n
c
y
e
st
im

a
ti
o
n

w
e
re

re
sc
a
le
d

to
c
e
n
ti
g
ra

m
s
(c
g
),

w
h
il
e

th
o
se

in
th

e
e
c
o
-e
ffi
c
ie
n
c
y

e
st
im

a
ti
o
n

w
e
re

re
sc
a
le
d

to
h
e
c
to

g
ra

m
s
(h

g
)
(r
e
fe
r
to

T
a
b
le

A
2

fo
r
d
e
ta

il
s)
.

‡
T
h
is

v
a
ri
a
b
le

is
a

c
o
m
p
o
si
te

o
f
b
o
th

o
rg

a
n
ic

a
n
d

in
o
rg

a
n
ic

in
se
c
ti
c
id
e
s.

T
h
e
a
v
e
ra

g
e
e
x
ch

a
n
g
e
ra

te
fo
r
th

e
re
fe
re
n
c
e
p
e
ri
o
d

(D
e
c
e
m
b
e
r
2
0
2
2
to

M
a
rc
h

2
0
2
3
)
w
a
s
K
E
S

1
2
8
U
S
D

−
1
.
S
o
u
r
c
e
:
S
u
rv

e
y

D
a
ta

(2
0
2
3
).

16



737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782

In line with expectations, the amount of carbon footprints and energy deficits
were highest among the intensive class, due to their higher output levels and input
consumption, respectively. It is well established that the level of output is positively
correlated with the level of emissions [97]. Intensive adopters recorded the highest rates
of nutrient depletion, which could be due to their high nutrient efficiency. Rapid uptake
of nutrients by high-yielding trees can lead to net nutrient mining if replenishment
does not fully match crop demand. Non-intensive adopters had the lowest nutrient
depletion rates.

On average, adopters achieved significantly higher mango yields (2,231kg ha−1

more than non-adopters) which translated into substantially higher net value added,
amounting to an additional KES 32,633 ha−1 (approx. USD 255.7 ha−1). On the
other hand, intensive and non-intensive adopters recorded 4036kg ha−1 (approx. KES
60,548 ha−1 or USD 473.8 ha−1) and 496.98 kg ha−1 (approx. KES 7,455 ha−1 or
USD 58.2 ha−1), respectively, more than non-adopters. This is in line with previous
studies that have reported increased net income from the adoption of sustainable fruit
fly management practices [29, 34–37].

3.3 Empirical results

3.3.1 Class-specific stochastic frontier elasticities

Table 3 shows the parameter estimates of the class-specific stochastic frontier for the
technical efficiency model. The coefficient of the gamma (γ) variable approximates
unity for all models, indicating that the proportion of total error variance due to
inefficiency is relatively high. This suggests that most of the deviation from the frontier
is due to inefficiency and justifies the use of the more complex stochastic frontier
procedure over simple alternatives such as ordinary least squares. Monotonicity holds
adequately for all inputs in all models (see Table A3 in Appendix A).

The LCSF identified two distinct groups among APM-adopting orchard managers.
Attempts to estimate models with additional classes failed to converge, indicating that
the two-class model was optimal and at saturation [78]. The average posterior prob-
abilities of class membership for intensive and non-intensive orchard managers based
on the LCSF is 84% and 89%, respectively. The coefficient of the separating variable
“number of APM practices adopted” is negative, implying that the likelihood of being
assigned to Class 1 decreases with higher uptake of APM practices. This suggests
that Class 1 and Class 2 can be broadly categorised as non-intensive and intensive
adopters, respectively. To validate this classification, we analysed the extent of adop-
tion of APM within each predicted class. The average intensity of APM adoption
among all adopters was 24.5%, equivalent to about four of the eighteen practices con-
sidered. The non-intensive class (Class 1) adopted slightly below this average at 22%,
about four practices, whereas the intensive class (Class 2) adopted above the average
at 27%, translating to approximately five practices.

Class-specific frontier estimates for technical efficiency show substantial het-
erogeneity in production technologies, which reveals the limitations of uniform,
one-size-fits-all assumptions of the production environment when promoting sustain-
able agriculture, particularly in smallholder systems. In fact, of the 28 elasticities,

17



783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828

Table 3: Estimates of class-specific stochastic frontier models (Translog) for
technical efficiency

Single class SF Two class LCSF

Class 1 Class 2
Non-adopters (Non-intensive) (Intensive)

Parameter Coef. (SE) Coef. (SE) Coef. (SE)

β0 0.048 (0.055) -0.100 (0.107) 0.357*** (0.061)
βLA -0.556*** (0.121) -0.643*** (0.134) -0.245* (0.144)
βFE 0.157** (0.068) 0.693*** (0.135) 0.052 (0.072)
βFU 0.191*** (0.041) -0.013 (0.129) -0.300*** (0.101)
βIN 0.275** (0.126) 0.263** (0.118) 0.240** (0.104)
βLB 0.047 (0.074) -0.609*** (0.138) 0.537*** (0.119)
βMA -0.044 (0.051) -0.816*** (0.160) -0.139** (0.070)
βLA2 -0.434 (0.321) 0.331 (0.238) 0.807*** (0.273)
βFE2 -0.115* (0.065) -0.504*** (0.125) -0.261*** (0.046)
βFU2 -0.208** (0.083) -0.321*** (0.098) -0.096 (0.069)
βIN2 -0.059 (0.128) -0.147** (0.067) -0.143*** (0.052)
βLB2 0.520*** (0.154) -0.623*** (0.231) -0.811*** (0.262)
βMA2 0.214** (0.087) 0.700** (0.288) -0.036 (0.065)
βLA × FE 0.402*** (0.052) -0.051 (0.102) 0.405*** (0.086)
βLA × FU -0.088 (0.084) -0.094 (0.126) -0.127 (0.102)
βLA × IN -0.033 (0.173) 0.504*** (0.143) -0.239* (0.123)
βLA × LB 0.104 (0.153) 0.012 (0.167) 0.239* (0.138)
βLA × MA -0.150 (0.128) 0.187 (0.221) 0.563*** (0.124)
βFE × FU 0.056 (0.043) 0.084 (0.062) 0.231*** (0.053)
βFE × IN -0.043 (0.032) -0.172** (0.077) 0.055 (0.038)
βFE × LB 0.210*** (0.065) -0.071 (0.153) -0.217** (0.106)
βFE × MA -0.013 (0.043) -0.037 (0.116) -0.024 (0.040)
βFU × IN 0.020 (0.082) 0.172*** (0.060) -0.044 (0.031)
βFU × LB -0.132 (0.104) -0.020 (0.138) 0.362*** (0.098)
βFU × MA 0.056 (0.038) 0.106 (0.107) 0.243*** (0.063)
βIN × LB 0.163 (0.127) 0.234** (0.103) -0.051 (0.089)
βIN × MA -0.175*** (0.062) 0.137* (0.074) -0.103** (0.044)
βLB × MA -0.255* (0.135) 0.974*** (0.182) -0.140 (0.114)

σu 0.340 —
σv 0.000 —
σ 0.583 —
γ 1.000 —
logL -11.627 -22.391

Separating variables
Constant 1.300** (0.534)
APM practices adopted -4.540** (1.993)
APM intensity tree−1 ha−1 -0.075 (0.149)

Posterior probability 1.000 0.891 0.836
APM intensity 0.000 0.223 0.267
Observations 173 125 120

Notes: *, **, and *** denote significance at the 10%, 5%, and 1% levels, respec-
tively. Abbreviations:—FE, fertilisers; FU, fungicides; IN, insecticides; LA, land;
LB, labour; LCSF, latent class stochastic frontier; MA, manure; and SF, stochastic
frontier. Values in parentheses are standard errors. Source: Survey Data (2023)

18



829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874

only 3 (insecticide, land-fungicide- and fertiliser-manure interactions) were consistent
in all three classes. The expected diminishing marginal returns mostly hold for the
non-adopter class, with orchard size (land) as the only input exhibiting an inverse
elasticity with respect to mango yield at the sample mean. Whereas initial increments
in land, labour and manure decrease the yields at the sample average for the non-
intensive class, all else equal, fungicides and manure appear to reduce yields among
the intensive class.

The results of the class-specific frontier parameters for the eco-efficiency model are
presented in Table 4. The results show a high degree of variation in the magnitude and
direction of the environmental impact categories in the three models. The squared and
interaction terms show strong non-linear relationships and interdependencies among
environmental indicators, suggesting that trade-offs and complementarities between
inputs critically affect eco-efficiency. As expected, the elasticity of net value added
with various environmental impact categories as well as their interactions are mostly
negative across the three classes at the sample mean. This suggests that these delete-
rious impact categories reduce the net value added and should therefore be minimised
to improve the sustainability of the orchards. Nutrient deficits exhibit the largest elas-
ticities. In particular, the non-intensive class is associated with large coefficients for
phosphorus and nitrogen deficits, suggesting complex nutrient management challenges
compared to the other classes.

3.3.2 Latent class stochastic metafrontier elasticities

Our primary focus was to assess how APM transitions and intensification create
technological heterogeneity and how the transition influences orchard-level efficiency
outcomes in smallholder contexts. To this end, we focus the ensuing discussion on
the metafrontier results. Table 5 presents the parameter estimates of the latent class
stochastic metafrontier model for both technical and eco-efficiency. The estimated
input elasticities suggest diminishing marginal returns across most inputs, consistent
with the quasi-concavity condition of the assumed production technology.

The results indicate that, at the sample mean, land exhibits a negative elasticity,
indicating a 0.43% decline in mango yield for every percentage increase in the orchard
area. However, beyond a certain point, output eventually increase with further land
expansion. An extensive review by Menzel and Lagadec [98] found an inverse relation-
ship between yields per tree and tree density, alongside a positive association between
total yields and tree density. Similarly, Zhang et al. [99] reported a positive relationship
between yield and tree density at levels comparable to those observed in Makueni and
noted diminishing marginal returns at higher densities. These findings suggest that
although individual tree productivity decreases with increasing density, overall land
productivity improves, implying that intensive land use through higher tree density can
enhance total yields. The observed non-linear pattern may also reflect initial inefficien-
cies and resource constraints faced by smallholders as orchard size increases, followed
by improved management once farms expand, when commercialisation becomes more
feasible and economies of scale can be exploited.

Labour exhibits an elasticity of 0.20 at the sample mean, implying that a
one-percent increase in man-days devoted to orchard tasks increases mango yield by

19



875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920

Table 4: Estimates of class-specific stochastic frontier models (Translog)
for eco-efficiency

Single class SF Two class LCSF

Class 1 Class 2
Non-adopters (Non-intensive) (Intensive)

Parameter Coef. (SE) Coef. (SE) Coef. (SE)

τ0 −2.564*** (0.803) 2.474*** (0.867) 0.207 (0.833)
τCF 2.516*** (0.786) −1.347 (1.168) 1.548 (1.411)
τND 8.087*** (2.376) 6.476 (7.118) −1.415 (3.666)
τPD −3.643*** (1.145) −10.877* (6.297) 0.109 (3.005)
τTOX 1.365*** (0.388) −0.498 (0.490) 0.914** (0.406)
τEN −2.233*** (0.526) 1.339** (0.544) −0.284 (0.667)
τSP −0.065** (0.032) −0.226*** (0.050) 0.149** (0.061)
τCF2 −1.413** (0.697) −1.107 (1.231) −2.335** (1.101)
τND2 −6.805** (2.964) −70.461** (28.799) 5.210*** (1.943)
τPD2 −0.981** (0.471) −47.625*** (17.376) −0.589 (0.879)
τTOX2 −0.126* (0.071) −0.204** (0.080) −0.108* (0.065)
τEN2 −1.397*** (0.280) −0.926*** (0.191) −0.304 (0.282)
τSP2 −0.003 (0.002) 0.001 (0.003) 0.005** (0.002)
τCF×ND −3.508*** (1.346) 3.603 (3.748) −0.592 (2.884)
τCF×PD 1.927** (0.908) −0.700 (2.770) 1.250 (2.113)
τCF×TOX −0.573* (0.316) −0.125 (0.232) −0.173 (0.252)
τCF×EN 0.786** (0.324) −0.388 (0.359) −0.153 (0.489)
τCF×SP −0.041 (0.029) 0.002 (0.056) −0.073** (0.037)
τND×PD 2.954*** (0.887) 61.216*** (22.883) −0.962 (1.146)
τND×TOX −1.475*** (0.563) 8.042*** (2.694) −0.378 (2.059)
τND×EN 2.799*** (0.519) −5.560* (2.844) −0.240 (1.429)
τND×SP 0.156*** (0.044) 0.603*** (0.146) −0.161 (0.232)
τPD×TOX 0.063 (0.200) −7.560*** (2.231) −0.585 (1.640)
τPD×EN −0.705*** (0.180) 4.869** (2.308) 0.602 (1.135)
τPD×SP −0.062*** (0.019) −0.346*** (0.107) 0.073 (0.188)
τTOX×EN 0.663*** (0.217) 0.310*** (0.107) 0.333*** (0.124)
τTOX×SP −0.004 (0.011) −0.008 (0.016) 0.008 (0.010)
τEN×SP 0.008 (0.016) −0.026 (0.025) 0.020 (0.021)

σµ 0.007 0.003 0.004
σν 0.000 0.000 0.000
σ 0.082 0.058 0.061
γ 0.984 0.999 0.994
logL 296.005 261.517 242.819
Observations 173 125 120

Notes: *, **, and *** denote significance at the 10%, 5%, and 1% levels, respec-
tively. Abbreviations:—CF, carbon footprint; LCSF, latent class stochastic frontier;
ND, nitrogen deficit; PD, phosphorus deficit; TOX, pesticide toxicity; EN, energy
balance; SF, stochastic frontier; SP, specialisation. Values in parentheses are stan-
dard errors. Source: Survey Data (2023)

0.20%. The marginal product of labour rises rather than falls within the observed
range, indicating convexity of the production function for labour use. In practical
terms, once a basic labour threshold is met, each extra man-day allows more thorough
canopy management, quicker detection and correction of pest or nutrient problems,
and more precise timing of cultural operations, all of which reinforce one another.

20



921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966

Table 5: Estimates of latent class stochastic metafrontier models
(Translog) for technical efficiency and eco-efficiency

Technical efficiency Eco-efficiency

Parameter Coef. (SE) Parameter Coef. (SE)

β0 0.271*** (0.030) τ0 1.284*** (0.150)
βLA -0.431*** (0.063) τCF 0.171 (0.281)
βFE 0.063* (0.036) τND -1.397*** (0.352)
βFU 0.144*** (0.043) τPD -1.109*** (0.231)
βIN 0.173*** (0.044) τTOX 0.515*** (0.090)
βLB 0.200*** (0.056) τEN 0.125 (0.133)
βMA -0.129*** (0.050) τSP -0.017 (0.012)
βLA2 0.338*** (0.125) τCF2 -1.348*** (0.303)
βFE2 -0.074** (0.030) τND2 3.005*** (0.277)
βFU2 -0.188*** (0.041) τPD2 -0.395*** (0.115)
βIN2 -0.026 (0.027) τTOX2 -0.103*** (0.024)
βLB2 0.506*** (0.079) τEN2 -0.684*** (0.087)
βMA2 0.201*** (0.051) τSP2 -0.002*** (0.001)
βLA×FE 0.173*** (0.038) τCF×ND 0.289 (0.357)
βLA×FU -0.112** (0.052) τCF×PD 0.862*** (0.213)
βLA×IN 0.118** (0.057) τCF×TOX -0.144** (0.072)
βLA×LB 0.177** (0.073) τCF×EN -0.085 (0.126)
βLA×MA 0.040 (0.052) τCF×SP -0.048*** (0.011)
βFE×FU 0.077*** (0.019) τND×PD 0.432*** (0.152)
βFE×IN -0.018 (0.021) τND×TOX -0.497*** (0.140)
βFE×LB 0.142*** (0.039) τND×EN 0.252 (0.164)
βFE×MA -0.035 (0.025) τND×SP 0.070*** (0.016)
βFU×IN 0.006 (0.025) τPD×TOX -0.073 (0.057)
βFU×LB -0.071 (0.049) τPD×EN -0.001 (0.092)
βFU×MA 0.031 (0.027) τPD×SP -0.018** (0.008)
βIN×LB 0.008 (0.066) τTOX×EN 0.256*** (0.038)
βIN×MA -0.017 (0.019) τTOX×SP -0.006* (0.003)
βLB×MA -0.292*** (0.057) τEN×SP 0.014** (0.006)

σu 1.135 0.070
σv 0.005 0.000
σ 1.068 0.266
γ 0.996 0.995
logL 33.365 946.908
Observations 418 418

Notes: *, **, and *** denote statistical significance at the 10%, 5%, and
1% levels, respectively. Elasticities are evaluated at the sample geomet-
ric mean. Abbreviations:—CF, carbon footprints; EN, energy deficits; FE,
fertilisers; FU, fungicides; IN, insecticides; LA, land; LB, labour; MA,
manure; ND, nitrogen deficits; PD, phosphorus deficits; SP, specialisation;
and TOX, toxicity. Values in parentheses are standard errors. Source:
Survey Data (2023).

These complementarities mean that the marginal product of labour rises as additional
hands are brought in, so that yield gains are incremental.

Yield response to manure application is U-shaped with an inverse elasticity of
yields with respect to manure at low application rates and increments in yields with
manure application at higher manure application rates beyond some threshold. Small
application rates often involve fresh or partially matured manure, which can introduce

21



967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012

high concentrations of ammonium and soluble salts into the feeder-root zone, causing
localised root scorch and transient micronutrient imbalances that redirect assimilates
toward vegetative flushes at the expense of floral initiation and fruit set. Over time,
with increased volumes and when manure is fully decomposed or has mineralised in
situ, its release of nutrients often aligns with the orchard’s nutrient demand throughout
flowering and fruit expansion with improved soil structure and microbial activity,
increasing yields.

Fertiliser use shows a positive elasticity of 0.06, suggesting that a 1% increase in
fertiliser use improves the mango yield by 0.06% at the sample mean. Diminishing
marginal returns are observed at higher fertiliser application rates beyond some thresh-
old. Our findings show that fertiliser application is more effective in larger orchards,
possibly due to a better nutrient distribution. These findings align with Zhang et al.
[99]’s evidence that fertilisation management is a key limiting factor in mango pro-
ductivity. In the same light, nutrient deficits show an inverse relationship to NVA. In
line with expectations, a percentage increase in nitrogen and phosphorus deficits per
hectare reduces NVA by 1.41% and 1.16%, respectively. Nitrogen and phosphorus are
crucial to maintaining soil fertility and crop productivity, so deficits directly translate
into lower economic output in the form of fewer harvested fruits. Although the detri-
mental impact of nitrogen deficits decreases at higher levels, possibly as orchards adapt
or change management strategies, the results show that phosphorus deficits intensify
this effect. Nitrogen deficits appear less harmful in highly specialised orchards with
higher tree densities.

Our results shows that each percentage point increase in the application of fungi-
cide improves mango yield by 0.14% at the sample mean, although the benefits
diminish at higher application levels. These findings are corroborated by El-Nasr et al.
[100] who obtained similar results using a randomised complete block design with ten
replications in an Egyptian mango orchard growing the Keitt variety. The study found
that the foliar application of sulphur significantly reduced the incidence and severity
of powdery mildew, particularly after the second and third sprayings, thus increasing
mango productivity and the physical and chemical characteristics of fruits.

In line with expectations, insecticide use is positively related to yields. A 1%
increase in insecticide use improves the output by 0.17% at the sample mean. This
finding is in line with several studies that have arrived as similar conclusions [29, 34].
The results also indicate that increased effectiveness of insecticides is realised in larger
orchards than on smaller orchards. Larger orchards are often correlated with higher
levels of expertise. Additionally, larger orchards permit adoption of several systematic
pest management strategies. When pesticides are part of a broad pest management
strategy, the overall effectiveness of pesticides is improved.

Pesticide-related toxicity is positively related to NVA. Within certain limits, the
use of toxic substances such as insecticides, fungicides and herbicides can increase
yields and thereby increase NVA, despite potential environmental and health costs.
However, excessive use of hazardous chemicals eventually becomes counterproductive,
as it kills beneficial organisms vital for ecosystem functioning [101] and poisons farm
workers when used improperly [17, 102], reducing productivity.

22



1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058

As expected, diminishing gains in NVA are observed from higher CO2 emissions,
toxicity, and energy deficits once CO2 emissions exceed a certain threshold, beyond
which emissions begin to undermine productivity and reduce the net contribution
to economic value added. This inverted U-shaped relationship is analogous to the
environmental Kuznets curve hypothesis, which suggests a positive correlation between
economic growth and deleterious environmental impacts, followed by a diminishing,
and eventually inverse relationship at higher levels of economic growth beyond some
threshold. Several studies have reached a similar conclusion. A review by Alae-Carew
et al. [103] found that fruit yields increased with higher concentrations of CO2. Kumar
et al. [104] also found strong positive correlations between carbon footprint and yields
in maize-wheat systems in India. In contrast, a macro-level study in Ethiopia by
Mulusew and Hong [105] reported a negative association between carbon emissions
and agricultural productivity.

Orchard-level energy consumption directly depends on the quantity of intermediate
inputs used. Consequently, it is inversely related to the net value added at higher
levels of economic output. Our findings corroborate this intuition and show an inverse
relationship with NVA at higher levels of energy use, suggesting that higher energy
inputs reduce the economic value added.

3.3.3 Distribution of efficiencies and technology gap ratios

Table 6 shows the distribution of efficiencies and technology gap ratios across the
three classes, along with the environmentally adjusted scores. Figure 4 presents the
distribution of adjusted TGR. For ease of comparison, we also present the distribution
of TGR (Figure 3a) and PTGR (Figure 3b) for various adoption classes. The average
TGR was 78% across all farmers. For non-adopters, this value was highest at 87%
while for adopters, the score was 69%. However, both groups had almost identical
PTGR (98%). The PTGRs remained relatively similar across all groups on average.
These results for PTGRs align with those of Weltin and Hüttel [9] who found that the
eco-efficiency technology gap was almost at the frontier (99%) of system technology.

The average TGRadj for the pooled sample was 77%, suggesting that the overall
technology used by all the orchard managers surveyed is relatively advanced although
not at the frontier (Figure 4). The adopters require more improvements (29.3%)
to attain the same level of output as the best available technology in the industry
compared to non-adopters (14.8%). On the other hand, non-intensive and intensive
classes had TGRadj averaging 68% and 74%, representing a moderate and relatively
close distance from the frontier technology, respectively. This indicates that conven-
tional farmers’ predominant use of synthetic pesticides in fruit fly management puts
them closer to the most efficient technology in the industry. This high TGRadj for
pesticide-reliant orchards is consistent with the well-documented yield gap favouring
conventional practices over low external input systems [12]. In the current case, this can
be attributed to the potent short-term efficacy of synthetic pesticides in suppressing
fruit fly, consequently reducing yield losses.

Turning to efficiency scores, our findings show an average MTE gap of 11% between
intensive (MTE = 70%) and the non-intensive (MTE = 59%) and conventional (MTE
= 59%) classes (Figure 5a). This finding agrees with the results of Rodrigues et al.

23



1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104

T
a
b
le

6
:
D
is
tr
ib
u
ti
o
n
o
f
effi

ci
en

ci
es

a
n
d
te
ch
n
o
lo
g
y
g
a
p
ra
ti
o
s

C
a
te

g
o
r
y

T
ec
h
n
ic
a
l
effi

ci
en

cy
m
et
ri
cs

E
co
-e
ffi
ci
en

cy
m
et
ri
cs

E
n
vi
ro
n
m
en

ta
ll
y
a
d
ju
st
ed

m
et
ri
cs

M
et
ri
c

M
ea

n
S
D

M
in
–
M
a
x

M
et
ri
c

M
ea

n
S
D

M
in
–
M
a
x

M
et
ri
c

M
ea

n
S
D

M
in
–
M
a
x

P
o
o
le
d

T
E

0
.7
0
2

0
.1
4
3

0
.2
4
3
–
0
.9
3
3

E
E

0
.9
4
7

0
.0
4
4

0
.7
0
1
–
0
.9
9
5

—
T
G
R

0
.7
8
1

0
.1
6
6

0
.2
4
3
–
0
.9
7
8

P
T
G
R

0
.9
8
2

0
.0
1
6

0
.8
5
6
–
0
.9
9
7

T
G
R

a
d
j

0
.7
6
7

0
.1
6
5

0
.2
3
8
–
0
.9
7
0

M
T
E

0
.6
2
3

0
.1
9
0

0
.1
1
0
–
0
.9
6
8

M
E
E

0
.9
3
4

0
.0
4
5

0
.6
7
7
–
0
.9
9
3

M
T
E

a
d
j

0
.5
8
6

0
.1
8
7

0
.0
7
5
–
0
.9
5
6

A
d
o
p
te
rs

T
E

0
.8
3
3

0
.0
8
1

0
.3
2
1
–
0
.9
6
5

E
E

0
.9
6
7

0
.0
2
4

0
.8
4
2
–
0
.9
9
3

—
T
G
R

0
.7
2
0

0
.1
8
4

0
.2
4
3
–
0
.9
7
8

P
T
G
R

0
.9
8
1

0
.0
1
5

0
.8
7
8
–
0
.9
9
7

T
G
R

a
d
j

0
.7
0
7

0
.1
8
2

0
.2
3
8
–
0
.9
7
0

M
T
E

0
.6
4
5

0
.1
8
8

0
.2
0
5
–
0
.9
6
8

M
E
E

0
.9
3
7

0
.0
3
8

0
.7
6
5
–
0
.9
9
3

M
T
E

a
d
j

0
.6
0
7

0
.1
8
4

0
.1
7
2
–
0
.9
5
6

N
o
n
-a
d
o
p
te
rs

T
E

0
.6
8
9

0
.2
2
3

0
.1
1
4
–
1
.0
0
0

E
E

0
.9
4
6

0
.0
5
0

0
.6
8
2
–
0
.9
9
5

—
T
G
R

0
.8
6
8

0
.0
7
9

0
.5
6
8
–
0
.9
6
9

P
T
G
R

0
.9
8
2

0
.0
1
6

0
.8
5
6
–
0
.9
9
4

T
G
R

a
d
j

0
.8
5
2

0
.0
7
8

0
.5
5
8
–
0
.9
5
8

M
T
E

0
.5
9
3

0
.1
8
8

0
.1
1
0
–
0
.9
3
4

M
E
E

0
.9
2
9

0
.0
5
2

0
.6
7
7
–
0
.9
8
3

M
T
E

a
d
j

0
.5
5
6

0
.1
8
9

0
.0
7
5
–
0
.9
1
2

N
o
n
-i
n
te
n
si
v
e

T
E

0
.8
6
8

0
.1
2
2

0
.4
2
1
–
0
.9
8
4

E
E

0
.9
5
5

0
.0
3
3

0
.8
6
6
–
0
.9
9
8

—
T
G
R

0
.6
9
0

0
.1
8
1

0
.2
5
3
–
0
.9
7
1

P
T
G
R

0
.9
8
0

0
.0
1
4

0
.8
9
4
–
0
.9
9
7

T
G
R

a
d
j

0
.6
7
7

0
.1
8
1

0
.2
5
1
–
0
.9
6
4

M
T
E

0
.5
9
4

0
.1
6
5

0
.2
4
1
–
0
.9
3
3

M
E
E

0
.9
3
6

0
.0
3
7

0
.8
4
3
–
0
.9
8
9

M
T
E

a
d
j

0
.5
5
6

0
.1
6
2

0
.2
2
6
–
0
.9
0
0

In
te
n
si
v
e

T
E

0
.9
3
5

0
.1
3
8

0
.2
4
1
–
1
.0
0
0

E
E

0
.9
5
6

0
.0
3
9

0
.7
7
3
–
0
.9
9
9

—
T
G
R

0
.7
5
2

0
.1
8
1

0
.2
4
3
–
0
.9
7
8

P
T
G
R

0
.9
8
2

0
.0
1
6

0
.8
7
9
–
0
.9
9
4

T
G
R

a
d
j

0
.7
3
9

0
.1
7
8

0
.2
3
8
–
0
.9
7
0

M
T
E

0
.6
9
9

0
.1
9
5

0
.2
0
5
–
0
.9
6
8

M
E
E

0
.9
3
8

0
.0
4
0

0
.7
6
5
–
0
.9
9
3

M
T
E

a
d
j

0
.6
6
0

0
.1
9
2

0
.1
7
2
–
0
.9
5
6

N
o
te
s
:
A
b
b
re
v
ia
ti
o
n
s:
—

T
E
,
te
ch
n
ic
a
l
effi

ci
en

cy
;
T
G
R
:
te
ch
n
o
lo
g
y
g
a
p
ra
ti
o
;
M
T
E
:
m
et
a
-t
ec
h
n
ic
a
l
effi

ci
en

cy
;
E
E
:
ec
o
-e
ffi
ci
en

cy
;

M
E
E
:
m
et
a
-e
co
-e
ffi
ci
en

cy
;
P
T
G
R
:
p
re
ss
u
re
-g
en

er
a
ti
n
g
T
G
R
;
S
D
:
st
a
n
d
a
rd

d
ev
ia
ti
o
n
.
B
o
ld

va
lu
es

h
ig
h
li
g
h
t
th
e
h
ig
h
es
t
fi
g
u
re
s

a
cr
o
ss

a
ll

ca
te
g
o
ri
es
.
T
E

a
n
d

E
E

a
re

d
er
iv
ed

fr
o
m

cl
a
ss
-s
p
ec
ifi
c
fr
o
n
ti
er
s
a
n
d

a
re
,
th
er
ef
o
re
,
n
o
t
d
ir
ec
tl
y
co
m
p
a
ra
b
le
;
a
ll

o
th
er

m
et
ri
cs

d
er
iv
e
fr
o
m

th
e
m
et
a
fr
o
n
ti
er
s
a
n
d
a
re

th
er
ef
o
re

co
m
p
a
ra
b
le

a
cr
o
ss

fa
rm

er
ca
te
g
o
ri
es
.
S
o
u
rc

e
:
S
u
rv
ey

D
a
ta

(2
0
2
3
).

24



1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

1

2

3

4

5

6

7

m
ean

=
0
.8
7m

ea
n
=

0
.6
9

m
ea
n
=

0.
7
5

TGR

K
er
n
el

d
en
si
ty

Non-adopters
Non-intensive

Intensive

(a)

0.86 0.88 0.9 0.92 0.94 0.96 0.98 1

0

10

20

30

40

50

m
ean

=
0.98

m
ea
n
=

0.
98

m
ea
n
=

0.
98

PTGR

K
er
n
el

d
en
si
ty

Non-adopters
Non-intensive

Intensive

(b)

Fig. 3: Distributions of (a) TGR and (b) PTGR across adoption categories. Source: Survey
Data (2023)

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

1

2

3

4

5

6

7

m
ea
n
=

0.
85

m
ea
n
=

0.
68

m
ea
n
=

0.
74

TGRadj score

K
er
n
el

d
en
si
ty

Non-adopters
Non-intensive

Intensive

Fig. 4: Distribution of adjusted TGR for various adoption classes. Source: Survey Data
(2023)

[51] who found that intensive adopters of biological pest control methods were more
technically efficient (86.3%) than non-intensive adopters (82.3%) in Brazilian agricul-
tural systems. On the other hand, our results show only slight differences in MEE
among the three classes of farmers. All classes had average MEE between 93–94%, with
intensive and non-intensive adopters only marginally ahead of the conventional class
(Figure 5b). Weltin and Hüttel [9] found that farmers practicing sustainable inten-
sification were associated with higher eco-efficiency (75%) than non-intensive group
(63%) in Italian farms.

25



1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.5

1

1.5

2

2.5

m
ean

=
0.59m

ea
n
=

0.
59

m
ea
n
=

0.
70

MTE score

K
er
n
el

d
en
si
ty

Non-adopters
Non-intensive

Intensive

(a)

0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

0

2

4

6

8

10

12

m
ea
n
=

0.
93

m
ea
n
=

0.
94

m
ea
n
=

0.
9
4

MEE score

K
er
n
el

d
en
si
ty

Non-adopters
Non-intensive

Intensive

(b)

Fig. 5: Distributions of (a) MTE and (b) MEE across adoption categories. Source: Survey
Data (2023)

The MTEadj incorporates ecological footprints, thus providing a more precise mea-
sure of productive efficiency from a sustainability perspective. On average, orchards
operate at 59% of their potential production once environmental constraints are taken
into account, leaving a shortfall of 41% relative to the frontier (Figure 5 and 6). This
indicates a significant efficiency gap, suggesting a considerable room for improvement
in achieving optimal performance at current input levels. A subgroup analysis indi-
cates that the distribution of efficiency scores is more dispersed among adopters than
among non-adopters. Adopters attain a higher average efficiency score (61%) com-
pared to non-adopters (56%), suggesting that although APM users are closer to the
frontier, both groups exhibit substantial efficiency gaps with the frontier potential.

The stark difference in MTE between intensive and non-intensive and non-adopter
classes persists even after accounting for ecological footprints. While the intensive
class achieves an average MTEadj of 66%, non-intensive and non-adopter classes are
10 percentage points below the intensive class efficiency. These findings suggest that,
under intensive management, mango yield could potentially be increased by approx-
imately 34% without consuming additional inputs, as opposed to a 44% increase for
the non-intensive class. However, there are no efficiency gains in non-intensive adop-
tion relative to conventional farmers. This could indicate high transition costs. For
example, there are risks of failure under limited adoption if the efforts not substan-
tially reduce pest pressure leading to pest resurgence. Similarly, APM is labour and
knowledge intensive, requiring additional labour and higher learning costs, reducing
net benefits. These transition costs present a key barrier for non-intensive users, requir-
ing supportive policies that reduce or offset these initial costs. Ultimately, effective
fruit fly suppression requires a coordinated and integrated set of practices rather than
isolated measures, which are easily undermined by re-infestation.

26



1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.5

1

1.5

2

2.5

m
ea
n
=

0.
56

m
ean

=
0.5

6 m
ea
n
=

0.
66

MTEadj score

K
er
n
el

d
en
si
ty

Non-adopters
Non-intensive

Intensive

Fig. 6: Distribution of adjusted MTE for various adoption classes. Source: Survey Data
(2023)

3.4 Treatment effects, sensitivity analysis and robustness
checks

Table 7 presents the treatment effect estimates from IPWRA (fractional probit), PSM
(linear regression), and RA (fractional probit). Despite differences in estimators, the
methods produce substantively similar point estimates, confirming our earlier finding
of negligible selectivity bias on observables. We further assess sensitivity to omitted
variables using Oster’s δ. In IPWRA and RA models, δ falls between 0.97 and 1.72, and
0.97 and 1.53, respectively, while in PSM models it ranges from 0.95 to 1.60, indicating
a relatively strong degree of robustness, since unobserved confounders would need to
be almost at least as influential as included covariates to reduce the estimated effects
to zero. This confirms the insignificant ρ initially observed in the frontier models.
The common-support diagnostics for PSM are displayed in Figure 7. Together, these
findings provide compelling evidence that the estimated treatment effects are robust
and unbiased. However, given the attractive doubly robust property of IPWRA, we
focus the subsequent discussion on the results from Columns 1–6 of Table 7.

The results indicate that the adoption of APM is associated with a positive ATE
of 0.32, although significant only at the 10% level, with a POMean of 0.575. This
indicates that in a counterfactual scenario in which no one in the sample adopted
APM, the average MTEadj score would be approximately 57.5%. In contrast, had
all orchard managers adopted the APM, the MTEadj scores would have increased by
3.2 percentage points on average. However, according to the ATT, orchard managers
who actually adopted APM improved their scores by 2 percentage points on average.
Non-intensive adoption produces a small, negative effect (ATE = –0.011) that is not
statistically significant, and its associated POMean of 0.564. This suggests that in
a counterfactual world where no orchard manager in the sample adopted APM, the

27



1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288

T
a
b
le

7
:
E
st
im

a
te
d
tr
ea
tm

en
t
eff

ec
ts

IP
W

R
A

P
S
M

R
A

T
re
a
tm

en
t

C
o
ef
.
(S

E
)

P
O
M
ea

n
O
st
er
’s

δ
C
o
ef
.
(S

E
)

O
st
er
’s

δ
C
o
ef
.
(S

E
)

O
st
er
’s

δ

A
T
E

N
o
n
-i
n
te
n
si
v
e

-0
.0
1
1
(0
.0
2
0
)

0
.5
6
4
*
*
*

1
.1
0
8

-0
.0
1
4
(0
.0
2
5
)

1
.0
1
3

-0
.0
1
4
(0
.0
2
1
)

0
.9
9
1

In
te
n
si
v
e

0
.0
8
1
*
*
*
(0
.0
2
6
)

0
.5
7
6
*
*
*

1
.6
6
0

0
.0
7
0
*
*
(0
.0
3
1
)

1
.6
2
1

0
.0
7
7
*
*
*
(0
.0
2
5
)

1
.5
3
0

In
te
n
si
v
e†

0
.0
9
1
*
*
*
(0
.0
2
2
)

0
.5
5
8
*
*
*

0
.9
7
3

0
.0
7
8
*
*
*
(0
.0
2
7
)

0
.9
5
4

0
.0
9
2
*
*
*
(0
.0
2
3
)

0
.9
7
1

A
d
o
p
te
r

0
.0
3
2
*
(0
.0
1
7
)

0
.5
7
5
*
*
*

1
.7
2
4

0
.0
4
9
*
(0
.0
2
5
)

1
.4
4
7

0
.0
3
4
*
(0
.0
2
0
)

1
.5
0
2

A
T
T

N
o
n
-i
n
te
n
si
v
e

-0
.0
1
5
(0
.0
2
4
)

0
.5
7
4
*
*
*

-0
.0
1
1
(0
.0
3
8
)

-0
.0
1
2
(0
.0
2
4
)

In
te
n
si
v
e

0
.0
5
6
*
*
(0
.0
2
5
)

0
.6
0
7
*
*
*

0
.0
6
1
*
*
(0
.0
2
9
)

0
.0
6
8
*
*
(0
.0
2
9
)

In
te
n
si
v
e†

0
.1
0
5
*
*
*
(0
.0
2
4
)

0
.5
5
7
*
*
*

0
.1
1
4
*
*
*
(0
.0
2
6
)

0
.1
0
4
*
*
*
(0
.0
2
4
)

A
d
o
p
te
r

0
.0
2
0
(0
.0
2
1
)

0
.5
8
8
*
*
*

0
.0
2
7
(0
.0
3
2
)

0
.0
2
8
(0
.0
2
1
)

N
o
te
s
:
*
,
*
*
,
a
n
d
*
*
*
d
en

o
te

st
a
ti
st
ic
a
l
si
g
n
ifi
ca
n
ce

a
t
th
e
1
0
%
,
5
%
,
a
n
d
1
%

le
v
el
s,

re
sp

ec
ti
v
el
y.

P
O
M
ea
n
re
p
re
se
n
ts

p
o
te
n
ti
a
l
o
u
tc
o
m
e
m
ea
n
fo
r
u
n
tr
ea
te
d
g
ro
u
p
.
O
st
er
’s

δ
va
lu
es

in
d
ic
a
te

ro
b
u
st
n
es
s
to

se
le
ct
io
n
o
n
u
n
o
b
se
rv
a
b
le
s—

va
lu
es

n
ea
r
o
r
a
b
ov
e
1
im

p
ly

th
a
t
u
n
o
b
se
rv
ed

co
n
fo
u
n
d
er
s
w
o
u
ld

n
ee
d
to

b
e
a
t
le
a
st

a
s
in
fl
u
en
ti
a
l
a
s
in
cl
u
d
ed

co
va
ri
a
te
s

to
n
u
ll
if
y
th
e
es
ti
m
a
te
d
eff

ec
ts
.
† T

h
es
e
es
ti
m
a
te
s
a
re

re
la
ti
v
e
to

th
e
n
o
n
-i
n
te
n
si
v
e
g
ro
u
p
;
th
e
re
st

a
re

re
la
ti
v
e
to

n
o
n
-

a
d
o
p
te
rs
.
V
a
lu
es

in
p
a
re
n
th
es
es

a
re

h
et
er
o
sk
ed

a
st
ic
it
y
ro
b
u
st

st
a
n
d
a
rd

er
ro
rs
.
S
o
u
rc

e
:
S
u
rv
ey

D
a
ta

(2
0
2
3
).

28



1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334

0 10.2 0.4 0.6 0.8 0 10.2 0.4 0.6 0.8

(a)

0 10.2 0.4 0.6 0.8
Propensity score

Treated (Non-intensive)
Untreated (Non-adopters)

0 10.2 0.4 0.6 0.8
Propensity score

Treated (Intensive)
Untreated (Non-adopters)

(b)

Fig. 7: Common support (CS) mirror bars: (a) before imposing CS and (b) after imposing
CS.

average efficiency score would be about 56.4%. Based on the ATT estimate, the non-
intensive class realised an insignificant decline of 1.5 percentage points in the efficiency
score on average.

In contrast, intensive use of APM is associated with a positive significant ATE
of 0.081, with a POMean of 0.576. This implies that if all farmers transitioned from
non-adoption to intensive APM use, their efficiency score would increase by roughly
8.1 percentage points on average, up from the baseline of 57.6% if no one adopts the
APM. According to ATT, intensive APM adopters were, on average, associated with
significant increments of 5.6 percentage points in the MTEadj score. The magnitude
of the ATE increases by a percentage when intensive and non-intensive classes are
compared, with a POMean of 0.558. Interestingly, the ATT based on this comparison
is higher than the ATE by 1.6 percentage points, indicating a potential selection bias.
However, the consistency of IPWRA, PSM and RA results, together with Oster’s δ,
provides strong confidence that this finding is not an artefact of selectivity or omitted
variable bias, therefore, a possible heterogeneity in treatment effects. We discuss this
shortly in Subsection 3.5. In essence, based on Oster’s δ coefficients, the selection on

29



1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380

observables would have to be at least 95.4% as influential as the observed covariates
to explain away these results.

Similar findings have been reported in extant literature. In Kenya, a comprehen-
sive IPM package for mango fruit flies disseminated by the ICIPE was found be to
associated with a 54.5% reduction in produce rejection and up to 22.4% increase in
farmers’ net income in pilot areas [36]. In a larger trial, mango farmers using vari-
ous components and combinations of the IPM package saw fruit loss drop by 30%,
pesticide expenditure nearly halved, and net income rise by 48% compared to non-
adopters on average [34]. In conformity to our findings, the study found that while
intensive adopters increased net income by 115%, those who used only one compo-
nent reported 7% decline in net income. Midingoyi et al. [29] found that while the
introduction of 1-2 IPM practices raised yields by between 6–27% and farm income
by roughly 9–33% relative to non-adoption, intensive users (those who implemented
three or more practices) recorded yield advantages of 95% and income improvements of
137% compared to non-adopters. Pecenka et al. [106] found that using only some IPM
options did not significantly reduce crop damage, whereas a full suite of IPM practices
nearly eliminated (up to 95%) chemical sprays through conservation of wild pollinators
and natural enemies while maintaining yields, improving profitability while enhancing
ecosystem services. These findings demonstrate that with sufficient intensification and
proper management, agro-ecological approaches can match or even outperform the
productivity of pesticide-reliant systems, lending weight to the idea that sustainable
intensification is achievable.

3.5 Heterogeneity in treatment effects

The CATE estimates reveal considerable heterogeneity in MTEadj attributable to the
intensity of APM adoption (Figure 8). Orchard managers who perceive the pest as
severe experience the highest treatment effects (Figure 8a). These farmers are likely
to be more conscious and intentional in their approach to suppressing the pest since
they already recognise its potential effects on yields. Participation in co-creation activ-
ities lifts the CATE well above the sample-wide ATE, with 95% confidence bands
that remain entirely above zero (Figure 8i). This reflects the central role of knowledge
sharing in labour-intensive agro-ecological systems. Farmer-to-farmer learning reduces
search and experimentation costs, enabling faster mastery of synergistic interactions
of APM options that lead to efficiency gains. The CATE curve slopes upwards beyond
upper primary schooling, attaining a marked upward shift at roughly 10 years of educa-
tion, indicating that better educated orchard managers experience the highest positive
effects from APM use (Figure 8d). The diminishing marginal return beyond secondary
schooling, however, suggests that basic agronomic literacy rather than advanced cre-
dentials is sufficient for sizeable gains. Educated farmers are more likely to process
technical information better, adjust input mixes and time operations precisely, thereby
capturing the synergistic effects of APM.

Other social-context variables such as gender, age, household size, and number
of neighbouring adopters all produce CATE ribbons that overlap zero almost every-
where (Figures 8e , 8b, 8f, and 8h). Among the adopters, gender and age of the
manager do not yield significant differences, illustrating that, conditional on access to

30



1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426

0

-0.05

0.05

0.1

0.15

0.2

0 1
Perceived severity (1 = severe)

(a)

0

-0.5

0.5

33 43 53 63 73
Age (years)

(b)

0

-0.4

-0.2

0.2

0.4

7 11 15 19 23
Age of rootstock (years)

(c)

0

-0.4

-0.2

0.2

0.4

En
vi

ro
nm

en
ta

lly
-a

dj
us

te
d 

ef
fic

ie
nc

y

6 8 10 12 14
Formal education (years)

(d)

0

.05

.1

.15

.2

0 1
Gender (1 = male)

(e)

0

-0.1

0.1

0.2

0.3

2 4 6 8
Household size (members)

(f)

0

-0.05

0.05

0.1

0.15

0.2

2.7 3.2 3.7 4.2 4.7
Log of tree density (trees/acre)

(g)

-1

0

1

-0.5

0.5

2 4 6 8 10 12 14 16 18 20
Neighbouring APM adopters (count)

(h)

0

0.05

0.1

0.15

0.2

0 1
Participation in co-creation (1 = yes)

(i)

Fig. 8: Doubly robust conditional average treatment effect (DRCATE) estimates illustrating
heterogeneous effects on environmentally adjusted efficiency. The dotted blue line depicts the
ATE, the solid line the CATE, and the olive-teal shaded band the 95% confidence interval.
Source: Survey Data (2023)

information and resources, women and older farmers can benefit equally from APM.
This suggests that the APM’s effectiveness is relatively gender-neutral and resilient to
variations in orchard-manager’s age, household labour endowment or peer adoption

31



1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472

density. Balogun et al. [107] reported similar results in a study of Nigerian pineap-
ple farms. This finding aligns with the principle of fairness in agroecology. Tailored
information transfer mechanisms can close any residual gaps.

In contrast, biophysical moderators exert a selective influence on MTEadj. Orchard
density displays a concave pattern, with moderate densities between 20–66 trees acre−1

yielding more positive CATEs, whereas poorly and overly dense orchards do not
(Figure 8g). Managing canopy competition and pest microclimates could be easier at
intermediate densities, allowing the APM intervention to reach full agronomic poten-
tial. Rootstock age exhibits wide confidence intervals that straddle zero throughout
(Figure 8c), implying negligible and highly uncertain moderation.

3.6 Drivers of environmentally adjusted inefficiency

Table 8 presents the average marginal effects from a bootstrap fractional probit
regression for the determinants of adjusted inefficiency. The Wald test statistic was sig-
nificant at the 1% level, confirming the joint significance of the predictors. A negative
coefficient shows that a variable reduces inefficiency and vice versa.

Table 8: Estimates of bootstrap fractional probit for the determinants
of environmentally adjusted inefficiency

Variable Coef. (SE) AME (SE)

Formal education (years) -0.016** (0.007) -0.006** (0.003)
Household size (count) -0.012 (0.009) -0.005 (0.003)
Gender (1 = male) -0.048 (0.054) -0.018 (0.021)
Intensity (semi-continuous) -1.400*** (0.505) -0.538*** (0.193)
Intensity squared 4.554*** (1.475) 1.748*** (0.564)
Orchard prospects (1 = positive) -0.200** (0.095) -0.077** (0.036)
Age of rootstock (years) -0.008** (0.003) -0.003** (0.001)
ln(Tree density (trees acre−1)) -0.049 (0.036) -0.019 (0.014)
Number of orchards (count) 0.013 (0.046) 0.005 (0.018)
Group membership (1 = yes) -0.125** (0.051) -0.048** (0.019)
Credit access (1 = yes) 0.153* (0.088) 0.059* (0.034)
Off-farm income (KES yr−1)† -0.007 (0.006) -0.003 (0.002)
Mango export quantity (kg) -0.001*** (0.000) -0.001*** (0.000)
Co-creation (1 = yes) -0.117** (0.048) -0.045** ( 0.018)
Extension access (1 = yes) -0.005 (0.051) -0.002 (0.020)
Distance to input market (meters) 0.007 (0.018) 0.003 (0.007)
Constant 0.650*** (0.227)

Log pseudo-likelihood -279.512
Wald χ2(16) 57.94***
Pseudo R2 0.14
Replications 1000
Observations 418

Notes: *, ** and *** denote significance at the 10%, 5%, and

1% levels, respectively. † This variable was transformed using an
inverse hyperbolic sine to reduce skewness and heteroscedasticity
while accommodating zero observations. AME denotes the average
marginal effect. Values in parentheses are bootstrapped standard
errors. Source: Survey Data (2023).

32



1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518

We found a non-linear relationship between APM adoption intensity and ineffi-
ciency. In the initial stages, greater adoption of APM practices reduces inefficiency, but
beyond some threshold, increased intensification begins to undermine efficiency. This
finding suggests a possible optimal level of APM adoption after which adding more
practices yields less benefit and may even strain the farmer’s management capacity.
The APM is a complex strategy with high labour demand for proper coordination of
practic