1 Introduction

This is an exploratory data analysis (EDA) of the brucellosis knowledge, attitudes, and practices (KAP) survey conducted across six locations in Isiolo County, Kenya including garbatulla_reserve,sericho_reserve,kina_reserve,kina_main,sericho_main,and garbatulla_main. The survey was administered to 409 livestock-keeping households.

The purpose of this document is not just to describe the data, but to check, whether the dataset can actually answer the four research objectives set out in the analysis plan:

Assess the level of awareness and knowledge of brucellosis among livestock-keeping households.
Evaluate management practices and behaviors related to brucellosis transmission and control.
Examine household perceptions and attitudes towards brucellosis transmission and prevention.
Determine the influence of household economic status and access to resources on knowledge and adoption of preventive

Rows (households): 409

Columns (variables): 185

Note on data missingness

A large share of “missingness” in this survey is not random ,it is structural, caused by the collection tool’s skip logic. For example, only respondents who answered “Yes” to “Are you aware of brucellosis?” (brucella_aware) were shown the entire knowledge section (Section B) and several subsequent sections. This means low coverage on a variable is often a feature of the questionnaire design, not a data quality failure. Each section below makes clear which kind of missingness is in play.

2 Dataset summary

Before looking at brucellosis-specific responses, it’s worth understanding who was surveyed, since every objective will use these demographic and livelihood variables as explanatory factors.

2.1 Where respondents were surveyed

Figure 1: Number of households surveyed by location

Table 1: Households surveyed by ward
ward	n_households	pct
Garbatulla	174	42.5
Kinna	144	35.2
Sericho	91	22.2

The survey spans three wards (Garbatulla, Kinna, Sericho), each split into a “main” town area and a “reserve” (more remote pastoral) area, giving six location strata in total. Garbatulla ward has the largest share of the sample, Sericho the smallest.

2.2 Demographic summary of respondents

Table 2: Demographic summary of respondents (n = 409)
Statistic	Value
Mean age	43.6
Median age	40.0
% Female	44.0
Mean household size	7.0
Mean years schooling	4.8
% No formal education	51.5

Figure 2: Age group and education category distribution

Figure 3: Age group and education category distribution

Figure 4: Primary income source of surveyed households

Explaination

The sample is mostly pastoralist, which fits the research context well since brucellosis is fundamentally a disease of livestock-keeping populations.

For Objectives 1 and 2 (factors associated with knowledge, and management practices), age_group, edu_cat, sex, income_source, and no_hh_members are all fully populated (100% or near-100% coverage) and ready to be used as explanatory variables. There is enough spread across age groups and education categories to support meaningful comparisons.

3 Objective 1 | Awareness and knowledge of brucellosis among livestock-keeping households

We are trying to answer how many households have heard of brucellosis, and among those who have, how much do they actually know about it?

3.1 Overall awareness

Figure 5: Brucellosis awareness among surveyed households

84.1% of respondents (344 of 409) reported having heard of brucellosis before. This is the key skip-logic gate for the rest of the knowledge section , only these 344 respondents went on to answer the detailed knowledge items below.

3.2 Where awareness comes from?

Figure 6: First source of brucellosis awareness, among aware respondents

3.3 Knowledge depth, what aware respondents actually know

Awareness (“have you heard of it”) is a much lower bar than knowledge (“do you know how it spreads, what it looks like, and how it affects people”). The questionnaire probed four separate knowledge domains, each captured as a set of multi-select items now expanded into 0/1 indicator columns:

Animal signs : abortion, stillbirth, weak calf, retained placenta, reduced milk, swollen testes, swollen joints, weight loss
Animal-to-animal transmission routes
Animal-to-human transmission routes
Human signs of brucellosis

Figure 7: Distribution of overall knowledge score (% of items correctly identified)

Table 3: Knowledge score (%) summary statistics, aware respondents only
Min	25th pct	Median	Mean	75th pct	Max
2	18	22	22.6	28	66

The distribution is right-skewed and clustered at the low end: most aware respondents correctly identify only a small fraction of the full set of signs and transmission routes. A mean knowledge score around the low-to-mid 20s (out of 100) suggests that simply being “aware” of brucellosis does not translate into detailed understanding of brucellosis.

3.4 Which knowledge domain is weakest?

Figure 8: Average number identified, by knowledge domain

3.5 Knowledge score by demographic group

A comparison of whether knowledge varies systematically by age, sex, or education.

Figure 9: Knowledge score (%) by age group, sex, and education category

Figure 10: Knowledge score (%) by age group, sex, and education category

Figure 11: Knowledge score (%) by age group, sex, and education category

Summary

Awareness itself is recorded for all 409 respondents with zero missingness, and the detailed knowledge items (40 indicator variables across four domains) are consistently populated for all 344 aware respondents (84% of the sample). The data clearly supports both halves of Objective 1: the descriptive half (how aware are people, and what do they actually know) is fully answerable with the figures above, and the inferential half (what predicts higher knowledge) has a complete, non-missing set of explanatory variables — age_group, sex, edu_cat, income_source, yrs_keep_livestock, rec_ext_services, grp_member — to regress knowledge_score_pct or aware_binary against.

Of the three demographic splits shown, education shows the clearest separation:

The tertiary-educated group’s median knowledge score (26%) sits visibly above the other three education categories, which cluster tightly together around 22% regardless of formal schooling level.
Age shows a smaller shift, with the 31–45 group scoring modestly higher than the other two age bands;
Sex shows the smallest difference of the three, with substantial overlap between male and female distributions.

This ordering education > age > sex gives a working hypothesis for the regression: education is the demographic variable most likely to retain a significant, independent effect on knowledge score once the other covariates are controlled for.

4 Objective 2 | Management practices and behaviours related to transmission an dcontrol

This objective asks what households actually do to prevent or manage brucellosis risk, as opposed to what they know. The questionnaire captured this through a multi-select list of specific prevention practices, plus a smaller sub-section on individual risk behaviours.

4.1 Which prevention practices are actually used?

Figure 12: Prevention practices reported, as a share of all 409 households

A genuine finding, not a data error

Eight of the twelve prevention-practice options in the questionnaire including restricting movement, farm sanitation, slaughtering positive animals, isolating animals during parturition, disposing of fetal material safely, disinfecting, public education, and seeking veterinary advice , were selected by zero respondents across all 409 households.

This suggests a real and large gap between the practices considered “textbook” prevention and what pastoralist households in this part of Isiolo are actually doing.

4.2 How many practices does a typical household report?

Figure 13: Number of prevention practices reported per household

Table 4: Households adopting at least one prevention practice (used as the binary outcome for Objective 2 modelling)
adopted_any_practice	n	pct
Adopted none	42	10.3
Adopted ≥1 practice	302	73.8
NA	65	15.9

4.3 Individual-level risk behaviours

A smaller sub-section of the questionnaire asked about specific risk behaviours during animal handling. This block was only shown to a subset of respondents (the human-illness branch of the questionnaire), so coverage is much lower at about 17% of the sample (60–69 households).

Figure 14: Self-reported risk behaviours, among the sub-sample asked (n ≈ 69)

4.4 Does knowledge translate into practice?

A key analytical question for Objective 2: do households with higher knowledge scores actually report more prevention practices?

Figure 15: Knowledge score vs. number of prevention practices adopted

What this means for Objective 2

The self-reported adoption question (adopt_prev_ctrl) turned out to have zero variance since every respondent who answered it said “Yes.” Therefore, it cannot serve as a regression outcome. The practice-count and practice-variables (prev_practice_count, adopted_any_practice, and the four practices with real variation: vaccination, testing, isolating infected animals, testing new animals) are the variables that should anchor this objective’s analysis instead. The individual risk-behaviour items (raw milk consumption, glove use, etc.) are real and usable, but only for a descriptive sub-analysis — at n ≈ 69 they are too small to support a separate regression model.

The boxplot above shows:

households with 0-3 practices show essentially indistinguishable medians (22–26%) with heavily overlapping IQRs, so partial adoption does not appear to track with knowledge score.
The 4-practices group sits higher (median ≈ 32%) but is also the smallest of the five groups (n = 18, versus n = 42–76 for groups 0–3) and shows no plotted outliers — consistent with too few observations for any point to fall outside 1.5×IQR, rather than a genuinely tighter distribution.

Because the data are cross-sectional, any association that does emerge cannot establish whether knowledge drives adoption or adoption (and the experience that comes with it) builds knowledge ,this directionality should be flagged as a limitation regardless of the regression result.

5 Objective 3 | Perceptions and attitudes towards brucellosis transmission and prevention

This objective covers Section E of the questionnaire: Likert-scale items capturing how serious, preventable, and threatening respondents perceive brucellosis to be. Unlike the knowledge section, this section has two separate blocks with very different coverage:

Perceived risk of transmission (a2a or a2h)
Perceived effectiveness of prevention

Table 5: Coverage of the two perception blocks in Section E
Block	Example items	n (coverage)	% of sample
seqb / seqc (transmission & prevention attitudes)	Risk from consuming milk; vaccination effectiveness	275	67.2
seq1–seq16 (general severity & risk perception)	Brucellosis is a serious threat to animals/humans	69	16.9
comm_* (community-level perception)	Shared grazing; shared water points	275	67.2

Why the seq1–16 block has only 17% coverage

This is a structural skip-logic split tied to which field team administered the questionnaire, not random missingness. The broader seq1–seq16 items were only shown to a subset of respondents.
This means seq1–seq16 should be treated as descriptive-only (n = 69 is too small for a stable regression), while the seqb/seqc/comm_* block (n = 275, 67% coverage) can be used for inferential analysis.

5.1 Attitudes toward transmission risk and prevention (seqb / seqc)

Figure 16: Distribution of responses across transmission and prevention attitude items

5.2 Community-level perception items

Figure 17: Community-level perception items

5.3 General severity perception (descriptive only, n = 69)

Figure 18: General severity/preventability perception (small sub-sample)

Summary

The seqb/seqc/comm_* block (16 items, n = 275) are enough to support both descriptive summaries and a perception-index regression (e.g. summing or averaging Likert scores and regressing against knowledge score and SES).
The seq1–16 block (n = 69) should be reported as a descriptive table but not enpugh for inferential analysis as stated earlier.

6 Objective 4 | Socioeconomic status and access to resources on knowledge and adoption of preventive practices

This objective characterizes household wealth and access as both a descriptive picture of the sample and as an explanatory variable for knowledge and practice in the inferential models.

6.1 The SES index

ses_index is a composite asset index built from ten components: seven binary assets (radio, bicycle, motorbike, car, house ownership, piped water, electricity), phone count (capped at 3), house wall material, and toilet type , each rescaled to a 0–1 contribution and averaged.

Figure 19: Distribution of the household SES index

Table 6: Missingness pattern in the SES block
Households missing SES index	All from brucella_aware = No?	% of full sample
59	FALSE	14.4

The 59 households missing an SES index are exactly the households who were not asked Section I (because they were routed past it after answering “No” to brucellosis awareness), this is structural missingness, not random, and should not be imputed.

6.2 Asset ownership breakdown

6.3 Livestock holdings (a wealth proxy specific to pastoralist households)

Figure 21: Distribution of total livestock holdings

6.4 Institutional access: extension services and group membership

Figure 22: Access to extension services and group membership

6.5 Does SES relate to knowledge?

Figure 23: SES index vs. knowledge score

Summary

The ten-component index, plus all of its raw inputs (assets, house construction, toilet type, phone count) and the separate livestock-holding variables, are populated for 353 of 409 households (86%), enough for both a descriptive wealth profile and use as an explanatory variable in the knowledge and practice regressions. The missingness pattern is fully understood and structural (tied to the brucellosis-aware skip gate), so it should be reported as such rather than imputed.

The scatter plot above shows whether wealthier households tend to know more about brucellosis (n = 344).

The bivariate association is weak , r = 0.2 (p = 0) , with the fitted trend rising only about 3 percentage points across the full SES range (0–0.85). The confidence band also widens noticeably above SES ≈ 0.6, where data are sparser, so the apparent upward tilt at the high end of the scale should be read cautiously rather than as a strong trend.

7 Will the final dataset help answer our objectives?

Table 7: Summary verdict by objective
Objective	Key variables	Coverage	Verdict
Obj 1 \| Awareness & knowledge	brucella_aware, knowledge_score_pct, 40 knowledge dummies	100% awareness / 84% knowledge (n=344)	Fully answerable
Obj 2 \| Management practices	prev_practice_count, adopted_any_practice, 4 practice dummies	100% practice dummies / 66%* self-report (unusable)	Answerable, reframed outcome variable
Obj 3 \| Perceptions & attitudes	seqb/seqc (16 items), comm_* (4 items); seq1-16 descriptive only	67% (n=275) main block / 17% (n=69) general block	Answerable in two tiers
Obj 4 \| SES & access	ses_index, 10 asset/housing inputs, livestock holdings	86% (n=353)	Fully answerable

8 Data and Methods

8.1 Data source

Data were collected via a household survey administered across three wards of Isiolo County (Garbatulla, Kinna, Sericho), yielding 409 household responses. The survey instrument covered household demographics, brucellosis awareness and knowledge, management practices, perceptions/attitudes (5-point Likert items), and socio-economic status and access to services.

8.2 Data cleaning and quality assurance

The raw dataset (409 respondents, 155 variables) required substantial cleaning: standardising skip-logic missing codes, recoding categorical and Likert-scale variables, expanding multi-select survey items into individual binary indicators, and constructing composite indices. Several non-trivial data-quality issues were identified and corrected during this process, summarised here for transparency:

Multi-select response coding. Several multi-select survey items (animal signs, transmission routes, prevention practices, information channels) were originally expanded using search terms that did not match the actual raw response, in others due to apparent data-entry typos in the survey tool itself (e.g. "isolate_partuition", "restrict_mvmnt").
Knowledge score construction. The “cause of brucellosis” item is 97.7% missing even among brucellosis-aware respondents.
Perception items. Two disjoint sets of Likert perception items were administered to non-overlapping groups of respondents (69 respondents interviewed in the first three weeks of fieldwork received one item set; the remaining 275 aware respondents received a different, more specific item set). Rather than forcing these into a single blended column, perception indices are reported separately by module (perception_index_early, perception_index_later).
Skip-logic-as-zero. Several derived indicators (multi-select-based knowledge domains, prevention-practice adoption) are structurally undefined for brucellosis-unaware respondents, since the underlying questions were never asked of them. These are coded NA, not 0, for unaware respondents throughout, to avoid conflating “not asked” with “answered no.”
Data entry errors. Two respondents had no_phones values in the tens of millions (evidently a phone number typed into a count field) and one respondent had an implausible school_years value (44); both were corrected to missing rather than left to distort summary statistics.

8.3 Variable construction

Knowledge score. A composite score was built from binary correct/incorrect indicators across four domains: affected species, animal-disease signs, animal-to-animal and animal-to-human transmission routes, and human-disease signs, restricted to brucellosis-aware respondents and expressed as a percentage of the maximum achievable score.

Socio-economic status (SES) index. Constructed as the equal-weighted average of two domains: an asset/housing index (radio, bicycle, motorbike, car, house ownership, piped water, electricity, phone count, wall material, toilet type) and an income index (four income-source categories, each rescaled 0–1). Access to credit was deliberately kept as a separate explanatory variable rather than folded into the index.

Perception indices. Five sub-indices were constructed from 5-point Likert items (risk perception, prevention effectiveness, severity/impact, behavioural risk, and prevention-practice attitude).

8.4 Statistical analysis

Two generalised linear models were specified. Both restricted to brucellosis-aware respondents.

8.4.1 Quasibinomial GLM for knowledge

The obvious default for a “knowledge score” outcome would be ordinary least squares (OLS) linear regression on knowledge_score_pct. This was deliberately not used, for two reasons specific to how the score is constructed:

The outcome is a bounded proportion, not an unbounded continuous measure. knowledge_score_pct is built by summing binary correct/incorrect answers across a fixed number of knowledge items and expressing the total as a percentage of the maximum achievable score. That is exactly the data-generating process a binomial model describes (k successes out of n known trials), and OLS on the resulting percentage ignores this structure — among other things, it can predict fitted values below 0% or above 100%, and it assumes constant variance across the whole range when variance in a proportion is naturally largest near 50% and smallest near the boundaries.
A plain binomial GLM would understate uncertainty. Binomial variance assumes each of the underlying “trials” (each knowledge item) is answered independently. In practice, a respondent’s answers on the different knowledge items are correlated with each other — someone who correctly identifies several animal-disease signs is more likely to also correctly identify the transmission routes, because both draw on the same underlying exposure/awareness. This within-respondent correlation overdisperses the variance relative to what a naive binomial model expects. The quasibinomial family keeps the binomial mean structure but estimates the dispersion parameter from the data instead of fixing it at 1, which widens the standard errors to a more honest level given that correlation. The fitted model’s estimated dispersion parameter was well above 1, confirming this adjustment was necessary rather than a purely precautionary choice.

8.4.2 Bias-reduced (Firth-type) logistic regression for adoption

The obvious default here is ordinary binomial logistic regression. This was tried first, and it failed in a specific, diagnosable way: R’s glm() reported fitted probabilities numerically 0 or 1 occurred, and several coefficients came back with implausible magnitudes (estimates around 17–21 with standard errors in the thousands). This is the textbook signature of separation: with roughly 15–20 parameters (several multi-level categorical predictors: ward, age group, education, marital status) estimated on a few hundred respondents, and the minority outcome class fairly thin (42 of 344 respondents in the primary model, and only 7 of 275 in the secondary self-report model), some combination of predictor categories perfectly or near-perfectly predicts the outcome for a subset of respondents. Ordinary maximum likelihood has no finite solution in that situation — the optimiser pushes the relevant coefficient toward infinity and the standard error along with it, and the resulting p-values and confidence intervals are not trustworthy.

Bias-reduced (Firth-type) estimation [@firth1993], implemented here via brglm2 [@kosmidis2023; @kosmidis2021], solves this by penalising the likelihood with the Jeffreys invariant prior, which keeps coefficient estimates finite even when ordinary MLE would diverge, while remaining a standard, well-established, epidemiologically-accepted approach for exactly this rare-outcome/separation scenario. Re-fitting both adoption models with this method removed the separation warning entirely and brought every coefficient and standard error back to a sane, interpretable magnitude (full comparison available on request, since the unpenalised fit is not reported in this document — presenting divergent estimates would be misleading).

8.4.3 Model diagnostics

For both models, variance inflation factors were checked to rule out problematic multicollinearity in the final models — see Table 14 and Table 15’s accompanying VIF exports, and a variable with excessive missingness (primary_role, 78.5% missing among aware respondents) was excluded from the multivariable models after confirming that including it collapsed the usable sample from ~340 to ~72 via listwise deletion. income_source was collapsed from 8 categories to 4 for modelling purposes only (three categories had 1–4 respondents each, which produced unstable, uninterpretable estimates for those specific levels).

9 Results

9.1 Descriptive Findings

9.1.1 Respondent characteristics

Table 8: Continuous respondent characteristics
variable	n_valid	n_missing	pct_missing	mean	sd	median	min	max
age	409	0	0.0	43.6	16.6	40	18	97
school_years	408	1	0.2	4.8	5.4	0	0	18
no_hh_members	409	0	0.0	7.0	3.0	7	0	21
yrs_keep_livestock	408	1	0.2	24.4	18.4	20	0	80

Table 9: Categorical respondent characteristics (valid %)
variable	level	n	pct_valid
sex	Female	180	44.0
sex	Male	229	56.0
age_group	18-30	112	27.4
age_group	31-45	132	32.3
age_group	>45	165	40.3
edu_cat	No formal education	210	51.5
edu_cat	Primary (1-7 yrs)	38	9.3
edu_cat	Secondary (8-13 yrs)	117	28.7
edu_cat	Tertiary (14+ yrs)	43	10.5
marital_status	Married	341	83.4
marital_status	Separated/Divorced	11	2.7
marital_status	Single	30	7.3
marital_status	Widowed	27	6.6
religion	Christian	3	0.7
religion	Muslim	406	99.3
rel_hhead	Herdsperson	10	2.4
rel_hhead	Household head	234	57.2
rel_hhead	Nephew/Niece	1	0.2
rel_hhead	Sibling	4	1.0
rel_hhead	Son/Daughter	14	3.4
rel_hhead	Spouse of HH head	146	35.7
ward	Garbatulla	174	42.5
ward	Kinna	144	35.2
ward	Sericho	91	22.2
location	garbatulla_main	86	21.0
location	garbatulla_reserve	88	21.5
location	kina_main	51	12.5
location	kina_reserve	93	22.7
location	sericho_main	38	9.3
location	sericho_reserve	53	13.0

9.1.2 Awareness and knowledge

84.1% of respondents were aware of brucellosis. Among aware respondents, the mean knowledge score was 22.6% (SD 9.9, median 22%, range 2–66%) — indicating generally low knowledge levels even among those aware of the disease.

Table 10: Proportion of aware respondents correctly identifying each knowledge item, by domain
domain	item	n_base	pct_yes
a2a_routes	sbq5b_trans_anim_anim__cont_feed_water	344	13.1
a2a_routes	sbq5b_trans_anim_anim__poor_sanitation	344	7.3
a2a_routes	sbq5b_trans_anim_anim__direct_contact	344	5.5
a2a_routes	sbq5b_trans_anim_anim__sexual_trans	344	4.7
a2a_routes	sbq5b_trans_anim_anim__admin_same_syringe	344	4.7
a2a_routes	sbq5b_trans_anim_anim__cont_milk	344	2.3
a2a_routes	sbq5b_trans_anim_anim__inf_placenta	344	2.0
a2a_routes	sbq5b_trans_anim_anim__lick_calves	344	1.5
a2a_routes	sbq5b_trans_anim_anim__lick_genitals	344	0.9
a2a_routes	sbq5b_trans_anim_anim__ai	344	0.3
a2h_routes	sbq6b_trans_anim_hum__drink_raw_milk	344	92.2
a2h_routes	sbq6b_trans_anim_hum__raw_milk_products	344	78.8
a2h_routes	sbq6b_trans_anim_hum__uncooked_meat	344	49.1
a2h_routes	sbq6b_trans_anim_hum__milking	344	5.5
a2h_routes	sbq6b_trans_anim_hum__assist_bare_hands	344	4.4
a2h_routes	sbq6b_trans_anim_hum__slaughter	344	3.8
a2h_routes	sbq6b_trans_anim_hum__handle_fetal	344	3.2
a2h_routes	sbq6b_trans_anim_hum__living_animals	344	3.2
a2h_routes	sbq6b_trans_anim_hum__inhale_bacteria	344	0.6
animal_signs	sbq4_anim_signs__abortion	344	27.9
animal_signs	sbq4_anim_signs__still_birth	344	20.9
animal_signs	sbq4_anim_signs__swollen_joints	344	11.0
animal_signs	sbq4_anim_signs__reduce_milk	344	6.4
animal_signs	sbq4_anim_signs__retention_fetal	344	5.8
animal_signs	sbq4_anim_signs__weak_calf	344	5.5
animal_signs	sbq4_anim_signs__weight_loss	344	5.5
animal_signs	sbq4_anim_signs__swollen_testes	344	0.6
animal_signs	sbq4_anim_signs__female_infertility	344	0.6
animal_signs	sbq4_anim_signs__male_infertility	344	0.3
human_signs	sbq6c_hum_signs__joint_pain	344	99.1
human_signs	sbq6c_hum_signs__back_pain	344	84.0
human_signs	sbq6c_hum_signs__fever	344	68.9
human_signs	sbq6c_hum_signs__fatigue	344	66.3
human_signs	sbq6c_hum_signs__headache	344	60.2
human_signs	sbq6c_hum_signs__malaise	344	42.2
human_signs	sbq6c_hum_signs__chills	344	26.5
human_signs	sbq6c_hum_signs__appetite_loss	344	25.9
human_signs	sbq6c_hum_signs__night_sweating	344	9.9
human_signs	sbq6c_hum_signs__nausea	344	4.7
human_signs	sbq6c_hum_signs__abortion_hum	344	2.6
human_signs	sbq6c_hum_signs__vomiting	344	2.0
human_signs	sbq6c_hum_signs__pain_scrotum	344	1.7
human_signs	sbq6c_hum_signs__diarrhoea	344	0.3
human_signs	sbq6c_hum_signs__blurred_vision	344	0.3
species	sbq3_affected_species__cattle	344	66.3
species	sbq3_affected_species__goats	344	53.8
species	sbq3_affected_species__camels	344	19.2
species	sbq3_affected_species__sheep	344	16.0

9.1.3 Management practices

Table 11: Prevention/control practices adopted, among aware respondents
item	n_base	pct_yes
scq3b_prev_control__vaccination	344	55.2
scq3b_prev_control__test_new	344	33.7
scq3b_prev_control__isolate_infected	344	30.5
scq3b_prev_control__testing	344	15.7
scq3b_prev_control__restrict_mvmt	344	14.2
scq3b_prev_control__dispose_fetal	344	11.3
scq3b_prev_control__seek_vet	344	11.0
scq3b_prev_control__farm_sanit	344	6.7
scq3b_prev_control__disinfect	344	2.0
scq3b_prev_control__pub_educ	344	1.5
scq3b_prev_control__slaughter_pos	344	1.2
scq3b_prev_control__isolate_parturition	344	0.9

87.8% of aware respondents had adopted at least one specific prevention or control practice.

9.1.4 Perceptions

Table 12: Perception index summary statistics (1-5 scale)
variable	n_valid	mean	sd	median
risk_perception_index	69	4.0	0.6	4.0
prevention_effectiveness_index	69	3.6	0.5	3.5
severity_perception_index	69	4.1	0.5	4.0
behaviour_risk_index	275	3.5	0.6	3.5
prevention_practice_attitude_index	275	3.8	0.3	3.9
perception_index_early	69	3.9	0.4	3.9
perception_index_later	275	3.6	0.4	3.7

The risk_perception_index, prevention_effectiveness_index, and severity_perception_index reflect the early-fieldwork questionnaire module (n ≈ 69), while behaviour_risk_index and prevention_practice_attitude_index reflect the later, larger module (n ≈ 275). These are not directly comparable to one another.

9.1.5 Economic characteristics

Table 13: SES index components
variable	n_valid	mean	sd	median
asset_index	350	0.2	0.0	0.2
income_index	353	0.1	0.1	0.1
ses_index	350	0.1	0.0	0.1

9.2 Inferential Findings

9.2.1 Objective 1: Factors associated with knowledge

Table 14: Knowledge model: odds ratios (quasibinomial GLM)
Term	OR	2.5%	97.5%	p-value
sexMale	0.972	0.846	1.117	0.689
age_group18-30	0.911	0.738	1.124	0.385
age_group31-45	1.043	0.889	1.226	0.606
edu_catPrimary (1-7 yrs)	0.847	0.688	1.038	0.113
edu_catSecondary (8-13 yrs)	1.024	0.878	1.194	0.762
edu_catTertiary (14+ yrs)	1.035	0.841	1.270	0.747
marital_statusSeparated/Divorced	1.074	0.759	1.494	0.680
marital_statusSingle	1.107	0.852	1.431	0.445
marital_statusWidowed	0.947	0.731	1.220	0.679
no_hh_members	1.005	0.984	1.026	0.664
income_source_grpAgro-pastoralist	1.155	0.903	1.473	0.250
income_source_grpEmployed/Business	0.989	0.763	1.273	0.933
income_source_grpOther	1.598	1.273	1.999	0.000
yrs_keep_livestock	0.996	0.992	1.000	0.077
rec_ext_servicesyes	1.254	0.984	1.589	0.065
grp_memberyes	0.927	0.812	1.058	0.265
ses_index	3.821	0.959	15.074	0.057
wardKinna	1.019	0.884	1.175	0.791
wardSericho	0.718	0.608	0.845	0.000

Ward of residence was the strongest and most consistent predictor of knowledge: respondents in Sericho ward had significantly lower odds of correctly identifying knowledge items than the reference ward (Garbatulla), holding demographic, livelihood, and access variables constant. Household income source (the “Other” category, relative to the Pastoralist majority) was also significant, though this category is not well-defined substantively. No other demographic, livelihood, or access variable — including extension service receipt and the SES index — reached conventional significance, though both trended in the expected (positive) direction.

9.2.2 Objective 2: Factors associated with adoption of preventive practices

Table 15: Adoption model: odds ratios (bias-reduced logistic GLM, primary outcome)
Term	OR	2.5%	97.5%	p-value
knowledge_score_pct	1.030	0.989	1.072	0.150
ses_index	273.406	0.057	1314369.114	0.195
is_credityes	3.323	1.374	8.033	0.008
rec_ext_servicesyes	4.048	0.229	71.721	0.340
sexMale	0.373	0.145	0.960	0.041
age_group18-30	1.921	0.407	9.056	0.409
age_group31-45	0.883	0.350	2.227	0.792
edu_catPrimary (1-7 yrs)	0.427	0.148	1.232	0.115
edu_catSecondary (8-13 yrs)	2.463	0.768	7.903	0.130
edu_catTertiary (14+ yrs)	0.980	0.271	3.537	0.975
marital_statusSeparated/Divorced	0.348	0.040	3.001	0.337
marital_statusSingle	0.571	0.111	2.923	0.501
marital_statusWidowed	0.627	0.136	2.884	0.548
wardKinna	0.874	0.410	1.863	0.726
wardSericho	17.569	1.286	240.111	0.032

Access to credit was the strongest and most reliable predictor of adopting at least one prevention practice (OR = 3.32, 95% CI 1.37–8.03). Sex was also significant, with male respondents showing lower odds of adoption than female respondents. Ward (Sericho) was nominally significant but with a very wide confidence interval and should not be over-interpreted; the SES index coefficient in this model was numerically unstable (extremely wide interval) despite bias reduction and is not interpretable as reported.

A secondary model using the self-reported adoption question (rather than the itemised practice checklist) is available in output/inferential/6b_adoption_model_secondary_coefficients.csv. It disagreed with the primary model in the direction of two effects (knowledge score and credit access); given that outcome’s severe class imbalance (7 of 275 respondents in the minority class), the primary, itemised-checklist model is preferred for substantive conclusions.

10 Discussion

10.1 Awareness and knowledge (Objective 1)

The gap between awareness and knowledge is the first thing worth naming plainly: 84.1% of respondents had heard of brucellosis, but among those aware, the mean knowledge score was only 22.6%. Being aware that a disease exists is evidently not the same as understanding how it spreads or how to recognise it.

Within the knowledge domains, recognition was uneven. Some individual items were identified by a reasonable share of aware respondents (e.g. cont feed water in the a2a_routes domain, at 13.1%), while others were recognised by only a handful of respondents (the three lowest: male infertility (0.3%), ai (0.3%), diarrhoea (0.3%), blurred vision (0.3%)).

10.2 Management practices (Objective 2)

87.8% of aware respondents had adopted at least one prevention or control practice,but the itemised breakdown (Table 11) shows this is concentrated in a small number of practices,mainly vaccination (55.2%), test new (33.7%), isolate infected (30.5%) while several others in the checklist were adopted by under 5% of respondents (4 of the 12 listed practices).

Knowledge score itself was not a significant predictor of adoption in the primary model (p = 0.15), despite trending in the expected positive direction.

10.3 Perceptions (Objective 3)

Because the perception module was administered as two non-overlapping questionnaire version. Within both, mean scores across the five sub-indices sat in the “agree” range (roughly 3.5–4.1 on the 1–5 scale; Table 12).

10.4 Economic status and access (Objective 4)

The broad SES index was not a reliable predictor of either knowledge or adoption (positive trends in both models, but not statistically significant, and outright unstable in the adoption model), while the narrower is_credit variable was the single strongest predictor in the adoption model. Extension service receipt trended positive for knowledge (p = 0.065) but didn’t reach significance either.