Selecting Therapy Dogs: What Studies Show About Suitability, Assessment and Retirement
Michael Sauerwein
Many US therapy dog organizations require an in-person behavioral assessment before visits, but the organizational survey does not establish an overall pass rate for individual dog-handler applicants. In a US survey designed to be representative, 22 of 24 responding regional therapy dog organizations required a formal in-person behavioral evaluation. For 2017, 71% of these organizations reported that no more than 10% of new registration applications were unsuccessful, and six of 24 reported no unsuccessful applicants. Regular behavioral re-evaluation was required by 14 of the 24 regional organizations (Serpell et al., 2020). Whether any of these evaluations predicts which dogs will cope with the work, and for how long, has not been tested in the studies considered here. The large prospective selection datasets come from guide and assistance dog programs, where questionnaires and behavior tests predicted training outcome only modestly better than simply predicting the most common outcome, and were better at flagging likely failures than at picking successes (Bray et al., 2019).
This article reviews what is known about the suitability of therapy and visiting dogs: how dogs are selected today, what the few studies of therapy dog questionnaires and behavior tests found, what research on guide, assistance and detection dogs shows as a different population, how age, sex, breed, experience and the handler relate to behavior during work, and why therapy dogs are retired. It closes with practical consequences for trainers and handlers and with the main research gaps.
1. Why Selection Matters, for the Dog First
1.1 Therapy, Visiting, Facility and Assistance Dogs
Therapy and visiting dogs are, in most programs, privately owned dogs that work together with their owner as handler, often on a volunteer basis, in animal-assisted activities (AAA), animal-assisted therapy (AAT) or animal-assisted education (AAE). They usually live as family dogs, start the work as adults and are assessed once or a few times by an organization before they are registered. Assistance and guide dogs are a different population: most are purpose-bred, raised by volunteer puppy raisers, trained by professional staff for specific tasks and placed with one person. Detection and military dogs differ again in breeding, housing and tasks (what studies show for assistance dogs and their handlers).
Facility dogs sit between these groups. They are trained in assistance dog programs but work in schools, hospitals or courthouses alongside a professional. In a study of six US and Canadian assistance dog organizations with 678 successful and 1,082 released dogs, facility dogs showed behavioral profiles distinct from those of assistance dogs placed with individual clients (Hilby & McCue, 2026). Even among dogs raised and trained in the same kind of program, different jobs appear to suit different dogs. A framework paper on dogs in animal-assisted therapy makes the same point for therapy work: suitability depends on the fit between a dog, its handler and a specific setting, not on a test passed once (Winkle et al., 2020).
1.2 What "Suitable" Means in the Literature
An early review of therapy dog welfare described the usual requirements: dogs complete special training and a temperament test, and "minimum common prerequisites include immaculate health status, the willingness to interact with unfamiliar people and the absence of behaviors that could jeopardize AAI recipients" (Glenk, 2017). The same review added a caution that remains relevant: "Although many dog owners would consider their animals as ideal companions, this does not necessarily make them good candidates for AAIs."
Behind the word "suitable" lie at least three different questions. Is the dog safe for the people it meets? Does it tolerate the work without lasting strain? And does it benefit from, or at least enjoy, the encounters? A follow-up review noted that dogs are "the least controlling participant" in animal-assisted interventions and that "it is questionable whether dogs give any form of free and informed consent" (Glenk & Foltin, 2021). A review of positive welfare argued that selection should look at sociability and playfulness, traits directly relevant to positive encounters, because "all animal behavioral displays and physiological responses are dependent on the dog's individual (and breed) temperament"; it concluded that "to date, no studies have measured all of these factors to assess therapy dog welfare" (Miller et al., 2022). A practice paper added that dogs "do express 'assent and dissent' through body language and behavior", while many people misread signs of dissent (Townsend & Gee, 2021).
How a study defines suitability shapes what it measures. A critical review of 116 papers on dog and horse welfare in animal-assisted services found ethical frameworks that were "highly diverse and at times insufficiently specified", and described a "conceptual conflation" when positive and negative welfare perspectives are mixed (Zenker et al., 2026). A test that screens for aggression answers the safety question; it says little about whether the dog will enjoy the work.
1.3 What Kind of Evidence Exists
The evidence on therapy dogs themselves is small. A review of therapy dog welfare studies published between 2017 and 2021 found that eight of twelve studies had fewer than ten dogs, with a median of 7.5 dogs (Glenk & Foltin, 2021). Studies that directly address selection are even fewer: among the studies considered here, there is one questionnaire study of dogs that had passed an aptitude test, one study combining a behavioral examination with a simulated intervention, one pilot of a new behavioral test, two cross-sectional comparisons of therapy and pet dogs, one field study with 63 dogs and one survey on retirement. Among the studies reviewed here, large prospective datasets linking early tests to training or career outcomes come from guide, assistance and detection dog programs. This article therefore uses them as clearly labeled parallels. Session-level stress physiology, such as cortisol, heart rate and behavior during visits, is covered in detail elsewhere (what animal-assisted interventions mean for the dog); here it appears only where it bears on selection.
2. How Therapy Dogs Are Selected Today
2.1 Evaluations, Minimum Age and Re-Evaluation in the United States
The most detailed data on selection practice come from a survey of US therapy dog organizations conducted between January and March 2018 (Serpell et al., 2020). Four of six national organizations and 24 of 33 randomly selected regional organizations responded. Among the regional organizations, 22 of 24 (92%) required a formal in-person behavioral evaluation, 20 of 24 (83%) a minimum age of one year, which all four national organizations also required, and 9 of 24 (38%) the American Kennel Club's Canine Good Citizen test, which none of the national organizations required. Eleven of 24 required that the dog had lived in its home for at least six months, five required spaying or neutering, and only one excluded particular breeds.
Two findings stand out for selection. First, many surveyed organizations reported low failure rates: 71% of regional organizations reported that no more than 10% of new applicants failed to complete screening and registration successfully in 2017, and 25% reported no such failures. The survey concerns new handler applicants seeking registration, not a dog-only behavioral outcome. These organization-level reports cannot be pooled into a pass rate for individual dog-handler applicants without their respective applicant numbers. Second, regular behavioral re-evaluation was required by only 14 of 24 regional (58%) and two of four national organizations, although hospital guidelines reviewed in the same paper require re-evaluation "at least every 2–3 years". About half of the organizations imposed no time limit on visits. The survey reported no data on the validity of any of these evaluations. A low failure rate can reflect lenient tests, owners who do not apply with unsuitable dogs, or both; the survey does not allow these explanations to be separated. The study was funded by Mars Petcare and WALTHAM, and one author was a paid consultant to Mars on unrelated issues.
2.2 Governance in Europe and Conflicts of Interest
Comparable figures for European countries were not found among the studies considered here. A narrative review of selected European and US frameworks, based on academic databases and grey literature from organizations such as IAHAIO, ISAAT and national therapy dog associations, described governance as "fragmented" (Shajid Pyari et al., 2026). The review describes Austria as an example of a formally regulated framework, requiring training of both dog and handler and a "positive expert assessment"; it also identifies a nationally coordinated guideline system in Italy. However, its English-language grey-literature search did not systematically cover every European country. The comparison is illustrative rather than comprehensive. For Germany, the review mentions the Teilhabestärkungsgesetz, a law not specific to therapy dogs; that mention cannot be treated as a complete account of German law. Italy's national guidelines require a veterinary assessment before a dog joins canine-assisted interventions (Meers et al., 2022). In France, a handler survey stated that "currently, there is no systematic selection of dogs or required training" (Mignot et al., 2022).
Shajid Pyari et al. (2026) identified "a need to move beyond dog-level pass–fail certification toward assessment of the human–dog team within specific intervention contexts" and described readiness as "a career-long process involving careful selection, reward-based preparation, assessment of the human–dog team, monitoring during active service, workload adjustment, and retirement planning". Winkle et al. (2020) were more direct about existing tests: they argued that the quality of most rating and behavior tests is unknown or poor, including their validity and reliability, and called for urgent scientific attention. They also pointed to a possible conflict of interest when organizations offer both the education and the evaluation of dogs, and noted that it is not known whether AAA evaluations "accurately and consistently measure how a dog works contextually in AAT". Both papers are reviews or frameworks without new data.
2.3 What Experts Agree On, and What They Do Not
A Delphi study on minimum standards for canine-assisted psychotherapy with adolescents began with 40 international experts; only 21 took part in its second round. Across the two rounds, 32 of 49 proposed items reached consensus (Jones et al., 2024). The experts supported flexible, individualized welfare assessment rather than fixed work schedules. Standardized tools such as the Canine Behavioral Assessment and Research Questionnaire (C-BARQ) or quality-of-life scales were not proposed by panelists as minimum standards; the authors discussed individualized ethograms as a potentially more practical approach. Winkle et al. (2020), by contrast, suggested validated instruments such as the C-BARQ, the Dog Impulsivity Assessment Scale or the Canine Frustration Questionnaire for repeated monitoring. Expert agreement and expert disagreement are both opinion, not evidence of validity; the Delphi study shows which practices experts find workable, not which ones predict outcomes.
3. What Studies on Therapy Dogs Actually Tested
3.1 Questionnaires: The C-BARQ in Dogs That Passed
In Japan, owners of dogs that had passed the aptitude test of the Hokkaido Volunteer Dog Association completed the C-BARQ, a questionnaire with 101 items in the version used (Sakurama et al., 2023). After four exclusions, 110 dogs from 30 breeds were analyzed, most commonly Labrador Retrievers, Golden Retrievers and Toy Poodles. Active dogs were 2 to 15 years old, with a mean of 8.53 years; nine retired and 27 deceased dogs were included, the latter rated by their owners from memory. Of the 101 questionnaire items, 98 entered the factor analysis; 51 items were ultimately assigned to 14 factors. The authors concluded: "Given these personality traits and the fact that breed and age did not influence aptitude, we believe that a variety of dogs have the potential to become therapy dogs."
The design limits that conclusion. Only dogs that had passed were included, so there was no comparison group of dogs that failed. The study can show that dogs who passed came from many breeds and ages; it cannot show which traits predict passing, or that breed and age have no influence on aptitude. The authors themselves recommended that "future studies should compare their C-BARQ scores with those of dogs that failed the therapy dog aptitude test". Ratings from memory for deceased dogs add recall bias. The study received no external funding.
3.2 Behavior Tests: Simulated Sessions, SuiTe and the Owners' View
In Italy, Mongillo et al. (2015) studied 60 dogs: 20 dogs already working in animal-assisted interventions and rated suitable by their handlers, 20 working dogs rated suitable with reservations, and 20 non-working companion dogs. Every dog underwent a behavioral examination followed by a simulated intervention that included unexpected social and environmental stimuli. Agreement between the two procedures was limited, whereas inter-rater agreement for the video ratings of a subset of 21 simulated sessions reached a Cohen's kappa of 0.807. Simulations can reveal specific responses, but the study did not establish whether its assessments predict later welfare or suitability in actual interventions (Mongillo et al., 2015).
At the University of Pisa, a research group developed a pre-screening test for animal-assisted interventions, the dog-suitability test SuiTe (Russo et al., 2025). 44 dog and caregiver pairs were recruited by word of mouth. Three were excluded because health issues emerged during or immediately after the test, and three because their caregivers did not complete the questionnaire. The final analysis comprised 38 dogs aged 1 to 14 years. Only six had previously been certified as therapy dogs; the others were companion dogs. The test exposed dogs to environmental and social stimuli, including unusual objects such as wheelchairs, sounds, handling by unfamiliar people and food, and two independent evaluators scored each reaction on a grid of valence and arousal. Caregivers also completed a questionnaire that included the Italian C-BARQ, and salivary cortisol was measured before and after the test.
The test classified 24 dogs as unsuitable, nine as suitable and five as pending. Inter-rater reliability was moderate for arousal (ICC = 0.605; 95% CI 0.563 to 0.645) and good for valence according to the authors' classification (ICC = 0.699; 0.665 to 0.731). Suitable and pending dogs had higher valence scores, and dogs classified as unsuitable scored higher on questionnaire items about separation, attachment and fear or anxiety. However, differences in valence between suitability groups defined by the same test are not independent external validation of that test. The change in cortisol did not differ significantly between groups (p = 0.42); with groups this small, that does not show that the test situation affected all dogs equally.
The most striking result concerned the caregivers. They rated 22 of their dogs as suitable, whereas the test found nine; they rated 11 as unsuitable, whereas the test found 24. Agreement between caregiver judgment and test result was only fair (Cohen's kappa 0.285). The authors concluded that their study "highlights the complexity of this assessment and the limited awareness of caregivers regarding the situations their dogs face every day." Whether the test or the owners were closer to the truth cannot be answered, because the dogs were not followed into real intervention work; the predictive validity of SuiTe is unknown. The authors describe the results as preliminary, and the groups had to be merged for the statistics (14 versus 24 dogs) (what behavioral assessments in dogs can and cannot measure).
3.3 Are Therapy Dogs Different from Pet Dogs?
Two studies compared dogs working in animal-assisted interventions with pet dogs and reached opposite conclusions about social orientation. In Milan, 45 dogs, 20 registered in Italy's national AAI registry and 25 controls, faced an unsolvable task after their owners had completed a personality questionnaire based on reinforcement sensitivity theory (Piotti et al., 2021). AAI dogs looked longer at their owner than controls (β = 17.90 ± 7.45, p = 0.021) and oriented less toward the task. Dogs with high scores on the behavioral inhibition scale also looked at their owner more.
In Hungary, 90 dogs in three groups of 30, balanced for size, sex and breed, worked on a food-retrieval apparatus: untrained family dogs, family dogs trained for recreational activities and certified assistance or therapy dogs living with their owners, with assistance and therapy dogs pooled (Carballo et al., 2020). A single "social strategy" score differed by group (F(2, 85) = 16.477, p < 0.001, partial η² = 0.275): untrained dogs relied most on humans, working dogs least, and the working dogs solved the task more independently. Age and sex had no effect.
The tasks, scoring methods and populations differ, so neither result is settled. Both studies share a basic limitation that the authors named: they are cross-sectional. Carballo et al. (2020) considered it possible that dogs that completed training as assistance or therapy dogs "had pre-existing characteristics that distinguished them from other dogs"; Piotti et al. (2021) noted that owners who believe their dog will enjoy a test are more likely to volunteer. Whether working dogs differ because they were selected that way or because the work changed them cannot be separated in these designs.
3.4 Temperament as the Missing Variable
Several reviews note that temperament is rarely measured in therapy dog research, even though it probably shapes every other result. A literature review organized along Welfare Quality categories, which assessed 423 articles in detail and selected 118, concluded that "the dog's temperament and individuality is seldom investigated, although it is likely that this influences physiological and behavioral outcomes" (Meers et al., 2022). In a review, Miller et al. (2022) recommended considering sociability and playfulness in assessments, but did not validate a new selection instrument. Among the studies considered here, no such instrument has been validated against later welfare or career length (how temperament and coping style differ between dogs).
In practice: For a school that trains visiting and therapy dog teams, it is worth assessing a candidate dog more than once, in at least two different settings and with people the dog does not know, and having a second trained observer score the dog independently of the owner. A sensible approach is to treat the first supervised visits as part of the assessment rather than as proof of a test already passed. This is a practice recommendation, not a study result.
4. Guide, Assistance and Detection Dogs: A Different Population
The studies in this section concern purpose-bred, professionally raised dogs with defined training outcomes, usually "qualified" versus "released" or "withdrawn". Their ages at testing, base rates and definitions of success differ from those of privately owned adult therapy dogs. They are included because they provide the large prospective datasets available within this review for examining whether tests predict training outcomes. They show what selection can achieve at best under favorable conditions, not what a single therapy dog test achieves.
4.1 Half or Fewer Qualify, Mostly for Behavioral Reasons
A systematic review of behavioral tests for working dogs cited estimates that on average only about half of working dogs become fully operational (Brady et al., 2018). At Canine Companions for Independence in the United States, program success averaged 43% over 13 years, and field-wide success typically ranges from 30% to 50% (Bray et al., 2019). In a Japanese guide dog study, the authors cited figures showing that behavior accounts for roughly 70% of disqualifications, reported as 77.3% in Australia, 65.5% in the United States and 69.5% in Japan (Arata et al., 2010). A review summarized that "even after undergoing rigorous selection and training, large numbers of dogs who enter training fail to complete these programs, largely for behavioral reasons", and that organizations with their own breeding programs report the highest success rates (Bray et al., 2021).
The closest parallel to therapy dogs comes from Belgium. Five assistance dog associations without their own breeding, which bought dogs from the general population, enlisted 537 dogs between 2001 and 2015 (Bogaerts et al., 2019). Of these, 92% remained after the host family phase, 75% after orthopedic screening, and 60% (324 dogs) became assistance dogs. The most frequent behavioral reason for rejection was "insecure and/or anxious" (41 dogs), followed by aggression (16), excitation (8), lack of motivation (5) and predatory behavior (3). Some dogs were given more than one reason, and the paper's percentages for these categories use different denominators (73 and 63), so the counts are the clearer figures. Among the guide and mobility assistance dogs compared, Golden Retrievers and Labrador Retrievers did not differ significantly in the probability of completing training successfully. The average financial loss per rejected dog was estimated at €10,524. Even dogs chosen from popular, family-friendly breeds were rejected most often for insecurity and anxiety.
4.2 How Good Is Prediction?
The largest prediction study used data from Canine Companions for Independence: puppy-raiser C-BARQ ratings for 3,569 dogs, a standardized temperament test at the start of professional training for 5,967 dogs, and both for 2,990 dogs (Bray et al., 2019). Models were trained on part of the data and validated on the rest. Overall accuracy was 64% to 68%. The C-BARQ model reached an accuracy of 0.68 and an area under the curve (AUC) of 0.71, but the no-information rate, the accuracy of simply predicting the most common outcome for every dog, was already 0.60; for the temperament test, 0.66 compared with 0.58. Combining the two instruments did not improve overall accuracy. The pattern at the extremes was more informative: among the 10% of dogs with the lowest predicted probability of success, predictions were 85% to 92% accurate, among the top 10% only 62% to 72%. The tests were better at identifying dogs likely to fail than at picking dogs likely to succeed. The study was funded by the Office of Naval Research and the AKC Canine Health Foundation.
Other programs show similar limits. At The Seeing Eye, 98 dogs were followed from birth: maternal behavior in the first three weeks predicted outcome with an AUC of 0.664, temperament and cognition tests at 14 to 17 months with 0.750, and the combination with 0.754 (Bray et al., 2017). A UK guide dog questionnaire completed by training supervisors at 5, 8 and 12 months was associated with outcome on all seven scales at one or more ages, but thresholds chosen to avoid false alarms identified only 8.4% of the dogs later withdrawn for behavior and 8.5% of qualified dogs, with specificities of 84% and 85% (Harvey et al., 2017). High specificity at this level means that most later failures were not flagged. In Japan, a trainer questionnaire after three months of training classified 80.6% of 144 Labradors correctly with one cut-off, and a stricter cut-off flagged 28.2% of later failures without any false alarms (Arata et al., 2010). The 80.6% accuracy was obtained within the study sample and was not confirmed in an independent external sample. The raters were the center's own trainers, who also contribute to the final decision, which may inflate agreement between rating and outcome.
Cognitive and neural measures have been tested as well. A 25-item cognition battery was examined in assistance and detection dogs. In an independent validation sample of 180 prospective assistance dogs, 77% graduated; predictions were on average 86% accurate for the quarter with the highest predicted chances of success, not for the full sample; successful assistance dogs made more eye contact with an experimenter in an unsolvable task or when a joint activity was interrupted (MacLean & Hare, 2018). Functional MRI responses to hand signals were examined in 49 dogs entering assistance-dog training. However, the prediction model used only 43 dogs with eligible training outcomes: 33 graduated and ten were released for behavioral reasons; four breeding dogs and two medical releases were excluded. The AUC was 0.91 in the analyzed sample and 0.80 under four-fold cross-validation; no independent external validation was performed. Two authors held equity in a company related to the method (Berns et al., 2017). In 990 dogs at the Israel Guide Dog Center, a model based on a standard behavior checklist reached an AUC of 0.730; adding trainers' free-text comments processed by a large language model raised it to 0.830 to 0.834 (Solomon et al., 2026). Here, "success" was defined specifically as becoming a guide dog or entering the breeding program. Dogs placed in other service roles were labeled "unsuccessful", along with dogs released from the program (314 positive versus 676 negative outcomes). The model thus predicts a program-specific outcome rather than suitability for working-dog roles in general. Because trainers' comments may already contain their judgment of the dog, part of that gain may reflect trainer expertise rather than a better test.
Very high positive predictive values can also mislead. In 3,969 Labradors at Guiding Eyes for the Blind, positive predictive values for favorable mean behavior checklist scores ranged from 91.5% to 99.9% (Hilby et al., 2025). What the study predicted was a favorable score on the same checklist at a later assessment, not completion of guide dog training. When favorable checklist scores are common, high positive predictive values can occur even if unfavorable developments are not reliably identified. The authors also noted that poor early scorers are often removed, so later data are "biased toward higher-scoring dogs". A review concluded that "at least a subset of measures in all of these studies had some predictive validity", but also that "current behavioral tests are imperfect" (Bray et al., 2021).
4.3 Why Early Tests Differ in Predictive Value
Behavior changes with development. A meta-analysis of 31 studies with 7 to 938 dogs found an overall personality consistency of r = 0.43; consistency was higher in adults (r = 0.51) than in puppies (r = 0.30), higher over shorter intervals and with identical tools, and in puppies, responsiveness to training, fearfulness and sociability were the least consistent traits (Fratkin et al., 2013). These are exactly the traits most relevant to therapy work. In Border Collies, 99 puppies were tested as neonates and 134 at 40 to 50 days; only 50 were assessed again as adults aged 1.5 to 2 years. There was "little correspondence" between developmental tests, and the authors concluded that "the predictive validity of early tests for predicting specific behavioural traits in adult pet dogs is limited" (Riemer et al., 2014) (what puppy temperament tests can and cannot predict).
Rearing conditions can be associated with later test scores, but these observational associations alone do not establish causation. In 978 prospective guide dogs, more experienced puppy raisers and living with at least one other dog were associated with better C-BARQ scores at 12 months, while puppies that had been frightened by a person or threatened by an unfamiliar dog had worse scores for fear of strangers or of dogs (Serpell & Duffy, 2016). The C-BARQ copyright is jointly owned by the corresponding author and the University of Pennsylvania, as the paper declared. In the Seeing Eye study, more maternal behavior was associated with a higher chance of later release (odds ratio 3.39 per standard-deviation increase in the maternal-behavior measure); the authors saw benefits of maternal care "in moderation", but mothers raised their own puppies, so genetic effects could not be separated (Bray et al., 2017).
Detection-dog studies show why it matters which part of a test battery is being discussed. Of 60 puppies from a university breeding program tested at 3, 5 and 11 months and at the end of training, 28 were selected and 32 rejected (Lazarowski et al., 2021). In the performance-test component, "Working Arousal" at three months predicted later selection (p = 0.048), while the corresponding predictors were not significant at five or eleven months. But other components of the same test battery told a different story: measures of emotional reactivity and environmental responses showed significant associations with selection at three, five and eleven months. It would therefore be incorrect to say that none of the test variables at five or eleven months predicted outcome. Individual significant findings from multiple analyses warrant caution, and this detection-dog study does not establish predictive validity for therapy dogs.
Single tests also capture some behaviors better than others. In a study of 123 dogs assessed in an Australian shelter before adoption, friendly, fearful and anxious responses were statistically associated with corresponding behavior one month after adoption. However, predictive validity for aggression, resource guarding and separation-related problems could not be adequately assessed: dogs showing these behaviors in the shelter were not rehomed and therefore not followed up (Clay et al., 2020). How well a single test detects these safety-relevant, context-specific problems remains unresolved.
4.4 Job-Specific Instruments and Career Fit
Job-specific instruments can capture behavioral traits omitted by general questionnaires, but this does not automatically establish greater predictive accuracy. In 359 guide dogs rated at eight months, a refined puppy walker questionnaire designed for the program was associated with training outcome on five scales, the C-BARQ on two, excitability and separation-related behavior (Hunt et al., 2020). The C-BARQ "was developed originally to study the prevalence of behavior problems in the pet dog population, and it therefore lacks the capacity to measure specific behavioral propensities" relevant to working dogs, according to a study that adapted it for 1,117 detection dogs; five of its 15 factors were new: playfulness, impulsivity, distractibility, basophobia (fear of falling) and food focus (Hare et al., 2024). That working-dog version has not yet been tested for outcome prediction, and among the studies considered here no therapy-dog-specific version exists.
Career fit also matters. In a retrospective analysis from six assistance dog organizations, each one-unit worsening in "unwillingness to settle" was associated with 58.7% higher odds of a response-dog career, and male dogs had 60.8% higher odds than females of that career category; facility dogs also had behavioral profiles distinct from other assistance dogs (Hilby & McCue, 2026). These percentages concern odds, not absolute probabilities of training success. The authors emphasize that checklist scoring is subjective and that their retrospective design cannot establish predictive capacity or show that behavior-based career assignment improves placements. Different jobs can place different demands on dogs.
4.5 What Transfers to Therapy Dogs
The transfer is limited. Therapy dogs are mostly adult family dogs of many breeds, assessed once or a few times by volunteers or instructors, and their "success" is rarely defined. Whether the lessons of working-dog research hold for them has not been tested in the studies considered here. Working-dog data show modest prediction overall and, in some samples, better identification of likely failures than of likely successes (Bray et al., 2019). Multiple time points and task-specific observations can provide additional information, but they have not been shown to be universally superior to a single general assessment. In the large study by Bray et al., combining two instruments expressly did not improve overall accuracy.
5. Age, Sex, Breed and Experience
5.1 Age and Experience
Most organizations set a minimum age of one year (Serpell et al., 2020). A review on retirement cited earlier work stating that "mature dogs are selected as therapy animals more often than young (less than two years of age) dogs because they are calmer and demonstrate fewer behavioral indicators of stress" (Ng & Fine, 2019). In the adult dog, personality is more consistent than in puppies (Fratkin et al., 2013), which supports assessing adults, but preferences still change: "things the dog enjoyed at 2 years old may be very different at 7 years old" (Winkle et al., 2020) (how behavior changes as dogs age).
The largest field dataset comes from the Netherlands, with 63 dogs, 837 sessions and 30 professionals across animal-assisted activities, education, coaching and therapy (Haven-Pross et al., 2026). Handlers scored 19 behaviors after a six-hour online training and a competency test. The following individual coefficients come from models stratified by session type, particularly animal-assisted activities (AAAs); they must not be generalized to education, coaching and therapy sessions as a whole. Older dogs had higher scores on a component labeled "detached" (B = 0.01, p < 0.001); dogs that started the work later in life had lower "playful" scores (B = −0.18, p < 0.001); and dogs with more years of experience had higher "thoughtful" (B = 0.19) and "playful" (B = 0.08) and lower "uncertain" scores (B = −0.06; all p < 0.001). The component labels in the abstract differ from those in the results tables; the labels here follow the results. The study is observational, and handlers mostly scored their own dogs. Experience effects may partly reflect survivorship: dogs that disliked the work may have stopped earlier and are therefore underrepresented among experienced dogs. In the Japanese questionnaire study, active therapy dogs averaged 8.5 years, which shows that many older dogs continue to work but does not test age effects (Sakurama et al., 2023).
5.2 Sex and Breed
In the Dutch field analysis for animal-assisted activities (AAAs), female dogs had lower "thoughtful" (B = −0.34) and "playful" (B = −0.11) and higher "uncertain" scores (B = 0.10; p < 0.001) (Haven-Pross et al., 2026). In the Hungarian problem-solving study, sex had no effect (Carballo et al., 2020), and in assistance dogs, males were more often found in response careers (Hilby & McCue, 2026). These results do not point consistently in one direction and are not a basis for selection rules.
Breed findings are similar. Dogs that passed the Japanese aptitude test came from 30 breeds (Sakurama et al., 2023). The French handler survey described 53 retrievers and 39 herding dogs among 165 dogs overall (Mignot et al., 2022). The percentages printed in the source do not consistently use 165 as their denominator, so the absolute counts are reported here. In some session-specific models of the Dutch dataset, breed groups differed from cross-breeds; for example, sheep and cattle dogs had higher "cautious" scores (B = 0.20) (Haven-Pross et al., 2026). This is not a general breed effect across every intervention type. Within the guide- and mobility-assistance groups studied in Belgium, Golden and Labrador Retrievers did not differ significantly in training success (Bogaerts et al., 2019), while in UK guide dogs, breed and sex shifted the odds of specific reasons for withdrawal (Caron-Lormier et al., 2016b). A review stated that "any breed of dog has the potential to be a therapy animal" (Ng & Fine, 2019). Breed describes probabilities at group level; it does not replace an assessment of the individual dog.
5.3 The Handler
The handler is part of selection, both as an observer and as half of the team. In a pilot with nine new teams at the Mayo Clinic, dogs with no previous visiting experience and a mean age of 3.7 years were studied during their first three visits; dog cortisol did not change significantly across visits, and handlers' ratings of their dog's stress correlated with the dog's cortisol (P = 0.02) (Clark, Smidt, et al., 2020). Handlers can be perceptive. In a French survey of 111 handlers about 165 dogs, however, handlers scored highest on stopping a session once the dog was stressed and low on stopping before stress appeared, and the authors noted "a potential for positive bias in the handlers' representations" (Mignot et al., 2022). A review on retirement described how "a handler may be blind to stress signals when there is a personal motive to continue volunteering", and how organizational rewards for volunteer hours may make retirement harder to accept (Ng & Fine, 2019).
Team fit cannot be reduced to characteristics of the dog alone. In the Dutch field study, working with a handler unfamiliar to the dog was associated with higher tension in animal-assisted coaching sessions (Haven-Pross et al., 2026). In guide dogs, a New Zealand study of 50 handlers and 118 pairings found that not every early return was attributed to a mismatch and that some handlers kept dogs they considered poorly matched (Lloyd et al., 2016). Suitability belongs to a team, not to the dog alone.
6. Willingness and Enjoyment During Work
6.1 Assent and Dissent
If dogs cannot consent in a human sense, their ongoing willingness has to be read from behavior. Townsend and Gee (2021) describe a "canine consent test": the person turns sideways and holds an open palm near the chest; a dog that comes close and seeks contact signals assent, while a dog that stays motionless or moves away signals dissent. Winkle et al. (2020) advised against using recipients, such as patients or children, for such consent tests, for safety reasons. Assent and dissent are concepts from practice and expert opinion; among the studies considered here, no measure of willingness has been validated against later welfare or withdrawal (what calming signals are and how to read them).
6.2 What Physiology and Behavior Show in Selected Dogs
Studies of dogs already working mostly report little physiological strain under favorable conditions. In 15 registered AAA dogs, a 60-minute activity with college students did not appear to cause significant activation of the stress axis, while cortisol at 30 minutes was higher in an unfamiliar room than during the activity or at home. Dogs also spent significantly more time standing and ambulating during the activity than at home; these activity differences are not automatically stress signs, and unchanged cortisol alone does not establish positive welfare (Ng et al., 2014). In 19 Mayo Clinic therapy dogs, heart rate was lower at the end of 20-minute outpatient sessions in which dogs were free to interact or not, and cortisol and oxytocin showed no significant change (Clark, Martin, et al., 2020); in 14 dog and handler pairs in inpatient stroke rehabilitation, heart rate fell and heart rate variability rose after 30-minute sessions, while cortisol, oxytocin and stress behavior showed no significant change (Shih et al., 2025). In the same study, right tympanic temperature nevertheless rose slightly but significantly (from 38.18 to 38.27 °C; p = 0.034), so the physiological indicators did not all point in the same direction. The authors of the latter noted that their highly trained teams "may not be representative for most therapy dogs". These three studies were funded by Nestlé Purina. For Shih et al. (2025), the authors additionally disclosed that the funder participated in data analysis and interpretation. In Rome, nine dogs observed for 174 hours showed no aggressive or stereotypic behavior and more attention, affiliation and play during sessions than before or after (Corsetti et al., 2019).
A scoping review found that only 9 of 27 studies of molecular stress markers in canine-assisted interventions measured dogs, that interventions "typically do not cause negative impacts on the studied canine stress markers", and that many studies did not time sampling to expected biomarker peaks (Gandenberger et al., 2022). In 56 therapy dogs, those working in Ukrainian military hospitals had lower urinary cortisol than German dogs in peaceful settings, and no salivary cortisol change was attributable to sessions; the authors concluded that "the environment rather than involvement in DAIs has an impact on therapy dogs' cortisol secretion" (Foltin et al., 2026). The study was cross-sectional with convenience samples.
A further finding concerns signs of positive welfare. Visser et al. (2026) asked 13 experts to set thresholds for 19 observable behaviors using hypothetical session scenarios and compared these thresholds with 837 recorded sessions involving 63 dogs. Stress-related behaviors were mostly within the ranges judged acceptable by the experts. In contrast, play, voluntarily lying down and wide, slow tail wagging were often observed less frequently than the proposed minimum thresholds for positive-affective behavior. Low levels of overt stress therefore do not, by themselves, demonstrate positive welfare. These thresholds are expert judgments rather than validated clinical cutoffs. Moreover, the dataset is the same 63 dogs and 837 sessions analyzed by Haven-Pross et al. (2026), not an independent replication (Visser et al., 2026).
For selection, these results have a specific meaning. They describe dogs that had already been selected, trained and kept in the work. They show that some measured stress markers were not elevated in selected working dogs under the studied conditions; this does not establish freedom from strain or positive welfare overall. They offer little information on dogs that were not accepted or on subsequent withdrawals. A best-practice paper from a long-running US hospital program warned that "well-trained and temperament tested dogs who are newly placed into a hospital setting may become stressed and less responsive to their handlers" (Barker & Gee, 2021). The details of session-level physiology are covered in the article on therapy dog welfare linked above.
6.3 Individual Dogs Who Say "No"
Even a previously selected dog may avoid contact or wish to end an interaction during a particular visit. Townsend and Gee (2021) describe approaching and initiating touch, on the one hand, and remaining still or moving away, on the other, as possible indications of willingness or unwillingness. This is practical guidance, not a prospectively validated test of future suitability. The field study by Haven-Pross et al. (2026) also found associations between dogs' behavioral responses and characteristics of the session and participating people. Passing a selection test therefore does not replace monitoring the dog during each visit.
In practice: Teams trained at our school have a written agreement with each facility allowing the handler to interrupt or end a visit at any time for the dog's sake. Visits have a maximum duration, rest breaks take place in a quiet room, the dog may leave contact without being called back, and a visit is stopped when stress signals such as lip licking, yawning or turning away accumulate. This describes practice, not a study result.
7. Why Dogs Leave the Work
7.1 Therapy Dogs: Age, Illness, and "No Longer Enjoyed It"
The only data found on why therapy dogs are retired come from an anonymous online survey of volunteers in animal-assisted services, conducted between January and June 2025 and recruited via social media and organization newsletters (Kogan et al., 2026). Of 247 responses analyzed, 227 answered the relevant question, and 83 of them (36.6%) had retired a therapy dog. Multiple reasons could be given: old age was named for 46 dogs (55.4%), medical issues for 36 (43.4%), the dog "didn't seem to enjoy it anymore" for 21 (25.3%), behavioral or stress-related issues for 14 (16.9%), the handler's health for five and lifestyle changes for four. Most respondents (91.6%) felt the timing had been right, 3.6% that it had been too early and 4.8% too late.
Roughly one in four retirements was attributed at least partly to the dog no longer enjoying the work, and about one in six to behavior or stress. The authors named the limits: an "online, self-selected sample ... may have introduced selection and expectancy biases". The reasons are the handlers' retrospective views, and handlers who notice a dog's reluctance may be more likely to answer such a survey. The figures are not population rates, but they show that loss of willingness is a recognized reason for ending the work.
7.2 When and How to Retire
Ng and Fine (2019) noted that decisions about when and how to retire therapy animals are usually left to handlers, while the scientific basis for those decisions remains largely unresolved. It proposed a Therapy Animal Quality of Life Scale, which has not been validated. Barker and Gee (2021) noted that handlers "may not recognize changes in their dogs over time that may indicate they are fatigued or no longer enjoying the hospital setting" and that "the CAI program must be willing to administratively retire a therapy dog if it is in the best interest of the dog". Winkle et al. (2020) found that therapy work "continues to lack formal processes such as annual registration evaluations ... including recognition of the dog's need to retire and retirement transition protocols". All three are expert opinion. Together with the survey data, they suggest that retirement is decided case by case, mostly by the handler.
7.3 Parallels: Guide and Military Dogs
Working-dog data show that behavior-related losses continue long after selection. In UK guide dogs that left service between 1994 and 2013, restricted to dogs retired for age or withdrawn for behavior, 6,465 of 7,770 (83%) were retired and 1,305 (17%) withdrawn for behavior (Caron-Lormier et al., 2016b). The main groups were environmental anxiety (321 dogs), willingness or confidence (311) and fear or aggression (226); the abstract names training rather than willingness among the main groups. Fear and aggression problems tended to appear early, whereas withdrawals for willingness and environmental anxiety "can occur at any age"; withdrawals for willingness occurred at a median of 1,617 days after qualification, when dogs were about 6.4 years old. Withdrawal decisions were made by staff together with the owner, weighing human safety, dog welfare and the dog's ability to work.
In the same population, including health withdrawals, 71.9% of about 8,994 working guide dogs were retired, 14.5% withdrawn for behavior and 13.5% for health (Caron-Lormier et al., 2016a). Dogs withdrawn for behavior had worked a mean of 1,152 days, those withdrawn for health 2,001 days and retired dogs 3,103 days, so behavioral withdrawals cost more working life than health withdrawals. In New Zealand, 43 of 118 guide dogs were returned before retirement age, 26 for behavioral, 14 for physical and three for handler-related reasons (Lloyd et al., 2016). Among 1,230 US military working dogs discharged in fiscal years 2019 to 2021, training accounted for 14.31% and fear-anxiety for 9.10% of discharges, and each additional service year was associated with 29.01% lower modeled odds of fear-anxiety rather than another reason for discharge among the dogs already discharged (Discepolo et al., 2025). This retrospective analysis of already discharged dogs does not establish that longer service reduces fear. Fear problems tended to surface early, loss of willingness at any age. Selection does not end with the first test.
In practice: Each visit by a team trained at our school is documented. It is worth reviewing these records at regular intervals for gradual changes, such as shorter contact phases, more frequent withdrawals or slower recovery after visits, and agreeing on criteria for reducing work or retiring the dog when a team starts, not when problems appear. This is a practice recommendation, not a study result.
8. Why the Evidence Stays Thin
8.1 Survivor Samples and Missing Follow-Up
Most therapy dog studies sample dogs that passed or are still working: the Japanese questionnaire study included only dogs that had passed (Sakurama et al., 2023), the Dutch field study included only working dogs (Haven-Pross et al., 2026), and the retirement survey relied on self-selected volunteers (Kogan et al., 2026). In guide dogs, poorly scoring dogs are removed early, which biases later data (Hilby et al., 2025). Among the studies considered here, no therapy dog evaluation has been followed prospectively to later welfare, career length or withdrawal (Sakurama et al., 2023; Russo et al., 2025; Serpell et al., 2020).
8.2 Base Rates and Accuracy
Accuracy figures only make sense relative to base rates. When 60% of dogs succeed, a "test" that predicts success for every dog is right 60% of the time (Bray et al., 2019). High specificity combined with low sensitivity (Harvey et al., 2017) can look impressive while missing many failures. In Hilby et al. (2025), very high positive predictive values concerned favorable scores on later behavior checklists, not graduation from training; their interpretation depends on how common favorable ratings are. In the Israeli dataset, dogs placed in other service roles were assigned to the negative class, so accuracy and AUC must be interpreted in light of both the imbalanced classes and this specific outcome definition (Solomon et al., 2026).
8.3 Reliability Rarely Reported
In the systematic review of 16 working-dog test papers, inter-rater reliability was reported in only four, within-session intra-rater reliability in three and test-retest reliability in four, with mixed results; the authors found "a widespread lack of information relating to the reliability and validity of measures to assess behaviour and inconsistencies in terminologies, study parameters and indices of success" (Brady et al., 2018). Inter-rater agreement for video scoring of a subset of 21 simulated animal-assisted sessions was high (kappa 0.807; Mongillo et al., 2015). The SuiTe pilot distinguished moderate agreement for arousal (ICC 0.605) from good agreement for valence (ICC 0.699; Russo et al., 2025). The working-dog cognition battery reported high coding agreement (kappa 0.96; MacLean & Hare, 2018). None of these figures, on its own, establishes prediction of later welfare in therapy dogs. Reliability of routine evaluations used in therapy dog practice was not reported in the US survey (Serpell et al., 2020).
8.4 Raters Who Also Decide
When the people who rate a dog also decide its future, or have a stake in it, ratings and outcomes are not independent. This applies to trainer questionnaires (Arata et al., 2010), to trainer comments in prediction models (Solomon et al., 2026), to handlers scoring their own dogs (Haven-Pross et al., 2026) and, in the other direction, to caregivers who judged their dogs much more suitable than a test did (Russo et al., 2025).
8.5 Reporting and Funding
A conceptual methodological analysis attributed mixed results in human-animal interaction research partly to variability in both human and animal participants (Rodriguez et al., 2021), and studies of interventions rarely report how their dogs were selected. A new reporting guideline for dog-assisted intervention trials lists the absence of manualized protocols, "including details relating to the animal partner", among common problems and states that trials "must consider the welfare and safety for the patient, dog, and therapist/handler" (Hall et al., 2026). This is reporting guidance, not a validation study of a dog-selection test. Several studies were funded by the pet food industry (Ng et al., 2014; Clark, Martin, et al., 2020; Shih et al., 2025; Serpell et al., 2020), guide dog studies often by the organization that also owns the data (Harvey et al., 2017; Hunt et al., 2020; Caron-Lormier et al., 2016a), and in the fMRI study two authors held equity in a related company (Berns et al., 2017). Funding is not proof of bias, but readers should know it.
9. Practical Consequences for Trainers and Handlers
The studies support several principles, within their limits. Assessing adult dogs rather than puppies is consistent with greater personality consistency in adults (Fratkin et al., 2013); Riemer et al. (2014) also documented the limited ability of very early tests to predict behavior in adult pet dogs. Repeated assessment rather than one test is consistent with the working-dog data on development and with behavioral withdrawals that occur years after selection (Caron-Lormier et al., 2016b; Hilby et al., 2025). In an assistance-dog program, tests offered some advantage in identifying dogs with especially low prospects of success (Bray et al., 2019). For aggression and resource guarding, Clay et al. (2020) could not adequately evaluate the shelter test’s predictive validity because of the preselection of dogs available for adoption. Owner judgment alone is not enough: caregivers and a structured test disagreed considerably (Russo et al., 2025), and handlers tend to view their dogs positively (Mignot et al., 2022).
The handler and the setting belong in the assessment. Reviews and frameworks favor assessing the team in the context where it will work (Shajid Pyari et al., 2026; Winkle et al., 2020), and handler training to recognize early stress signals is a recurring recommendation (Mignot et al., 2022; Townsend & Gee, 2021). Regular re-evaluation, which hospital guidelines require at least every two to three years, was not universal in the US survey (Serpell et al., 2020). Retirement should be planned from the start, because loss of willingness is among the reasons handlers give for retiring dogs (Kogan et al., 2026), and because handlers may not notice gradual changes (Barker & Gee, 2021). Breed and sex are not useful selection rules on the present evidence (Haven-Pross et al., 2026; Ng & Fine, 2019). For specific settings such as hospitals, additional requirements apply (what studies show for dogs in hospitals).
In practice: A sensible approach is to re-assess each working dog at least once a year, and sooner after illness, a move or a noticeable change in behavior, with the handler present and an observer who does not own the dog. It is also worth discussing openly with handlers at the start that a dog may outgrow the work, and that ending it is part of the job, not a failure. This is a practice recommendation, not a study result.
10. Summary at a Glance
In the US survey, 22 of 24 regional therapy dog organizations required a formal behavioral evaluation. 71% reported that no more than 10% of new registration applicants were unsuccessful in 2017, while six of 24 reported that all applicants succeeded. Only 58% of these organizations required regular re-evaluation (Serpell et al., 2020). Two studies reported inter-rater reliability measures: Mongillo et al. (2015) for video scoring of a subset of simulated interventions and Russo et al. (2025), with moderate agreement for arousal and good agreement for valence. In the SuiTe pilot, caregivers judged their dogs suitable far more often than the test did. A questionnaire study of dogs that had passed an aptitude test found 30 breeds but could not test which traits predict passing (Sakurama et al., 2023), and two comparisons of therapy and pet dogs pointed in opposite directions (Piotti et al., 2021; Carballo et al., 2020).
In guide and assistance dogs, a different population, tests predicted training outcome with accuracies of 64% to 68% compared with 58% to 60% from simply predicting the most common outcome, and were better at identifying likely failures (Bray et al., 2019). Puppy behavior is less consistent than adult behavior (Fratkin et al., 2013). In therapy dogs, about one in four retirements in a self-selected survey was attributed at least partly to the dog no longer enjoying the work (Kogan et al., 2026), and in guide dogs, withdrawals for willingness occurred years after qualification (Caron-Lormier et al., 2016b).
11. Limitations and Research Gaps
11.1 No Prospective Validation of Therapy Dog Tests
Among the studies considered here, no prospective study has tested whether a therapy or visiting dog evaluation predicts later welfare, career length or withdrawal. The Japanese questionnaire study included only dogs that had already passed an aptitude test (Sakurama et al., 2023). Mongillo et al. (2015) compared a behavioral examination with a simulated intervention but not with later outcomes in real work. Russo et al. (2025) likewise did not follow assessed dogs into actual interventions. The US survey also reported no data on predictive validity (Serpell et al., 2020).
11.2 Pass Rates and Practice Outside the United States
Failure rates and re-evaluation practices are documented for the United States (Serpell et al., 2020). Among the studies considered here, no comparable data were found for Germany, Austria or Switzerland; the European review describes Austria's legal framework but not pass rates or outcomes (Shajid Pyari et al., 2026).
11.3 Retirement and Career Length
Reasons for retirement come from one self-selected survey (Kogan et al., 2026). Among the studies considered here, there are no data on the typical career length of therapy dogs, the age at retirement or whether a loss of willingness can be detected earlier, and the proposed quality-of-life scale has not been validated (Ng & Fine, 2019).
11.4 Willingness and Enjoyment
Assent and dissent are well described as concepts (Townsend & Gee, 2021), and positive behaviors have been observed during sessions (Corsetti et al., 2019; Haven-Pross et al., 2026). Visser et al. (2026) compared such observations with expert-derived thresholds for positive- and negative-affective behavior. These thresholds have not been independently validated, and none of the studies considered here prospectively linked a measure of willingness or enjoyment with later outcomes.
11.5 Handlers and Transfer from Working Dogs
Handler perception was studied in one small pilot (Clark, Smidt, et al., 2020) and one survey (Mignot et al., 2022); no study considered here tested how handler training affects selection or the timing of withdrawal. The prediction research concerns purpose-bred dogs tested by professional staff, and whether its lessons hold for privately owned adult dogs is untested (Bray et al., 2021).
12. Conclusion
Selecting therapy and visiting dogs is common practice, but the evidence that current evaluations identify the right dogs is thin. In the US organization survey, many organizations reported low failure rates; only some required regular re-evaluation, and none of the therapy dog tests examined here was prospectively linked to later career outcomes (Serpell et al., 2020; Mongillo et al., 2015; Russo et al., 2025). The best data come from guide and assistance dogs, a different population: even with large samples, professional raters and repeated tests, prediction was modest and better at spotting likely failures than likely successes (Bray et al., 2019; Bray et al., 2021).
For trainers and handlers, this means treating selection as a process rather than an event: assessing adult dogs repeatedly and in context, including the handler in the assessment, watching for loss of willingness throughout the dog's career and planning retirement from the start (Shajid Pyari et al., 2026; Kogan et al., 2026). Better research needs comparison groups of dogs that did not pass, prospective follow-up and reported reliability (Brady et al., 2018; Hall et al., 2026).
Key Insights (Takeaways)
In a representative US survey, 71% of regional therapy dog organizations reported failure rates of 10% or less among new applicants, six of 24 reported no unsuccessful registration applicants in 2017, and only 14 of 24 required regular re-evaluation (Serpell et al., 2020).
In a pilot of 38 dogs, caregivers judged 22 suitable, while a structured test found nine; agreement was only fair (kappa 0.285), and the test's predictive validity is unknown (Russo et al., 2025).
A questionnaire study of 110 therapy dogs that had passed an aptitude test found 30 breeds, but without dogs that failed it cannot show which traits predict suitability (Sakurama et al., 2023).
In assistance dogs, a different population, questionnaires and temperament tests predicted training outcome with 64% to 68% accuracy, compared with about 60% from predicting the most common outcome, and identified likely failures better than likely successes (Bray et al., 2019).
Personality consistency was higher in adult dogs (r = 0.51) than in puppies (r = 0.30), and sociability and fearfulness were among the least consistent puppy traits (Fratkin et al., 2013).
In a self-selected survey, handlers named the dog no longer seeming to enjoy the work as a reason for about one in four retired therapy dogs (Kogan et al., 2026).
In UK guide dogs, withdrawals for willingness occurred at a median of about 4.4 years after qualification, showing that selection does not end with the first test (Caron-Lormier et al., 2016b).
References
Arata, S., Momozawa, Y., Takeuchi, Y., & Mori, Y. (2010). Important behavioral traits for predicting guide dog qualification. Journal of Veterinary Medical Science, 72(5), 539–545. https://doi.org/10.1292/jvms.09-0512
Barker, S. B., & Gee, N. R. (2021). Canine-assisted interventions in hospitals: Best practices for maximizing human and canine safety. Frontiers in Veterinary Science, 8, 615730. https://doi.org/10.3389/fvets.2021.615730
Berns, G. S., Brooks, A. M., Spivak, M., & Levy, K. (2017). Functional MRI in awake dogs predicts suitability for assistance work. Scientific Reports, 7(1), 43704. https://doi.org/10.1038/srep43704
Bogaerts, E., Moons, C. P. H., Van Nieuwerburgh, F., Peelman, L., Saunders, J. H., & Broeckx, B. J. G. (2019). Rejections in an non-purpose bred assistance dog population: Reasons, consequences and methods for screening. PLOS ONE, 14(6), e0218339. https://doi.org/10.1371/journal.pone.0218339
Brady, K., Cracknell, N., Zulch, H., & Mills, D. S. (2018). A systematic review of the reliability and validity of behavioural tests used to assess behavioural characteristics important in working dogs. Frontiers in Veterinary Science, 5, 103. https://doi.org/10.3389/fvets.2018.00103
Bray, E. E., Levy, K. M., Kennedy, B. S., Duffy, D. L., Serpell, J. A., & MacLean, E. L. (2019). Predictive models of assistance dog training outcomes using the Canine Behavioral Assessment and Research Questionnaire and a standardized temperament evaluation. Frontiers in Veterinary Science, 6, 49. https://doi.org/10.3389/fvets.2019.00049
Bray, E. E., Otto, C. M., Udell, M. A. R., Hall, N. J., Johnston, A. M., & MacLean, E. L. (2021). Enhancing the selection and performance of working dogs. Frontiers in Veterinary Science, 8, 644431. https://doi.org/10.3389/fvets.2021.644431
Bray, E. E., Sammel, M. D., Cheney, D. L., Serpell, J. A., & Seyfarth, R. M. (2017). Effects of maternal investment, temperament, and cognition on guide dog success. Proceedings of the National Academy of Sciences, 114(34), 9128–9133. https://doi.org/10.1073/pnas.1704303114
Carballo, F., Cavalli, C. M., Gácsi, M., Miklósi, Á., & Kubinyi, E. (2020). Assistance and therapy dogs are better problem solvers than both trained and untrained family dogs. Frontiers in Veterinary Science, 7, 164. https://doi.org/10.3389/fvets.2020.00164
Caron-Lormier, G., Harvey, N. D., England, G. C. W., & Asher, L. (2016a). A new metric for quantifying the relative impact of risk factors on loss of working life illustrated in a population of working dogs. PLOS ONE, 11(11), e0165414. https://doi.org/10.1371/journal.pone.0165414
Caron-Lormier, G., Harvey, N. D., England, G. C. W., & Asher, L. (2016b). Using the incidence and impact of behavioural conditions in guide dogs to investigate patterns in undesirable behaviour in dogs. Scientific Reports, 6(1), 23860. https://doi.org/10.1038/srep23860
Clark, S. D., Martin, F., McGowan, R. T. S., Smidt, J. M., Anderson, R., Wang, L., Turpin, T., Langenfeld-McCoy, N., Bauer, B. A., & Mohabbat, A. B. (2020). Physiological state of therapy dogs during animal-assisted activities in an outpatient setting. Animals, 10(5), 819. https://doi.org/10.3390/ani10050819
Clark, S. D., Smidt, J. M., & Bauer, B. A. (2020). Therapy dogs' and handlers' behavior and salivary cortisol during initial visits in a complex medical institution: A pilot study. Frontiers in Veterinary Science, 7, 564201. https://doi.org/10.3389/fvets.2020.564201
Clay, L., Paterson, M. B. A., Bennett, P., Perry, G., & Phillips, C. J. C. (2020). Do behaviour assessments in a shelter predict the behaviour of dogs post-adoption? Animals, 10(7), 1225. https://doi.org/10.3390/ani10071225
Corsetti, S., Ferrara, M., & Natoli, E. (2019). Evaluating stress in dogs involved in animal-assisted interventions. Animals, 9(10), 833. https://doi.org/10.3390/ani9100833
Discepolo, D., Farr, B., Broach, D., Henderson, A., Gerardo, A., Dyer, J., & Best, E. (2025). United States military working dogs from 2019 to 2021: Analysis of causes of service discharge and decreased service life. Frontiers in Veterinary Science, 12, 1580628. https://doi.org/10.3389/fvets.2025.1580628
Foltin, S., Kostenko, S., Hartwig, A.-D., & Glenk, L. M. (2026). War exposure and canine cortisol responses: Country differences in cortisol profiles of therapy dogs. Animals, 16(3), 381. https://doi.org/10.3390/ani16030381
Fratkin, J. L., Sinn, D. L., Patall, E. A., & Gosling, S. D. (2013). Personality consistency in dogs: A meta-analysis. PLoS ONE, 8(1), e54907. https://doi.org/10.1371/journal.pone.0054907
Gandenberger, J., Flynn, E., Moratto, E., Wendt, A., & Morris, K. N. (2022). Molecular biomarkers of adult human and dog stress during canine-assisted interventions: A systematic scoping review. Animals, 12(5), 651. https://doi.org/10.3390/ani12050651
Glenk, L. M. (2017). Current perspectives on therapy dog welfare in animal-assisted interventions. Animals, 7(2), 7. https://doi.org/10.3390/ani7020007
Glenk, L. M., & Foltin, S. (2021). Therapy dog welfare revisited: A review of the literature. Veterinary Sciences, 8(10), 226. https://doi.org/10.3390/vetsci8100226
Hall, S. S., Shoesmith, E., Riga, E., Mills, D. S., Gibsone, S., McMillan, D., Wu, Q., Clarke, C., & Ratschen, E. (2026). Reporting of dog-assisted intervention trials: Extension of the SPIRIT 2025 and CONSORT 2025 statement. BMC Medical Research Methodology, 26(1), 113. https://doi.org/10.1186/s12874-026-02848-7
Hare, E., Essler, J. L., Otto, C. M., Ebbecke, D., & Serpell, J. A. (2024). Development of a modified C-BARQ for evaluating behavior in working dogs. Frontiers in Veterinary Science, 11, 1371630. https://doi.org/10.3389/fvets.2024.1371630
Harvey, N. D., Craigon, P. J., Blythe, S. A., England, G. C. W., & Asher, L. (2017). An evidence-based decision assistance model for predicting training outcome in juvenile guide dogs. PLOS ONE, 12(6), e0174261. https://doi.org/10.1371/journal.pone.0174261
Haven-Pross, S. C., Jens, A. L., Maarleveld, K. N., van Honk, P., de Kort, M., & Visser, E. K. (2026). Dogs' behavioural responses to dog-assisted interventions: A field study. Animals, 16(7), 1063. https://doi.org/10.3390/ani16071063
Hilby, E. K., Rendahl, A., Russenberger, J., Zimmermann, M., Mickelson, J. R., & McCue, M. E. (2025). Temporal consistency of behavior trait measurements in guide dogs. Frontiers in Veterinary Science, 12, 1549360. https://doi.org/10.3389/fvets.2025.1549360
Hilby, E., & McCue, M. (2026). Decoding the path to success: Unveiling how behavior drives assistance dog career outcome. Frontiers in Veterinary Science, 13, 1777631. https://doi.org/10.3389/fvets.2026.1777631
Hunt, R. L., England, G. C. W., Asher, L., Whiteside, H., & Harvey, N. D. (2020). Concurrent and predictive criterion validity of a puppy behaviour questionnaire for predicting training outcome in juvenile guide dogs. Animals, 10(12), 2382. https://doi.org/10.3390/ani10122382
Jones, M. G., Filia, K., Rice, S. M., & Cotton, S. M. (2024). Guidance on minimum standards for canine-assisted psychotherapy in adolescent mental health: Delphi expert consensus on health, safety, and canine welfare. Animals, 14(5), 705. https://doi.org/10.3390/ani14050705
Kogan, L. R., Currin-McCulloch, J., Packman, W., & Bussolari, C. (2026). Navigating loss in animal-assisted services: Volunteer experiences and implications for programs following therapy dog death or retirement. Animals, 16(2), 202. https://doi.org/10.3390/ani16020202
Lazarowski, L., Rogers, B., Krichbaum, S., Haney, P., Smith, J. G., & Waggoner, P. (2021). Validation of a behavior test for predicting puppies' suitability as detection dogs. Animals, 11(4), 993. https://doi.org/10.3390/ani11040993
Lloyd, J., Budge, C., La Grow, S., & Stafford, K. (2016). An investigation of the complexities of successful and unsuccessful guide dog matching and partnerships. Frontiers in Veterinary Science, 3, 114. https://doi.org/10.3389/fvets.2016.00114
MacLean, E. L., & Hare, B. (2018). Enhanced selection of assistance and explosive detection dogs using cognitive measures. Frontiers in Veterinary Science, 5, 236. https://doi.org/10.3389/fvets.2018.00236
Meers, L. L., Contalbrigo, L., Samuels, W. E., Duarte-Gan, C., Berckmans, D., Laufer, S. J., Stevens, V. A., Walsh, E. A., & Normando, S. (2022). Canine-assisted interventions and the relevance of welfare assessments for human health, and transmission of zoonosis: A literature review. Frontiers in Veterinary Science, 9, 899889. https://doi.org/10.3389/fvets.2022.899889
Mignot, A., de Luca, K., Servais, V., & Leboucher, G. (2022). Handlers' representations on therapy dogs' welfare. Animals, 12(5), 580. https://doi.org/10.3390/ani12050580
Miller, S. L., Serpell, J. A., Dalton, K. R., Waite, K. B., Morris, D. O., Redding, L. E., Dreschel, N. A., & Davis, M. F. (2022). The importance of evaluating positive welfare characteristics and temperament in working therapy dogs. Frontiers in Veterinary Science, 9, 844252. https://doi.org/10.3389/fvets.2022.844252
Mongillo, P., Pitteri, E., Adamelli, S., Bonichini, S., Farina, L., & Marinelli, L. (2015). Validation of a selection protocol of dogs involved in animal-assisted intervention. Journal of Veterinary Behavior, 10(2), 103–110. https://doi.org/10.1016/j.jveb.2014.11.005
Ng, Z. Y., & Fine, A. H. (2019). Considerations for the retirement of therapy animals. Animals, 9(12), 1100. https://doi.org/10.3390/ani9121100
Ng, Z. Y., Pierce, B. J., Otto, C. M., Buechner-Maxwell, V. A., Siracusa, C., & Werre, S. R. (2014). The effect of dog–human interaction on cortisol and behavior in registered animal-assisted activity dogs. Applied Animal Behaviour Science, 159, 69–81. https://doi.org/10.1016/j.applanim.2014.07.009
Piotti, P., Albertini, M., Trabucco, L. P., Ripari, L., Karagiannis, C., Bandi, C., & Pirrone, F. (2021). Personality and cognitive profiles of animal-assisted intervention dogs and pet dogs in an unsolvable task. Animals, 11(7), 2144. https://doi.org/10.3390/ani11072144
Riemer, S., Müller, C., Virányi, Z., Huber, L., & Range, F. (2014). The predictive value of early behavioural assessments in pet dogs – A longitudinal study from neonates to adults. PLoS ONE, 9(7), e101237. https://doi.org/10.1371/journal.pone.0101237
Rodriguez, K. E., Herzog, H., & Gee, N. R. (2021). Variability in human-animal interaction research. Frontiers in Veterinary Science, 7, 619600. https://doi.org/10.3389/fvets.2020.619600
Russo, G., Borrelli, C., Riggio, G., Rosson, E., Bentivoglio, M., & Mariti, C. (2025). A pre-screening tool to assess dog suitability for animal-assisted interventions: Preliminary results for dog-suitability tests (SuiTe). Veterinary Sciences, 12(12), 1110. https://doi.org/10.3390/vetsci12121110
Sakurama, M., Ito, M., Nakanowataru, Y., & Kooriyama, T. (2023). Selection of appropriate dogs to be therapy dogs using the C-BARQ. Animals, 13(5), 834. https://doi.org/10.3390/ani13050834
Serpell, J. A., & Duffy, D. L. (2016). Aspects of juvenile and adolescent environment predict aggression and fear in 12-month-old guide dogs. Frontiers in Veterinary Science, 3, 49. https://doi.org/10.3389/fvets.2016.00049
Serpell, J. A., Kruger, K. A., Freeman, L. M., Griffin, J. A., & Ng, Z. Y. (2020). Current standards and practices within the therapy dog industry: Results of a representative survey of United States therapy dog organizations. Frontiers in Veterinary Science, 7, 35. https://doi.org/10.3389/fvets.2020.00035
Shajid Pyari, M., Guo, K., & Kargas, N. (2026). Preparing therapy dogs for animal-assisted services: Governance, welfare-centred preparation, and a lifecycle research agenda across selected European and United States frameworks. Animals, 16(16), 2557. https://doi.org/10.3390/ani16162557
Shih, H.-Y., Martin, F., Ness, D., Romine, W., Peck, T. L., Turpin, T., Horoschak, R., Steeby, C., Phillips, H., Claypool, M., Theuer, A., Herbeck, G. M., Sexton, J., Pittman, E., Bellamkonda, E., Mohabbat, N. M. L., Lyn, S. A., Bauer, B. A., & Mohabbat, A. B. (2025). Behavioral and physiological responses of therapy dogs to animal-assisted treatment in an inpatient stroke rehabilitation program. Animals, 15(2), 121. https://doi.org/10.3390/ani15020121
Solomon, A., Shamis, E., Racah, H., Toren, A., Weiss, C., Braun, O., & Zamansky, A. (2026). Predicting guide dog career success using machine learning and large language models. Scientific Reports, 16(1), 19877. https://doi.org/10.1038/s41598-026-48238-3
Townsend, L., & Gee, N. R. (2021). Recognizing and mitigating canine stress during animal assisted interventions. Veterinary Sciences, 8(11), 254. https://doi.org/10.3390/vetsci8110254
Visser, E. K., Jens, A. L., van Honk, P., van Asselt, M., & Haven-Pross, S. C. (2026). Establishing behavioural thresholds for dogs in animal-assisted services: Expert-derived thresholds and field study comparison. Animals, 16(7), 1078. https://doi.org/10.3390/ani16071078
Winkle, M., Johnson, A., & Mills, D. (2020). Dog welfare, well-being and behavior: Considerations for selection, evaluation and suitability for animal-assisted therapy. Animals, 10(11), 2188. https://doi.org/10.3390/ani10112188
Zenker, F., Thodberg, K., & Sandøe, P. (2026). Partners or tools? Ethical framings shape welfare research on dogs and horses in animal-assisted services. Frontiers in Veterinary Science, 13, 1960144. https://doi.org/10.3389/fvets.2026.1960144