Animal-Assisted Interventions and Research Design
Puppy therapy studies are harder to run cleanly than most people realize. The reason comes down to how you define your experimental unit, which is the level at which you randomly assign interventions and the level at which outcomes are actually independent. Get this wrong and your p-values are meaningless. In most puppy-assisted therapy research, the experimental unit is the individual participant — the person receiving the intervention. You randomize people to either a puppy contact group or a control group, then measure outcomes like anxiety scores, blood pressure, or depression inventories. That part is straightforward. The complication starts when interventions happen in groups. If you run a puppy visitation program inside a single nursing home wing where ten residents participate together in the same session, those ten people are not independent. Their outcomes correlate because they share the same dog, the same environment, the same therapist, and the same time slot. Treating each person as an independent experimental unit in that scenario inflates your sample size artificially. You might have twenty participants on paper but effectively only two experimental units — the two sessions or the two rooms they experienced.
I ran into this exact problem a few years back while reviewing a study that claimed a 0.03 p-value for reduced agitation in dementia patients exposed to therapy puppies. The intervention was delivered in six-week residential care units. Each unit received the puppy visits once a week. There were twelve units total, but the analysis treated every patient as an independent data point. The real experimental unit was the unit, not the person. That study's effective N was twelve, not the two hundred-plus patients listed in the results. We had to reanalyze the data using a mixed-effects model with unit-level random effects, and the effect dropped to non-significance. The puppy visits probably still help patients. The study just couldn't prove it with that design.
When the Dog Is the Experimental Unit
Sometimes the experimental unit is actually the dog itself, which catches people off guard. If you're comparing two different training protocols for therapy dogs, or testing whether one dog rotated through multiple wards produces different outcomes than a different dog on each ward, the dog is the unit of assignment. A single therapy dog might visit five different rooms over a month, but those five visits are pseudoreplication if you analyze them as five independent data points. The dog is the thing you randomized, not the room. This matters because dogs vary enormously in temperament, energy level, and patient preference. Using one dog as its own control across multiple conditions can work, but you need enough dogs to absorb that individual variation. Three dogs visiting eight rooms each is not the same design as eight dogs visiting three rooms each. The first gives you more room-level data points but leaves you with almost no ability to generalize beyond those three animals. The second lets you say something about dogs in general but gives you thinner data per animal.
Get the Full Details
Cluster Randomization Complications
The cleanest design for puppy therapy research is cluster randomization, where you randomize entire sites — a hospital ward, a school classroom, a senior center — to either receive the intervention or serve as a control. The cluster is your experimental unit. This avoids contamination, which happens constantly in this field. A control group patient who watches the treatment group interact with a puppy every Tuesday is no longer a true control. They're exposed to the intervention through osmosis. Cluster designs require more clusters than you might think. With individual randomization, fifteen participants per group can give you decent power. With cluster randomization, you often need ten to twenty clusters per arm, depending on the intraclass correlation coefficient. That ICC — the degree to which people within the same cluster resemble each other — is usually around 0.05 to 0.15 in these kinds of studies. Higher ICC means you need substantially more clusters to detect the same effect size. Most puppy therapy papers report ICC values nowhere near low enough to justify their sample sizes.
Practical Tips That Actually Matter
If you're designing a study, stop thinking about how many people you can recruit and start thinking about how many independent assignments you can make. Every constraint that ties multiple people together — same room, same dog, same time slot, same therapist — collapses independent data points into a single experimental unit. Use hierarchical linear modeling or generalized estimating equations rather than t-tests and ANOVAs when your data has this kind of nested structure. Standard statistical tests assume independence, and violations of that assumption are the single most common error I see in this literature. It's not complicated to fix once you know what to look for. A couple of lines of R code with the lme4 package will give you the correct standard errors, whereas a standard t-test on the same data will hand you a falsely narrow confidence interval and a headline-grabbing but unreliable result. The bottom line is that puppy therapy research is real and useful, but the field has a structural problem with experimental unit confusion. It shows up in nearly every paper I read that claims strong evidence from a small number of sites. The intervention might work. The statistics usually don't support the strength of the claim being made.
