In order to address this question we will consider an example. We want to evaluate the effectiveness of adding adjunctive immunosuppressive medication (intervention) to prednisolone (the conventional treatment) in preventing death (outcome) in patients with MUO (Figure 1). In our example we manage to recruit 200 dogs to our RCT and these patients are split equally into two groups; group A will receive the intervention (prednisolone + adjunct) and group B is our active control (prednisolone only). We will evaluate outcome after 3 months.

It is important to realize that the nature of our study design is that we have a short period of 72 h between randomization (i.e. point of diagnosis) and intervention. At the 3-month follow-up, 30 dogs in group A have died (the study's primary outcome measure). Fifteen of these dogs died within 72 h of enrolment whereas the 15 other dogs died 72 h−3 months later. A similar outcome is seen with group B (Figure 1).

Now consider that adding this immunosuppressive medication (intervention) does not prevent our primary outcome (death). We can consider this as the “truth.” Our work should therefore be to perform randomized trials to determine this “truth.” The way we choose to analyse the data will determine if we can arrive at an unbiased conclusion (finding the truth). Using the right analysis, we should determine that the adjunctive immunosuppression is unsuccessful, whereas in the wrong analysis, we will come to a false and biased assumption that additional immunosuppression improves mortality.

The first step in our analysis is to consider which dog received the protocol-allotted intervention. This is a method known as per-protocol analysis (5). This may seem an odd thing to do as all dogs in the intervention group could be assumed to have received the intervention. The reality, however, is that only 85 dogs in group A received the intervention; 15 dogs died before it could be administered. Based on this method of analysis, the mortality rate is 0.18 or 18% (15/85) whereas the death risk for our control group is 0.3 or 30% (30/100).

The relative risk of death refers to the ratio of the intervention group's death risk to the control group's death risk. In this case, the risk ratio is 0.59 (0.18 divided by 0.3). To determine the relative risk reduction of death we must subtract this value from 1 to give 0.41 or 41%. In other words, data analysis using a per-protocol method may mean that a veterinarian would conclude that the adjunctive immunosuppression reduces a dog's death risk by 41%.

We know, however, that the adjunct has no effect on reducing mortality. Analyzing the intervention in this manner would be to grossly misinterpret and inaccurately misrepresent the results. Clinical practice based on such an inaccurate interpretation would result in patients suffering from receiving an additional immunosuppressive medication no benefit.

There is another method of analyzing this data that considers the dogs based on the treatment they had administered rather than the group they were originally designated. This method, called as-treated analysis (3), also suffers from the introduction of selection bias and thus removes any advantage gained by randomization.

Using this strategy, dogs receiving the adjunct within the control sample (whatever the reason), are considered to be participants of the intervention arm, and those allocated within the adjunctive treatment group that do not receive the adjunct are considered to be participants of the control arm. When assessing an RCT's findings, the use of as-treated and per-protocol methods increases the possibility of bias. To avoid this trap, there is a strategy to analyse data from an RCT that will not result in this kind of erroneous conclusion. Intention-to-treat (ITT) analysis is the name of this technique.

Regardless of all the mishaps that can happen during a study, ITT means that one should analyse all dogs allocated randomly to one of the treatments collectively as representing that treatment, whether they received only some or even none of the treatment. In doing this it ensures that randomization remains intact and confounding variables are reduced preserving the prognostic balance. Therefore, using the ITT principle in our study requires uniformly complete follow-up of all dogs (as far as possible) and assigns credit for all positive or negative events and outcomes to the randomized treatment group, regardless of the changes in treatment actually received by those dogs.

Looking back at our study, a dog in the intervention group faces a 30% chance of dying. This calculation includes the 15 dogs that died (the study's primary outcome) before the adjunctive medication was administered. Control dogs have a 30% (0.3 or 30/100) mortality. Compared to the control group, dogs receiving the intervention have a relative risk of death of 1 (0.3/0.3) with the relative risk reduction being zero.

Accordingly, the ITT principle is correct to conclude that the immunosuppressive adjunct is ineffective. It may be contested that it is unreasonable to include the 15 dogs that died before being administered the immunosuppressive agent. However, this is an entirely reasonable practice. It disturbs the prognostic balance achieved by randomization to remove dogs from one arm or the other. A RCT that excludes dogs is more likely to be biased (6).

Studies investigating treatment protocols in MUO often fail at this first hurdle. It is well-recognized that MUO carries a high risk of death within the first 72 h (7–9). Patients randomized after this time frame will have a better prognosis. The period between randomization and group treatment allocation should also always be included in analysis. It is all too frequent that patients with MUO die in these initial hours before treatment is administered and researchers exclude them from analysis; this is a particular problem when retrospective studies gather information on outcome (10, 11). Timing of accession and randomization in MUO studies will directly affect the perceived efficacy of the treatment in question. Those patients lost before randomization will not be prognostically identical to those who remain. The omission of patients that die before the adjunctive treatment is given would cause fewer deaths to be counted and consequently the as-treated analysis would be biased in favor of adjunctive immunosuppression. An example of this is shown in a retrospective study evaluating procarbazine as adjunctive therapy for MUO (12). The median reported survival time in 21 dogs was 14 months according to an as-treated analysis. However, this study is to be commended as it reports that 11 dogs failed to survive to receive treatment giving the study a 34% mortality rate (11/32). Including these deaths in an ITT analysis would significantly decrease the headline median survival time. Failure to acknowledge these deaths sets a dangerous precedent. Studies reporting as-treated analysis for MUO are often conspicuous by their relatively longer median survival times than those published in prospective reports. Once randomization has been performed, dogs (or specifically owners) that fail to adhere to the group's protocol can vary in many regards, not just in the failure to receive the allotted treatment. It is well-known that adherent owners tend to own dogs that have a better outcome than owners of dogs that are non-adherent. This phenomenon is coined the “healthy adherer” effect (13). An ITT analysis maintains the prognostic balance achieved from randomization, thereby minimizing the risk of bias being introduced by unwittingly creating and comparing groups with potentially confounding variables.

The application of ITT allows us to gain an unprejudiced evaluation of the efficacy of a treatment on preventing the study outcome (death). When an intervention decreases the chance of death but adherence to the study protocol is reduced, an ITT analysis underestimates the benefit that medication will give to those dogs that adhere to the treatment schedule. As with a per-protocol or as-treated analysis, ITT yields a more accurate and unbiased efficacy although the magnitude of its effect may be underestimated.