Intention to treat analyses

Author: Dr Simon Moss

Overview

In many studies, participants do not always adhere completely to the protocol or procedure. To illustrate:

• Researchers might want to examine whether ingesting some drug, ingested every morning, is more effect than a placebo in enhancing concentration. Some of the participants, however, might not follow the instructions and ingest the drug every morning.
• Researchers might want to assess whether daily meditation improves wellbeing. Again, however, some of the participants might not have adhered to the prescribed protocol of daily meditation, but instead meditated only once.
• Researchers might want to assess whether some surgery improves health relative to no surgery. Some of the participants, however, might have died before the surgery began.

Typically, in these instances, researchers will often exclude the responses from individuals who did not adhere to the protocol. They might, for example, discard participants who did not adhere to the prescribed drug regime, did not meditate as often as stipulated, or died before they received the surgery. Nevertheless, this approach can generate several problems. For example, the sample in one condition is biased-sometimes comprising only the individuals who adhere to difficult regimes. This bias might distort the findings.

Intention to treat analyses, sometimes called intent to treat analyses, can be conducted to overcome these limitations (see Fisher, Dixon, Herson, Frankowski, Hearron, & Peace, 1990;; Horvitz-Lennon, O'Malley, Frank, & Normand, 2005;; Lachin, 2000). Specifically, the responses of all participants who were assigned to the various conditions, regardless of whether or not these individuals adhered to the protocol, are subjected to conventional analyses. That is, the researcher undertakes a typical analysis, except all participants who were allocated to a condition or protocol are included.

Problems with conventional analyses

Randomization is not preserved

In most of these studies, participants are randomly allocated to conditions. For example, every second person who agrees to participate might be assigned the drug, and the remaining individuals might be assigned the placebo. Alternatively, and preferably, every person receives a number, generated randomly from a computer. Every person who is assigned an even number is asked to ingest the drug, and the remaining individuals are assigned he placebo. Such randomization ensures that participants in both conditions-the drug and the placebo, for example-are the same on all key variables, such as age, weight, heart rate, and so forth.

If individuals who do not adhere to the protocol are excluded, the randomization is not preserved, and the results might be misleading. For example:

• Perhaps individuals who did not concentrate when the instructions were presented did not ingest the drug each morning. If these individuals are excluded from the analyses, only participants who concentrate well will remain. Hence, individuals who received the drug will tend to perform better on measures of concentration that individuals who did not receive the drug, even if this treatment is actually futile
• Perhaps individuals who feel anxious did not meditate everyday. If these individuals are excluded from the analysis, the participants who did meditate will be relatively calm-regardless of whether meditation is effective.
• Perhaps someone might die before receiving surgery. Researchers will usually decide to exclude these individuals from subsequent analyses. However, perhaps waiting for surgery did increase the likelihood of death.

In some instances, however, this problem diminishes if the variables that differed between the conditions can be measured and controlled statistically. The researcher, for example, could measure concentration or wellbeing before the treatment was administered. They could then conduct an ANCOVA to control baseline measures of concentration or wellbeing.

When the key variables cannot be controlled, however, intention to treat analyses should be undertaken. That is, the researcher would examine whether concentration or wellbeing differs between the conditions, and every participant, even the individuals who did not adhere to instructions, is included.

Rate of discontinuation differs across conditions

A related, but distinct problem arises when some individuals are not available to be measured rather than fail to adhere to the protocol (for a detailed discussion, see Little & Yau, 1996;; Nich & Carroll, 2002). For example:

• Some individuals might ingest the drug each morning, but nevertheless not arrive at the location in which their concentration will be tested later.
• Some individuals might meditate each day, but refuse to complete questionnaires that assess their wellbeing.

Again, when these individuals are excluded, the results can be misleading. To illustrate, perhaps the drug reduces concentration in some individuals but increase concentration in other individuals. Conceivably, when concentration is reduced appreciably, the participants might be so distracted they forget to arrive at the designated location. Hence, only the individuals whose concentration was enhanced by the drug are subsequently tested. The drug will seem effective, but might be destructive.

Several forms of intention to treat analysis can be conducted in these instances, although none of these approaches are infallible. Specifically, researchers can:

• Examine whether the proportion of individual who dropped out-that is, did not complete the final measures-is the same in each condition. If these proportions are not significantly different from one another, and indeed seem almost the same, the analyses might be able to continue as usual
• Attempt to substitute the missing values for some score that seems reasonable. If the concentration of someone after the protocol was completed was not measured, the last assessment of concentration could be considered instead-called the Last Observation Carried Forward. Alternatively, more sophisticated methods, such as regression extrapolation, could be applied. Once these values are estimated, all participants are included in the analysis.

These approaches are not optimal, however. Researchers cannot really be certain what the missing values would have been, and hence the results can be biased.

The results might not generalize to real life

If conventional analyses are conducted, difficulties with compliance or adherence are disregarded. For example, suppose most individuals feel ill every time they ingest some drug and, therefore, do not adhere to the recommended regime. Accordingly, in real life, the drug will be futile because individuals will not ingest this substance. If the analysis is confined only to the individuals who do adhere to the schedule, this problem will be overlooked.

To overcome this problem, intention to treat analysis is conducted. That is, the responses of all participants, regardless of whether they adhered to the protocol, are included in subsequent analyses.

Nevertheless, intention to treat analysis does present some logical complications. In short, intention to treat analysis explores a different issue to conventional analyses. In other words, intention to treat analysis, in some senses, changes the research question. In particular:

• Intention to treat analyses answers a different question to conventional analyses. Conventional analyses examine the question "If everyone adhered to the regime, would this protocol-this drug, exercise, surgery, and so forth-be effective?" Intention to treat analyses examine the issue "Do individuals adhere to this protocol sufficiently to ensure the regime is effective". That is, intention to treat analyses confound the question of efficacy and adherence.
• Sometimes, however, researchers are interested only in the question of efficacy, not whether adherence is likely.
• The reason that adherence is not always a relevant question is that compliance to some protocol partly depends on the perceived efficacy of a treatment. If the efficacy of some drug is established and publicized, adherence might rise, even if the substance is unpleasant.

Protection against biased exclusion of participants

When researchers undertake conventional analyses, they often exclude participants for a variety of reasons: outliers, excessive missing data, suspicion of deliberate distortion, and so forth. The problem, however, is that researchers might, either intentionally or inadvertently, exclude participants to increase the likelihood of a significant result.

Perhaps, one of the individuals who undertook meditation experienced severe back problems. Whether this individual should be excluded or not is a complex issue.

Intention to treat analyses circumvents this problem. Nevertheless, many other approaches, some as seeking the services of someone blind to the hypotheses and assignment of individuals to conditions could decide who to exclude.

Common practices

In practice, many researchers conduct both conventional analyses and intention to treat analyses (e.g., Fredrickson, Cohn, Coffey, Pek, & Finkel, 2008). If both analyses generate the same outcome, the conclusion is straightforward. If only the conventional analyses, but not the intention to treat analyses, generate significant findings, the researcher might need to include a caveat such as "The benefits of this treatment cannot be generalized to individuals who do not adhere to the protocol".

References

Fisher, L. D., Dixon, D. O., Herson, J., Frankowski, R. K., Hearron, M. S. & Peace, K. E. (1990). Intention-to-treat in clinical trials. In K. E. Peace (Ed.), Statistical issues in drug research and development (pp. 331-350.) Marcel Dekker: New York.

Frangakis, C. & Rubin, D. B. (1999). Addressing complications of intent-to-treat analysis in the combined presence of all-or-none treatment-non-compliance and subsequent missing outcomes. Biometrika, 86, 365-379.

Fredrickson, B. L., Cohn, M. A., Coffey, K. A., Pek, J., & Finkel, S. M. (2008). Open hearts build lives: Positive emotions, induced through loving-kindness meditation, build consequential personal resources. Journal of Personality and Social Psychology, 95, 1045-1062.

Horvitz-Lennon, M., O'Malley, A. J., Frank, R. G., & Normand, S. T. (2005). Improving traditional intention-to-treat analyses: A new approach. Psychological Medicine, 35, 961-970.

Kleinman, K. P., Ibrahim, J. G., & Laird, N. M. (1998). A Bayesian framework for intent-to-treat analyses with missing data. Biometrics, 54, 265-278.

Lachin, J. M. (2000). Statistical considerations in the intention-to-treat principle.  , 21, 167-189.

Little, R. & Yau, L. (1996). Intent-to-treat analysis for longitudinal studies with drop-outs. Biometrics, 52, 1324-1333.

Mazumdar, S., Liu, K. S., Houck, P. R. & Reynolds, C. F. (1999). Intent-to-treat analysis for longitudinal clinical trials: coping with the challenge of missing values. Journal of Psychiatric Research, 33, 87-95.

Nich, C., & Carroll, K. (2002). Intention to treat meets missing data: Implications of alternate strategies for analyzing clinical trials data. Drug and Alcohol Dependence, 68, 121-130.