Treatment Assignment and Causal Questions

From a Research Question to a Randomized Treatment Question

A question such as “Does a reminder improve appointment attendance?” is causal: it asks what would change if people were assigned one condition rather than another. A randomized study can address this question by assigning eligible people to a reminder condition or a comparison condition and then measuring attendance. To make the question precise, specify the conditions being assigned, the people the question concerns, and the period in which the outcome is measured.

The key is to define the effect of assignment. Assignment is the condition a study allocates to a participant. Treatment received is what the participant actually does or experiences. These can differ: someone assigned a reminder may not read it, while someone assigned no reminder may receive one through another route. A causal question about assignment remains well-defined even when compliance is imperfect.

Definition: A randomized treatment study assigns participants to specified conditions using a chance mechanism. For participant $i$, let $Z_i=1$ indicate assignment to the treatment condition and $Z_i=0$ assignment to the comparison condition. Let $R_i$ describe the treatment actually received; $R_i$ need not equal $Z_i$. A causal effect of assignment compares the outcome that would occur under assignment to one condition with the outcome that would occur under assignment to the other.

For example, “the effect of offering a text reminder” is about assignment to an offer, not necessarily about the effect of reading the text. The distinction matters because randomization applies to $Z_i$, the assigned condition. It does not automatically randomize $R_i$, what each participant receives or uses.

Define the Conditions, Population, and Outcome Period

A useful causal question names both assignment conditions in enough detail that they are distinguishable. “Treatment versus control” is often too vague by itself. State what the treatment group is assigned, what the comparison group is assigned, and any relevant duration or delivery details. Then identify the target population and when the outcome is assessed.

An outcome period is part of the question, not a minor reporting detail. The effect of an offer on attendance within 30 days is not automatically the effect on attendance within a year. Likewise, measuring a score immediately after a course answers a different question from measuring it six months later.

Definition: For a specified population of $N$ units, let $Y_i(1)$ be the outcome for unit $i$ if assigned to condition 1, and let $Y_i(0)$ be the outcome for that same unit if assigned to condition 0. The unit-level causal effect is $Y_i(1)-Y_i(0)$. The average treatment effect in this population is $$ \tau=\frac{1}{N}\sum_{i=1}^{N}\left[Y_i(1)-Y_i(0)\right]. $$ The outcomes and effect refer to the outcome definition and assessment period specified in the question.

The two potential outcomes describe alternatives for the same unit, but only one is observed: a participant cannot simultaneously be assigned to both conditions in the same study. Random assignment makes the assigned groups comparable on average, enabling their outcome means to estimate the average effect of assignment. It does not reveal both potential outcomes for each person.

Worked Example: Specifying a Reminder Study

Suppose a clinic plans a hypothetical randomized study among 300 adults scheduled for a routine appointment. Participants are assigned either an offer of one text reminder 48 hours before the appointment or no study text reminder. The outcome is whether the participant attends the scheduled appointment within 30 days.

Assignment conditions. Condition 1 is assignment to the offer of a text reminder 48 hours before the appointment. Condition 0 is assignment to no study text reminder. The question is about the effect of these assigned conditions, not the effect of actually reading a reminder.

Population and outcome period. The directly defined study population is the 300 adults enrolled in this hypothetical study; the outcome period is the 30 days after the scheduled appointment. A broader target, such as all adults attending the clinic, would need to be stated separately and supported by how participants were recruited.

Causal question. “Among the enrolled adults scheduled for a routine appointment, what is the average effect of assignment to an offer of a text reminder 48 hours beforehand, compared with assignment to no study text reminder, on attendance within 30 days?”

If some people assigned a reminder do not read it, that does not change the assignment question. It does mean the answer concerns the effect of being assigned the reminder offer under the study’s actual pattern of uptake, rather than the effect of reading it.

Why a Randomized Comparison Estimates an Assignment Effect

Before interpreting a difference between groups causally, check that the study’s procedures match the question. The assignment must actually be made by a chance mechanism, and the conditions must be specified. Outcomes should be defined and measured for the same period in each group. The interpretation also relies on participants’ outcomes corresponding to their assigned conditions and on one participant’s assignment not changing another participant’s outcome. These last conditions are often called consistency and no interference.

In this setup, the observed outcome is $Y_i=Y_i(Z_i)$: it is the potential outcome under the condition to which participant $i$ was assigned. If the assignment is randomized, then $Z_i$ does not systematically favor participants with larger or smaller potential outcomes. Thus, across the random assignment mechanism, the mean outcome in each assigned group estimates the mean potential outcome under that assignment.

Justification: Suppose $n_1$ of the $N$ units are assigned to condition 1 by simple random assignment, and $n_0=N-n_1$ to condition 0. Let $\bar{Y}_{Z=1}$ and $\bar{Y}_{Z=0}$ be the observed outcome means in those groups. Each unit has probability $n_1/N$ of assignment to condition 1, so the expected treated-group mean over the random assignments is $$ E(\bar{Y}_{Z=1})= \frac{1}{n_1}\sum_{i=1}^{N}\frac{n_1}{N}Y_i(1) =\frac{1}{N}\sum_{i=1}^{N}Y_i(1). $$ Similarly, $E(\bar{Y}_{Z=0})=\frac{1}{N}\sum_{i=1}^{N}Y_i(0)$. Therefore, $$ E(\bar{Y}_{Z=1}-\bar{Y}_{Z=0})= \frac{1}{N}\sum_{i=1}^{N}[Y_i(1)-Y_i(0)] =\tau. $$ So the difference in assigned-group means is an unbiased estimate of the finite-population average effect of assignment under this randomization scheme. The equality is about the average over possible assignments; in one particular study, the observed difference can still vary by chance.

This justification concerns the effect of assignment among the units randomized, under the stated conditions. Random assignment is not the same as random sampling. Random assignment supports a causal comparison within the study; it does not by itself make the participants representative of a broader population. Extending a result beyond the randomized participants requires a separate argument about the target population and recruitment.

Worked Example: Calculating an Assignment-Group Difference

In a hypothetical reminder study, 40 enrolled patients are randomly assigned, 20 to an offer of a text reminder and 20 to no study reminder. Attendance within 30 days is recorded for everyone. Seventeen of the 20 assigned a reminder attend; 13 of the 20 assigned no study reminder attend.

The outcome is binary, coded 1 for attended and 0 for did not attend. Thus, each group’s mean outcome is its attendance proportion: $$ \bar{Y}_{Z=1}=\frac{17}{20}=0.85,\qquad \bar{Y}_{Z=0}=\frac{13}{20}=0.65. $$ The estimated effect of assignment is $$ \hat{\tau}=\bar{Y}_{Z=1}-\bar{Y}_{Z=0}=0.85-0.65=0.20. $$

In this hypothetical sample, attendance was 20 percentage points higher in the group assigned the reminder offer than in the group assigned no study reminder. This is a difference between assigned groups, not a claim that the reminder caused exactly 20 additional attendances per 100 people in every population. Because assignment was randomized, the difference estimates the average effect of assignment for the enrolled study population, subject to the design and measurement conditions. The calculation alone does not establish how precise the estimate is.

Suppose 18 people assigned the offer actually read the text. The calculation still uses all 20 people assigned to that group, including the two who did not read it. Restricting the comparison to readers would no longer be the randomized assigned-group comparison.

Assignment Is Not the Same as Receipt

A study may record both $Z_i$, assignment, and $R_i$, treatment received. Keeping them separate helps match the analysis to the question. If the research question concerns the effect of assignment to an offer, the relevant comparison is by assigned group, regardless of whether each participant follows the assigned condition. This is often called an intention-to-treat comparison.

The distinction is especially important when participants do not comply, receive treatment outside the study, or experience only part of the assigned intervention. The difference in outcomes by assignment captures the effect of implementing the assignment policy as it occurred in the study. It is not automatically the effect of receiving the treatment. A direct comparison of people who received treatment with people who did not may lose the protection of random assignment, because receipt itself was not necessarily allocated at random.

Worked Example: A Workshop Offer and Attendance

A hypothetical employer randomly assigns 50 employees to be offered a voluntary time-management workshop and 50 to receive the usual staff information. The outcome is the number of self-reported late arrivals during the following 8 weeks. In the offer group, 35 employees attend the workshop; 15 do not. The mean number of late arrivals is 1.6 for all 50 employees assigned the offer, and 2.1 for all 50 assigned usual information.

The estimated assignment effect, measured as offer group minus comparison group, is $$ \hat{\tau}=1.6-2.1=-0.5 $$ late arrivals per employee over 8 weeks.

The negative value means that, in this hypothetical study, the group assigned an offer averaged 0.5 fewer late arrivals per employee than the group assigned usual information during the specified period. This answers a question about the workshop offer as implemented, including the fact that 15 assigned employees did not attend.

It would not be valid to replace the offer-group mean of 1.6 with the mean among only the 35 attendees and call the resulting comparison randomized. Attendance was voluntary, while assignment to the offer was randomized. The causal question and the comparison must stay aligned: assignment-based question, assignment-based groups.

Common Mistakes and a Practical Checklist

Calling the assigned condition the treatment received. Assignment to an offer does not guarantee uptake. Name both assignment and receipt when both matter.
Leaving the population vague. “Does it work?” is incomplete. State whether the claim concerns enrolled participants, a defined eligible group, or another target population.
Omitting the outcome period. A causal effect on an outcome at 8 weeks need not be the effect at 6 months. Include the assessment window in the question.
Assuming random assignment means random sampling. Assignment supports causal interpretation of a randomized comparison; it does not alone justify generalizing to everyone beyond those studied.
Comparing recipients instead of assigned groups without justification. When receipt is voluntary or incomplete, a received-treatment comparison is not protected by the original randomization in the same way.

Before describing a randomized study as causal, write down the assigned conditions, target population, outcome, and outcome period. Then verify that the allocation was genuinely randomized and that outcomes correspond to the defined conditions and are measured comparably. State the conclusion as an effect of assignment unless the design and analysis justify a different treatment-receipt question.

Check Your Understanding

For each question, keep the assignment, population, and outcome period explicit.

A study randomly assigns patients to an offer of a dietitian consultation or usual care. Some patients assigned the offer do not attend. Which groups should be compared to estimate the effect of the offer?
Write a causal question for a hypothetical randomized study of a study-skills program, naming the assigned conditions, population, outcome, and assessment period.
In a randomized study, the outcome mean is 12 points in the assigned program group and 15 points in the comparison group. What is the estimated assignment effect if it is defined as program minus comparison, and what does its sign mean?
Why does random assignment alone not guarantee that a result generalizes to every person in the country?
Explain why comparing only participants who actually received a treatment may not preserve the original randomized comparison.