Response Variables And Realized Outcomes

Why the Distinction Matters

A study records a value for each response: perhaps whether a patient improved, how many adverse events occurred, or a measured blood pressure. Before data are collected, however, that value is uncertain. A probability model describes this uncertainty; the data set contains the value that was actually recorded. Keeping these two stages distinct is essential for reading a model correctly.

The notation makes the distinction visible. An uppercase letter such as \(Y\) denotes a response random variable. A lowercase letter such as \(y\) denotes a particular value. When the response is observed, write \(y_{\mathrm{obs}}\) for its realized value. The random variable represents what could be recorded under the study process; the realized value is what was recorded on this occasion.

Definition: A response random variable \(Y\) is a measurable function from the outcomes of a random experiment to response values. Its possible outcomes are the values that the model allows \(Y\) to take. If the experiment produces an actual outcome \(\omega_{\mathrm{obs}}\), the realized response is \(y_{\mathrm{obs}}=Y(\omega_{\mathrm{obs}})\). Before observation, \(Y\) is random; after observation, \(y_{\mathrm{obs}}\) is a recorded value.

More formally, a probability model has a sample space \(\Omega\), a collection of events \(\mathcal{F}\), and a probability measure \(P\). A real-valued response random variable is a measurable function \(Y:\Omega\to\mathbb{R}\): for every Borel set \(A\subseteq\mathbb{R}\), the set \(\{\omega:Y(\omega)\in A\}\) is an event in \(\mathcal{F}\). Measurability ensures that probabilities about the response can be assigned. For the examples below, it is enough to keep in mind that each experimental outcome leads to a response value.

The sample-space outcome \(\omega\) and the response value \(Y(\omega)\) need not be the same thing. The outcome \(\omega\) may encode all the details of a trial; the response is the particular feature of that trial that the analysis records. For example, \(\omega\) could describe everything that happened to a patient during follow-up, while \(Y(\omega)\) records only whether a specified clinical event occurred.

Possible Values and Observed Values

A response can have different possible values in the model, but a single observed unit contributes one realized value. In a data table, this distinction appears across rows: \(Y_i\) denotes the response random variable for unit \(i\), while \(y_i\) denotes the value recorded for that unit. Once the study data are in hand, the observed responses \(y_1,\ldots,y_n\) are fixed entries in the data set. The variables \(Y_1,\ldots,Y_n\) describe the responses before their values are observed.

A probability statement refers to the random variable. For instance, \(P(Y=1)=0.08\) says that, under the specified model, the response has probability \(0.08\) of taking value 1. It does not say that the observed response must be 1, nor that a particular patient has already experienced the event. After observing \(y_{\mathrm{obs}}=0\), the response recorded for this trial is known to be 0; the model statement about possible trials remains a different kind of statement.

Worked Example: A Binary Clinical Response

Suppose \(Y\) records whether a patient experiences a specified event within 30 days: \(Y=1\) if the event occurs and \(Y=0\) otherwise. For a simple model, suppose

P(Y=1)=0.08,\qquad P(Y=0)=0.92.

These are probabilities for the response before its value is observed. They account for all the response values in this model because \(0.08+0.92=1\). If one patient's record shows no event, then the realized response is \(y_{\mathrm{obs}}=0\). The response variable is still \(Y\); the recorded value is 0. The value 0 is not itself a probability, and the probability statement \(P(Y=1)=0.08\) is not a report that this patient had the event.

If the event did occur for another patient, that patient's realized response would be \(y_{\mathrm{obs}}=1\). The same response definition and model can therefore be used for different patients while the realized values differ.

Worked Example: A Count Response

Suppose a monitoring protocol allows at most two episodes to be recorded during a defined observation window. Let \(Y\) be the number recorded for a participant. A model for this example assigns

P(Y=0)=0.20,\qquad P(Y=1)=0.50,\qquad P(Y=2)=0.30.

The possible response values in this stated model are 0, 1, and 2. The assigned probabilities sum to \(0.20+0.50+0.30=1.00\), as required for these exhaustive possibilities. If the participant's record contains one episode, the realized value is \(y_{\mathrm{obs}}=1\). It is not the random variable \(Y\), and it does not change what \(Y\) means: \(Y\) still denotes the count that the protocol would record for a participant before that count is known.

The number 1 can appear in two different roles in this example. In \(Y=1\), it names one possible response value; in \(y_{\mathrm{obs}}=1\), it is the actual count observed for a participant. The surrounding notation tells you which role it plays.

Worked Example: A Measurement Response

Suppose a response is a measurement of a biomarker, in specified units, and the model for an individual measurement is \(Y\sim N(10,4)\). Under the course notation, the variance is 4, so the standard deviation is \(\sqrt{4}=2\). A laboratory records \(y_{\mathrm{obs}}=11.3\).

Here \(Y\) is the measurement that is uncertain before it is taken, and 11.3 is the realized measurement. Its difference from the model mean is \(11.3-10=1.3\) units, or \(1.3/2=0.65\) standard deviations. These calculations describe the recorded value relative to the model's stated center and spread; they do not turn 11.3 into a random variable. The notation \(Y\sim N(10,4)\) describes the model for possible measurements, whereas \(y_{\mathrm{obs}}=11.3\) reports this individual's observed result.

A measurement is not required to equal a convenient round number for it to be a realized response. The observed value can be recorded to the instrument's precision, while the model describes uncertainty in the response before it is observed.

Two Useful Consequences

The distinction can be made precise by connecting response values to events and by tracking what happens when a response is expressed on a different scale. These facts are elementary, but they prevent common notation errors in later model calculations.

Theorem: Let \(Y\) be a measurable response random variable and let \(A\) be a Borel set of response values. The event that the response lies in \(A\) is \(\{\omega:Y(\omega)\in A\}\), and its probability is \(P(Y\in A)\). If the observed response is \(y_{\mathrm{obs}}\), the event occurred exactly when \(y_{\mathrm{obs}}\in A\). This observed membership does not by itself imply \(P(Y\in A)=1\).

Proof. By definition, a response value lies in \(A\) precisely when it is the image under \(Y\) of an experimental outcome in the set

\{\omega\in\Omega:Y(\omega)\in A\}=Y^{-1}(A).

Measurability makes this preimage an event, so its probability is defined and is denoted \(P(Y\in A)\). For the actual outcome \(\omega_{\mathrm{obs}}\), the response is \(y_{\mathrm{obs}}=Y(\omega_{\mathrm{obs}})\). Thus \(\omega_{\mathrm{obs}}\) belongs to the event exactly when \(Y(\omega_{\mathrm{obs}})\in A\), which is exactly when \(y_{\mathrm{obs}}\in A\). This tells us whether the event occurred on the observed trial, not that every possible trial lies in the event. Therefore it does not establish that its probability is 1. \(\square\)

For example, after observing \(y_{\mathrm{obs}}=1\), the event \(\{Y=1\}\) occurred on that trial. It does not follow that \(P(Y=1)=1\): the binary example assigned that event probability \(0.08\). A realized event and a probability statement answer different questions.

Result: If \(Y\) is a response random variable and \(g\) is a measurable function, the transformed response \(Z=g(Y)\) is also a random variable. When \(Y\) is observed as \(y_{\mathrm{obs}}\), the transformed response is observed as \(z_{\mathrm{obs}}=g(y_{\mathrm{obs}})\).

Proof. For any Borel set \(B\) in the range of \(Z\), the event that \(Z\) lies in \(B\) can be written as

\{\omega:g(Y(\omega))\in B\} = \{\omega:Y(\omega)\in g^{-1}(B)\}.

Because \(g\) is measurable, \(g^{-1}(B)\) is a Borel set; because \(Y\) is measurable, the event on the right belongs to \(\mathcal{F}\). Hence \(Z=g(Y)\) is measurable and is a random variable. At the observed outcome, \(z_{\mathrm{obs}}=Z(\omega_{\mathrm{obs}})=g(Y(\omega_{\mathrm{obs}}))=g(y_{\mathrm{obs}})\), as claimed. \(\square\)

This result formalizes a simple practical rule: changing units or recoding a response changes its values according to a specified transformation. It does not convert a previously observed value back into an uncertain quantity.

Worked Example: Changing Measurement Units

Suppose \(Y\) is a temperature measurement in degrees Celsius and the recorded value is \(y_{\mathrm{obs}}=20\). To express the same response in degrees Fahrenheit, use \(g(y)=\frac{9}{5}y+32\). The transformed random variable is \(Z=g(Y)\), and the observed value is

z_{\mathrm{obs}}=g(20)=\frac{9}{5}(20)+32=36+32=68.

Thus \(Z\) is the response expressed in Fahrenheit before observation, while 68 is the realized Fahrenheit value for this measurement. The value 20 remains the observed Celsius value. The transformation changes the scale and the numerical representation, not the distinction between a response variable and its realization.

Common Mistakes

Calling the observed value the random variable. Writing “\(Y=11.3\) is the data” blurs the distinction. Write \(Y\) for the response variable and \(y_{\mathrm{obs}}=11.3\) for the recorded measurement.
Treating a probability as a recorded result. \(P(Y=1)=0.08\) describes the model across possible trials. It does not say that the observed response is 0.08, or determine the outcome of a particular trial.
Confusing a study outcome with the response value. An experimental outcome \(\omega\) may include much more information than the response \(Y(\omega)\). The response is the feature the study defines and records.
Assuming an observed event must be certain under the model. Observing \(y_{\mathrm{obs}}\in A\) shows that the event occurred on this occasion. It does not imply that \(P(Y\in A)=1\).

Key takeaway: Use \(Y\) for the response before its value is known, name the model's possible response values separately, and use \(y_{\mathrm{obs}}\) for the value actually recorded. Probability statements concern the random variable; a data table contains realizations.

Check Your Understanding

For each question, keep separate the response variable, its possible values, and the value actually observed.

A study records whether a participant completed follow-up, coded 1 for yes and 0 for no. What does \(Y\) represent, and what does \(y_{\mathrm{obs}}=0\) report?
A count response has possible values 0, 1, and 2 under a stated model. Explain the difference between \(Y=2\) as a possible response and \(y_{\mathrm{obs}}=2\).
If \(P(Y=1)=0.15\) and the observed response is 1, does this imply \(P(Y=1)=1\)? Explain.
A response is measured in kilograms and then converted to pounds by a specified function \(g\). What are the random variable and observed value on the converted scale?
In the formal definition, how can an experimental outcome \(\omega\) contain more information than the response value \(Y(\omega)\)? Give a study example.