Angeramo CA. Regresión Logísꢀca: lo que autores y revisores no pueden ignorar. Rev Argent Cir. 2025;117(2):e-cymcaa
3
conꢁanza (IC), generalmente del 95%; si este intervalo Conclusión
incluye el valor 1, no se considera que la asociación re-
1
3
presente un riesgo signiꢁcaꢀvo . En nuestro ejemplo,
La regresión logísꢀca es una herramienta esen-
el consumo de alcohol no se asoció signiꢁcaꢀvamente cial en la invesꢀgación médica, que permite no solo
con el evento estudiado (OR = 1,01; IC 95%: 0,83.1,23; idenꢀꢁcar factores de riesgo o de protección, sino tam-
p = 0,90), dado que el intervalo de conꢁanza incluye el bién esꢀmar posibilidades o chances: es hacer predic-
valor nulo (OR = 1). En la tabla 1 se informan los OR, IC ciones úꢀles para la prácꢀca clínica. Sin embargo, su co-
9
5% y valor p de nuestro ejemplo. rrecta aplicación requiere una cuidadosa selección de
Los supuestos (assumpꢀons) y la evaluación variables y una interpretación críꢀca de los resultados,
del modelo de regresión logísꢀca también exceden el siempre considerando que, aunque úꢀl para idenꢀꢁcar
límite de este arꢂculo. asociaciones, no implica causalidad por sí misma.
■
ENGLISH VERSION
“
Facts do not cease to exist because they are
ignored”.
Aldous Huxley, Brave New World.
Introducꢀon
The objecꢀve of logisꢀc regression is to analyze
the effect of one or more predictor (independent)
3
When analyzing data, it is common pracꢀce variables on the dependent variable , allowing for:
to determine whether there is a relaꢀonship between 1) Idenꢀꢁcaꢀon of risk factors or protecꢀve factors.
two variables. The uꢀlizaꢀon of staꢀsꢀcal tools such as
correlaꢀon and regression is opꢀmal for achieving this
objecꢀve. Despite their similariꢀes, these tools serve
disꢀnct purposes.
For instance, to determine whether a clinical
characterisꢀc (e.g., age or smoking habits) increases
or decreases the likelihood or probability of
presenꢀng a disease.
Correlaꢀon measures the strength and 2) Esꢀmaꢀng the likelihood or probability of an event
direcꢀon of the relaꢀonship between two variables. It
is a symmetrical measure, which means the order of
the variables does not affect the result. The correlaꢀon
between A and B is the same as the correlaꢀon between
occurring. such as the likelihood of developing a
disease based on comorbidiꢀes. These possibiliꢀes
can be represented visually in nomograms or
implemented in on-line calculators to clearly and
5
B and A. However, correlaꢀon does not imply causaꢀon,
accessibly communicate the risk of an event .
1
nor does it allow for the making of predicꢀons . On the 3) Making predicꢀons: the primary objecꢀve is to
other hand, regression also analyzes the relaꢀonship
between variables, but not symmetrically. In regression
analysis, it is essenꢀal to disꢀnguish between a
dependent variable, which is to be predicted or
predict the most likely category of the dependent
variable for a given paꢀent (e.g., classifying a paꢀent
3
as high or low risk of developing a disease) .
We speak of likelihood or odds because logisꢀc
explained, and one or more independent variables, regression provides odds raꢀos and not relaꢀve risks
which act as predictor variables. This disꢀncꢀon which are esꢀmates of probabiliꢀes.
enables the modeling of how the independent variables
Selecꢀon of independent variables: how to
influence the dependent variable. This allows us to avoid overꢁꢃng and underꢁꢃng.
make predicꢀons and evaluate the relaꢀve importance
A mulꢀple logisꢀc regression model allows for
of each predictor variable. However, regression, like the evaluaꢀon of the effect of two or more independent
2
correlaꢀon, does not imply causaꢀon .
variables on an event (dependent variable). Unlike
the model with a single predictor variable, this model
adjusts for confounding factors and esꢀmates more
precise associaꢀons. One of the most important aspects
of construcꢀng the model is to appropriately select the
When is logisꢀc regression analysis used?
Logisꢀc regression is commonly used in clinical independent variables in terms of both number and
research when the dependent variable is dichotomous, relevance. Including too many variables can lead to
meaning it has only two possible categories (e.g., overꢁꢃng, which causes the model to lose its ability to
3
“
presence” or “absence” of a disease) . This is the generalize to new data. Conversely, excluding relevant
binomial or binary logisꢀc distribuꢀon. There are other variables can lead to underꢁꢃng, resulꢀng in a lack of
distribuꢀons beyond the scope of this arꢀcle, such as precision. For example, including irrelevant variables
4
the mulꢀnomial and ordinal distribuꢀons .
(such as eye color) in a study to predict the risk of