Hi,
Sometimes a negative intercept makes no sense at all (i.e. weight, height, score etc...). And this is completely dependent on the research problem.
How can it be forced to be above zero in R (not just forced to the origin)? and what are the possible implications on the regression model?
Some software (e.g. MLwiN) has the capacity to do this usually under the term 'constraints' .
I would however caution against its routine use. My experiences is that obtaining impossible negative values is a sign that is something is wrong. Usually outliers titling the slope to unrepresentative values , or predicting outside of the range of the data. The intercept is the mean estimated value of the response when all the predictors are zero, and in many situations such values do not exist in the observed data. Consequently and routinely I center all continuous predictor variables around their mean or median value. Thus if demographic Age is included in the model if there is no centering then the estimated response represents a baby 0 years old, but if I include (Age-43), the intercept represents a typical person.
See my previous response to this question
https://www.researchgate.net/post/A_negative_coefficient_for_a_constant_in_a_linear_regression
Some software (e.g. MLwiN) has the capacity to do this usually under the term 'constraints' .
I would however caution against its routine use. My experiences is that obtaining impossible negative values is a sign that is something is wrong. Usually outliers titling the slope to unrepresentative values , or predicting outside of the range of the data. The intercept is the mean estimated value of the response when all the predictors are zero, and in many situations such values do not exist in the observed data. Consequently and routinely I center all continuous predictor variables around their mean or median value. Thus if demographic Age is included in the model if there is no centering then the estimated response represents a baby 0 years old, but if I include (Age-43), the intercept represents a typical person.
See my previous response to this question
https://www.researchgate.net/post/A_negative_coefficient_for_a_constant_in_a_linear_regression
I agree with Kelvyn.
This also commonly arises when you have a model that doesn't capture key aspects of the structure and context of the data. This may be a sign that you need to transform one or more variables or switch to a generalised linear model (e.g., logistic, Poisson) or some other approach. The latter are particularly useful when the response is bounded to e.g., 0 to 1 or between 0 and infinity.
All advice on this topic points to the elimination of errors before resorting to constraints like this. If you simply must do it, the lavaan package in R allows for specifying a wide variety of constraints, including inequality constraints.
If you had an outcome variable y with predictor variables x1 and x2 in data.frame d, you could specify a positive intercept b0 like so:
library(lavaan)
model
Kelvyn Jones indeed, sample size is small and there are two non-extreme outliers. Maybe I should exclude outliers and run the regression to get the positive intercept.
Thom Baguley I will consider this approach first.
W. Joel Schneider is it a SEM package?
Thanks all, you helped a lot.
Yes, lavaan does SEM, but also multiple regression, path analysis, and confirmatory factor analysis.
I agree with Kelvyn Jones Do not drop the outliers since they are not "extreme". See the attached. These procedures (which I recommend) are available in R package MASS as rlm function. This would be better than dropping IMO. Best, David Booth
If you simply want to change the intercept (let's say using the lm function in base R), you can run the model first and get the intercept (let's say -5). Then add a value larger than this to your DV (e.g., +6) and rerun the model. This will make the intercept positive, while the slopes for your IVs would remain the same (because adding a constant to DV doesn't affect the variance or covariance in the model).
Thanks David Eugene Booth . I discovered, late, that it is a nonlinear trend line that may explain the negative intercept.
Okan Bulut it would be good idea in some cases.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression