If we have two variables, one dependent and one independent and correlation between them is almost 0. Can we apply method or technique to develop a model??
Yes: plot the data and get an impression. If you see pattern, you may think about some model that would produce such a pattern. You may then think harder to find out how you could experimentally show that this is not an approriate model (or at least that the key feature of the model is not a good explanation for what you see). Then you can actually perform the experiment and "test that hypothesis".
a correlation close to 0, may mean that the variables are uncorrelated, then the model would not respond to anything.
Obviously, this is, if the correct model applies !!
I have often seen that linear correlations calculated for models that are nonlinear
- search trend, through a scatterplot.
- type of variables and scale,
This helps to select the type of model,
To complement Jochen's point, the other direction you have to work is from model to data. What process generated the data? I've been working this past couple of weeks on data from an epilepsy study, where there should be two distinct processes at work. One process determines whether seizures will occur, while the other process determines how many seizures will occur each day. So we're expecting to see an excess of zero seizures. Once I understood the process that gave rise to the data, I knew what I should see when I explored it, and how I might model it.
You cannot tell simply from looking at the data what process gave rise to it. To fit the right model, you need to understand (or think you understand!) the process that gave rise to the data. And if there are several candidate processes, you need to understand how each might leave its 'fingerprint' on the data.
Lov Kumar.
of course, but what you should look, are the variables that theoretically are slated to be related.
Evaluate models, as Jochen says, according to the hypothesis
but always var. Dependent should correlate with one or more var. independent.
After that, work according to the model and investigate the requirements. among others, and according to the model
Normality of the errors., Homoscedasticity, independence from errors. Minimum variance, No multicollinearity, autocorrelation No,, etc.
but it depends on the model and variables ..
for example, in economics, there are autocorrelated models.
if there is no linear correlation, you should try to observe:
More variables for multiple linear model, or other behaviors, such as
nonlinear models, or there are more dependent variables and can build a multivariate model.
according to information you write, it may not be a model that responds linearly.
how much data you have ?, which is the patron of the scatterplot ?.
Often the presence of extraneous variables and generate outliers dirty models.
You have to check the presence of anomalies in a data set.
See, for example,
Prykhodko S.B. Statistical anomaly detection techniques based on normalizing transformations for non-Gaussian data // Computational Intelligence (Results, Problems and Perspectives): Proceedings of the International Conference, May 12-15, 2015, Kyiv-Cherkasy, Ukraine / Ministry of Education and Science of Ukraine, Taras Shevchenko National University of Kyiv and [etc]; Vitaliy Ye. Snytyuk (Editor). – Cherkasy: editor July Chabanenko, 2015. – p.286-287. – ISBN 978-966-493-975-8
https://www.researchgate.net/publication/282329599_Statistical_anomaly_detection_techniques_based_on_normalizing_transformations_for_non-Gaussian_data
Conference Paper Statistical anomaly detection techniques based on normalizin...
I believe that some of the answers above get you going in the right direction -- what are the theoretical underpinnings of the model, what does the theory tell you about which variables to include, and are the assumptions of OLS working.
Dear Carlos Jimenez-Gallardo,
currently we are doing this analysis on 500 different PLC program. For all program, we calculate 9 different type of software metrics and the change proneness.
After getting all value, we try to find the relation between change proneness and software metrics.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression