In life, prediction and causality are two different concepts. In multiple regression prediction and causal analyses are also different. In a prediction analysis, the goal is to develop a formula for making predictions about the dependent variable, based on values of the independent variables. In causation analysis, the independent variables are regarded as causes of the dependent variable.

This is explained by Kraemer et al in the context of medicine and risk factors. They distinguish what they call a proxy risk factor as something that is correlated with an outcome (or disease) and can predict its occurrence over time. For example, fingers stained yellow is related to future development of lung cancer (borrowing Illari & Russo's, 2014 example). The predictor is not necessarily the cause, but might be related to the cause, in this case yellow staining is due to smoking. They refer to a causal risk factor as one that when manipulated will affect the outcome. So if you get people to stop smoking, their chances of developing lung cancer declines. Of course, finding that a manipulation, even in a laboratory experiment, results in an outcome is not a guarantee that the manipulated variable is the real cause. It is possible that there is confounding, or unintended effects (e.g., demand characteristics) which can affect even the most well designed experiment). For example, suppose you do a randomized trial in which a new medication is shown to be helpful in treating a particular condition. It might be that the medication itself is the causal factor, but it is possible that there is something else going on. Perhaps the medication makes people thirsty, so they drink more water, and it is hydration, not the medication, that is the real cause. If we suspect this and conduct a trial in which the "treated group" is assigned to drinking more water, we would determine that water alone is the cause of the outcome. But is it really water, or something associated with the water?

Illari and Russo are two philosophers of science whose book on causality is a nice introduction to the concepts of causality from a variety of philosophical perspectives.

Illari, P., & Russo, F. (2014). Causality: Philosophical theory meets scientific practice. Oxford, UK: Oxford University Press.

Kraemer, H. C., Stice, E., Kazdin, A., Offord, D., & Kupfer, D. (2001). How do risk factors work together? Mediators, moderators, and independent, overlapping, and proxy risk factors. The American Journal of Psychiatry, 158(6), 848-856. doi:http://dx.doi.org/10.1176/appi.ajp.158.6.848

Thank you so very much Dr. Mahfuz Judeh for your reply. Can you please mention some research based examples for Regression based Prediction and Causality separately, especially from management discipline?

Best Regards

Thank you so very much Prof. Paul E. Spector for your detailed reply. However, sure because of my lack of understanding, it seems to me that Causality and Prediction concept is more confusing in the stained fingers and lungs cancer as well as new medicine for a certain condition examples.

I quote your words to make my point more clear "They distinguish what they call a proxy risk factor as something that is correlated with an outcome (or disease) and can predict its occurrence over time. For example, fingers stained yellow is related to future development of lung cancer (borrowing Illari & Russo's, 2014 example)".

Can we theoretically (absolute reason vs probable reason) as well as methodically (experimental vs non-experimental) separate Prediction and Causality when we know that "confounding, or unintended effects (e.g., demand characteristics) which can affect even the most well designed experiment)"?

I would be grateful if you explain it again.

Best Regards

Maybe one way to look at it is that prediction is a description of a phenomenon, that is, we can establish that variable X can predict the occurrence of variable Y in a given population under certain conditions. It is possible that X only predicts for certain populations (e.g., men but not women), and under certain conditions (e.g., in the laboratory but not the field). So there can be boundary conditions to our prediction.

Causality subsumes prediction--if X causes Y, then X will predict Y. Part of the idea of causality is that you have prediction. But causality is more than prediction. It says that X is the reason for Y, e.g., yellow fingers are the reason people get cancer. In practice, causality is an inference that we make, and goes beyond description. We look at our evidence and draw the conclusion that X causes Y. Kraemer et al. define a causal risk factor in terms of manipulation--if you manipulate X and get Y, then you conclude that X is a causal risk factor. Although showing the effect of a manipulation is relevant to building a causal case, it isn't sufficient because the manipulation of X can lead to Y even if X is not the real cause--like the yellow finger example. If your intervention to remove yellow fingers is to have people stop smoking, you will erroneously conclude that yellow fingers are the cause. Like theories, we cannot prove causality--we merely collect evidence for it and compile a case based on the weight of that evidence. We can conduct a well-controlled experiment, but there is always some possibility that our interpretation of the nature of the X and Y variable and why X led to Y are incorrect. Sometimes we think we are measuring something, but our measures reflect something else or are biased. Sometimes we think we know why X led to Y, but there is some other factor at work that we are unaware of.

Dear Prof. Paul E. Spector thanks a lot for your concise but comprehensive explanation regarding Causality and Prediction.

Best Regards

A simple technical+theoretical difference that distinguish causality from prediction is the time variable. In theory, causality can be only confirmed in retrospective meaning that any effect now had a cause sometime in the past. Prediction on the other hand address prospective outcome of a current causal factor. In practice both are tested using regression analysis techniques. Causality is practically possible to establish using controlled conditions and a time gap in measurement of cause and effect. Perdition on the other hand is not possible to statistically test as one can't measure variance of a variable in future so we do a compromise when testing prediction and try to explain/understand variance of a variable in a cross-sectional design. Hence in my opinion technical difference between causality and prediction is based on research design (causality require time difference in measurement) while prediction can be made in cross-sectional studies both using regression analysis techniques.

Thank you Jamil Ahmad Malik for your useful input in understanding the Causality-Prediction puzzle :)

You can look at our papers then you see that they are different. In science we just have probabilities that a phenomenon contributes to a behavior, we cannot state with 100% certainty that it causes the behavior:

Thank You S. Béatrice Marianne Ewalds-Kvist I will read the material and will get back if need guidance. Best Regards

Please read the following paper for details.

It gives two differences between prediction and explanation.

P1. Explanation is not the same as prediction. Predictions don’t require explanations but explanations generate predictions. Explanation requires a deeper level of understanding.

P2. (Statistically speaking) to say that one observation ‘explains’ another observation, is not, theoretically speaking, an explanation.

Reference

Whetten, D. A. (2009). An examination of the interface between context and theory applied to the study of Chinese organizations. Management and Organization Review, 5(1), 29–55.

Dear Mr Shushanta,

Though it made the concept little more confusing to me :), I am really thankful for your feedback and will try to read the reference you provided for my better understanding of the topic.

Best Regards

At best, one can say "perhaps, X causes Y" and not beyond that, I guess.

Prediction is simply the estimation of an outcome based on the observed association between a set of independent variables and a set of dependent variables. Its main application is forecasting.

Causality is the identification of the mechanisms and processes through which a certain outcome is produced. Its main application is control. Of course this yields predictive capacity as well, but through an explanation of the forces that rule the phenomenon in question.

A predictive association between variables means that the behavior of one set of variables is useful to forecast the behavior of another set, but the intentional modification of the variables in the first set will not necessarily produce a specific change in the second set. A causal relation is one where interfering in the behavior of one set of variables will indeed have an effect on the behavior of another set.

As an example, consuming artificial sweetners has a positive association to the probability of being obese, but such behavior does not cause obesity, therefore, avoiding them will not make one lose weight. On the other hand, the consumption of a large number of calories does have a causal relation to body weight, so that cutting back on calories will actually tend to make one lose weight.

Statistical analysis per se cannot distinguish whether an association between variables is merely of predictive value or if it is indicative of a causal relationship between them. The only way to determine that in any reliable way is through an experimental design for the research, preferably with a control group.

Thank you Dr Bruno Campello de Souza for your detailed explanation on the matter. Best Regards

If it is true that "x causes y", we should necessarily find that "x predicts y" in a regression analysis. ok?

" If it is true that "x causes y", we should necessarily find that "x predicts y" in a regression analysis. ok? "

One would expect so, but it is not always the case. There might be covariances that "mask" the effect.

If increases in X cause increases in Y, but increases in Z cause decreases in Y and there is a positive association between X and Z, then the "net" effect observed might be zero. Things can be further complicated by the existence of still other variables that might mediate the interactions between X, Y and Z.

A good example of causality is smoking causing lung cancer. In this case, current smoking is also a predictor of later lung cancer. At first sight you would think that causal factors should always have some degree of ability to predict an outcome, because these causal factors have a direct effect on the outcome; thus the very existence of the causal effect must make the effect on the outcome more likely to happen. In practice, however, causal factors may vary in their ability to be an accurate predictor, because the strength of association with the outcome will vary. For example, we may show that X causes an increase in subsequent outcome (after adjustment for all plausible confounders) but that the effect on Y is small. This means that X is likely to be one of many causal factors influencing Y. If there are many causal factors the predictive value of just one of them is not likely to be very strong. However if a causal factor is the main cause of the outcome then, like smoking, it will be a very good predictor.

On the other hand you can have good predictors that are not causal factors at all. These are often referred to as 'markers'. These are usually variables that are correlated with a true causal factor. For example, heavy alcohol consumption is not a confirmed causal factor for lung cancer but because of the correlation between heavy drinking and smoking there an association between heavy drinking and lung cancer. This association is enough to allow heavy drinking to have some predictive power with regards to later lung cancer.

So causality and prediction are rather different, though they do inter-relate.

Thanks a lot Mark Christopher Perry for explaining the difference between causal factors (e.g., Smoking to Lung Cancer) and markers (e.g., Drinking to Lung Cancer).

This is an old discussion and many good answers, comprehensive and covering all aspects, are already provided.

However, as one that I have dealt from time to time with the same issue, I would emphasize two aspects.

One is the temporal relation between the events. If one even is consequent to the other, it is impossible for the latter to be the cause (at least in our understanding of how entropy works).

The second aspect is that regression can do predictions as well as retrodictions (if one may use the word). What I mean is that if there is an association (correlation) between two variables, either one can be predicted based on the other, regardless of which one is the cause (if there is a causal relation between them) or which one came first.

I am only saying this because once the data crunching is done, the interpretation is more than a simple "ah, anovas are (usually/commonly) used for experiments, thus the IV is the cause", or "ah, but I chose this as my predictor, so it can't be the effect".

Once the statistical results are in, and we need to move to interpreting them, we are entering, by necessity, or going back for that matter, to the basic theoretical assumptions and the dogmatic views one has when choosing a research design, and conceptualizing the main constructs, in the first place.

Best regards,

Thanks Dorin Stanciu for your thoughtful response.

Yes, because Causality is looking at how one event leads to another, while prediction is concerned about forecasting future events based on past data or observations.

Casuality refers to the relationship between cause and effect, while prediction is the process of estimating the likelihood of a future event based on past observations.