This question has a lot of caveats, but without getting into all of them I offer you the following opinion: My preferred method when using education as an independent variable is always to use years as a continuous variable. It makes the interpretation of the final regression (assuming ols and not logistic) a bit more "comfortable" for non-quants. When I'm stuck with education as a categorical variable in a data set, I first try and determine where the education level was attained. If it is an American data set you can be reasonably certain that a 12th grade level of education is equivalent to 12 years; an Associate's Degree is equivalent to 14 years' a B.S./B.A. is equivalent to 14 years, etc., etc. Thus, you can recode the categories to match the appropriate  years giving you a continuous variable. 

If using this approach isn't feasible you could always recode the categories to represent traditional breaks in levels of education. This would reduce your categories from 7 to say 3:  less than high school, high school grad, college grad. This way your final regression table isn't muddled up with so many different categories and interpretation is a bit more straightforward.

If you're using logistic regression, I might suggest you recode your categories into a binary variable. In my discipline of criminal justice we often use high school as a benchmark, so I might recode to 0 = non high school grad/GED 1= high school grad/GED. These categories may not make sense for you, but you get the point.

Can you elaborate on what your data set entails, what other IVs you are using and what your outcome variable is as well?

cheers,

Dr. Mathews

Dr. Mathews, 

Thanks a lot for your informative answer. I am trying to see effects of parent-child relationship (close and conflict) along with child's temperament on children's self-regulation skills. So My intention to control for mother's education and child ethnicity (hispanic -non-hispanic). Since it is my control variable, I am trying to simplify it into meaningful categories.  It is an American data set. 

What are the other variables in your model?

I have parent-child relationship (close and conflict) along with child's temperament on children's self-regulation skills as an outcome; and mother's education and child ethnicity (hispanic -non-hispanic), gender, and age as control variables. 

Is the response variable a continuous random variable?

Yes it is a continuous random variable. My initial mother reported levels of education are: 

1 = 'Eighth grade or less'

2 = 'Some hs no diploma or ged'

3 = 'HS diploma or GED'

4 = 'Some College'

5 = 'HS diploma, GED and tech training or certificate'

6 = 'AA, AS, two year degree'

7 = 'Bachelor''s'

8 = 'Master''s'

9 = 'Doctoral Degree'

10 = 'Other' ;

How do you think I can group these to make sense out of it by simplifying them? 

Thanks 

Ibrahim,

Looking at your actual levels of the education variable it doesn't make sense to try and make them continuous since you have "some college" and "other" included.

You could use less than high school, high school/GED/Tech training, and college graduate as levels. However you'd have to deal with the "other" category and determine why they are not placed elsewhere. If you can't delineate the "other" category you would end up with the following four levels:

1 = less than high school

2 = High School/GED/Tech School

3 = College Graduate

4 = Other

These seem to be natural breaks in the existing categories and take them down from 10 to four. You'll lose some nuance between type of college grad, but you can always do some post hoc work to investigate if you really need to know what the difference between having an associates, masters or a bachelors means.

Remember that when you drop these categories into your stats package one of them will be the reference against which all others will be measured. So try and identify a category that makes some kind of theoretical and practical sense to use for that purpose. I'm not familiar with your dependent variable and previous connections between it and level of education, so I can't offer you any opinion on which category should be used as the reference.

Ibrahim,

Looking at your actual levels of the education variable it doesn't make sense to try and make them continuous since you have "some college" and "other" included.

You could use less than high school, high school/GED/Tech training, and college graduate as levels. However you'd have to deal with the "other" category and determine why they are not placed elsewhere. If you can't delineate the "other" category you would end up with the following four levels:

1 = less than high school

2 = High School/GED/Tech School

3 = College Graduate

4 = Other

These seem to be natural breaks in the existing categories and take them down from 10 to four. You'll lose some nuance between type of college grad, but you can always do some post hoc work to investigate if you really need to know what the difference between having an associates, masters or a bachelors means.

Another option is to make the variable dichotomous. You could recode them in the following manner:

0 = Less than college graduate

1 = College graduate

or

0 = Less than high school

1 = More than high school

Again, you lose some nuance in your levels, but it's another way to get your categories to a manageable level.

Remember if you use more than two categories when you drop them into your stats package one of them will be the reference against which all others will be measured. So try and identify a category that makes some kind of theoretical and practical sense to use for that purpose. I'm not familiar with your dependent variable and previous connections between it and level of education, so I can't offer you any opinion on which category should be used as the reference.

If one's measure is "amount of X" then it is ordered, even if it has only two levels. Treating ordered variables as being qualitative in statistical analysis, while omnipresent in the literature, is always incorrect.

Combining groups with respect to qualitative or ordered variables can (and often does) induce paradoxical confounding, and/or result in lower predictive accuracy.

Both of these topics have been covered in many articles. Many articles on both topics are available in a free eJournal: www.ODAJournal.com . Articles on these topics in pay-for-view journals are listed on this site, in the Publications tab.

A book covering these topics is in press, and is expected in late February or early March. The title is "Maximizing Predictive Accuracy" (DOI: 10.13140/RG.2.1.1368.3286), and the Table of Contents is available on my RG page (see Featured Publications).

Though Dr. Yarnold is technically correct, the idea that it is "always incorrect" is a bit too definitive for the social sciences. A decision is clearly going to be made and he was asking for some guidance as to what might be "best" given his circumstances. So, indicating it is "incorrect" doesn't help him much as he is clearly dealing with a social science issue and is at a decision point.

The suggestions that have been made are by no means perfect; however either is his data which are categorized in a messy way. Of course recategorizing the raw data will remove the nuance he might otherwise get by leaving them as they are; however does that help him explain the results to the target audience? Does it make it clearer or muddy it up more? Additionally, there are plenty of post-hoc analyses that can be used to identify confounding that may or may not arise. 

As a primarily criminal justice researcher I deal with this kind of data on a near daily basis and these types of decisions have to be made in at least 8 out of 10 analyses I do (yes that is anecdotal) as a result of the sets being used, the way data points are collected, and the nature of the statistical tests traditionally used in the field. 

It appears Ibrahim is dealing with some kind of behavioral/social science (likely education or social work) so he is likely safe recategorizing his set using some of the theory from his field to identify more accurate groupings.

Though Dr. Yarnold is technically correct, the idea that it is "always incorrect" is a bit too definitive for the social sciences. A decision is clearly going to be made and he was asking for some guidance as to what might be "best" given his circumstances. So, indicating it is "incorrect" doesn't help him much as he is clearly dealing with a social science issue and is at a decision point.

The suggestions that have been made are by no means perfect; however either is his data which are categorized in a messy way. Of course recategorizing the raw data will remove the nuance he might otherwise get by leaving them as they are; however does that help him explain the results to the target audience? Does it make it clearer or muddy it up more? Additionally, there are plenty of post-hoc analyses that can be used to identify confounding that may or may not arise. 

As a primarily criminal justice researcher I deal with this kind of data on a near daily basis and these types of decisions have to be made in at least 8 out of 10 analyses I do (yes that is anecdotal) as a result of the sets being used, the way data points are collected, and the nature of the statistical tests traditionally used in the field. 

It appears Ibrahim is dealing with some kind of behavioral/social science (likely education or social work) so he is likely safe recategorizing his set using some of the theory from his field to identify more accurate groupings.

There is no theoretical imperative for "technically incorrect" methods in science: the objective of science is to identify technically correct answers. Technically incorrect is the reason that social science models are inaccurate and often non-reproducible. The difference between successful and unsuccessful disciplines lies in the quality of theory, measures, and statistical methods. There are solutions to statistical inadequacy, but they are new--not broadly known. Investigating the resources I cited will empower all researchers to overcome well-known limitations of legacy statistical methods. It is the professional imperative of an academic (or, minimally, of a statistician) to understand and empirically evaluate new analytic approaches. I have used these new methods in various criminal justice applications, and the models worked so well that they are used to determine which prisoners to release, and the level of probation monitoring assigned to each individual. The models were much more accurate and parsimonious than alternative models, and had superior cross-generalizability. Other laboratories had similar experience. Some of these studies (those I was able to find) are listed in the Publications tab.

Based on previous research it is found that the relation between a variable measured as a categorical variable (such as age, income, education)  and the same variable reported as continuous  variable is approximately 0.99 if and only if:

1- The categorical variable  has categories with approximately equal intervals

2-  we don't have a lot of observations at the lowest and highest categories.

If the two conditions are not satisfied then, don't treat a categorical variable as a continuous 

I did not read all the discussion above.

What I can say that you must consider education level as a categorical variable.

If you have enough data you can make a stratified analysis through the levels of education, or you can consider interactions.

I expressed my view, but I have "a bit" of advice for young researchers.

Improper "scientific" analysis can sometimes produce unfortunate ends: failure of O rings on rocket boosters, innumerable adverse drug interactions, and invading countries on the basis of analysis indicating WMDs, are obvious examples.

Clinging to the troubled past and ignoring promising new methods is not a fruitful scientific modus operandi. Set your course towards the future.

Without getting into the weeds, it is clear that Dr. Yarnold is a statistical theorist. Though useful for moving the field of statistics forward, rigid textbook statistical theory is often incongruent with the realities of social science research where data are messy and unobservable variables persist. The argument that the categories in question must be ordered ignores the reality as to whether or not the discipline where the categories are being applied has an empirical basis for such an assumption.

Only Ibrahim can tell us whether his discipline says that the education "level" (really nominal categories) should be viewed as ordered. If that discipline has an empirical basis to say that having a master's degree over a bachelor's degree or a GED over 8th grade education, etc. is important as it pertains to his dependent variable(s), then yes they should be treated as ordered. However, if no such basis exists than they are simply nominal categories, like colors of the rainbow or seasons of the year. Remember, he is using this variable as a control so there should be some guidance in the literature here.

This is the difference between applied field research and statistical theory. The question that was being asked had nothing to do with statistical theory and everything to do with what appears to be a messy field data collection that resulted in an unwieldy amount of categories for a specific variable. Options provided to the poster all have their uses and limitations. Ultimately it will be up to him to make the determination of which route to go based upon the conventions in his discipline, previous empirical research involving the same variable, etc.

Without getting into the weeds, it is clear that Dr. Yarnold is a statistical theorist. Though useful for moving the field of statistics forward, rigid textbook statistical theory is often incongruent with the realities of social science research where data are messy and unobservable variables persist. The argument that the categories in question must be ordered ignores the reality as to whether or not the discipline where the categories are being applied has an empirical basis for such an assumption.

Only Ibrahim can tell us whether his discipline says that the education "level" (really nominal categories) should be viewed as ordered. If that discipline has an empirical basis to say that having a master's degree over a bachelor's degree or a GED over 8th grade education, etc. is important as it pertains to his dependent variable(s), then yes they should be treated as ordered. However, if no such basis exists than they are simply nominal categories, like colors of the rainbow or seasons of the year. Remember, he is using this variable as a control so there should be some guidance in the literature here.

This is the difference between applied field research and statistical theory. The question that was being asked had nothing to do with statistical theory and everything to do with what appears to be a messy field data collection that resulted in an unwieldy amount of categories for a specific variable. Options provided to the poster all have their uses and limitations. Ultimately it will be up to him to make the determination of which route to go based upon the conventions in his discipline, previous empirical research involving the same variable, etc.

Yes, I am indeed a double-time uncompensated theoretical statistician--I do this for God and humanity, and am grateful this is obvious. And, I am presently a part-time compensated applied statistician--I am fortunate do this to survive, to pay numerous Caesars, and to have some intellectual fun with a few brilliant colleagues. I spent three decades as a university professor, and when most of my dear mentors, colleagues, friends and family died, I decided to devote what is left of my time to the greatest good.

My laboratory and numerous independent labs provide a cornucopia of applied articles in a myriad of substantive fields, that demonstrate the ecological (and analytic) superiority of new methods. Success in this endeavor is imperfectly quantified by one's impact score, that (unlike the RG popularity score) is based on publications in indexed journals--a function of both quantity and quality. Citations are another imperfect index of "success". But, I am not greedy: last year I had several applied articles accepted at prominent journals, and since there were too many authors to be listed, I stepped down to help propel the career of youngsters on the team who met the criteria for authorship. I have done this innumerable times over my career, including articles in high-impact journals such as NEJM, JAMA, and Lancet. I find sharing the wealth to be intrinsically rewarding, in part because I have an alternative outlet for my academic ambition.

Most of my theoretical work (which is most of my work) is not included in my impact score, which otherwise would be at least three times higher. Much of this work is available to anyone without charge via the ODA eJournal. As of this writing a total of N = 5,092 researchers in 135 countries have read more than 14,000 articles in the last two years: the average number of articles read per visitor, or per country, or per visitor per country, isn't representative of actual behavior. Discoveries happen regularly and rapidly, so I wrote the book to make sense of all that is known as of the end of 2015.

Data--the foundation of an applied researcher's career--are only (i.e., primarily) "messy" because legacy methods require that the data meet the geometric and distributional assumptions of the method. The most frequently used legacy methods provide performance indices that aren't chance-corrected, and that can't be maximum-corrected because perfect prediction induces division by zero. Cross-generalizability of legacy-based models is infrequently reported. Deep understanding of the methods requires a strong foundation in math, as there are many complex equations. Many books must be mastered to understand what is known for each legacy method, and there are many legacy methods.

New methods require that models exactly conform to the hypothesis and empirical geometry for each individual application, without making distributional assumptions. These new methods explicitly maximize predictive accuracy (normed against chance) and parsimony, and perfect classifiers are the objective. Performance measures are chance- and maximum-corrected, and the emphasis is on cross-generalizability. Deep understanding of the methods doesn't require a strong foundation in math, as there are few (elementary algebra) equations. One book (in press) must be mastered to understand all of what is known about the new paradigm.

Which sounds better?

**

SUMMARY

The new paradigm resolves all known limitations of legacy methods--for which no resolution is possible owing to inherent deficiencies: which is better?

The new paradigm is transparent, with models expressed in the original units of measurement, while the legacy paradigm-based models are expressed in opaque-to-dark units unrelated to the original metrics of variables: which is better?

Models developed using the new paradigm explicitly maximize predictive accuracy, whereas no legacy models maximize predictive accuracy as their objective function (variance explained, and value of the likelihood function, are prevalent in the literature): which is better?

Models developed using the new paradigm are more parsimonious (and more accurate) than models developed using legacy methods: which is better?

Models developed using the new paradigm are invariant over any monotonic transformation of the variables, legacy models are sensitive to transformations used to try to change data to meet model assumptions: which is better?

Models developed using the new paradigm require no distributional assumptions, so Type I error is never compromised, whereas legacy-based models require that assumptions are met to ensure validity of estimated p-values (and results of this crucial requirement is rarely discussed in articles): which is better?

Empirical research in many areas shows that models developed using the new paradigm cross-generalize better than legacy-based models: which is better?

Learning the new paradigm requires reading one current book, learning the legacy methods requires reading a library of books: which is better?

For the new paradigm running any statistical analysis, no matter how complex, requires learning one statistical analysis system that for most analyses typically involves a handful of commands that are transparent--easily understood. In contrast, numerous legacy software packages need to be mastered to cover the enormous number of different legacy methods, and the syntax (within and between software systems) can be a puzzle worthy of the New York Times: which is better?

Many examples of problems that could NOT be analyzed using legacy methods have been published using the new paradigm to identify accurate, parsimonious models that have strong face validity: which is better?

What arguments are there against learning the new paradigm?

About parenting education level, do you all know how to combine mother's and father's eduation level given as categorical variable into one categorical variable? For example, should I use higher level for relevant?

Since your are interested in " My intention to control for mother's education and child ethnicity (hispanic -non-hispanic). "  then,  you better re code variables in K-1 vectors  if your sample is large enough.  For instance , education in your study is broken into 10 categories and you may re code this variables into  10 - 1  = 9  vectors of "1"  and "0" . You may combine  neighboring categories to end with less than 9 vectors. Check  frequency before combining categories.

Note that you can't "combine mother's and father's education level given as categorical variable into one categorical variable"  but you can create an interaction between the dummy variables measuring both