Gender Pay Gap - Not Actually that Big?

[Loren Pechtel]

Wrong. The gender pay gap is an empirical fact and the OP article confirms it. The only question is why the gap exists.
The claim that it does not much exist is a purely an inferential, assumption-ridden conclusion reached by woefully inadequate efforts to explain away the empirical reality by abuse of inferential statistical techniques. Note, "inferential" is not my personal view of their technique, it is the standard definitional term for what they are doing.

It exists only in the use of a flawed yardstick. When you control for the obvious, easy to measure things 90% of the gap disappears. I find it very unlikely that the remaining gap isn't due to things that aren't so easy to measure. (For example: Willingness to take extended business trips--fathers are much more likely to do so than mothers.)

 

[ronburgundy]

We both agree that causality must be argued for outside a correlational model. Yet you continue to regard gender discrimination as the better account of the variance without evidencing it. An appeal to "parsimony" fails.

It doesn't fail. We need more data to establish with confidence what the causal paths are, but there are also a priori probabilities in which parsimony factors and favors a direct impact of gender on decisions made by people who are well aware of the gender of the people they are deciding about.

The most parsimonious explanation for the observed covariance between gender and pay is a causal influence of one on the other and pay cannot cause gender.

Your making my case for me. You do not believe gender causes pay differences; you believe gender discrimination causes pay differences. You already have mediating effects.

Incorrect. Gender Discrimination is not a variable, it is a label for a causal relationship between the variable of gender and cognitive acts where a person is making a discrimination/distinction between options.
Deciding to pay person X more than Y is inherently and by definition an act of discrimination which is merely "to recognize a distinction". Thus, all variability in pay results from a person discriminating between employees and which deserve what pay. All accounts of the average pay difference between gender and pay must include both gender and the cogntive act of discrimination (unless you assume all employers make pay decisions at complete random).
The gender discrimination model presumes only those variables and a direct relation between them. All non-discrimination models presume 1 to many extra variables that happen to relate to gender and are the real casual influence on how employers discriminate/differentiate between employees and what they will be paid.

And among the countless ways in which they differ is likely to be the different responses they evoke from their employers related to pay.

It's the effect size of that response that we ought to be concerned with, not that it exists.\

The effect size is the simple bivariate difference in average pay of men and women without any other variables in the model.

You don't just have to assume that gender relates to other variables, you must assume that those particular variables it relates to also causally impact pay and are the only reason gender relates to pay.

One can look for a 'gender discrimination of the gaps' all she likes -- but it makes no sense at all to not control for variables known to impact pay (such as industry and hours worked).

What you are actually adding is assumptions to your model. Calling them control variables requires the assumption that they have only on particular possible causal relation to the other variables, namely that they cause the outcome (pay), and are impacted by gender in some way. IOW, that they are mediators that are not themselves impacted by the outcome variable or other variables correlated with the outcome which are in turn caused by main predictor of interests (gender).
It is a common misuse of regression to just toss any variable that might impact the outcome into the model as a "control". Every variable has to be carefully considered for its potential relationships to the other variables that would create shared variance and thus alter the results by entering it into the model.
The question is whether A -> C, and your are considering only the possibility that A -> B -> C, so you want to enter B as a control and when the A -> B relationship gets reduced you infer that B is the real cause of A. That is wrong. IF it is possible (and in this case it is just as possible) that A -> C ->B, then "controlling" for B is does not make sense and your interpretation of the reduction in the relevance of A is wrong. Both causal chains mean the the A and B share much of their variance that relates to C, and yet in the latter situation that shared variance in no way reduces the reality that A is a cause of B.
Pay is not only a major aspect of how employees are treated at their job, but it is correlated with other aspects of workplace treatment, and their is a mountain evidence showing that how people are treated by others will causally impact how much time they spend around those other (hours per week at work), and how long they maintain that relationship (change employers, change fields, leave the workforce, etc.). There is nothing magical about the workplace. IT is just another form of social interaction subject to to all the same factors known to impact human interactions in general.

BTW, gender is just as much a "control" variable on those other predictors as the other ways around, and those other predictors becomes less related to pay when gender is in the model. IT is just as valid (or just as invalid) to interpret the results in terms of how much of the seeming effect of hours and time at job are accounted for by gender.
It is entirely a matter of assumption-based framing.

Here is the false "logic" of your interpretations, and the sole difference in my example is a priori theoretical assumptions. We have a correlation between smoking and lung cancer. We decide to "control" for things related to lung cancer, so we control for medical bills. Low and behold, the relation between smoking and cancer greatly reduces. You come in and say, "Therefore, there is no real difference in cancer rates between smokers and non-smokers, because its attributable to other factors like medical bills.
There is absolutely nothing in the empirical data and the logic of the analysis that is different in that case and ours. The only difference is that in that case we both share the assumption that cancer would impact medical bills, so its stupid to treat it as simply a control variable and not as partly an indirect outcome of smoking caused cancer. your
With pay, you assume that effort and time at a job can only possibly be the cause of pay, whereas I consider the actual relevant evidence that they both casually impact each other, and thus a large portion of their shared variance reflects variance caused by other factors such as gender on pay that then impact effort and time at a job. Thus, it is false to treat all that shared variance as though effort and time at work get to claim causal credit for it.

Unless other patterns of causality can be ruled out a priori as impossible on principle, the reduction in predictiveness of a variable after entering other variables cannot be interpreted as evidence against the importance or causal impact of that variable on the outcome or as evidence that the other variables are the real cause. It can only be interpreted as evidence that whatever the relationship of the focus variable to the outcome, it also has a shared relationship to the additional variables added. The data is agnostic on what the nature of those relationships are.

First, let's be clear that it is the OP and its supporters claiming that all or near all the wage gap is due to something other than discrimination. I am not saying that all of it is due to discrimination, just that there is no good evidence favoring other explanations for any part of the gap, let alone most or near all of it.

No good evidence? Are we on the same planet?

Correct. No good evidence, and I've explained why numerous time. Evidence requires that one the theory is more consistent with the data. It isn't. None of the data presented here favors a non-discrimination model over any theoretically plausible discrimination model that has been proposed. The only model made less likely by the data is a model in which gender impacts pay without having any relationship to any of the other variables in the analysis. That is a strawman that ignores the mountain of general psychological evidence that people's perceptions of mistreatment (less pay than deserved) would certainly impact nearly all of the other "control" variables. Such an account is not more presumptive that that a non-gender discrimination model, and only differs in the directions of the causal arrows.

That aside, NO, your analogy is a total fail and a discrimination account for pay is pretty much the direct opposite of what it would be for height. In fact, assuming that their is no pay discrimination is more like assuming that gender is only related to height because women do not try hard enough to be tall.

No, I am pointing out a clear biological difference between men and women that is unrelated to 'discrimination' by employers or anyone else.

No, you are pointing to your purely faith based assumption that it is unrelated to employer decisions, an assumption made extremely implausible by its contradiction to experimentally demonstrated impacts on interpersonal judgments and preferences of people in general (FYI: employers are people) and specifically job related judgments.

A direct effect of genes on height will always be more parsimonious that any other route, including discrimination or any other account involving people's behaviors (such as successful nutrition acquisition).

Gender does not equal genes. There will always be mediating effects (even if they are all 'biological').

Zeno's paradox. Technically, there are infinite steps between the most direct causal factor and its outcome, so that is a red herring. It is a matter of introducing new variables in the middle that have only highly contextually dependent relations to either gender or pay. You are increasing the mediating factors by at least double or much more for every single "control" variable you presume is a mediator in the relationship.
Each of those require the same or more number of steps as the more direct effect of gender, just to get from gender those variables, then they require at least as many steps again to get from those variables to pay decisions. You see a guy with a bloody fist standing over a guy with a bloody nose. Is is more plausible he punched the guy or that he punched someone else who then fell into that guy? (hint: your theory is the latter).

In contrast to height variance, pay determinations are not a biological trait but rather inherentl the product of an employer's mental processes in which they distinguish between employees in deciding who gets what level of pay. IOW, by definition, all pay variance is the product of the mental act of discriminating between options.
Thus, all possible models accounting for pay variance share the same presumption that discrimination in the general cognitive sense is happening, and they differ in what factors are assumed to impact those acts of discrimination.

To use the language of 'discrimination' here is to muddy the waters. I am not worth as much to my organisation as the big boss is, so our employers have 'discriminated' based on service value and paid her more. That is not the kind of discrimination anyone ought be concerned with.

No. it is to clarify that actual variables. You are muddying by bring issues of morality and politics and "concerns" into it. Ignore at that. Pretend you are being objective and not trying to abuse science to support an political conclusion. Gender discrimination might be politically negative, but it is also just a potential scientific process in which an employers act of differentiating things (pay levels for employees) is impacted by the variable gender of the employees. That is all gender discrimination is on a scientific level, without any moral or political bias. It isn't a concern, just a potential cognitive process.

Thus, the outcome to be explained is why are employers making discrimination in pay that related to the gender of the employees? Well, the most parsimonious account is the one that simply presumes only the variables in the question itself, gender and discrimination.

No: that is not the most parsimonious because assuming gender discrimination is a mediating variable.

Wrong yet again. Gender discrimination is not a variable. Where in the regression model is the measured variable gender discrimination? No where. Gender is in the model and discrimination is just a word for the differentiation in pay levels. All models presume that basic cognitive processes of employers give rise to the pay variance, so that isn't a factor in relative parsimony. Gender discrimination is nothing but a label for the model that presumes the variable gender impacts pay discrimination judgments without assuming it is being entirely causally mediated by other variables that are logically relevant to job performance and impact pay discrimination judgments just as much when they vary within as between gender. Gender discrimination is analogous to death-by-cancer, which isn't a variable by label to refer to the variable of having cancer causing the variable of death.

The employer is factoring gender into their process of discriminating between pay levels for various employees.

The employer is doing no such thing. You are assuming facts not in evidence. Factoring in seniority is not factoring in gender. Factoring in sales volumes is not factoring in gender.

Aarg!!! I am explicating what the theory of gender discrimination is, not what is proven to be true. And you are assuming that employers factor these things in regardless of gender. You actually have very little valid evidence demonstrating the employer is doing that most of the time. IT is just an unquestioned assumption in your implausibly rational model of employer decision making.
The question is when gender varies are employers only considering those non-gender factors and doing so in an identical fashion no matter the gender of the applicant? You do not have evidence that is the case. Part of the problem is that all the data you try to use is incapable of answering that question because the variables of hours, time on job, etc. are only collected after numerous pay decisions have already been made. Thus, you cannot infer that the covariance in those other factors is the cause rather than the effect of pay decisions, or a mixture of both (which is what the most reasonable assumption consistent with more general psychological principles would suggest).

In sum, discrimination is the fat more parsimonious explanation for pay differences for the same reasons that it would be a far less parsimonious explanation for height differences.

First, it isn't more parsimonious, as I've already explained. You are proposing a mediating effect of gender discrimination and it's the gender discrimination that causes the pay difference.

First, wrong again. Gender discrimination is not a mediator because it is not even a variable. It is a label for the relation of gender on pay decisions when not mediated by all the extra assumed factors you claim are responsible for why gender correlates with pay.

Second, if you believe paying a toilet cleaner who works twenty hours a week is 'discriminating' against cleaners who work ten hours a week, you are deliberately conflating the kind of discrimination we ought to be concerned with (paying someone less merely because of their gender) and the kind of discrimination we ought not be concerned with (paying someone less because they produced less work).

Second, you are the only one conflating meanings by injecting unscientific concepts into the discussion. The primary and scientifically relevant meaning to the current objective empirical question is employers discriminating between employees to determine variable pay. The question is what factors impact those discrimination judgments and why is the variable of gender correlated with those judgments. Gender discrimination, absent the emotional/moral evaluations you are trying to make about them, is merely the idea that the correlation is partly the result of gender (a variable that nearly every employer is aware of when making these judgment) impacts those judgments without assuming it is completely mediated by job-relevant factors that impact the judgments the same way no matter the gender of the employees.

And yet it is just as theoretically plausible as your alternative and at least as parsimonious that your account. Also, that is not data directly relevant to the gender pay gap.

Of course it is relevant. How could you imagine something that directly affects labour market attachment as irrelevant?

It is not empirically relevant to the variable of gender and therefore not to the phenomena to be explained which is how gender related to pay differences. And again, it is equally explicable by the same assumptions underlying the discrimination model as it is by your assumed model. Your dismissal of it for your model is purely ideological and not empirical or based in established psychological theory.

For example, relevant data would be why do gender pay gaps at time 1 predict the amount of child care related employment reduction for both parents at time 2? The timing rules out the possibility that these employment reductions are the cause of the pay difference, and support the reverse causality that pay differences impact employment reduction choices.

You are simply pushing back the locus of explanation. Why assume the gender pay gap at time 1 is due to discrimination?

I don't assume it is gender discrimination, I just recognize that it contradicts your assumption that such choices are only causes that precede and impact pay differentials and cannot be effects that result from pay differentials. This is your assumption that your entire interpretation of the OP and the other results require. Without it, nothing meaningful regarding the gender discrimination model can be inferred from the fact that the relationship for gender reduced when employment reductions variables are entered in the model.

IT is no more a "just-so" explanation than all the non-discrimination accounts. In fact, there is massive evidence, including from controlled experiments) that female attractiveness causally impacts job-related treatment by others including compensation, and that perceived attractiveness declines from 18 to average childbearing years.
This is more of an empirical fact than any a data you have or could offer to support your non-gender account. Incorporating established facts is not adding assumptions but increasing coherence with other knowns.

Is it not an established fact that seniority in a role is related to pay?

Actually, a real causal impact of seniority has little more established than gender's role in determining pay. IT is mostly an unquestioned and untested assumption based on anecdote. Show me the data that seniority causes variance in pay independent of all other variables that seniority is in any way related to (i.e, the same standard you require for evidence of gender impacts).

Also, gender discrimination easily incorporates the relationship between seniority and pay. People that get higher pay relative to their peers have more motive to stick around, and gender is merely one thing that impacts pay and thus seniority.

 

[ronburgundy]

Wrong. The gender pay gap is an empirical fact and the OP article confirms it. The only question is why the gap exists.
The claim that it does not much exist is a purely an inferential, assumption-ridden conclusion reached by woefully inadequate efforts to explain away the empirical reality by abuse of inferential statistical techniques. Note, "inferential" is not my personal view of their technique, it is the standard definitional term for what they are doing.

It exists only in the use of a flawed yardstick. When you control for the obvious, easy to measure things 90% of the gap disappears. I find it very unlikely that the remaining gap isn't due to things that aren't so easy to measure. (For example: Willingness to take extended business trips--fathers are much more likely to do so than mothers.)

No. You, Metaphor, dismal (and the OP) are misusing and misunderstanding the severely limited and flawed yardstick that is multiple regression analyses.
Read my reply to Metaphor, but I will summarize a critical point here, because its critical and worth stating with multiple phrasings.

Such regression analyses are purely correlational and do as much to mislead as to clarify the nature of simple bivariate correlations. They do not allow you to make any of the inferences you are making and say nothing about whether the gender discrepancy is causal, and they do not in fact "control" variables.
The only way to actually control variables in any way relevant to causal questions is to very literally and physically via experimental methods manipulate the variance in specific variables independent of all the others by random assignment to different manipulated conditions.
Statistic "control" of variables is an unfortunate misnomer that is not real control and does not at all mean the same things or allow for similar inferences. The results of regression are NOT data! The results are assumption-laden inferential predictions of what the correlations between each predictor and the outcome would roughly estimated (with lots of error) to look like in a fictional world where whatever causal relationships that give rise the pattern of covariance among all the variables did not exist resulting in the various predictors no longer being correlated with each other. It is NOT comparing actual men and women with the identical status on all other variables in the model. Such men and women likely do not exist and whether they do is irrelevant to the underlying math of how the model arrives at its hypothetical estimates. The same results being reported in the OP tell you what non-existent people who are neither men nor women are predicted to get paid, and what people not born yet or 1000 years old are predicted to get paid, and what people with absurd non-possible values on every other variable would be predicted to get paid. This should clue you in that the results are not reality, they are not data, they do not reflect what is true about the actual observations or anyone in the sample.

The results are partialed into shared and unshared variance, and their is no reliable connection between those categories and causal patterns. The gap does not "disappear", but rather that gap has 90% shared variance with other variables, and the same is true of those other variables. The seeming effect of time on job partly "disappears" when you control for gender, and using the same incorrect phrasing "Women who work less than men, wouldn't lose as much pay when working less if they were men." How you phrase which variable is accounting for the others relationship is purely a matter of assumption and in this case ideology.

In contrast, the real observed data is that within the vast majority of job-fields that have been measured, female employees make notably less than men. That is not inferred. It is a simple empirical reality. IT is also an empirical reality that people with more seniority have higher salaries. When you don't have an true experiment, real data exists at the simple two-variable level, and as soon as you start putting multiple correlated variables into the same analysis you leave the world of data and what was observed and enter a purely hypothetical universe of presumption. That doesn't mean it should be done, just that you have to never loose sight that it is a purely hypothetical world and will rarely tell you anything more than you already can infer from just looking at the pattern of how each variable is correlated with each other one in the set (which usually leaves most causal questions unanswered).

BTW, what we are talking about here and the lack of understanding you guys are showing is a rampant problem in the social sciences who rely on such correlational data in making their casual claims. It is what make the soft science so soft, so ripe for bias, and so slow to progress. It also why professional mathematician who really and deeply grasp the mathematics and logic of that tests have so little regard for the social sciences and how these tests are abused to both assert and to deny causality, when the results are almost always near agnostic about such issues.

 

[ronburgundy]

Metaphor, Loren, dismal, and all other's who think their is any validity to the OP,

Before you respond to my latest replies to your individual posts, please address this simple hypothetical scenario because it gets the the very heart of my disagreement with most of what each of you has said and with research like in the OP.

A researcher collects data on 3 variables of smoking, lung cancer, and medical bills. Put away all of your a priori assumptions of causality. The issue is what is actually demonstrated by the empirical data and the results of regression. All 3 variables show simple correlations with the other 2. So, a regression analysis uses both smoking and medical bills as predictors of cancer (again, stop your a piori causal assumptions, this about data).
The result (and this would very likely happen) shows that 90% of the relationship between smoking and lung cancer in only 1/3 as large as it is when medical bills are not "controlled" for.

What do you conclude about the likely causal impact of smoking on lung cancer?
Does this mean that true difference in lung cancer rates between smokers and non smokers is only 1/3 as big as thought?
Do you now have evidence that most of the relationship is really caused by medical bills?

 

[Metaphor]

Metaphor, Loren, dismal, and all other's who think their is any validity to the OP,

Before you respond to my latest replies to your individual posts, please address this simple hypothetical scenario because it gets the the very heart of my disagreement with most of what each of you has said and with research like in the OP.

A researcher collects data on 3 variables of smoking, lung cancer, and medical bills. Put away all of your a priori assumptions of causality. The issue is what is actually demonstrated by the empirical data and the results of regression. All 3 variables show simple correlations with the other 2. So, a regression analysis uses both smoking and medical bills as predictors of cancer (again, stop your a piori causal assumptions, this about data).
The result (and this would very likely happen) shows that 90% of the relationship between smoking and lung cancer in only 1/3 as large as it is when medical bills are not "controlled" for.

What do you conclude about the likely causal impact of smoking on lung cancer?
Does this mean that true difference in lung cancer rates between smokers and non smokers is only 1/3 as big as thought?
Do you now have evidence that most of the relationship is really caused by medical bills?

This doesn't get to the heart of anything because you've made assumptions that I've made assumptions that I haven't made.

I understand multiple regression in the social sciences. I've taught it to university undergraduates. I understand what shared variability is, exogenous and endogenous variables and reciprocal causation. I understand that a correlation of .3 between x and y means x and y 'explain' the same amount of variability in the other variable, and that 'explain' is a term of convenience.

 

[Metaphor]

It doesn't fail. We need more data to establish with confidence what the causal paths are, but there are also a priori probabilities in which parsimony factors and favors a direct impact of gender on decisions made by people who are well aware of the gender of the people they are deciding about.

Okay, sure. The direct impact is that women are less competent workers and the employers know it. Gender gap explained.

Incorrect. Gender Discrimination is not a variable, it is a label for a causal relationship between the variable of gender and cognitive acts where a person is making a discrimination/distinction between options.

Okay, sure. Employers are gender discriminating because women are less competent and employers know it.

Deciding to pay person X more than Y is inherently and by definition an act of discrimination which is merely "to recognize a distinction". Thus, all variability in pay results from a person discriminating between employees and which deserve what pay. All accounts of the average pay difference between gender and pay must include both gender and the cogntive act of discrimination (unless you assume all employers make pay decisions at complete random).

Your insistence of using the broad use of the term 'discrimination', which means choosing one option over another in the context of a discussion about the gender pay gap (where 'discrimination' is used to mean unfair and prejudiced allocation of wages and salaries by employers based on gender) needs to stop.

What you are actually adding is assumptions to your model. Calling them control variables requires the assumption that they have only on particular possible causal relation to the other variables, namely that they cause the outcome (pay), and are impacted by gender in some way. IOW, that they are mediators that are not themselves impacted by the outcome variable or other variables correlated with the outcome which are in turn caused by main predictor of interests (gender).

With pay, you assume that effort and time at a job can only possibly be the cause of pay, whereas I consider the actual relevant evidence that they both casually impact each other, and thus a large portion of their shared variance reflects variance caused by other factors such as gender on pay that then impact effort and time at a job. Thus, it is false to treat all that shared variance as though effort and time at work get to claim causal credit for it.

You're right. Women are less competent than men and employers know it. This is consistent with the model.

Unless other patterns of causality can be ruled out a priori as impossible on principle, the reduction in predictiveness of a variable after entering other variables cannot be interpreted as evidence against the importance or causal impact of that variable on the outcome or as evidence that the other variables are the real cause. It can only be interpreted as evidence that whatever the relationship of the focus variable to the outcome, it also has a shared relationship to the additional variables added. The data is agnostic on what the nature of those relationships are.

I know this. Please stop repeating it. I know that a set of shared covariances in columns in SPSS cannot tell you the direction of the causal arrow.

No, you are pointing to your purely faith based assumption that it is unrelated to employer decisions, an assumption made extremely implausible by its contradiction to experimentally demonstrated impacts on interpersonal judgments and preferences of people in general (FYI: employers are people) and specifically job related judgments.

No, I agree. The data are consistent with employers knowing that women are less competent and they therefore pay them less.

Each of those require the same or more number of steps as the more direct effect of gender, just to get from gender those variables, then they require at least as many steps again to get from those variables to pay decisions. You see a guy with a bloody fist standing over a guy with a bloody nose. Is is more plausible he punched the guy or that he punched someone else who then fell into that guy? (hint: your theory is the latter).

No, I agree. The data are consistent with employers knowing that women are less competent and they therefore pay them less.

No. it is to clarify that actual variables. You are muddying by bring issues of morality and politics and "concerns" into it. Ignore at that. Pretend you are being objective and not trying to abuse science to support an political conclusion. Gender discrimination might be politically negative, but it is also just a potential scientific process in which an employers act of differentiating things (pay levels for employees) is impacted by the variable gender of the employees. That is all gender discrimination is on a scientific level, without any moral or political bias. It isn't a concern, just a potential cognitive process.

No, I agree. The data are consistent with employers knowing that women are less competent and they therefore pay them less.

Wrong yet again. Gender discrimination is not a variable. Where in the regression model is the measured variable gender discrimination? No where. Gender is in the model and discrimination is just a word for the differentiation in pay levels.

No, I agree. The data are consistent with employers knowing that women are less competent and they therefore pay them less.

The question is when gender varies are employers only considering those non-gender factors and doing so in an identical fashion no matter the gender of the applicant? You do not have evidence that is the case

There is, of course, evidence that non-gender factors are considered in identical fashion. For example, regressing pay on seniority produces the same slope for men and women.

But I have a different concern with the OP than you in its use of multiple regression. An ordinary least squares regression can account for only linear relationships between variables, and pay is not normally distributed and right-skewed.

. Part of the problem is that all the data you try to use is incapable of answering that question because the variables of hours, time on job, etc. are only collected after numerous pay decisions have already been made. Thus, you cannot infer that the covariance in those other factors is the cause rather than the effect of pay decisions, or a mixture of both (which is what the most reasonable assumption consistent with more general psychological principles would suggest).

I recognise that and I've never denied it.

First, wrong again. Gender discrimination is not a mediator because it is not even a variable. It is a label for the relation of gender on pay decisions when not mediated by all the extra assumed factors you claim are responsible for why gender correlates with pay.

Take the top 100 mens and top 100 womens finishing times for the 100m sprint over the modern Olympic games. Label the relation of finishing time difference when not mediated by all the extra assumed factors you claim are responsible for why gender correlates with 100m sprint times as 'gender discrimination'.

It is not empirically relevant to the variable of gender

You know that how?

Actually, a real causal impact of seniority has little more established than gender's role in determining pay. IT is mostly an unquestioned and untested assumption based on anecdote. Show me the data that seniority causes variance in pay independent of all other variables that seniority is in any way related to (i.e, the same standard you require for evidence of gender impacts).

Do you mean like in the jobs of academics, where remaining alive and employed is sufficient to advance to the next 'step' on the pay scale?

 

[dismal]

Metaphor, Loren, dismal, and all other's who think their is any validity to the OP,

Before you respond to my latest replies to your individual posts, please address this simple hypothetical scenario because it gets the the very heart of my disagreement with most of what each of you has said and with research like in the OP.

A researcher collects data on 3 variables of smoking, lung cancer, and medical bills. Put away all of your a priori assumptions of causality. The issue is what is actually demonstrated by the empirical data and the results of regression. All 3 variables show simple correlations with the other 2. So, a regression analysis uses both smoking and medical bills as predictors of cancer (again, stop your a piori causal assumptions, this about data).
The result (and this would very likely happen) shows that 90% of the relationship between smoking and lung cancer in only 1/3 as large as it is when medical bills are not "controlled" for.

What do you conclude about the likely causal impact of smoking on lung cancer?
Does this mean that true difference in lung cancer rates between smokers and non smokers is only 1/3 as big as thought?
Do you now have evidence that most of the relationship is really caused by medical bills?

This doesn't get to the heart of anything because you've made assumptions that I've made assumptions that I haven't made.

I understand multiple regression in the social sciences. I've taught it to university undergraduates. I understand what shared variability is, exogenous and endogenous variables and reciprocal causation. I understand that a correlation of .3 between x and y means x and y 'explain' the same amount of variability in the other variable, and that 'explain' is a term of convenience.

We can argue about whether working fewer hours cause women to be paid less, or whether being paid less causes women to work fewer hours, I suppose. (One of those would seem to me make more sense than the other given the presence of hourly pay schemes and the fact it's generally lack of money that tends to cause people to want to work, but hey, argue away...)

But what we can't do is argue women do equal work when there is evidence women work less.

 

[JonA]

Sorry, Ron, but you're clearly trying very hard to miss the point of this thread or see the significance of the research presented in the OP's article.

Too tied up in some weird word game it seems.

Whatever the case, I personally haven't the time to set you straight. Best of luck; I hope others are interested in a discussion on this topic because I find it interesting.

 

[Loren Pechtel]

It exists only in the use of a flawed yardstick. When you control for the obvious, easy to measure things 90% of the gap disappears. I find it very unlikely that the remaining gap isn't due to things that aren't so easy to measure. (For example: Willingness to take extended business trips--fathers are much more likely to do so than mothers.)

No. You, Metaphor, dismal (and the OP) are misusing and misunderstanding the severely limited and flawed yardstick that is multiple regression analyses.
Read my reply to Metaphor, but I will summarize a critical point here, because its critical and worth stating with multiple phrasings.

Such regression analyses are purely correlational and do as much to mislead as to clarify the nature of simple bivariate correlations. They do not allow you to make any of the inferences you are making and say nothing about whether the gender discrepancy is causal, and they do not in fact "control" variables.
The only way to actually control variables in any way relevant to causal questions is to very literally and physically via experimental methods manipulate the variance in specific variables independent of all the others by random assignment to different manipulated conditions.
Statistic "control" of variables is an unfortunate misnomer that is not real control and does not at all mean the same things or allow for similar inferences. The results of regression are NOT data! The results are assumption-laden inferential predictions of what the correlations between each predictor and the outcome would roughly estimated (with lots of error) to look like in a fictional world where whatever causal relationships that give rise the pattern of covariance among all the variables did not exist resulting in the various predictors no longer being correlated with each other. It is NOT comparing actual men and women with the identical status on all other variables in the model. Such men and women likely do not exist and whether they do is irrelevant to the underlying math of how the model arrives at its hypothetical estimates. The same results being reported in the OP tell you what non-existent people who are neither men nor women are predicted to get paid, and what people not born yet or 1000 years old are predicted to get paid, and what people with absurd non-possible values on every other variable would be predicted to get paid. This should clue you in that the results are not reality, they are not data, they do not reflect what is true about the actual observations or anyone in the sample.

The results are partialed into shared and unshared variance, and their is no reliable connection between those categories and causal patterns. The gap does not "disappear", but rather that gap has 90% shared variance with other variables, and the same is true of those other variables. The seeming effect of time on job partly "disappears" when you control for gender, and using the same incorrect phrasing "Women who work less than men, wouldn't lose as much pay when working less if they were men." How you phrase which variable is accounting for the others relationship is purely a matter of assumption and in this case ideology.

In contrast, the real observed data is that within the vast majority of job-fields that have been measured, female employees make notably less than men. That is not inferred. It is a simple empirical reality. IT is also an empirical reality that people with more seniority have higher salaries. When you don't have an true experiment, real data exists at the simple two-variable level, and as soon as you start putting multiple correlated variables into the same analysis you leave the world of data and what was observed and enter a purely hypothetical universe of presumption. That doesn't mean it should be done, just that you have to never loose sight that it is a purely hypothetical world and will rarely tell you anything more than you already can infer from just looking at the pattern of how each variable is correlated with each other one in the set (which usually leaves most causal questions unanswered).

BTW, what we are talking about here and the lack of understanding you guys are showing is a rampant problem in the social sciences who rely on such correlational data in making their casual claims. It is what make the soft science so soft, so ripe for bias, and so slow to progress. It also why professional mathematician who really and deeply grasp the mathematics and logic of that tests have so little regard for the social sciences and how these tests are abused to both assert and to deny causality, when the results are almost always near agnostic about such issues.

Alternately, lower wages come from lower work hours and you have no basis for figuring that gender has anything to do with it. Your argument works both ways.

In a situation like this where a controlled experiment can't be done all you can do look at what factors remain predictive when you fix the other ones.

By this standard gender isn't very predictive and thus isn't a likely causative factor.

 

[Don2 (Don1 Revised)]

Interesting study abstract:

Prior research has suggested that gender differences in physicians’ salaries can be accounted for by the tendency of women to enter primary care fields and work fewer hours. However, in examining starting salaries by gender of physicians leaving residency programs in New York State during 1999–2008, we found a significant gender gap that cannot be explained by specialty choice, practice setting, work hours, or other characteristics. The unexplained trend toward diverging salaries appears to be a recent development that is growing over time. In 2008, male physicians newly trained in New York State made on average $16,819 more than newly trained female physicians, compared to a $3,600 difference in 1999.

 

[Kusa]

Women that can't move up due to their job proficiency usually have alternative methods of moving up.

 

[ronburgundy]

Women that can't move up due to their job proficiency usually have alternative methods of moving up.

You do realize how much this actually supports the claim of gender discrimination against women, right?

The same men who advance a women for giving him a blowjob, are likely to refuse to advance them unless they give him a blowjob, no matter how proficient they are at their job.

 

[Don2 (Don1 Revised)]

This is a clear case of reverse discrimination because men can do the jobs, too.

 

[ronburgundy]

No. You, Metaphor, dismal (and the OP) are misusing and misunderstanding the severely limited and flawed yardstick that is multiple regression analyses.
Read my reply to Metaphor, but I will summarize a critical point here, because its critical and worth stating with multiple phrasings.

Such regression analyses are purely correlational and do as much to mislead as to clarify the nature of simple bivariate correlations. They do not allow you to make any of the inferences you are making and say nothing about whether the gender discrepancy is causal, and they do not in fact "control" variables.
The only way to actually control variables in any way relevant to causal questions is to very literally and physically via experimental methods manipulate the variance in specific variables independent of all the others by random assignment to different manipulated conditions.
Statistic "control" of variables is an unfortunate misnomer that is not real control and does not at all mean the same things or allow for similar inferences. The results of regression are NOT data! The results are assumption-laden inferential predictions of what the correlations between each predictor and the outcome would roughly estimated (with lots of error) to look like in a fictional world where whatever causal relationships that give rise the pattern of covariance among all the variables did not exist resulting in the various predictors no longer being correlated with each other. It is NOT comparing actual men and women with the identical status on all other variables in the model. Such men and women likely do not exist and whether they do is irrelevant to the underlying math of how the model arrives at its hypothetical estimates. The same results being reported in the OP tell you what non-existent people who are neither men nor women are predicted to get paid, and what people not born yet or 1000 years old are predicted to get paid, and what people with absurd non-possible values on every other variable would be predicted to get paid. This should clue you in that the results are not reality, they are not data, they do not reflect what is true about the actual observations or anyone in the sample.

The results are partialed into shared and unshared variance, and their is no reliable connection between those categories and causal patterns. The gap does not "disappear", but rather that gap has 90% shared variance with other variables, and the same is true of those other variables. The seeming effect of time on job partly "disappears" when you control for gender, and using the same incorrect phrasing "Women who work less than men, wouldn't lose as much pay when working less if they were men." How you phrase which variable is accounting for the others relationship is purely a matter of assumption and in this case ideology.

In contrast, the real observed data is that within the vast majority of job-fields that have been measured, female employees make notably less than men. That is not inferred. It is a simple empirical reality. IT is also an empirical reality that people with more seniority have higher salaries. When you don't have an true experiment, real data exists at the simple two-variable level, and as soon as you start putting multiple correlated variables into the same analysis you leave the world of data and what was observed and enter a purely hypothetical universe of presumption. That doesn't mean it should be done, just that you have to never loose sight that it is a purely hypothetical world and will rarely tell you anything more than you already can infer from just looking at the pattern of how each variable is correlated with each other one in the set (which usually leaves most causal questions unanswered).

BTW, what we are talking about here and the lack of understanding you guys are showing is a rampant problem in the social sciences who rely on such correlational data in making their casual claims. It is what make the soft science so soft, so ripe for bias, and so slow to progress. It also why professional mathematician who really and deeply grasp the mathematics and logic of that tests have so little regard for the social sciences and how these tests are abused to both assert and to deny causality, when the results are almost always near agnostic about such issues.

Alternately, lower wages come from lower work hours and you have no basis for figuring that gender has anything to do with it. Your argument works both ways.

Yes, it works both ways. That has been my entire point. Which is why all the results in the OP are meaningless with regard to the existence or magnitude of a causal impact of gender on pay decisions.
I am not drawing causal conclusions from the results, I am saying that no valid conclusions can be drawn. It is you, dismal, Metaphor, and the OP who are the one's making causal assertions based on the OPs causally uninterpretable correlational data.

In a situation like this where a controlled experiment can't be done all you can do look at what factors remain predictive when you fix the other ones.

You can look at that all you want, you just cannot draw any rational scientific conclusions about causality, which is what you are doing.
That is the problem. Researchers is such areas who have no valid data that speaks to causality, take the best tools they have and pretend they can determine causality, because no one would give a shit about what they are doing if they can't say anything about causality. This is why this kind of abuse of these methods is so pervasive as to be standard practice in fields like economics. They let each other get away with such pseudo-science tripe, because their social relevance of their field depends on claims that go well beyond what their very limited data allows.

By this standard gender isn't very predictive and thus isn't a likely causative factor.

Completely wrong and utterly unsupported inference from the data for all the reasons I explained. The probability that it is a causal factor is not reduced by the results, because the results are as consistent with a model in which it is causal and with any other model. The results are a wash and tell us nothing useful beyond what the simple correlation between gender and pay told us.
Each of the predictors, including gender and work hours, have their relation to pay reduced when all the other predictors are in the model. All that tells us is that whatever the various causal relations are, they lead to lots of shared variance between gender, pay, and the various employment reduction variables.
That does not count against the causality of any of the variables. The only thing we know now about the causality is what we already knew before this data was even collected, which is that gender is a cause of at least some and potentially all factors in the model, because a priori knowledge largely rules it out as being an effect of any of them, or being an effect of any plausible unmeasured variable that is the cause of the others.

 

[ronburgundy]

Sorry, Ron, but you're clearly trying very hard to miss the point of this thread or see the significance of the research presented in the OP's article.

Too tied up in some weird word game it seems.

Whatever the case, I personally haven't the time to set you straight. Best of luck; I hope others are interested in a discussion on this topic because I find it interesting.

No, I am trying hard to actually apply my expertise in Multi-variate analytic techniques to show why all efforts in the thread to draw conclusions about the existence of gender discrimination from the OP research are scientifically invalid nonsense.

I realize that to someone with no grasp of what such analyses really are, my analyses may sound like "word games", but that is your problem. These are inherently complex issues and it isn't my job to make up for your lack of education.

 

[Kusa]

Women that can't move up due to their job proficiency usually have alternative methods of moving up.

You do realize how much this actually supports the claim of gender discrimination against women, right?

The same men who advance a women for giving him a blowjob, are likely to refuse to advance them unless they give him a blowjob, no matter how proficient they are at their job.

No no no. This is an example of how women have more opportunities then men. How often does a man get a promotion for eating some piss smelling twat?

 

[ronburgundy]

Metaphor, Loren, dismal, and all other's who think their is any validity to the OP,

Before you respond to my latest replies to your individual posts, please address this simple hypothetical scenario because it gets the the very heart of my disagreement with most of what each of you has said and with research like in the OP.

A researcher collects data on 3 variables of smoking, lung cancer, and medical bills. Put away all of your a priori assumptions of causality. The issue is what is actually demonstrated by the empirical data and the results of regression. All 3 variables show simple correlations with the other 2. So, a regression analysis uses both smoking and medical bills as predictors of cancer (again, stop your a piori causal assumptions, this about data).
The result (and this would very likely happen) shows that 90% of the relationship between smoking and lung cancer in only 1/3 as large as it is when medical bills are not "controlled" for.

What do you conclude about the likely causal impact of smoking on lung cancer?
Does this mean that true difference in lung cancer rates between smokers and non smokers is only 1/3 as big as thought?
Do you now have evidence that most of the relationship is really caused by medical bills?

This doesn't get to the heart of anything because you've made assumptions that I've made assumptions that I haven't made.

I understand multiple regression in the social sciences. I've taught it to university undergraduates. I understand what shared variability is, exogenous and endogenous variables and reciprocal causation. I understand that a correlation of .3 between x and y means x and y 'explain' the same amount of variability in the other variable, and that 'explain' is a term of convenience.

If you understand all that, then you understand that the OP analysis have no implications for whether gender is a causal factor on pay.
IF you don't realize that, then your ideological faith about this issue is undermining your ability to apply your own knowledge to understanding what the data actually imply and what they do not.

IF you understand what you claim then you know that all of the claims made about the OP results showing evidence against gender discrimination in this thread are identical to claiming that my example results "show evidence against smoking causing cancer", or that the smoking effect "disappers once you control for medical bills", or that "the lung cancer gap between smokers and non-smokers is not that big".

BTW, I teach multi-variate analyses and causal path-modelling in the social sciences to graduate students, and I use these methods in my work and review research using them. I know that most people who use them and many who teach them don't know what the hell they are doing and misuse and misrepresent what their results actually allow in terms of inferences. I know that anyone who makes or agrees with the thread title or any claims that the results support the "control" variables as being the real causes rather than gender has only just enough superficial knowledge of these methods to be able to abuse them.

 

[laughing dog]

No, I am trying hard to actually apply my expertise in Multi-variate analytic techniques to show why all efforts in the thread to draw conclusions about the existence of gender discrimination from the OP research are scientifically invalid nonsense.

I realize that to someone with no grasp of what such analyses really are, my analyses may sound like "word games", but that is your problem. These are inherently complex issues and it isn't my job to make up for your lack of education.

If you want people to be convinced, it is up to you to overcome their lack of education. Around here, that is pretty much a herculean task in most cases.

 

[Loren Pechtel]

Alternately, lower wages come from lower work hours and you have no basis for figuring that gender has anything to do with it. Your argument works both ways.

Yes, it works both ways. That has been my entire point. Which is why all the results in the OP are meaningless with regard to the existence or magnitude of a causal impact of gender on pay decisions.
I am not drawing causal conclusions from the results, I am saying that no valid conclusions can be drawn. It is you, dismal, Metaphor, and the OP who are the one's making causal assertions based on the OPs causally uninterpretable correlational data.

So you're admitting there's no basis for saying women are being underpaid?

In a situation like this where a controlled experiment can't be done all you can do look at what factors remain predictive when you fix the other ones.

You can look at that all you want, you just cannot draw any rational scientific conclusions about causality, which is what you are doing.
That is the problem. Researchers is such areas who have no valid data that speaks to causality, take the best tools they have and pretend they can determine causality, because no one would give a shit about what they are doing if they can't say anything about causality. This is why this kind of abuse of these methods is so pervasive as to be standard practice in fields like economics. They let each other get away with such pseudo-science tripe, because their social relevance of their field depends on claims that go well beyond what their very limited data allows.

Then you can't learn much of anything about things you can't bring into the lab.

By this standard gender isn't very predictive and thus isn't a likely causative factor.

Completely wrong and utterly unsupported inference from the data for all the reasons I explained. The probability that it is a causal factor is not reduced by the results, because the results are as consistent with a model in which it is causal and with any other model. The results are a wash and tell us nothing useful beyond what the simple correlation between gender and pay told us.

You have accepted that your argument works both ways--that means that there is at most a 50% chance that gender is relevant. Before publishing a positive conclusion you normally need a minimum of a 95% confidence in the result. Since we have only 50:50 the error margin includes the null and that automatically makes it a "no influence shown" result.

Thank you for proving that gender is irrelevant.

 

[coloradoatheist]

Isn't the problem is that we are relying on an argument from authority and we have to assume that the people who adjusted the values did it right? Are the numbers and the specific tables available for anybody to use?