Model

IN ADDITION TO THE DEMOGRAPHIC, company/team, tasks, and tools variables used to create the linear model, we include two additional variables: bargaining skills and employability. In both cases, respondents are prompted to self-evaluate their skill based on a 1 (poor) to 5 (excellent) scale. We know the results are likely subjective and imprecise (for example, bargaining skills are difficult to quantify and likely inconsistent) but we think the results and model would suffer if we ignore the topic. It seems fairly clear, for example, that if someone routinely accepts lowball offers without putting up any fight, they will earn less than someone who bargains more aggressively (even if they are just as talented, experienced, and hardworking as the latter). Most respondents (66%) self-reported a bargaining score of 3 or 4, with 12% reporting a score of 5. Unsurprisingly, each bargaining “point” corresponded to a higher salary.

The other question, also on a five-point scale, was the ease of finding a new position, assuming “that the new job is more or less equivalent to your current one, in terms of compensation, workload, and your interest in the work.” Again, this is highly hypothetical and subjective, but is similarly hard to ignore.

If someone knows (and crucially, if their employer knows) that they could walk out today from their at-will employment and find equivalent work elsewhere, the employer is likely to be willing to pay more. In this sense, ease of finding work is a predictor of salary, so we introduced it into the model. However, it should be noted that this variable could also be considered a dependent variable that we could try to predict with other variables (i.e., alongside salary, not “below” it).

Now, for the model.¹ From the 65 questions, we extracted 202 variables (multiple-choice questions usually have almost as many variables as there are answer choices), and from the available variables, the algorithm selected the most significant. The algorithm favors a model with a smaller number of variables (to keep the model simpler) and assigns a coefficient (a positive or negative number) to each. To get your salary estimate, start with the constant ($51K), and for each variable that applies to you add (or subtract) the corresponding coefficient (or for years of experience, add the coefficient times the number of years you have worked in your field). Here is the model:

constant (everyone starts with this amount):
+$51,617

country = United States: +$12,257
world region = Europe (except UK/I): –$12,561
US region = California: +$17,390
US region = Northeast: +$2,956
gender = female: –$6,363
per 1 year of experience: +$1,292
academic speciality in mathematics, statistics or
physics: +$1,158
academic specialty in graphic design: –$5,178

title = “Graphic Designer”: –$1,848

industry = education: –$7,522
company size = 1: +$6,403
company size > 1,000: +$1,543
team size = 2: –$6,052
team size = 8 to 20: +$6,661
team size = More than 20: +$12,224
works with product managers: +$7,021
works with programmers: +$3,633
works on services: -$896
tasks: brainstorming = no involvement or minor involvement:
+$1,039
tasks: managing people = major involvement:
+$18,409
tasks: programming = I never perform this task:
–$16,682
tasks: pitching = major involvement: +$3,669
tasks: user research = no involvement: –$1,723
meetings: 9 - 20 hours / week: +$14,168
meetings: over 20 hours / week: +$15,547
work week < 30 hours / week: –$20,963
work week = 30 to 35 hours / week: –$16,834
work week > 55 hours / week: +$23,472

tools: Slack: +$6,378
tools: User Zoom: +$5,403
tools: PHP: +$5,147
tools: Alexa: +$3,661
tools: Keynote: +$3,198
tools: Dropbox: +$1,434
tools: Windows: –$3,740

bargaining skills = 1 or 2: –$1,259
bargaining skills = 4: +$3,288
bargaining skills = 5: +$4,164

ease of finding new work = 1, 2 or 3: –$5,444
 

This model gives us a more rigorous assessment of how much variables predict salary, allowing us to revisit some of the differences in salary between various groups of respondents. Since the model only includes a subset of the available variables, it shows us which ones are likely to have true correlations with salary and which are more likely to be insignificant (even if they might correlate with the significant ones).

The gap in salary between the US and Europe is reduced to +$25K (UK/Ireland and Canada are in between), but the coefficient for California is still quite large at +$17K. Aside from this and the +$3K value for the Northeast (from NY/NJ), no US region was significant enough to be kept in the model (so the large salaries in the Mid-Atlantic appear to have been somewhat of an accident).

The gender gap was reduced to –$6,403, although this doesn’t really make the previously observed $14K difference any better. In some ways, it makes it worse since the model is telling us that if we keep everything else constant—experience, skills, role—a woman will likely earn $6K less than a man.

One variable that is conspicuously absent is age, which seems to be completely replaced by other variables, most notably years of experience. For each year of experience, UX designers are predicted to gain $1,292: a rather modest raise, although this would be on top of expected salary changes from other variables (e.g., new tools or different tasks).

Education remained an important feature, the academic backgrounds of math/stats/physics giving a +$1,158 increase and graphic design a –$5,178 decrease. The further penalty of –$1,848 for “Graphic Designer” as a title means that if there are two people who both do graphic design, one called “Graphic Designer” and one called something else, the former will be expected to earn almost $2K less than the latter. The lack of other nonzero coefficients for title variables means that with the exception of “graphic designer,” salary is not based on what you are called but rather what you do.

For industry, company size, and team size there were few differences suggested by the model from the single-variable salary comparisons observed above. The only industry variable kept in the model is education (–$7,522); hardware/computers did not make it into the model. The boost in salary for large company employees is (only) +$1,543, while at the other end of the company size scale, self-employed respondents (i.e., company size = 1) had a large, positive coefficient of +$12,455.

The correlation between team size and salary that we noted earlier is made clear in the model: those with a team size of at least 8 earned $6,661 more than smaller teams; those with a team larger than 20 people earned an additional $5,263; and teams of only 1 or 2 people had a further penalty of –$6,052. The latter penalty also explains part of why the single-personcompany coefficient is so high: the negative effect of having a small team does not apply to them, so the negative, small-team coefficient needs to be cancelled out. Not only did larger teams predict a larger salary, but working with people in certain roles—namely, product managers and programmers— had positive effects on salary (+$7,021 and +$3,633, respectively).

As we would expect, managing people (major involvement) is accompanied with a large boost in salary: +$18,409. Almost as significant, however, is programming: those who do not program at all are expected to earn $16,682 less, according to the model. Major involvement (versus minor) does not change the salary prediction. While this does not mean that if you learn how to program you are guaranteed $16K more, the possibility that there is some causative aspect in this relationship (which seems somewhat intuitive) should be enough to motivate some UX designers to learn how to code (at least a little).

Given the huge discrepancies between median salaries of groups split by meeting hours, it is no surprise to see similarly large coefficients in the model. A ray of hope (for those of us who are not overly fond of spending half the day on the phone) might be found in the fact that while spending at least 9 hours/week in meetings bumps predicted salary up by $14,168, spending more than 20 hours/week gives only an additional $1,379. Workweek hours were also associated with large coefficients: +$23,472 for more than 55 hours; –$16,834 for less than 35 (and a further $4,129 for a workweek less than 30 hours).

None of the major design-specific tools were present in the model, although Slack is (+$6,378) and, as previously noted, it is highly correlated to Sketch. Aside from Keynote (+$3,198) and Dropbox (+$1,434), the only tools with high positive coefficients were tools each used by less than 6% of the sample: User Zoom (+$5,403), PHP (+$5,147), and Alexa (+$3,661). One tool had a negative coefficient: Windows (–$3,740).

The two self-evaluating scores gave the last coefficients. For each point (out of five) for one’s bargaining skills, predicted salary increases by about $1K. (This seems like a low estimate, but then again, it’s an average, and perhaps it doesn’t matter very much for some people in certain situations.) Those who would not rate their ease of finding new work higher than a 3 (out of 5) are penalized –$5,444.

Even after numbers are plugged in, there is still a fair amount of error in the estimation by this model. Among new data points—such as yours, if you are in the mood for a little math—half should be within $20K of the estimate, and three-quarters within $30K, but some larger outliers can be expected. Of course, there are many factors, some impossible to accurately assess in a survey, much less an anonymous one, that will undoubtedly have important effects on salary and throw off the estimate, but on average this model should be more or less correct.