Are wages rising, falling, or stagnating?

A finished steel coil is marked with its information by a worker at the Novolipetsk Steel PAO steel mill in Farrell, Pennsylvania, U.S., March 9, 2018. REUTERS/Aaron Josefczyk - RC1293A04C30

What is really happening to wages in America? Over the past 12 months, average hourly wages rose 3.2 percent, according to the latest jobs report from the Bureau of Labor Statistics. But the longer-term story is contested. Many analysts and commentators lament the situation of stagnating wages, while others celebrate wage growth. To take just two of hundreds of examples, our colleagues in the Hamilton Project here at Brookings report “long-run wage stagnation for lower-wage workers”, while Michael Strain over at AEI writes that “the wages of a typical worker have increased by 32% over the past three decades. That’s a significant increase in purchasing power”. Though we would be remiss if we did not point out that this corresponds to less than a one percent increase per year.

The honest but boring answer to the question of what is happening to wages is: It depends. Specifically, it depends on how you measure it. As so often, methodology really, really, really matters. In the case of wage growth, four analytical decisions bear heavily on the results: which time period, which deflator, which workers (by gender), and which workers (in terms of position). We discuss each of these four below and show how they influence the wage story. [1]

1. Time Period

To measure a trend, you need to select a time interval. Those who find evidence of wage stagnation are typically comparing wages today to those in the 1970s—usually 1973 or 1979. 1973 is a favored starting point because this is when many analysts see a gap opening up between productivity and pay. But 1979 is also frequently used because it is the first year in a good-quality data series. But going back this far, some argue, makes it more difficult to accurately compare the purchasing power of wages. Partly for this reason, other analysts examine wage trends since the 1990s. Many select July 1990—a business cycle peak—as a starting point. But it is important to note that wages were in fact quite low in 1990 (lower than in three out of four years for the whole period from 1979 to 2018).

The choice of starting year therefore makes a big difference to the reported wage trend. Median hourly wages (adjusted using the CPI-R-US deflator) have risen by just three percent since 1979, but by seven percent since 1990.

2. Inflation measure

Comparing wages at different points in time requires adjusting for inflation. Some researchers prefer the Consumer Price Index research series (CPI-U-RS), and their studies typically conclude that wages have been relatively flat. They argue that CPI-U-RS best captures out-of-pocket expenses, the appropriate concept when talking about wages. PCE, in contrast, attempts to capture all spending, most notably for all health care services. This concept is too broad when evaluating wage trends, given that most health care spending flows through insurance rather than out-of-pocket spending.

However, other researchers favor the Personal Consumption Expenditures (PCE) price deflator. They argue that CPI-U-RS overstates inflation by failing to account for “upper-level substitution,” the ability of consumers to switch between products in response to price changes. (Chained CPI does account for this but is only available back to 2000). Analyses using the PCE generally indicate significant wage growth.

The choice of deflator is not trivial. Since 1979 (extra points for those of you who spotted the decision to select that earlier date), hourly median earnings have risen by just three percent using the CPI-U-RS, but by 15 percent using the PCE:

3. Men, women, or both?

Another big factor is gender. There is broad agreement that the wages of men and women have been on opposite trajectories over recent decades (at least in parts of the distribution, but we’ll get to that). Women have seen a wage increase of 25 percent since 1979, reflecting increased educational attainment: in the spirit of this paper, we must report of course that this is median wages, adjusted using CPI-U-RS, since 1979. Men’s earnings, measured on the same basis, have fallen by 10 percent. The declines have been steeper for Black and Hispanic men. (It is important to note in passing that the gender pay gap is shrinking at the median in part because men are earning less than before).

What this means is that the gloomiest stories about wages tend to be those focused on men rather than women, or even all workers—who have seen a wage increase of three percent since 1979:

4. Median, Average, Poorest, All?

Workers at different points in the distribution may see very different trends. It therefore matters a great deal whose wages are tracked. The worst picture over recent decades has usually been found in the middle of the distribution. Many analyses of the wages of median workers, for example, find a trend line that has barely budged over time. A more upbeat story can be told by looking at the tails rather than the middle. Wages at the top of the distribution have been rising rapidly, especially for women. (It hardly needs saying that the selection of average or median is therefore consequential, too).

In order to remain consistent with our earlier analyses, let’s select 1979 as the starting year and CPI-U-RS as a deflator. As we’ve shown above, hourly median wages are up by three percent. But the story for wages at the 20th percentile is twice as good, with a rise of six percent:

Choose your story, choose your method

Just by varying the four decisions outlined here, 24 different wage time trends can be constructed. To dramatize the implications of these selections, in the following graph, we show the “best case” and the “worst case” stories that can be generated making what we consider to be at least defensible choices. If we begin in 1990, use PCE, include women and men, and look at the 20th percentile of wages, we can report that wages grew at a cumulative rate of 23 percent—corresponding to an annual increase of less than one percent. In contrast, if we begin in 1979, use CPI-U-RS, focus on men, and look at the 20th percentile of wages, we see wages decline by 13 percent.

Depending on the methodological choices made, very different stories can be told about what is happening to wages in America. This means that any statement about pay or wage trends has to be treated carefully. Since when? Using what inflation measure? Which workers? Here we have considered just four of these choices. We have not considered other consequential decisions, for example, the relative merits of measuring wages only, or total compensation, which includes employer-provided benefits like health care. Some analysts strongly prefer compensation to wages. They argue that total compensation best reflects the ultimate purchasing power workers receive for their labor, citing the fact that for the median wage earner, wages represent just two-thirds of their total pay. It is not that there is one right answer. But the choice of method should ideally be driven by a clearly articulated question at hand. Wherever possible, multiple results using different methodologies should be shown.

The danger is that researchers select methods that deliver results in line with their own priors. This is the empirical equivalent of what philosopher Bernard Williams warned against: “smuggling your answer into your question”. When it comes to a question as important to the quality of life of the middle class, the central concern of our initiative, the need for caution and transparency is especially great.

[1] The following analyses are restricted to all civilian prime-age workers (age 25-54). The wage series, retrieved from the Center for Economic and Policy Research, is based on the Current Population Survey Outgoing Rotation Group. The wage series includes “usual hourly earnings for hourly workers (excluding overtime, tips, and commission) and non-hourly workers (including overtime, tips, and commission).” Note that these two methodological choices—important in themselves—are not addressed in this short paper.