3
$\begingroup$

I know that when we standardize a predictor, "one unit change" becomes one standard deviation in the predictor, but what if we only center the data on the mean (i.e. only subtract all values by the mean) ?

  • edit below: bounty

This example is from Winter (2019: 139):

mutate(SER_c = SER - mean(SER, na.rm = T)
term  estimate
intercept 0.66
SER_c     0.11
POSVerb   0.72
SER_c:POSVerb 0.50

Obs: SER is a continuous predictor and POSVerb is a categorical predictor with two levels (0: Noun, 1: Verb)

The author concludes the following: "POSVerb is the difference between nouns and verbs for words with average SER, rather than some arbitrary 0". All right, that makes sense. The intercept should be, then, the value of the outcome variable for the mean of SER. Ok.

But, still, what does one-unit change in SER mean? If it wasn't centered, I'd say "one unit change in SER implies + 0.11 points to the outcome variable for the noun category", but I now that the intercept means the value at SER mean, how do I interpret this?

this post was helpful, but not enough. Thanks in advance!

  • edit : bounty:

Hello, again. Now I have a follow-up question:

  • 1: My model has a 2-level categorical predictor and a continuous one centered around its mean (no interaction). Should I say "one-unit change in the continuons predictor" or "one-unit change in the mean of the continous predictor" adds X to the mean of each categorical level?
$\endgroup$
8
  • 3
    $\begingroup$ Centering alone does not change the interpretation of the non-intercept parameters. $\endgroup$ Sep 7, 2022 at 19:15
  • $\begingroup$ I am confused by your provided notation: since POSVerb does not appear in the command mutate(SER_c = SER - mean(SER, na.rm = T): how can an estimate of its effect on the dependent variable be in the output? Also why is there no effect of SER on the dependent variable in the output? $\endgroup$
    – Alexis
    Sep 7, 2022 at 19:21
  • $\begingroup$ @Alexis , the model is: lm(OutComeVariable ~ SER * POS). SER is continuos and POS is a categorical with 2 predictors which are dummy-coded: 1: Noun, 0: Verb), the mutate function was how the author centered SER into SER_c , is it clearer now? $\endgroup$ Sep 7, 2022 at 19:44
  • $\begingroup$ In a science experiment you could measure temperature in degrees Kelvin, or recenter those to degrees Centigrade, or even recenter them to a difference relative to any reference temperature you like. Would that change the meaning or magnitude of one degree? $\endgroup$
    – whuber
    Sep 7, 2022 at 20:49
  • $\begingroup$ @LarissaCury I think editing a question to add a second part, awarding a bounty, after accepting an answer is gauche. You should probably ask a different question. $\endgroup$
    – AdamO
    Dec 14, 2022 at 19:28

2 Answers 2

3
$\begingroup$

As mentioned in the article that you linked, centering a feature does not impact its scale, so it should not change the interpretation of your model coefficients.

For example, consider the following linear model, where $x_1$ is continuous and $\delta_1$ is a dummy variable {0, 1}:

$$𝑦=B_0 + B_1x_1 + B_2\delta_1 + B_3 x_1 \delta_1 + u$$ In the equation above, the partial effect of $x_1$ on $y$ is

$$ \frac{\Delta y}{\Delta x_1} = B_1 + B_3\delta_1 $$ So, one unit of change in $x_1$ leads to an increase of $B1+B3$ in $y$ if $d_1 = 1$ and an increase of $B_1$ otherwise.

Suppose that we now center $x_1$ around its mean $\bar{x_1}$, let's call it $\hat{x}_1$, with the following operation:

$$ \hat{x_1} = x_1 - \bar{x_1} $$

Notice that we can replace $x_1$ in the initial model by $\hat{x_1} + \bar{x_1}$, such that $$ 𝑦=B_0 + B_1(\hat{x_1} + \bar{x_1}) + B_2\delta_1 + B_3 (\hat{x_1} + \bar{x_1}) \delta_1 + u $$ Which you can expand to

$$ 𝑦=B_0 + B_1\hat{x_1} + B_1\bar{x_1} + B_2\delta_1 + B_3\hat{x_1}\delta_1 + B_3\bar{x_1}\delta_1 + u $$ So now, the partial effect of $\hat{x_1}$ on $y$ is:

$$ \frac{\Delta y}{\Delta \hat{x_1}} = B_1 + B_3\delta_1 $$

Which is the same as before. In your particular case:

$$ \frac{\Delta y}{\Delta SER_c} = \frac{\Delta y}{\Delta SER} = 0.11 + 0.50 POSVerb $$

In the case of noun ($POSVerb = 0$) one unit change of $SER$ or $SER_c$ yields a 0.11 increase to the outcome. In the case of a verb ($POSVerb = 1$), then one unit of change in $SER$ or $SER_c$ yields a 0.11 + 0.50 = 0.61 increase to the outcome.

Hope this helps!

$\endgroup$
13
  • $\begingroup$ Can you expand your answer to address the interaction term in Larissa Cury's question? $\endgroup$
    – Alexis
    Sep 7, 2022 at 19:46
  • $\begingroup$ Thanks for your comment @Alexis! I would love to elaborate, but I wasn't able to track down the source. Googling for Winter 2019 didn't bring anything up. From what I can understand of Larissa's problem, I'd say the statement "one unit change in SER implies +0.11 points to the outcome variable for the noun category" is still valid, but having the full source would probably help provide a richer interpretation. $\endgroup$
    – privard
    Sep 7, 2022 at 19:58
  • 1
    $\begingroup$ Ah, yes I get your point now @Alexis, thanks. Let me try to adapt the answer to include interaction effects. $\endgroup$
    – privard
    Sep 7, 2022 at 20:37
  • 1
    $\begingroup$ @Alexis, here's an expanded answer with interaction terms. Hopefully, I didn't make any mistakes, but I would appreciate a quick review. Thanks for the feedback! $\endgroup$
    – privard
    Sep 7, 2022 at 21:13
  • 2
    $\begingroup$ Thanks @Alexis! @LarissaCury I think the best way to intuit the full model is to see as two separate models: 1 model for nouns and 1 model for verbs. The model for nouns is y = 0.66 + 0.11 SER_c = 0.66 + 0.11 (SER - mean(SER)) The model for verbs is y = 1.38 + 0.61 SER_c = 1.38 + 0.61 (SER - mean(SER)) So, when SER = mean(SER), y = 0.66 for nouns, and y = 1.38 for verbs. Then any changes to SER will result in changes of y around mean(SER). $\endgroup$
    – privard
    Sep 7, 2022 at 23:13
1
$\begingroup$

I know that when we standardize a predictor, "one unit change" becomes one standard deviation in the predictor, but what if we only center the data on the mean (i.e. only subtract all values by the mean) ?

If you don't standardize then one unit change will relate to whatever the units are that are used to measure the predictor variable.

For instance if the predictor is temperature measured in Kelvin, then on unit change means an increase by one Kelvin.

In your example you have 'sensory experience ratings' as predictor and one-unit change will relate to whatever the scale is used to obtain those ratings.

Possibly your sensory experience rating is the one from

Juhasz, B.J., Yap, M.J. Sensory experience ratings for over 5,000 mono- and disyllabic words. Behav Res 45, 160–168 (2013). https://doi.org/10.3758/s13428-012-0242-9

Please rate each word on a 1 to 7 scale, with 1 meaning the word evokes no sensory experience for you, 4 meaning the word evokes a moderate sensory experience, and 7 meaning the word evokes a strong sensory experience.

In that case the ratings are based on averages of questionnaires with a 7 point scale.

The meaning of one-unit is the change of the score on that questionnaire by one.

$\endgroup$
1
  • $\begingroup$ hi, @Sextux Empiricus, thanks! but how could we expand that to non-Likert-scale results, tho I'm also a linguistic as Winter, I'm from a completly different area, I've never worked with this 'iconicity' measure. I usially work with language exams and proficiency, for example. Right now I have proficiency as a predictor (exam from 0 to 120) and I've centered that. I have SCORE ~ Proficiency_c * type of exam (A or B) $\endgroup$ Dec 19, 2022 at 21:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.