I have data about investment preferences 1 year before the Covid and during the Covid lockdown.

Some changes appear using simple T-Test. I want to be able to assess if these changes are particularly strong for some specific demographics (e.g., older individuals ($X_1$), individuals with lower income ($X_2$), etc...).

Should I use the initial level of my dependant variable in the regressions? Basically, if I want to use OLS regressions to investigate which independant variable correlate with the change in my dependant variable, which model is preferrable?

Model 1 (apparently called Change Score Method): $(Y_2-Y_1)= \beta_1 . X_1+ \beta_2 . X_2 $

Model 2 (apparently called Regressor Variable Method) Score Method): $Y_2= \beta_1 . X_1+ \beta_2 . X_2 + \beta_3 . Y_1 $

Thank you so much for your help - Any reference would also be much appreciated!


Both methods have been used. See here for example. It depends what question you want to answer. If you want to talk mostly about "change" you can use

(Y2-Y1) ~ X1 + X2            # (1)

Basal (Y1) should not be added to above equation as it will always be correlated with difference (Y2-Y1) - see comments below by @EdM and here.

On the other hand, if you want to discuss factors affecting "final value", you can use

Y2 ~ X1 + X2 + Y1            # (2)

However, since repeated measurements (Y1,Y2 at 2 times) have been done on same subject, hence mixed model is also often used. (including interactions as commented by @dbwilson below):

Y ~ X1 + X2 + time + X1*time + X2*time + (1|subject)

Following simplified version of formula is effectively same as above:

Y ~ X1*time + X2*time + (1|subject)            # (3)

There is another method commonly used, especially in biomedical literature: "Percent change", i.e.

(100*(Y2-Y1)/Y1) ~ X1 + X2            # (4)

It is not correct to keep Y1 as a predictor variable in this last method as there will be strong correlation between baseline and percent change.

I think this last method (percent change) is most understandable.

See here for more information on this topic.

| cite | improve this answer | |
  • $\begingroup$ Thank you so much for this detailed answer. In the end, given that I was mostly interested in change, I used (Y2-Y1) ~ X1 + X2 It is however interesting to see the last two methods you propose. Thank you again! $\endgroup$ – L. M. Jul 21 at 10:30
  • $\begingroup$ Regressing the difference against the initial value is not a good idea. See this answer and its links and this answer to the question "What are the worst (commonly adopted) ideas/principles in statistics?" $\endgroup$ – EdM Jul 21 at 11:21
  • $\begingroup$ I have added a note regarding this in answer above. $\endgroup$ – rnso Jul 21 at 11:30
  • $\begingroup$ In the mixed-model, the interaction between X1*time and X2*time are estimating the same effect as the X1 and X2 effects in the change score model. The code, however, should be Y~X1+X2+time+X1*time+X2*time + (1|subject). $\endgroup$ – dbwilson Jul 21 at 11:55
  • $\begingroup$ I have added this in answer above with your reference. $\endgroup$ – rnso Jul 21 at 12:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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