Let's say I want to regress crop yield as a function of rainfall and temperature.

I collected data across 20 locations and 10 years and construct a model regressing yield on rainfall and temperature with location and year as random effects in R.

  mdl <- lmer(yield ~ rainfall + temperature + (1|location) + (1|year) 

However, I know irrigation is quite important for crop yield. However, I have irrigation data only for one year across all the 20 locations. So I can construct a single model for a year using this:

mdl.yld <- lmer(yield ~ rainfall + temperature + irrigation + (1|location) # for a single year which has irrigation data

Can I use the regression coefficients of irrigation from mdl.yld and add to the regression equation inmdl so that I have an equation that predicts yield as a function of rainfall, temperature and irrigation?

  • $\begingroup$ You can't simply throw in the irrigation here, because it'll be missing in other years. The MLE estimation should work here $\endgroup$
    – Aksakal
    Apr 13, 2018 at 18:45
  • $\begingroup$ What does "irrigation" measure? Is it a yes/no variable? Is it time dependent? $\endgroup$
    – AdamO
    Apr 13, 2018 at 18:57
  • $\begingroup$ @Aksakal just to be sure are you thinking about this as a missing data problem? Would it be appropriate to impute irrigation? They would produce the same estimates I think. I don't know if there's enough detail to recommend a missing-data approach... $\endgroup$
    – AdamO
    Apr 13, 2018 at 19:50
  • $\begingroup$ @AdamO, no, I wonder if OP runs the model as in the given code, whether the model will treat it as missing data, and impute. I thought maybe if you explicitly write MLE for the model when irrigation is optional whether this would produce a better fit $\endgroup$
    – Aksakal
    Apr 13, 2018 at 20:05
  • $\begingroup$ @Aksakal lmer does not impute data, unless I'm mistaken. Mixed models do have important connections with imputed analyses. I'm not aware that REML produces different estimates than ML in terms of how missing data are handled if that's what you're talking about. Maximum likelihood could also mean marginalizing the likelihood over missing values which is a complex likelihood. $\endgroup$
    – AdamO
    Apr 13, 2018 at 20:08

1 Answer 1


Visually check the correlation between irrigation and the other variables for that year. If they are correlated, which I imagine they are (different locations probably have different irrigation patterns based on demographics, etc...maybe not though!) then no. If there seems to be little correlation and irrigation just seems to increase yield by a fixed value...sure, you could make that assumption so long as it’s clearly stipulated and interpreted that way. This is where subject matter expertise really comes into play.


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.