[R] car::deltaMethod() fails when a particular combination of categorical variables is not present

Michael Cohn non @end|ng |rom mcohn@net
Fri Oct 6 12:52:34 CEST 2023


Thank you very much, John. This has allowed us to move forward on several
fronts and better understand our data.

- Michael Cohn

On Tue, Sep 26, 2023 at 8:39 AM John Fox <jfox using mcmaster.ca> wrote:

> Dear Michael,
>
> My previous response was inaccurate: First, linearHypothesis() *is* able
> to accommodate aliased coefficients by setting the argument singular.ok
> = TRUE:
>
>  > linearHypothesis(minimal_model, "bt2 + csent + bt2:csent = 0",
> +                  singular.ok=TRUE)
>
> Linear hypothesis test:
> bt2  + csent  + bt2:csent = 0
>
> Model 1: restricted model
> Model 2: a ~ b * c
>
>    Res.Df    RSS Df Sum of Sq      F Pr(>F)
> 1     16 9392.1
> 2     15 9266.4  1    125.67 0.2034 0.6584
>
> Moreover, when there is an empty cell, this F-test is (for a reason that
> I haven't worked out, but is almost surely due to how the rank-deficient
> model is parametrized) *not* equivalent to the t-test for the
> corresponding coefficient in the raveled version of the two factors:
>
>  > df$bc <- factor(with(df, paste(b, c, sep=":")))
>  > m <- lm(a ~ bc, data=df)
>  > summary(m)
>
> Call:
> lm(formula = a ~ bc, data = df)
>
> Residuals:
>      Min      1Q  Median      3Q     Max
> -57.455 -11.750   0.439  14.011  37.545
>
> Coefficients:
>              Estimate Std. Error t value Pr(>|t|)
> (Intercept)    20.50      17.57   1.166   0.2617
> bct1:unsent    37.50      24.85   1.509   0.1521
> bct2:other     32.00      24.85   1.287   0.2174
> bct2:sent      17.17      22.69   0.757   0.4610  <<< cf. F = 0.2034, p
> = 0.6584
> bct2:unsent    38.95      19.11   2.039   0.0595
>
> Residual standard error: 24.85 on 15 degrees of freedom
> Multiple R-squared:  0.2613,    Adjusted R-squared:  0.06437
> F-statistic: 1.327 on 4 and 15 DF,  p-value: 0.3052
>
> In the full-rank case, however, what I said is correct -- that is, the
> F-test for the 1 df hypothesis on the three coefficients is equivalent
> to the t-test for the corresponding coefficient when the two factors are
> raveled:
>
>  > linearHypothesis(minimal_model_fixed, "bt2 + csent + bt2:csent = 0")
>
> Linear hypothesis test:
> bt2  + csent  + bt2:csent = 0
>
> Model 1: restricted model
> Model 2: a ~ b * c
>
>    Res.Df    RSS Df Sum of Sq      F Pr(>F)
> 1     15 9714.5
> 2     14 9194.4  1    520.08 0.7919 0.3886
>
>  > df_fixed$bc <- factor(with(df_fixed, paste(b, c, sep=":")))
>  > m <- lm(a ~ bc, data=df_fixed)
>  > summary(m)
>
> Call:
> lm(formula = a ~ bc, data = df_fixed)
>
> Residuals:
>      Min      1Q  Median      3Q     Max
> -57.455 -11.750   0.167  14.011  37.545
>
> Coefficients:
>              Estimate Std. Error t value Pr(>|t|)
> (Intercept)   64.000     25.627   2.497   0.0256
> bct1:sent    -43.500     31.387  -1.386   0.1874
> bct1:unsent  -12.000     36.242  -0.331   0.7455
> bct2:other   -11.500     31.387  -0.366   0.7195
> bct2:sent    -26.333     29.591  -0.890   0.3886 << cf.
> bct2:unsent   -4.545     26.767  -0.170   0.8676
>
> Residual standard error: 25.63 on 14 degrees of freedom
> Multiple R-squared:  0.2671,    Adjusted R-squared:  0.005328
> F-statistic:  1.02 on 5 and 14 DF,  p-value: 0.4425
>
> So, to summarize:
>
> (1) You can use linearHypothesis() with singular.ok=TRUE to test the
> hypothesis that you specified, though I suspect that this hypothesis
> probably isn't testing what you think in the rank-deficient case. I
> suspect that the hypothesis that you want to test is obtained by
> raveling the two factors.
>
> (2) There is no reason to use deltaMethod() for a linear hypothesis, but
> there is also no intrinsic reason that deltaMethod() shouldn't be able
> to handle a rank-deficient model. We'll probably fix that.
>
> My apologies for the confusion,
>   John
>
> --
> John Fox, Professor Emeritus
> McMaster University
> Hamilton, Ontario, Canada
> web: https://www.john-fox.ca/
>
> On 2023-09-26 9:49 a.m., John Fox wrote:
> > Caution: External email.
> >
> >
> > Dear Michael,
> >
> > You're testing a linear hypothesis, so there's no need to use the delta
> > method, but the linearHypothesis() function in the car package also
> > fails in your case:
> >
> >  > linearHypothesis(minimal_model, "bt2 + csent + bt2:csent = 0")
> > Error in linearHypothesis.lm(minimal_model, "bt2 + csent + bt2:csent =
> > 0") :
> > there are aliased coefficients in the model.
> >
> > One work-around is to ravel the two factors into a single factor with 5
> > levels:
> >
> >  > df$bc <- factor(with(df, paste(b, c, sep=":")))
> >  > df$bc
> >   [1] t2:unsent t2:unsent t2:unsent t2:unsent t2:sent   t2:unsent
> >   [7] t2:unsent t1:sent   t2:unsent t2:unsent t2:other  t2:unsent
> > [13] t1:unsent t1:sent   t2:unsent t2:other  t1:unsent t2:sent
> > [19] t2:sent   t2:unsent
> > Levels: t1:sent t1:unsent t2:other t2:sent t2:unsent
> >
> >  > m <- lm(a ~ bc, data=df)
> >  > summary(m)
> >
> > Call:
> > lm(formula = a ~ bc, data = df)
> >
> > Residuals:
> >      Min      1Q  Median      3Q     Max
> > -57.455 -11.750   0.439  14.011  37.545
> >
> > Coefficients:
> >              Estimate Std. Error t value Pr(>|t|)
> > (Intercept)    20.50      17.57   1.166   0.2617
> > bct1:unsent    37.50      24.85   1.509   0.1521
> > bct2:other     32.00      24.85   1.287   0.2174
> > bct2:sent      17.17      22.69   0.757   0.4610
> > bct2:unsent    38.95      19.11   2.039   0.0595
> >
> > Residual standard error: 24.85 on 15 degrees of freedom
> > Multiple R-squared:  0.2613,    Adjusted R-squared:  0.06437
> > F-statistic: 1.327 on 4 and 15 DF,  p-value: 0.3052
> >
> > Then the hypothesis is tested directly by the t-value for the
> > coefficient bct2:sent.
> >
> > I hope that this helps,
> >   John
> >
> > --
> > John Fox, Professor Emeritus
> > McMaster University
> > Hamilton, Ontario, Canada
> > web: https://www.john-fox.ca/
> >
> > On 2023-09-26 1:12 a.m., Michael Cohn wrote:
> >> Caution: External email.
> >>
> >>
> >> I'm running a linear regression with two categorical predictors and
> their
> >> interaction. One combination of levels does not occur in the data, and
> as
> >> expected, no parameter is estimated for it. I now want to significance
> >> test
> >> a particular combination of levels that does occur in the data (ie, I
> >> want
> >> to get a confidence interval for the total prediction at given levels of
> >> each variable).
> >>
> >> In the past I've done this using car::deltaMethod() but in this dataset
> >> that does not work, as shown in the example below: The regression model
> >> gives the expected output, but deltaMethod() gives this error:
> >>
> >> error in t(gd) %*% vcov. : non-conformable arguments
> >>
> >> I believe this is because there is no parameter estimate for when the
> >> predictors have the values 't1' and 'other'. In the df_fixed dataframe,
> >> putting one person into that combination of categories causes
> >> deltaMethod()
> >> to work as expected.
> >>
> >> I don't know of any theoretical reason that missing one interaction
> >> parameter estimate should prevent getting a confidence interval for a
> >> different combination of predictors. Is there a way to use
> >> deltaMethod() or
> >> some other function to do this without changing my data?
> >>
> >> Thank you,
> >>
> >> - Michael Cohn
> >> Vote Rev (http://voterev.org)
> >>
> >>
> >> Demonstration:
> >> ------
> >>
> >> library(car)
> >> # create dataset with outcome and two categorical predictors
> >> outcomes <- c(91,2,60,53,38,78,48,33,97,41,64,84,64,8,66,41,52,18,57,34)
> >> persontype <-
> >>
> c("t2","t2","t2","t2","t2","t2","t2","t1","t2","t2","t2","t2","t1","t1","t2","t2","t1","t2","t2","t2")
> >> arm_letter <-
> >>
> c("unsent","unsent","unsent","unsent","sent","unsent","unsent","sent","unsent","unsent","other","unsent","unsent","sent","unsent","other","unsent","sent","sent","unsent")
> >> df <- data.frame(a = outcomes, b=persontype, c=arm_letter)
> >>
> >> # note: there are no records with the combination 't1' + 'other'
> >> table(df$b,df$c)
> >>
> >>
> >> #regression works as expected
> >> minimal_formula <- formula("a ~ b*c")
> >> minimal_model <- lm(minimal_formula, data=df)
> >> summary(minimal_model)
> >>
> >> #use deltaMethod() to get a prediction for individuals with the
> >> combination
> >> 'b2' and 'sent'
> >> # deltaMethod() fails with "error in t(gd) %*% vcov. : non-conformable
> >> arguments."
> >> deltaMethod(minimal_model, "bt2 + csent + `bt2:csent`", rhs=0)
> >>
> >> # duplicate the dataset and change one record to be in the previously
> >> empty
> >> cell
> >> df_fixed <- df
> >> df_fixed[c(13),"c"] <- 'other'
> >> table(df_fixed$b,df_fixed$c)
> >>
> >> #deltaMethod() now works
> >> minimal_model_fixed <- lm(minimal_formula, data=df_fixed)
> >> deltaMethod(minimal_model_fixed, "bt2 + csent + `bt2:csent`", rhs=0)
> >>
> >>          [[alternative HTML version deleted]]
> >>
> >> ______________________________________________
> >> R-help using r-project.org mailing list -- To UNSUBSCRIBE and more, see
> >> https://stat.ethz.ch/mailman/listinfo/r-help
> >> PLEASE do read the posting guide
> >> http://www.R-project.org/posting-guide.html
> >> and provide commented, minimal, self-contained, reproducible code.
> >
> > ______________________________________________
> > R-help using r-project.org mailing list -- To UNSUBSCRIBE and more, see
> > https://stat.ethz.ch/mailman/listinfo/r-help
> > PLEASE do read the posting guide
> > http://www.R-project.org/posting-guide.html
> > and provide commented, minimal, self-contained, reproducible code.
>
>

	[[alternative HTML version deleted]]



More information about the R-help mailing list