Best Smoothing Method for Visualizing Long-Term Trend in Mortality Rates

[Tim Wallner]

Hello,

I have a dataset containing monthly aggregated mortality rates for a specific (not very dangerous) procedure over a 10-year period. My goal is to create a plot that effectively illustrates the general trend over time, without displaying the individual data points, as the final visualization will be used for a lecture (but I have included the scatter plot for clarity here).

I have considered using locally weighted regression (LOWESS) or local polynomial smoothing (lpoly), but I’ve noticed some irregularities at the tails when using lpoly. Given my objective, would either of these methods be appropriate, or is there a better approach for producing a smooth, reliable trend line?

I appreciate any insights or alternative recommendations.

Thank you!

clear
input int admitmonth float monthly_mortality
622         0
623         0
624         0
625         0
626         0
627         0
628         0
629         0
630         0
631         0
632 .16666667
633  .2857143
634         0
635         0
636         0
637         0
638         0
639         0
640         0
641         0
642         0
643         0
644         0
645         0
646         0
647         0
648         0
649         0
650         0
651         0
652         0
653         0
654      .125
655         0
656         0
657         0
658         0
659         0
660         0
661         0
662         0
663         0
664         0
665         0
666         0
667         0
668         0
669 .14285715
670         0
671         0
672         0
673         0
674         0
675         0
676         0
677         0
678         0
679         0
680         0
681         0
682 .16666667
683         0
684         0
685         0
686         0
687         0
688         0
689         0
690         0
691         0
692         0
693         0
694      .125
695         0
696         0
697         0
698         0
699         0
700       .25
701         0
702         0
703         0
704         0
705         0
706         0
707         0
708         0
709         0
710         0
711         0
712 .16666667
713 .14285715
715         0
716         0
717         0
718         0
719         0
720         0
721         0
722         0
723         0
724         0
725         0
726         0
727         0
728         0
729         0
730         0
731         0
732         0
733         0
734         0
735         0
736         0
737         0
738       .25
739         0
740        .2
741         0
742         0
743         0
end
format %tm admitmonth

twoway lpoly monthly_mortality admitmonth, color(maroon%10) degree(4) || ///
       scatter monthly_mortality admitmonth ///
       , ///
       scheme(uncluttered) ///
       xtitle("Year") ytitle("In-hospital mortality [%]", margin(vsmall)) ///
       tlabel(624(12)744, format("%tmCY")) ///
       xlabel(, labsize(small)) ///
       ylabel(0(0.05)0.3, labsize(small)) ///
       legend(off) ///
       graphregion(margin(l-1 r+3)) ///
       title("Monthly mortality - lpoly") ///
       name(lpoly, replace)
      
      
twoway lowess monthly_mortality admitmonth, color(maroon%10) || ///
       scatter monthly_mortality admitmonth ///
       , ///
       scheme(uncluttered) ///
       xtitle("Year") ytitle("In-hospital mortality [%]", margin(vsmall)) ///
       tlabel(624(12)744, format("%tmCY")) ///
       xlabel(, labsize(small)) ///
       ylabel(0(0.05)0.3, labsize(small)) ///
       legend(off) ///
       graphregion(margin(l-1 r+3)) ///
       xtitle("Year", margin(medsmall) size(medium)) ///
       ylabel(0(0.05)0.3, labsize(small)) ///
       ytitle("In-hospital mortality [%]", margin(medsmall) size(medium)) ///
       title("Monthly mortality - lowess") ///
       name(lowess, replace)

[Nick Cox]

I've had good results with lpoly but I would never try to fit quartics and I've found the default choices of kernel type and especially bandwidth often unhelpful. That is why I wrote localp at SSC, which has some default choices I like, which I depart from if the result is too smooth or too rough (or, sometimes, I give up).

lowess is not flexible enough for my purposes. A very specific problem is that lowess or loess is far from standardized and over statistical software the results are not guaranteed identical.

In principle I like splines. In practice -- but that is a story too long for me at this time of night.

Not at all the answer you probably seek, but in your case (1) I would not suppress the data! (2) I would plot annual averages. The data don't seem particularly suited to any kind of smoothing.

Here are some small ideas on visualization. A monthly date display format sometimes just messes up something else I am doing, which is a known quirk (bug) at StataCorp.

clear
input int admitmonth float monthly_mortality
622         0
623         0
624         0
625         0
626         0
627         0
628         0
629         0
630         0
631         0
632 .16666667
633  .2857143
634         0
635         0
636         0
637         0
638         0
639         0
640         0
641         0
642         0
643         0
644         0
645         0
646         0
647         0
648         0
649         0
650         0
651         0
652         0
653         0
654      .125
655         0
656         0
657         0
658         0
659         0
660         0
661         0
662         0
663         0
664         0
665         0
666         0
667         0
668         0
669 .14285715
670         0
671         0
672         0
673         0
674         0
675         0
676         0
677         0
678         0
679         0
680         0
681         0
682 .16666667
683         0
684         0
685         0
686         0
687         0
688         0
689         0
690         0
691         0
692         0
693         0
694      .125
695         0
696         0
697         0
698         0
699         0
700       .25
701         0
702         0
703         0
704         0
705         0
706         0
707         0
708         0
709         0
710         0
711         0
712 .16666667
713 .14285715
715         0
716         0
717         0
718         0
719         0
720         0
721         0
722         0
723         0
724         0
725         0
726         0
727         0
728         0
729         0
730         0
731         0
732         0
733         0
734         0
735         0
736         0
737         0
738       .25
739         0
740        .2
741         0
742         0
743         0
end

* format %tm admitmonth

gen year = year(dofm(admitmonth))
egen annave = mean(monthly_mortality), by(year)

forval y = 2012/2021 {
    local end = ym(`y', 12) + 0.5  
    local ends `ends' `end'
    local middle = ym(`y', 6) + 0.5 
    local call `call' `middle' "`y'"
}

twoway spike monthly admitmonth || line annave admitmonth, c(J) ///
ytitle(Monthly mortality (%)) legend(order(2 "Annual average") pos(12)) ///
xtick(`ends', tlength(*4)) xla(`call', tlength(*0.5) tlc(none)) xtitle("")

The idea is to suppress the zeros and to make more emphatic when you have positive values. But the onus is then on you to explain that the averages are based on the zero too. See https://journals.sagepub.com/doi/pdf...867X0800700410 for the trickery with time axis ticks and labels.

[Nick Cox]

On second thoughts, I guess most people would prefer that the zeros were shown as such. Here's my suggestion now:

twoway spike monthly admitmonth || ///
scatter monthly admitmonth if monthly == 0, ms(oh) mc(stc1) || ///
line annave admitmonth, c(J) lc(stc2) ///
ytitle(Monthly mortality (%)) legend(order(3 "Annual average") pos(12)) ///
xtick(`ends', tlength(*4)) xla(`call', tlength(*0.5) tlc(none)) xtitle("")