Automatic donor data selection

[stefvanbuuren]

Ideally, the donordata chooser should disappear and the choice of donordata should be made in an automatic fashion.

Probably the best way to automate donor data selection is: define 0-4 donor data (lollypop+smocc) and 4-18 donor data (terneuzen+pops). We need

• well-fitting brokenstick models for the combined data
• columns holding the age range per id
• a method to pre-select cases based on the age range of the observed data & number of data points

[stefvanbuuren]

Break points per source

Time	lollypop	pops	smocc	terneuzen	nederland
0	L	L	L	L	L
4w	+		+		+
8w	+		+		+
3m	+		+	+	+
4m	+		+
5m		+
6m	+		+	+	+
7.5m	+		+
8m		+
9m	+		+		+
11m	+		+		+
12m
14m	+	+	+	+	+
15m
18m	+		+
24m	+		+	+	+
26m		+
36m	+		R
45m	+
4y				+	+
5y	R	+
6y					+
10y				+	+
10.25y		+
14y					+
14.25y		+
18y				+	+
19.25y		+
20.20y		R
29y				R	R

[stefvanbuuren]

Residual variance estimates

Brokenstick model (lmer) for breakpoints specified above. Unstructured covariance matrix of random effects.

Fitting brokenstick models for wfh_z requires specification of break points relative to child height (instead of child age), and has not yet been tried.

Measure	lol.preterm	lol.term	lollypop	pops	smocc	terneuzen	nederland
hgt_z	0.092	0.099	0.096	0.234	0.085	0.119	0.207
wgt_z	0.044	0.046	0.045	0.155	0.034	0.097	0.166
hdc_z	0.083	0.048	0.080	0.238	0.053		0.151
bmi_z	0.197	0.092	0.188		0.123	0.215	0.385
dsc_z					0.485

[stefvanbuuren]

Question 1: Can we generalise the prediction methodology such that it becomes agnostic to gestational age?

• For height and weight, separate and combined models in lollypop appear similar.
• For head circumference and bmi, the term model fits better than the pre-term model.
• Postnatal growth in hdc is extremely rapid and variable, but the correlations of the random effect are quite similar, which suggests that combining is possible. Both fits are below 0.10, but a possible cause of their difference is the low corresponds between the lollypop.term hdc and the nl1997 references (fixed effects are -0.41, +0.28, +0.26, +0.11 for ages 0-3 months).
• For bmi_z the lower fit could be caused by the lack of an appropriate bmi reference for preterms, which increases the variation in bmi_z in the first 3 months significantly beyond 1.
• The difference in fit for bmi_z for the three lollypop subset suggests that fit for preterms might improve if measurements are converted with preterm bmi references (which do not yet exist).

[stefvanbuuren]

Question 2: Can we combine the donor data into one set that is good for all purposes?

• The smocc and lollypop studies fit the brokenstick model equally well, closely followed by the terneuzen cohort. By comparison, the pops study has an unusual bad fit for all outcomes.
• Combining all studies into a nederland study yields low fits.
• If we remove pops from the combined data, we get slightly better - but still unsatisfactory - estimates: 0.156 (hgt_z), 0.125 (wgt_z), 0.146 (hdc_z), 0.273 (bmi_z).
• Alternatively, we could group lollypop and smocc into a 0-4 model, and terneuzen and pops into a long term model (0-18y or 4-18y).

[stefvanbuuren]

JAMES 1.6.3 automates the choice of the donor data and removes the menu for setting the donordata from the interactive interface. The choice is steered by the future age specified by the user.